admin(aimetrics): /v1/admin/aimetrics AI-metrics read view (Langfuse totals/latency/models + usage + eval scores) + table-canonical SQL tests

Replaces the pseudo-version pin (v1.49.2-0.20260719024505-24ff20a68f52, the
Bug-A iterator-leak fix only) with the released v1.49.2, which also carries the
Bug-B guard: org.Resolve skips the doomed GetById for a legacy all-digit cached
id (IAM Valkey's stale 1772587477 for 'hanzo') and resolves by name, so the
(*Query).ById legacy-numeric path that hot-looped in v1.801.95 is never taken.
Keeps ai v1.826.4 (in-proc TierReader). go.mod+go.sum only.

.github/workflows/containment.yml

@@ -89,18 +89,50 @@ jobs:
         with:
           go-version-file: go.mod
 
-      - name: go env for private modules (matches Dockerfile — zap-proto is direct+authenticated)
+      - name: go env for private modules
         env:
           GH_PAT: ${{ secrets.GH_PAT }}
+        # GOPRIVATE names exactly the namespace that is private. github.com/hanzoai/*
+        # is: ai, account, commerce, orm, xorm, beego, csqlite and ~30 more are
+        # private repos, so they must resolve direct+authenticated and skip a sumdb
+        # that cannot see them. Everything else stays on the public proxy + checksum
+        # db, which is what makes a module hash immutable: zap-proto (all 55 repos)
+        # and luxfi (all 37 deps here) are public and proxy-served.
+        #
+        # This previously named zap-proto — public, and never the reason anything
+        # here was direct — and then set GOSUMDB=off to compensate for hanzoai/*
+        # being absent, which disabled checksum verification for EVERY module in the
+        # build, public ones included. Naming the private namespace is what the off
+        # switch was standing in for.
         run: |
           git config --global url."https://x-access-token:${GH_PAT}@github.com/".insteadOf "https://github.com/"
           {
-            echo "GOPRIVATE=github.com/zap-proto/*"
-            echo "GONOSUMDB=github.com/zap-proto/*"
-            echo "GOSUMDB=off"
+            echo "GOPRIVATE=github.com/hanzoai/*"
             echo "GOPROXY=https://proxy.golang.org,direct"
           } >> "$GITHUB_ENV"
 
+      - name: zen streaming-fix floor — go.mod must pin github.com/hanzoai/zen >= v1.4.1
+        # The SSE body-close fix (zen commit 50328b8, first released in zen v1.4.1)
+        # is what makes streaming completions return a body instead of an empty
+        # stream. A stale-branch merge that reverts go.mod's zen pin below the floor
+        # silently re-breaks streaming, and `next build`'s ignoreBuildErrors hides
+        # the runtime break — so no image may be cut on a regressed pin. This is the
+        # durable root-cause guard: it reads the EFFECTIVE module version (post-MVS,
+        # exactly what the build links) and fails the PR/push below the floor.
+        run: |
+          set -euo pipefail
+          FLOOR="v1.4.1"
+          V="$(go list -m -f '{{.Version}}' github.com/hanzoai/zen)"
+          echo "effective github.com/hanzoai/zen = ${V} (floor ${FLOOR})"
+          # semver-correct compare: the lowest of {V, FLOOR} under `sort -V` must be
+          # the FLOOR, i.e. V >= FLOOR. (sort -V orders v1.4.2 above v1.4.10 too.)
+          low="$(printf '%s\n%s\n' "$V" "$FLOOR" | sort -V | head -1)"
+          if [ "$low" != "$FLOOR" ]; then
+            echo "::error::github.com/hanzoai/zen is pinned at ${V}, below the streaming-fix floor ${FLOOR} — this re-breaks SSE streaming (empty completions). Re-pin zen to >= ${FLOOR} in go.mod before merging."
+            exit 1
+          fi
+          echo "OK: zen ${V} is at or above the streaming-fix floor ${FLOOR}"
+
       - name: positive proof — clients/controlplane is unreachable from the default build
         run: |
           set -euo pipefail

.github/workflows/release.yml

@@ -96,9 +96,9 @@ jobs:
           # an image (even from a run that died before tagging) is never reused.
           cont_max=""
           if command -v gh >/dev/null 2>&1; then
-            cont_max="$(GH_TOKEN="$GH_PAT" gh api --paginate \
-              '/orgs/hanzoai/packages/container/cloud/versions' \
-              --jq '.[].metadata.container.tags[]' 2>/dev/null \
+            cont_max="$(GH_TOKEN="$GH_PAT" gh api \
+              '/orgs/hanzoai/packages/container/cloud/versions?per_page=100' \
+              --jq '.[].metadata.container.tags[]?' 2>/dev/null \
               | sed 's/^v//' | grep -E '^[0-9]+\.[0-9]+\.[0-9]+$' | sort -V | tail -1 || true)"
           fi
 
@@ -124,20 +124,6 @@ jobs:
           echo "sha_short=$(git rev-parse --short "$GITHUB_SHA")" >> "$GITHUB_OUTPUT"
           echo "Next release: v${version} (git_max='${git_max:-none}' container_max='${cont_max:-none}')"
 
-          # Console-embed cachebust. The console clone+build layer is keyed on this;
-          # prefer hanzoai/console main HEAD so a CONSOLE-ONLY change re-embeds without
-          # needing a cloud commit (cloud-sha alone froze the embed between cloud pushes).
-          # git ls-remote must CLEAR the extraheader actions/checkout installs (it carries
-          # THIS repo's GITHUB_TOKEN, which 404s the cross-repo console lookup); gh is not
-          # on the runner. If resolution yields nothing, fall back to the cloud sha — still
-          # unique per cloud commit, so the embed is never frozen. Either way THIS build
-          # busts (new value) and re-clones console main fresh.
-          console_head="$(git -c 'http.https://github.com/.extraheader=' ls-remote \
-            "https://x-access-token:${GH_PAT}@github.com/hanzoai/console.git" refs/heads/main 2>/dev/null | cut -f1 || true)"
-          cachebust="${console_head:-$GITHUB_SHA}"
-          echo "cachebust=${cachebust}" >> "$GITHUB_OUTPUT"
-          echo "console cachebust: ${cachebust} (console_head='${console_head:-none}')"
-
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
         with:
@@ -160,8 +146,12 @@ jobs:
           # Direct credential first (repo/org secret — works on private repos,
           # where the Free plan hides org KMS secrets); KMS kubeconfig fallback.
           if [ -n "${REGISTRY_USER:-}" ] && [ -n "${REGISTRY_PASSWORD:-}" ]; then
-            echo "$REGISTRY_PASSWORD" | docker login registry.hanzo.ai -u "$REGISTRY_USER" --password-stdin
-            echo "MIRROR_OK=1" >> "$GITHUB_ENV"; exit 0
+            if echo "$REGISTRY_PASSWORD" | docker login registry.hanzo.ai -u "$REGISTRY_USER" --password-stdin; then
+              echo "MIRROR_OK=1" >> "$GITHUB_ENV"
+            else
+              echo "mirror login failed (registry.hanzo.ai unreachable) — mirror skipped, release continues"
+            fi
+            exit 0
           fi
           [ -z "${KMS_CLIENT_ID:-}" ] && { echo "no KMS creds — mirror skipped"; exit 0; }
           TOKEN=$(curl -sf "$KMS_ENDPOINT/v1/kms/auth/login" -H 'Content-Type: application/json'             -d "{\"clientId\":\"$KMS_CLIENT_ID\",\"clientSecret\":\"$KMS_CLIENT_SECRET\"}" | jq -r '.accessToken // empty')
@@ -180,8 +170,11 @@ jobs:
           UP=$(echo "$CFG" | jq -r '.auths["registry.hanzo.ai"].auth // empty' | base64 -d)
           [ -z "$UP" ] && { echo "no registry auth — mirror skipped"; exit 0; }
           echo "::add-mask::${UP#*:}"
-          echo "${UP#*:}" | docker login registry.hanzo.ai -u "${UP%%:*}" --password-stdin
-          echo "MIRROR_OK=1" >> "$GITHUB_ENV"
+          if echo "${UP#*:}" | docker login registry.hanzo.ai -u "${UP%%:*}" --password-stdin; then
+            echo "MIRROR_OK=1" >> "$GITHUB_ENV"
+          else
+            echo "mirror login failed (registry.hanzo.ai unreachable) — mirror skipped, release continues"
+          fi
 
       - name: Log in to ghcr.io (GH_PAT — writes the cloud package despite its ai-repo linkage)
         uses: docker/login-action@v3
@@ -190,6 +183,39 @@ jobs:
           username: hanzo-dev
           password: ${{ secrets.GH_PAT }}
 
+      - name: Resolve decomplection artifact digests (the Go-only build's prebuilt inputs)
+        id: artifacts
+        run: |
+          set -euo pipefail
+          # cloud compiles ONLY Go; it pulls three prebuilt artifacts (console SPA,
+          # agent-skills catalog, native flags staticlib). Resolve each published
+          # :latest to an IMMUTABLE digest so THIS release is reproducible (pinned,
+          # not floating :latest) AND a console/skills/flags change is picked up —
+          # its CI republished :latest, so this resolves to the NEW digest. A MISSING
+          # artifact FAILS the release HERE, before build/smoke/push/tag: the receipt
+          # invariant means we never tag an image that couldn't embed the real console.
+          command -v crane >/dev/null 2>&1 || {
+            mkdir -p "$HOME/.local/bin"
+            curl -fsSL "https://github.com/google/go-containerregistry/releases/latest/download/go-containerregistry_Linux_x86_64.tar.gz" \
+              | tar -xz -C "$HOME/.local/bin" crane
+          }
+          export PATH="$HOME/.local/bin:$PATH"
+          resolve() {
+            local repo="$1" d
+            d="$(crane digest "ghcr.io/hanzoai/${repo}:latest" 2>/dev/null || true)"
+            [ -n "$d" ] || { echo "::error::decomplection artifact ghcr.io/hanzoai/${repo}:latest is not published — refusing to cut a release that would embed a stale/placeholder ${repo}"; return 1; }
+            printf 'ghcr.io/hanzoai/%s@%s' "$repo" "$d"
+          }
+          CONSOLE_IMAGE="$(resolve console-embed)" || exit 1
+          SKILLS_IMAGE="$(resolve agent-skills)" || exit 1
+          FLAGS_IMAGE="$(resolve cloud-flags)" || exit 1
+          {
+            echo "console_image=${CONSOLE_IMAGE}"
+            echo "skills_image=${SKILLS_IMAGE}"
+            echo "flags_image=${FLAGS_IMAGE}"
+          } >> "$GITHUB_OUTPUT"
+          echo "resolved: console=${CONSOLE_IMAGE} skills=${SKILLS_IMAGE} flags=${FLAGS_IMAGE}"
+
       - name: OCI labels
         id: meta
         uses: docker/metadata-action@v5
@@ -216,11 +242,13 @@ jobs:
           load: true
           tags: cloud:smoke
           labels: ${{ steps.meta.outputs.labels }}
-          # Bust the console clone+build layer every release (the cloud commit sha is
-          # unique per push) so the embed re-fetches console main HEAD fresh — never the
-          # frozen snapshot the persistent BuildKit cache would otherwise serve forever.
+          # cloud compiles ONLY Go: pull the three prebuilt artifacts pinned to the
+          # digests resolved above (reproducible, and fresh — a console/skills/flags
+          # change is a new digest). No node/python/rust toolchain in this build.
           build-args: |
-            CONSOLE_CACHEBUST=${{ steps.ver.outputs.cachebust }}
+            CONSOLE_IMAGE=${{ steps.artifacts.outputs.console_image }}
+            SKILLS_IMAGE=${{ steps.artifacts.outputs.skills_image }}
+            FLAGS_IMAGE=${{ steps.artifacts.outputs.flags_image }}
           # GIT_AUTH_TOKEN: BuildKit secret the Dockerfile consumes to fetch private
           # cross-org Go modules (hanzoai/*, luxfi/*) over authenticated git.
           secrets: |
@@ -293,6 +321,67 @@ jobs:
 
           echo "SMOKE PASS: cloud:smoke booted to \"listening\" with no crash signature"
 
+      # ── Functional smoke — authenticated per-subsystem probe (the REAL gate) ─────
+      # The boot check above proves the process REACHES "listening"; this proves the
+      # mounted HTTP surface actually WORKS. /smoke (cmd/smoke, baked into the image)
+      # hits ONE side-effect-free read per core subsystem and FAILS the release on any
+      # broken code — above all a 402 on a READ (the balance-gate-over-blocks-reads
+      # regression) or a 5xx (a crash, e.g. the /v1/billing/usage self-dispatch 500).
+      # So a release can never ship with chat/billing/projects/kms/... down.
+      - name: Functional smoke — per-subsystem probe (fails the release if a core endpoint is broken)
+        env:
+          KMS_CLIENT_ID: ${{ secrets.KMS_CLIENT_ID }}
+          KMS_CLIENT_SECRET: ${{ secrets.KMS_CLIENT_SECRET }}
+          KMS_ENDPOINT: ${{ vars.KMS_ENDPOINT || 'https://kms.hanzo.ai' }}
+          # A KMS-provisioned short-lived smoke bearer, injected as a secret (NEVER
+          # hardcoded). Absent → the anonymous matrix still gates public/authed and
+          # catches every 402-on-read / 5xx.
+          SMOKE_TOKEN: ${{ secrets.SMOKE_TOKEN }}
+        run: |
+          set -euo pipefail
+          IMAGE=cloud:smoke
+          CID=""
+          cleanup() { [ -n "$CID" ] && docker rm -f "$CID" >/dev/null 2>&1 || true; }
+          trap cleanup EXIT
+
+          KEY="$(head -c 32 /dev/urandom | base64 | tr -d '\n')"
+          CID="$(docker run -d \
+            --tmpfs /data:rw,size=64m \
+            -e CLOUD_DATA_DIR=/data -e CLOUD_ENV=smoke -e CLOUD_KMS_MASTER_KEY_REF="$KEY" \
+            "$IMAGE")"
+
+          # Wait for the HTTP listener to bind (or the process to die).
+          up=0
+          for _ in $(seq 1 60); do
+            lg="$(docker logs "$CID" 2>&1 || true)"
+            printf '%s' "$lg" | grep -q '"message":"listening"' && { up=1; break; }
+            [ "$(docker inspect -f '{{.State.Running}}' "$CID" 2>/dev/null || echo false)" != "true" ] && break
+            sleep 1
+          done
+          if [ "$up" != 1 ]; then
+            echo "::group::boot logs"; docker logs "$CID" 2>&1 || true; echo "::endgroup::"
+            echo "FUNCTIONAL SMOKE INFRA FAIL: image never reached \"listening\""
+            exit 1
+          fi
+
+          # Token: prefer the injected secret; else mint from KMS (a provisioned smoke
+          # identity); else run the anonymous matrix. Never hardcoded.
+          if [ -z "${SMOKE_TOKEN:-}" ] && [ -n "${KMS_CLIENT_ID:-}" ]; then
+            KT=$(curl -sf "$KMS_ENDPOINT/v1/kms/auth/login" -H 'Content-Type: application/json' \
+              -d "{\"clientId\":\"$KMS_CLIENT_ID\",\"clientSecret\":\"$KMS_CLIENT_SECRET\"}" | jq -r '.accessToken // empty' || true)
+            [ -n "$KT" ] && SMOKE_TOKEN=$(curl -sf "$KMS_ENDPOINT/v1/kms/orgs/hanzo/secrets/smoke/TOKEN?env=prod" \
+              -H "Authorization: Bearer $KT" | jq -r '.secret.value // empty' || true)
+          fi
+          if [ -n "${SMOKE_TOKEN:-}" ]; then echo "::add-mask::$SMOKE_TOKEN"; echo "smoke: AUTHENTICATED matrix"; else echo "smoke: ANONYMOUS matrix (no SMOKE_TOKEN wired)"; fi
+
+          # /smoke is baked into the image (Dockerfile) — exec it INSIDE the container,
+          # so it probes the real mounted surface at localhost:8080 with no port/network
+          # plumbing. A non-zero exit here fails the release BEFORE any image is pushed.
+          docker exec \
+            -e SMOKE_BASE_URL=http://127.0.0.1:8080 \
+            -e SMOKE_TOKEN="${SMOKE_TOKEN:-}" \
+            "$CID" /smoke
+
       # ── Migration smoke — the gate the v1.800.1 crashloop would have tripped ─────
       # The plain smoke above boots on a FRESH /data, so every subsystem's migrate()
       # takes its CREATE-TABLE path and no forward-migration is exercised — which is
@@ -401,10 +490,12 @@ jobs:
             ghcr.io/hanzoai/cloud:sha-${{ steps.ver.outputs.sha_short }}
             ghcr.io/hanzoai/cloud:latest
           labels: ${{ steps.meta.outputs.labels }}
-          # SAME cachebust as the smoke build → every layer is a cache hit from step 1
-          # and the pushed image is byte-identical to the one the smoke test proved.
+          # SAME artifact digests as the smoke build → every layer is a cache hit from
+          # step 1 and the pushed image is byte-identical to the one smoke proved.
           build-args: |
-            CONSOLE_CACHEBUST=${{ steps.ver.outputs.cachebust }}
+            CONSOLE_IMAGE=${{ steps.artifacts.outputs.console_image }}
+            SKILLS_IMAGE=${{ steps.artifacts.outputs.skills_image }}
+            FLAGS_IMAGE=${{ steps.artifacts.outputs.flags_image }}
           secrets: |
             GIT_AUTH_TOKEN=${{ secrets.GH_PAT }}
 
@@ -438,12 +529,25 @@ jobs:
             # newest-first and version tags are monotonic, so the highest version
             # is always among the most-recent versions; paginating the WHOLE
             # registry history is what livelocked this step as tags accumulated.
+            # Fail-CLOSED. An ORPHANED container tag — image pushed by a run that
+            # died or was cancelled after imagetools-create but before its git tag —
+            # MUST raise the floor, or a later run reassigns that same number to a
+            # different image (an ambiguous mutable prod tag; the v1.801.50 flip). A
+            # git-only floor can't see the orphan, so if the container-tag lookup
+            # ERRORS (vs legitimately returning no tags) we retry the whole attempt
+            # rather than silently proceeding — a version with a pushed image is never
+            # reused. (Reordering git-tag before imagetools-create is the WRONG fix: it
+            # reintroduces the phantom "tag ⇔ no image" this workflow exists to prevent.)
             cont_max=""
             if command -v gh >/dev/null 2>&1; then
-              cont_max="$(GH_TOKEN="$GH_PAT" gh api \
-                '/orgs/hanzoai/packages/container/cloud/versions?per_page=100' \
-                --jq '.[].metadata.container.tags[]?' 2>/dev/null \
-                | sed 's/^v//' | grep -E '^[0-9]+\.[0-9]+\.[0-9]+$' | sort -V | tail -1 || true)"
+              if cont_raw="$(GH_TOKEN="$GH_PAT" gh api \
+                  '/orgs/hanzoai/packages/container/cloud/versions?per_page=100' \
+                  --jq '.[].metadata.container.tags[]?' 2>/dev/null)"; then
+                cont_max="$(printf '%s\n' "$cont_raw" \
+                  | sed 's/^v//' | grep -E '^[0-9]+\.[0-9]+\.[0-9]+$' | sort -V | tail -1 || true)"
+              else
+                echo "  container-tag lookup failed — retry so an orphaned tag can't be reused (attempt $attempt)"; sleep 3; continue
+              fi
             fi
             max="$(printf '%s\n%s\n%s\n' "1.786.0" "$git_max" "$cont_max" \
               | grep -E '^[0-9]+\.[0-9]+\.[0-9]+$' | sort -V | tail -1)"
@@ -473,8 +577,12 @@ jobs:
                 export PATH="$HOME/.local/bin:$PATH"
               }
               for MT in "${V}" "${VER}" "${major}.${minor}"; do
-                crane copy "$SHA_IMG" "registry.hanzo.ai/hanzoai/cloud:${MT}" \
-                  || echo "::warning::mirror registry.hanzo.ai/hanzoai/cloud:${MT} failed"
+                # Bounded: registry.hanzo.ai can *hang* (not just fail), and this
+                # is best-effort — an unbounded crane copy once livelocked the whole
+                # tag step and held the serialized release lane. timeout makes the
+                # mirror truly best-effort so the git-tag receipt below always runs.
+                timeout 120 crane copy "$SHA_IMG" "registry.hanzo.ai/hanzoai/cloud:${MT}" \
+                  || echo "::warning::mirror registry.hanzo.ai/hanzoai/cloud:${MT} failed or timed out"
               done
             fi
             git tag -a "$V" -m "release $V — image ghcr.io/hanzoai/cloud:$V (retagged from sha-${{ steps.ver.outputs.sha_short }}, smoke-passed ${GITHUB_SHA})"
@@ -491,25 +599,86 @@ jobs:
           echo "::error::could not acquire a free version tag after 8 attempts"
           exit 1
 
-  # Notify universe so the GitOps pipeline rolls the new image to prod — same
-  # image-update contract every service uses (gateway, iam, …). Runs ONLY after
-  # build-amd64 succeeds, i.e. only for a version whose image is proven pushed and
-  # tagged. A failed release never reaches here, so universe is never asked to
-  # deploy a phantom tag.
-  notify-universe:
+  # ── Promote: the declared-tag bump that makes the release DEPLOY ─────────────
+  # The tag minted above is the receipt for a pushed, smoke-passed image; THIS job
+  # records it as the desired state Hanzo CD reconciles. The universe-crs ArgoCD
+  # Application (ns hanzo-cd, `automated` sync + selfHeal) syncs
+  # infra/k8s/operator/crs/*.yaml → cluster and the operator rolls the Deployment,
+  # so a tag bump committed here reaches api.hanzo.ai with NO hand-dispatch and NO
+  # hand-edit of the CR.
+  #
+  # This is the SAME yq-bump → `deploy(<svc>): <tag>` universe commit the hanzoai/ci
+  # reusable (build.yml deploy step) does for every other service. cloud owns it
+  # HERE because its image is built by this workflow, not the ci reusable — its
+  # hanzo.yml carries no main `images:` entry and `# NO deploy`, so the shared
+  # deploy step never bumps cloud's CR. A direct in-cluster CR patch is NOT enough:
+  # ArgoCD selfHeal reverts any live edit not also recorded in git within ~45s.
+  # The retired notify-universe repository_dispatch had no receiver after the
+  # image-update.yml deploy hub was deleted in the Hanzo CD cutover; the git commit
+  # IS the sanctioned path now.
+  promote:
     needs: build-amd64
+    # Only a real release promotes: build+smoke+push+tag all succeeded, so a
+    # proven v* image exists. A failure earlier leaves version_v empty → skipped.
+    if: ${{ needs.build-amd64.outputs.version_v != '' }}
     runs-on: [hanzo-build-linux-amd64]
     steps:
-      - name: Repository dispatch (image-update)
-        uses: peter-evans/repository-dispatch@v3
-        with:
-          token: ${{ secrets.UNIVERSE_DISPATCH_TOKEN }}
-          repository: hanzoai/universe
-          event-type: image-update
-          client-payload: |
-            {
-              "service": "cloud",
-              "image":   "ghcr.io/hanzoai/cloud:${{ needs.build-amd64.outputs.version_v }}",
-              "sha":     "${{ github.sha }}",
-              "env":     "all"
-            }
+      - name: Record the proven tag in universe crs/cloud.yaml (Hanzo CD rolls it)
+        env:
+          # GH_PAT already pushes this repo's git tags above (contents:write on the
+          # hanzoai org), so it writes hanzoai/universe too — the SAME token the ci
+          # reusable falls back to for the universe deploy commit.
+          GH_PAT: ${{ secrets.GH_PAT }}
+          VERSION_V: ${{ needs.build-amd64.outputs.version_v }}
+        run: |
+          set -euo pipefail
+          [ -n "${GH_PAT:-}" ] || { echo "::error::no GH_PAT — cannot record the declared-tag bump in universe"; exit 1; }
+
+          # Bare arc runners ship no yq — provision the static binary (sudo-free,
+          # same pattern the ci reusable and this workflow's kubectl/crane fetches use).
+          if ! command -v yq >/dev/null 2>&1; then
+            mkdir -p "$HOME/.local/bin"; export PATH="$HOME/.local/bin:$PATH"
+            curl -fsSL https://github.com/mikefarah/yq/releases/latest/download/yq_linux_amd64 \
+              -o "$HOME/.local/bin/yq" && chmod +x "$HOME/.local/bin/yq"
+          fi
+
+          git clone -q --depth 1 \
+            "https://x-access-token:${GH_PAT}@github.com/hanzoai/universe.git" \
+            "$RUNNER_TEMP/universe"
+          CR="$RUNNER_TEMP/universe/infra/k8s/operator/crs/cloud.yaml"
+          [ -f "$CR" ] || { echo "::error::crs/cloud.yaml not found in universe"; exit 1; }
+
+          CUR="$(yq -r '.spec.image.tag // ""' "$CR")"
+          echo "cloud CR: ${CUR:-<empty>} → ${VERSION_V}"
+
+          # Monotonic guard: never roll the CR BACKWARD. Release runs finish under a
+          # serialized lane but a slow older run must never overwrite a newer promote.
+          # Skip iff the CR already holds a semver >= the version we just cut.
+          CURN="${CUR#v}"; NEWN="${VERSION_V#v}"
+          if printf '%s' "$CURN" | grep -qE '^[0-9]+\.[0-9]+\.[0-9]+$'; then
+            top="$(printf '%s\n%s\n' "$CURN" "$NEWN" | sort -V | tail -1)"
+            if [ "$top" = "$CURN" ] && [ "$CURN" != "$NEWN" ]; then
+              echo "::notice::cloud CR already at v${CURN} (≥ ${VERSION_V}) — not rolling back"; exit 0
+            fi
+          fi
+
+          yq -i ".spec.image.tag = \"${VERSION_V}\"" "$CR"
+          if git -C "$RUNNER_TEMP/universe" diff --quiet; then
+            echo "::notice::crs/cloud.yaml already at ${VERSION_V} — nothing to record"; exit 0
+          fi
+          git -C "$RUNNER_TEMP/universe" -c user.name=hanzo-ci -c user.email=dev@hanzo.ai \
+            commit -qam "deploy(cloud): ${VERSION_V} (${GITHUB_REPOSITORY}@$(echo "${GITHUB_SHA}" | cut -c1-7))"
+
+          # Rebase-safe push: universe main advances on every service's deploy, so a
+          # concurrent commit must not make cloud's promote lose the whole roll. Retry
+          # a few times, rebasing between attempts.
+          for attempt in $(seq 1 5); do
+            if git -C "$RUNNER_TEMP/universe" push -q origin HEAD:main; then
+              echo "recorded deploy(cloud): ${VERSION_V} — Hanzo CD (universe-crs) will roll it to api.hanzo.ai"
+              exit 0
+            fi
+            echo "  universe push lost the race — rebasing (attempt ${attempt})"
+            git -C "$RUNNER_TEMP/universe" pull -q --rebase origin main || true
+            sleep 3
+          done
+          echo "::error::could not record the cloud tag bump in universe after 5 attempts"; exit 1

Dockerfile

@@ -2,133 +2,50 @@
 #
 # This image is a SINGLE artifact that serves BOTH the /v1 API AND the console
 # UI from one process: the console is compiled into the Go binary via
-# //go:embed (see webui.go). The pipeline is:
+# //go:embed (see webui.go). The final `/cloud` binary already carries the UI —
+# no separate console Service, no second origin; the embedded console calls /v1
+# on its own host.
 #
-#   1. console  stage → build the hanzoai/console static bundle
-#   2. (copied) → into  webui/dist/  of the Go build context
-#   3. build    stage → `go build` bakes webui/dist into the binary (go:embed)
-#
-# so the final `/cloud` binary already carries the UI. No separate console
-# Service, no second origin — the embedded console calls /v1 on its own host.
-#
-# ── console UI stage ─────────────────────────────────────────────────────────
-# Builds the console SPA and emits a STATIC bundle at /out. console is fetched
-# at a pinned ref (CONSOLE_REF) using the same GIT_AUTH_TOKEN BuildKit secret the Go
-# build uses for private modules.
-#
-# console exposes `npm run build:embed` (scripts/build-embed.mjs): it prunes the
-# Next server route handlers (BFF proxies — they collapse to the cloud /v1/* the
-# SPA calls same-origin), wraps the client catch-all pages for output:'export',
-# and neutralizes the root layout's request-time headers() read (the per-host
-# <title>, resolved client-side in the embed) so the STATIC export prerenders
-# clean — emitting out/. This stage runs it and copies out/ into /out, which the
-# Go build drops into webui/dist so //go:embed bakes the FULL @hanzo/gui console
-# into the ONE binary. This stage FAILS HARD: the prod image MUST carry the real
-# console — a missing/broken build:embed is a build ERROR, never a silent degrade
-# to the placeholder shell. The one escape hatch is --build-arg ALLOW_PLACEHOLDER=1
-# (pure-Go dev image with no Node console), which is NEVER set for prod.
-FROM public.ecr.aws/docker/library/node:24-alpine@sha256:a0b9bf06e4e6193cf7a0f58816cc935ff8c2a908f81e6f1a95432d679c54fbfd AS console
-ARG CONSOLE_REPO=https://github.com/hanzoai/console.git
-ARG CONSOLE_REF=main
-# CONSOLE_CACHEBUST busts this stage's BuildKit layer cache every build. WHY it must
-# exist: the clone+build layer's cache key is derived from the RUN text + build args.
-# With only a static `git clone --branch main`, the key NEVER changes, so on the
-# persistent ARC dind BuildKit cache every cloud image re-embedded the SAME frozen
-# console snapshot — new console work (the native Tracker, …) silently never shipped,
-# even on a freshly-built+deployed image. release.yml feeds this the cloud commit sha
-# (unique per push) so the clone RUN re-runs each build and re-fetches console
-# ${CONSOLE_REF} (main HEAD) fresh. Correctness over cache reuse: the console stage
-# rebuilds every time, but the embed is never stale.
-ARG CONSOLE_CACHEBUST=none
-RUN apk add --no-cache git
-WORKDIR /console
-# The static export prerenders every page (webpack compile + export prerender);
-# give the heap headroom so a large @hanzo/gui build never OOMs into the stub.
-ENV NEXT_TELEMETRY_DISABLED=1 NODE_OPTIONS=--max-old-space-size=8192
-# Hanzo Analytics: the console's <HanzoAnalytics/> (env-gated) renders the one
-# native analytics.hanzo.ai tag only when a website-id is baked in. Default to the
-# console.hanzo.ai property (7dce54ee, public per-site) so console+team track on
-# the next cloud build. GA4/Pixel stay off (unset). Public id, not a KMS secret.
-ARG NEXT_PUBLIC_ANALYTICS_WEBSITE_ID=7dce54ee-41f6-4751-96bf-fe005067c7c7
-ENV NEXT_PUBLIC_ANALYTICS_WEBSITE_ID=$NEXT_PUBLIC_ANALYTICS_WEBSITE_ID
-RUN --mount=type=secret,id=GIT_AUTH_TOKEN \
-    if [ -s /run/secrets/GIT_AUTH_TOKEN ]; then \
-      git config --global url."https://x-access-token:$(cat /run/secrets/GIT_AUTH_TOKEN)@github.com/".insteadOf "https://github.com/"; \
-    fi && \
-    echo ">> embedding console ${CONSOLE_REF} (cachebust ${CONSOLE_CACHEBUST})" && \
-    git clone --depth 1 --branch "${CONSOLE_REF}" "${CONSOLE_REPO}" . && \
-    echo ">> console @ $(git rev-parse HEAD)" && \
-    npm install --no-audit --no-fund --fetch-retries=5 --fetch-retry-mintimeout=20000 --fetch-timeout=120000
-# FAIL-HARD. build:embed MUST emit a REAL bundle — a non-empty out/index.html AND
-# an out/_next/ chunk dir — and /out then carries it into the Go embed path. If the
-# target is absent, the export fails, or the output is the placeholder shape, this
-# is a build ERROR (exit 1): the prod image can NEVER silently ship the committed
-# fallback shell. Escape hatch: --build-arg ALLOW_PLACEHOLDER=1 leaves /out empty
-# (Go build keeps the committed shell) for a pure-Go dev image — NEVER set in prod.
-ARG ALLOW_PLACEHOLDER=0
-RUN mkdir -p /out; \
-    ok=0; \
-    if npm run 2>/dev/null | grep -q ' build:embed'; then \
-      if npm run build:embed && [ -s out/index.html ] && [ -d out/_next ]; then \
-        cp -r out/. /out/; \
-        echo ">> embedded REAL console static bundle: $(wc -c < out/index.html)-byte index.html, $(du -sh out/_next | cut -f1) _next/"; \
-        ok=1; \
-      else \
-        echo ">> console build:embed produced NO real bundle (missing/empty out/index.html or out/_next)"; \
-      fi; \
-    else \
-      echo ">> console exposes no build:embed target"; \
-    fi; \
-    if [ "$ok" != "1" ]; then \
-      if [ "$ALLOW_PLACEHOLDER" = "1" ]; then \
-        echo ">> ALLOW_PLACEHOLDER=1 — keeping committed fallback shell (DEV image only; NEVER prod)"; \
-      else \
-        echo ">> FATAL: refusing to ship the placeholder console. Fix the console build:embed, or pass --build-arg ALLOW_PLACEHOLDER=1 for a pure-Go dev image."; \
-        exit 1; \
-      fi; \
-    fi
+# ── prebuilt decomplection artifacts (cloud compiles ONLY Go) ────────────────
+# The console SPA, the agent-skills catalog, and the native flags staticlib are
+# each built by THEIR OWN CI as a versioned immutable image and PULLED here,
+# instead of rebuilding node + python + rust from scratch every cloud release.
+# The heavy one (console: a cold `npm install` + full Next.js static export,
+# force-cache-busted every build) used to dominate the ~20-min build; it is now
+# a registry pull.
+#   console-embed (hanzoai/console Dockerfile.embed)  → /dist             → webui/dist                  (go:embed)
+#   agent-skills  (hanzoai/openapi Dockerfile.skills) → /catalog          → clients/agentskills/catalog (go:embed)
+#   cloud-flags   (native/flags    Dockerfile)        → /libhanzo_flags.a → CGO link (clients/featureflags)
+# Pinned to ghcr.io so BOTH buildx lanes (release.yml + platform arcbuild) pull
+# it directly; the SAME tags are mirrored to registry.hanzo.ai (S3-backed) for
+# GET-flow consumers (docker/kaniko/crane). Override any pin with
+# --build-arg <NAME>_IMAGE=…  — release.yml resolves CONSOLE_IMAGE to a fresh
+# console-embed digest, exactly as CONSOLE_CACHEBUST re-fetched console before.
+ARG CONSOLE_IMAGE=ghcr.io/hanzoai/console-embed:latest
+ARG SKILLS_IMAGE=ghcr.io/hanzoai/agent-skills:latest
+ARG FLAGS_IMAGE=ghcr.io/hanzoai/cloud-flags:latest
 
-# ── Go build stage (CGO=1 + SQLCipher — REAL at-rest encryption) ─────────────
-# The unified binary embeds IAM (clients/iam) whose per-org store is SQLCipher-
-# encrypted (orgIsolation=sqlite), and commerce's per-tenant money DBs likewise.
-# A CGO=0 modernc build SILENTLY SHIPS PLAINTEXT. So this builds CGO=1 against
-# system libsqlcipher — hanzoai/iam's proven recipe: the `libsqlite3` tag + a
-# libsqlcipher symlink + -DSQLITE_HAS_CODEC, with the modernc double-registration
-# guard, TestEncryptionProof, and the cek.go golden-vector KAT baked in — so a
-# build that fails to link REAL SQLCipher, or that would decrypt existing stores
-# differently, produces NO image. alpine3.22 MATCHES the runtime base so the
-# libsqlcipher soname the binary links is the SAME one present at runtime. ECR
-# Public mirror avoids Docker Hub's 429 rate-limit on shared CI runners.
-# ---- agent-skills stage: regenerate the FULL /.well-known/agent-skills catalog
-# from the hanzoai/openapi SOT (skills.py) and carry it into the Go embed path
-# BEFORE `go build`, the SAME way the console bundle is produced. The committed
-# catalog is only the tiny `ai` fallback; prod must embed the full set. FAIL-HARD:
-# if the clone/generation can't produce the master index, the image is not built.
-FROM public.ecr.aws/docker/library/python:3.12-alpine AS skills
-ARG OPENAPI_REPO=https://github.com/hanzoai/openapi.git
-ARG OPENAPI_REF=main
-RUN apk add --no-cache git && pip install --no-cache-dir pyyaml
-WORKDIR /openapi
-RUN --mount=type=secret,id=GIT_AUTH_TOKEN \
-    if [ -s /run/secrets/GIT_AUTH_TOKEN ]; then \
-      git config --global url."https://x-access-token:$(cat /run/secrets/GIT_AUTH_TOKEN)@github.com/".insteadOf "https://github.com/"; \
-    fi && \
-    git clone --depth 1 --branch "${OPENAPI_REF}" "${OPENAPI_REPO}" . && \
-    python3 skills.py --no-services --out /catalog && \
-    test -s /catalog/hanzo/index.json
+# ── toolchain base images: the golang + alpine FROMs below pull from our own
+# GHCR mirror (ghcr.io/hanzoai/mirror/*), pinned by digest. WHY: public.ecr.aws
+# rate-limits anonymous pulls (HTTP 429) on shared CI runners and a 429 on ANY
+# base pull aborts the release. The mirror packages are 1:1 amd64 copies of the
+# upstream public images, digest-pinned for immutability; release.yml logs the
+# build into ghcr.io (GH_PAT) before building so they resolve. REFRESH on a
+# toolchain bump: crane/regctl copy the new upstream into
+# ghcr.io/hanzoai/mirror/<name>:<tag> and repoint the digest below. Canonical
+# long-term home is registry.hanzo.ai/hanzoai/mirror/* — repoint once the runners
+# carry its IAM pull credentials (follow-up).
 
-# ── Native flags evaluator — hanzo-flags (Rust staticlib, FFI'd into the Go
-# binary by clients/featureflags). Stateless PostHog-compatible evaluation;
-# definitions live in the per-org SQLite stores. musl staticlib links clean
-# against the alpine cgo build below.
-FROM public.ecr.aws/docker/library/rust:1-alpine3.22 AS flagslib
-RUN apk add --no-cache musl-dev
-WORKDIR /src/native/flags
-COPY native/flags/Cargo.toml native/flags/Cargo.lock ./
-COPY native/flags/src ./src
-RUN cargo build --release --locked
+# ── console SPA static export (prebuilt → /dist) ─────────────────────────────
+FROM ${CONSOLE_IMAGE} AS console
 
-FROM public.ecr.aws/docker/library/golang:1.26-alpine3.22@sha256:727cfc3c40be55cd1bc9a4a059406b28a059857e3be752aa9d09531e12c20c56 AS build
+# ── agent-skills catalog (prebuilt → /catalog) ──────────────────────────────
+FROM ${SKILLS_IMAGE} AS skills
+
+# ── native flags evaluator staticlib (prebuilt → /libhanzo_flags.a) ──────────
+FROM ${FLAGS_IMAGE} AS flagslib
+
+FROM ghcr.io/hanzoai/mirror/golang:1.26-alpine3.22@sha256:47d47cb5cc3c7dac409dcb6c3a98a6263571218046cd02d709527feef804a77c AS build
 # CIPHER-FORMAT FREEZE (cek depends on this). The data-plane stores are
 # SQLCipher pages in a fixed on-disk format (cipher_compatibility 4). An at-open
 # compat pin is infeasible (mattn keys via URI before any pragma), so the format
@@ -149,18 +66,21 @@ RUN set -eux; \
     ln -sf "$SC" /usr/lib/libsqlite3.so; \
     ln -sf "$SC" /usr/lib/libsqlite3.so.0
 WORKDIR /src
-# hanzoai/* and luxfi/* are PUBLIC and resolve via the IMMUTABLE public proxy +
-# sumdb — go.sum pins those canonical hashes, so a force-re-pointed tag can never
-# break the build. GOSUMDB stays ON (a money image must not blanket-disable the
-# checksum database); only zap-proto/* is exempt (first-party-direct via GOPRIVATE,
-# authenticated git over gh_token). -mod=readonly means the committed go.sum is the
-# SOLE source of truth: any drift (a needed hash not present) FAILS the build
-# instead of being silently re-recorded. CGO_CFLAGS/LDFLAGS enable the SQLCipher
-# codec + URI keying.
+# zap-proto/* (all 55 repos) and luxfi/* (all 37 deps here) are PUBLIC and resolve
+# via the IMMUTABLE public proxy + sumdb — go.sum pins those canonical hashes, so a
+# force-re-pointed tag can never break the build. GOSUMDB stays ON (a money image
+# must not blanket-disable the checksum database); github.com/hanzoai/* is the
+# exempt namespace — ai, account, commerce, orm, xorm, beego, csqlite and ~30 more
+# are PRIVATE repos, so they resolve direct+authenticated (git over gh_token) and
+# skip a sumdb that cannot see them. GOPRIVATE named zap-proto until now, which is
+# public and was never the reason anything was direct; the private namespace it
+# stood for went unnamed and worked only on the GOPROXY `direct` fallback.
+# -mod=readonly means the committed go.sum is the SOLE source of truth: any drift
+# (a needed hash not present) FAILS the build instead of being silently
+# re-recorded. CGO_CFLAGS/LDFLAGS enable the SQLCipher codec + URI keying.
 ENV CGO_CFLAGS="-DSQLITE_HAS_CODEC -DSQLITE_USE_URI=1 -I/usr/include/sqlcipher" \
     CGO_LDFLAGS="-lsqlcipher" \
-    GOPRIVATE=github.com/zap-proto/* \
-    GONOSUMDB=github.com/zap-proto/* \
+    GOPRIVATE=github.com/hanzoai/* \
     GOPROXY=https://proxy.golang.org,direct \
     GOFLAGS=-mod=readonly
 COPY go.mod go.sum ./
@@ -180,13 +100,13 @@ RUN --mount=type=secret,id=GIT_AUTH_TOKEN \
 COPY . .
 # Drop the console static bundle into the embed path BEFORE `go build`, so
 # //go:embed all:webui/dist bakes it into the binary (same-origin console).
-COPY --from=console /out/ /src/webui/dist/
+COPY --from=console /dist/ /src/webui/dist/
 # Overlay the FULL agent-skills catalog before `go build` so //go:embed all:catalog
 # bakes the complete set (all services × brands), not the committed `ai` fallback.
 COPY --from=skills /catalog/ /src/clients/agentskills/catalog/
 # The native flags staticlib at the exact ${SRCDIR}-relative path the cgo
 # directive in clients/featureflags/engine.go links.
-COPY --from=flagslib /src/native/flags/target/release/libhanzo_flags.a /src/native/flags/target/release/libhanzo_flags.a
+COPY --from=flagslib /libhanzo_flags.a /src/native/flags/target/release/libhanzo_flags.a
 # RED gate — modernc double-registration guard: 0 modernc under CGO=1, else the
 # "sqlite" driver is registered twice (mattn + modernc) → panic at init.
 RUN --mount=type=cache,id=cloud-gomod-v4,target=/go/pkg/mod,sharing=locked \
@@ -215,12 +135,18 @@ RUN --mount=type=cache,id=cloud-gomod-v4,target=/go/pkg/mod,sharing=locked \
 RUN --mount=type=cache,id=cloud-gomod-v4,target=/go/pkg/mod,sharing=locked \
     --mount=type=cache,id=cloud-gobuild-v4,target=/root/.cache/go-build,sharing=locked \
     CGO_ENABLED=1 go build -tags "libsqlite3 sqlite_fts5" -ldflags="-s -w" -o /cloud ./cmd/cloud
+# The functional smoke prober (cmd/smoke) — a stdlib-only, static binary shipped
+# alongside /cloud so the release gate can `docker exec` it against the freshly-built
+# image (and any deployment can be smoked via `docker run --entrypoint /smoke ...`).
+RUN --mount=type=cache,id=cloud-gomod-v4,target=/go/pkg/mod,sharing=locked \
+    --mount=type=cache,id=cloud-gobuild-v4,target=/root/.cache/go-build,sharing=locked \
+    CGO_ENABLED=0 go build -ldflags="-s -w" -o /smoke ./cmd/smoke
 # Prove the SHIPPED binary binds sqlite3_* to libsqlcipher, not a plaintext libsqlite3.
 RUN readelf -d /cloud | grep -qE 'NEEDED.*(sqlcipher|sqlite3)' || { echo "FATAL: /cloud links no sqlite/sqlcipher .so"; exit 1; }; \
     ! ldd /cloud 2>/dev/null | grep -E 'libsqlite3' | grep -vq 'libsqlcipher' || { echo "FATAL: /cloud resolves a NON-sqlcipher libsqlite3 (plaintext risk)"; exit 1; }
 
 # ── final image (alpine, NOT scratch — CGO needs libc + libsqlcipher) ─────────
-FROM public.ecr.aws/docker/library/alpine:3.22@sha256:14358309a308569c32bdc37e2e0e9694be33a9d99e68afb0f5ff33cc1f695dce
+FROM ghcr.io/hanzoai/mirror/alpine:3.22@sha256:7c8cb692ae09657cbc4a3f3cbd0e8d5a2690ba38386aaaf252dbb060bf5eb2e6
 ARG REVISION=unknown
 LABEL org.opencontainers.image.revision="${REVISION}" \
       org.opencontainers.image.source="https://github.com/hanzoai/cloud"
@@ -246,6 +172,7 @@ COPY --from=build /usr/share/zoneinfo /usr/share/zoneinfo
 COPY --from=build /etc/passwd /etc/passwd
 COPY --from=build /etc/group /etc/group
 COPY --from=build /cloud /cloud
+COPY --from=build /smoke /smoke
 EXPOSE 8080 9090 9653
 USER 65532:65532
 ENTRYPOINT ["/cloud"]

LLM.md

@@ -6,31 +6,78 @@ artifact serves `api.hanzo.ai`, `api.lux.cloud`, `api.zoo.cloud`, `api.osage.clo
 and every white-label reseller. Brand, enabled subsystems, and org scope are
 deployment configuration.
 
+## Open Cloud planes
+
+Spec home: HIP-0129 `hip-0129-open-cloud-planes` (hips repo). This section is a
+map, not the spec. One noun, one owner, one route family. No plane reads another
+plane's store; imports flow custody-ward only (channels -> integrations, never
+reverse).
+
+| Route | Noun | Owner | Tier |
+| --- | --- | --- | --- |
+| `/v1/connectors` | Custody: per-user BYO external accounts | `clients/integrations` (extends; user scope new) | In flight (branch `feat/connectors`) |
+| `/v1/channels` | Transport: portable message envelope, DM pairing, send + inbox | `clients/channels` (new) | Planned (branch `feat/channels` reserved; no transport code yet) |
+| `/v1/sync` | Data: bidirectional sync engine | `clients/sync` | Shipped |
+| `/v1/automations` | Workflows: flows/runs, goja piece runtime | `clients/automations` | Shipped |
+| `/v1/compute/bots` | Hosting: `@hanzo/bot` Node containers | `clients/bots` | Shipped |
+| `/v1/tasks` | Durable engine | `clients/tasks` | Shipped |
+| `/v1/gpus` + fleet | BYO GPU presence | `clients/fleet` + `clients/visor` | Shipped |
+| IAM | Identity: users, orgs, roles | IAM | Shipped |
+| KMS | Secret custody: sealed secrets | `clients/kms` | Shipped |
+
+Custody invariants: secrets sealed in KMS at
+`/orgs/{org}/users/{user}/connectors/{provider}/{label}`, never in SQLite rows;
+verify before store. Refresh is single-flight with rotation resealing; the CLI
+does local browser PKCE and posts the bundle to
+`POST /v1/connectors/:provider/credential`; cloud owns device-code flows.
+
+Transport invariants: typed actions (`command|url|select|approval`), no raw
+string sniffing; pairing codes 8 chars, 1h TTL, max 3 pending per account,
+owner bootstrap on first approval.
+
+Container boundary is permanent for native-module, host-filesystem, loop-state,
+and vendor-Node work (agent loop, exec/PTY, harnesses, browser, voice, codecs,
+Node-bound channels, plugin SDK/loader). The Node plugin SDK is never ported to
+Go; cloud extensibility is connectors/automations/tools.
+
+Port roadmap (P1-P15) lives in HIP-0129; do not restate it here. Every claim
+carries its tier: Shipped (on main, named package/route), In flight (named
+pre-main branch), Planned (backlog id or named reservation).
+
 ## Framework doctrine
 
 One way to do everything. Composable, orthogonal, DRY. A new subsystem is a
 package under `clients/<name>` that obeys these seams — nothing more.
 
 - **Subsystem shape.** A subsystem exposes `func Mount(app *zip.App, deps cloud.Deps) error`
-  and self-registers at init with `cloud.Register("<name>", <order>, cloud.Typed(Mount))`
-  (or `RegisterWithShutdown`). `Mount` wires that subsystem's `/v1/<name>/*` routes
-  onto the shared `*zip.App`; `cloud.Deps` carries the process-wide handles
-  (Logger, DataDir, the subsystem `Client` seams). No subsystem reaches into
-  another's internals.
+  and is listed in `apps.Wire()` as a `cloud.MountSpec{Name, Mount: cloud.Typed(Mount)}`
+  (plus `Shutdown`/`OwnsHealth` where it owns them). `Mount` wires that subsystem's
+  `/v1/<name>/*` routes onto the shared `*zip.App`; `cloud.Deps` carries the
+  process-wide handles (Logger, DataDir, the subsystem `Client` seams). No
+  subsystem reaches into another's internals. There is no init()-registry and no
+  `cloud.Register` — subsystems do NOT self-register.
 - **Client seams.** Cross-subsystem calls go through a narrow in-process interface
   published in `types` and aliased at the provider, e.g. `commerce.Client =
   types.CommerceClient` (`GetOrgConfig` + `CheckEntitlement`). Consumers depend on
   the interface, never the implementation; the seam rides zap-proto/zip. Keep each
   interface minimal — add a method only when a consumer needs it.
-- **Composition root.** `apps/apps.go` blank-imports every subsystem
-  (its init runs `cloud.Register`), populating `cloud.Registry`. `MountAll`
-  (build.go) sorts the registry by `Order` and calls `Mount` on each ENABLED
-  subsystem (`cfg.Enabled`). That ordered blank-import set IS the wiring — there
-  is no separate `Wire()` function; to add a subsystem you add one import line.
-- **Route precedence is a framework guarantee.** The router is zap-proto/fiber
-  (zip v1.3.0). Most-specific route wins regardless of mount order; a genuine
-  route CONFLICT panics at mount rather than resolving ambiguously. Subsystems may
-  therefore mount in any order and still compose deterministically.
+- **Composition root.** `apps/apps.go:Wire()` returns `[]cloud.MountSpec` — every
+  linked subsystem, in mount order, as ONE explicit slice read top-to-bottom.
+  Slice position IS the order: there is no `Order` field and `MountAll`
+  (build.go) does NOT sort; it iterates as-given and mounts each ENABLED spec
+  (`cfg.Enabled`). To add a subsystem you add one line to `Wire()`.
+  `apps/wire_test.go` freezes the sequence, so a reorder/drop/add fails there.
+- **Route precedence.** The router is zap-proto/fiber (zip v1.8.3). Most-specific
+  route wins regardless of mount order, so subsystems may mount in any order and
+  still compose deterministically. But precedence is NOT a conflict guard: two
+  registrations of a byte-identical pattern do NOT panic — fiber MERGES them into
+  ONE route with both handlers chained, resolving by first-registration. That is
+  invisible to a `GetRoutes()` entry count (see the bots note below), and it is
+  NOT distinguishable from a legitimate middleware chain: `app.Post(path, mw1,
+  mw2, mw3, handler)` is one registration with four handlers (apps/commerce.go:151),
+  and the whole `/v1/store/*` surface is that shape. A high handler count is
+  therefore evidence of nothing on its own; only a subsystem that never chains
+  middleware (bots/visor/runtime) can read `len(Handlers) > 1` as a collision.
 - **Per-org data.** The ONE way any subsystem opens a per-org SQLite file is
   `cloud.OrgDB(dataDir, org, project, sub)` — or the cached `cloud.OrgStore[T]`
   (`NewOrgStore` + `For(org, project)`). Path convention:
@@ -42,6 +89,51 @@ package under `clients/<name>` that obeys these seams — nothing more.
   the SOLE driver (blank-imported once, in orgdb.go); subsystems never import a
   SQLite driver themselves. The caller owns its schema/migration and Close.
 
+## The route table has three projections, and the router is the source
+
+`serve.go` composes ONE route table and projects it three ways, all after
+`MountAll` so each sees a complete table: `/zap` REPLAYS the /v1 handlers
+(zapface), the console RENDERS them, and `GET /v1/openapi.json` DESCRIBES them
+(`openapi.Mount`). None holds a second copy of anything; none can drift.
+
+- **The spec IS the router.** `openapi.Live(app)` reads
+  `app.Fiber().GetRoutes(true)` — fiber's own filter drops `Use()` middleware —
+  and every other function in `openapi/` is a pure function of that `[]Route`.
+  There is NO checked-in spec file to hand-maintain and no second registry. The
+  drift guard is `cmd/cloud/openapi_test.go`: a BIJECTION over the fully-mounted
+  `apps.Wire()` (983 operations / 692 paths / 109 products) — every live route
+  appears as an operation, every operation is backed by a live route. It is the
+  only test whose failure means the document lies.
+- **Reading the LIVE router is the only total source.** `POST /v1/kms/auth/login`
+  is registered as `Group("/v1/kms/auth").Post("/login")` — no grep can find that
+  path; only the assembled router knows it. And the route set is a function of
+  deployment config (`cfg.Enabled`, plus internal gates like kms's `if kc != nil`),
+  so **the spec VARIES PER DEPLOYMENT** — correctly: a deployment that does not
+  mount admin does not advertise it. That is why the document is generated
+  per-process at request time, not built once in CI.
+- **The product axis is mechanical.** The first path segment after `/v1/` IS the
+  product (`openapi.Product`), tagged onto each operation so a CLI can build
+  `hanzo <product> <resource> <verb>` with no judgment. It is deliberately NOT the
+  subsystem name: `clients/billing` also serves `/v1/finance/*`.
+- **What the router CANNOT tell you — do not try to fix this in the generator.**
+  Method, path, path params, and product are derivable; request/response schemas,
+  query/header params, status codes, and auth are NOT. The router holds a
+  `func(*zip.Ctx) error`; the request type is a LOCAL inside the handler
+  (`var req secretPutRequest; json.Unmarshal(ctx.Body(), &req)`), and Go cannot
+  reflect from a func value into its body. `cloud.Handle[S]` does not help — `S`
+  is the SERVICE (service.go:90), not the payload; `cloud.Typed` is an
+  `any→*zip.App` mount adapter. The ONE path to schemas is zip's typed ops
+  (`zip.Get[In,Out]`), which carry the In/Out types and also yield an MCP tool
+  from the same registry (zip/openapi.go, zip/mcp.go — today `len(a.ops) == 0`,
+  so zip's own generator emits nothing here). `GetRoutes()` is a superset of
+  `app.ops`, so migrating a handler to a typed op adds schema without changing
+  this pipeline.
+- **Catch-alls are opaque, by construction.** `app.Post("/v1/billing/*")` proxies
+  to another service, so `POST /v1/billing/deposit` is NOT a route in this process
+  and cannot appear. Measured on the live table: 3 products are wholly opaque
+  (`bot`, `licensing`, `sentry` — the catch-all IS the product) and 12 more mix
+  concrete ops with a catch-all hiding an unknown remainder.
+
 ## Cross-subsystem seams that are values, not places
 
 - **The per-principal MCP plane is callable in-process.** `clients/automations`
@@ -130,3 +222,39 @@ Import path (already-incorporated orgs): Google Drive → data room, a Google Sh
 captable, via the `google` OAuth provider now completed in `clients/integrations`
 (token custodied in KMS; the automations `google` connector shares the same token).
 Runbook: `docs/company-dogfood.md`.
+
+## Deploy plane (`clients/deploy`, `/v1/deploy`)
+
+Native ArgoCD-grade GitOps console over the operator-managed fleet, parallel to
+`/v1/git`: each `hanzo.ai/v1` App CR IS the Application, and the plane OBSERVES the
+operator's reconcile — `GET /v1/deploy/applications` (fleet list), `/{name}/tree`
+(ownerRef resource tree + per-node health/sync), `/{name}/resource/{ref}` (live
+manifest + desired-vs-live diff), `/{name}/logs`; `POST /{name}/rollback` pins the CR
+image to a prior semver and `/{name}/sync` requests a reconcile. SUPERADMIN-only on
+`c.IsAdmin()`, fail-closed; Secret nodes are never surfaced. `engine.go` embeds the argo
+`gitops-engine` (`hanzoai/deploy/gitops-engine` v0.7.2, no replace) in-process for the
+reconcile half behind `DEPLOY_ENGINE_ENABLED` (default off), with a prune-safety fuse.
+
+## The `hanzo` CLI targets THIS binary — one contract, one IAM login
+
+The `hanzo` CLI (`cli/`) is the same unified binary; its control-plane verbs speak the
+routes THIS process serves, authorized off a plain `hanzo login` (the IAM access token is
+the final bearer fallback — no `--platform-token`). The ONE contract, no TS-Dokploy drift:
+
+- `hanzo apps list|get`  → `GET /v1/paas/apps[/{app}]`  (`clients/paas` fleet drift board)
+- `hanzo deploy <app>`   → `POST /v1/paas/apps/{app}/deploy` — a zero-downtime ROLLING
+  RESTART (stamps the Deployment pod-template `hanzo.ai/restartedAt` annotation; never
+  changes the declared TAG — that stays a git commit CD reconciles). `--env` picks the ns.
+- `hanzo clusters list|get` → `GET /v1/clusters`  (`clients/visor`, tenant-scoped)
+- `hanzo build`          → `POST /v1/runner`  (native buildkit fabric)
+
+`/v1/paas/*` auth mirrors `/v1/runner` (`clients/platform/runner.go`): the `guard` admits a
+validated principal who is SuperAdmin OR OrgAdmin, then each handler CONFINES a non-super
+caller to the platform namespaces its own validated org owns (`scopedNamespaces`, keyed on
+`principal.Org` — a tenant admin can never observe/restart another org's, or a platform,
+app; `?org=` cannot widen it). The rolling restart needs `patch` on `apps/deployments`
+(ClusterRole/cloud, universe `infra/k8s/cloud/rbac.yaml`). There is NO `/v1/apps`,
+`/v1/org/{org}/cluster`, or `/v1/platform/projects` CLI path — the first two never existed
+here (TS-Dokploy contract, 404), and `/v1/platform/*` needs a co-resident IAM store this
+deployment does not fold in (IAM runs as a separate svc) so it 500s; the live apps backend
+is `/v1/paas`, whose board reads k8s directly with no IAM-store dependency.

Makefile

@@ -24,7 +24,7 @@ OPENAPI_DIR    ?= ../openapi
 # forces the fork to modernc too so the whole binary registers "sqlite" once.
 CGO_ENABLED     ?= 0
 
-.PHONY: help native webui agentskills build build-standalone hanzo run smoke test test-cgo vet tidy docker docker-push clean
+.PHONY: help native webui deploy-ui agentskills build build-standalone hanzo run smoke test test-cgo vet tidy docker docker-push clean
 
 help: ## Show this help.
 	@awk 'BEGIN{FS=":.*##";printf "\nUsage: make <target>\n\nTargets:\n"} /^[a-zA-Z_-]+:.*##/{printf "  \033[36m%-16s\033[0m %s\n", $$1, $$2}' $(MAKEFILE_LIST)
@@ -40,6 +40,17 @@ webui: ## Build the real console static bundle into webui/dist (go:embed source)
 	cp -r "$(CONSOLE_DIR)/out/." webui/dist/
 	@echo ">> embedded real console bundle into webui/dist (index.html $$(wc -c < webui/dist/index.html) bytes)"
 
+deploy-ui: ## Build the monochrome ArgoCD dashboard bundle into clients/deploy/webui/dist (go:embed source). DEPLOY_DIR=<path to hanzoai/deploy>.
+	@command -v yarn >/dev/null 2>&1 || { echo "yarn is required to build the deploy dashboard bundle"; exit 1; }
+	@test -f "$(DEPLOY_DIR)/ui/package.json" || { echo "deploy checkout not found at $(DEPLOY_DIR) — set DEPLOY_DIR=<path to hanzoai/deploy on rebrand/hanzo-monochrome>"; exit 1; }
+	@test -d "$(DEPLOY_DIR)/ui/node_modules" || (cd "$(DEPLOY_DIR)/ui" && yarn install --frozen-lockfile)
+	cd "$(DEPLOY_DIR)/ui" && NODE_OPTIONS=--max-old-space-size=8192 yarn build
+	# Overlay the fresh bundle, keeping only the tracked fallback (.gitignore +
+	# index.html shell); the real 43MB bundle is build-time-only (gitignored).
+	find clients/deploy/webui/dist -mindepth 1 -maxdepth 1 ! -name .gitignore -exec rm -rf {} +
+	cp -r "$(DEPLOY_DIR)/ui/dist/app/." clients/deploy/webui/dist/
+	@echo ">> embedded monochrome ArgoCD bundle into clients/deploy/webui/dist (index.html $$(wc -c < clients/deploy/webui/dist/index.html) bytes)"
+
 agentskills: ## Regenerate the FULL agent-skills catalog into clients/agentskills/catalog (go:embed source) from the openapi SOT. OPENAPI_DIR=<path to openapi>.
 	@test -f "$(OPENAPI_DIR)/skills.py" || { echo "openapi checkout not found at $(OPENAPI_DIR) — set OPENAPI_DIR=<path> or clone hanzoai/openapi"; exit 1; }
 	# skills.py rewrites the whole catalog dir; the .gitignore keeps only the tiny

apps/apps.go

@@ -34,6 +34,7 @@ package apps
 import (
 	"context"
 	"fmt"
+	"os"
 
 	"github.com/hanzoai/cloud"
 	"github.com/zap-proto/zip"
@@ -45,14 +46,15 @@ import (
 	"github.com/hanzoai/authz"
 	"github.com/hanzoai/licensing"
 	"github.com/hanzoai/metrics"
-	o11ymod "github.com/hanzoai/o11y"
 
 	// In-repo subsystem packages (clients/*). Each exports a Mount (and, where it
 	// owns process-lifetime resources, a Shutdown); Wire references them directly.
 	"github.com/hanzoai/cloud/clients/account"
 	"github.com/hanzoai/cloud/clients/admin"
+	"github.com/hanzoai/cloud/clients/admission"
 	"github.com/hanzoai/cloud/clients/ads"
 	"github.com/hanzoai/cloud/clients/affiliates"
+	"github.com/hanzoai/cloud/clients/agent"
 	"github.com/hanzoai/cloud/clients/agents"
 	"github.com/hanzoai/cloud/clients/agentskills"
 	"github.com/hanzoai/cloud/clients/analytics"
@@ -64,21 +66,19 @@ import (
 	"github.com/hanzoai/cloud/clients/bots"
 	"github.com/hanzoai/cloud/clients/captable"
 	"github.com/hanzoai/cloud/clients/catalogsync"
-	"github.com/hanzoai/cloud/clients/chat"
+	"github.com/hanzoai/cloud/clients/cloudflare"
 	"github.com/hanzoai/cloud/clients/code"
 	"github.com/hanzoai/cloud/clients/company"
-	"github.com/hanzoai/cloud/clients/connectorruntime"
 	"github.com/hanzoai/cloud/clients/content"
 	"github.com/hanzoai/cloud/clients/crm"
-	"github.com/hanzoai/cloud/clients/cron"
 	"github.com/hanzoai/cloud/clients/dataroom"
 	"github.com/hanzoai/cloud/clients/deploy"
+	"github.com/hanzoai/cloud/clients/dns"
 	"github.com/hanzoai/cloud/clients/do"
 	"github.com/hanzoai/cloud/clients/entitlements"
 	"github.com/hanzoai/cloud/clients/eval"
 	"github.com/hanzoai/cloud/clients/exec"
-	"github.com/hanzoai/cloud/clients/featureflags"
-	"github.com/hanzoai/cloud/clients/featuregate"
+	"github.com/hanzoai/cloud/clients/flags"
 	"github.com/hanzoai/cloud/clients/framework"
 	"github.com/hanzoai/cloud/clients/functions"
 	"github.com/hanzoai/cloud/clients/gateway"
@@ -86,6 +86,7 @@ import (
 	"github.com/hanzoai/cloud/clients/graph"
 	"github.com/hanzoai/cloud/clients/guide"
 	"github.com/hanzoai/cloud/clients/iam"
+	"github.com/hanzoai/cloud/clients/iam2"
 	"github.com/hanzoai/cloud/clients/ingress"
 	"github.com/hanzoai/cloud/clients/integrations"
 	"github.com/hanzoai/cloud/clients/kafka"
@@ -115,6 +116,7 @@ import (
 	"github.com/hanzoai/cloud/clients/sign"
 	"github.com/hanzoai/cloud/clients/social"
 	"github.com/hanzoai/cloud/clients/storage"
+	"github.com/hanzoai/cloud/clients/sync"
 	"github.com/hanzoai/cloud/clients/tasks"
 	"github.com/hanzoai/cloud/clients/team"
 	"github.com/hanzoai/cloud/clients/templates"
@@ -162,6 +164,22 @@ func init() {
 	})
 }
 
+// identitySpec selects the ONE identity backend that owns /v1/iam/* (+ /login/oauth/*)
+// for this boot. CLOUD_IAM_IMPL=iam2 picks the clean-room iam2 (zip+orm, beego-free);
+// anything else — including unset, the production default — keeps the legacy beego
+// Casdoor embed, byte-for-byte today's behavior. The two impls register the SAME
+// absolute prefixes and therefore cannot co-mount, so selection (this func) stays
+// separate from activation (cfg.Enabled): exactly one spec occupies the identity slot
+// in Wire, preserving mount order either way. os.Getenv (not the unexported
+// cloud.getenv, which is unreachable from package apps) is the read — CLOUD_IAM_IMPL is
+// the deliberate, off-by-default opt-in that keeps iam2 inert until a canary flips it.
+func identitySpec() cloud.MountSpec {
+	if os.Getenv("CLOUD_IAM_IMPL") == "iam2" {
+		return cloud.MountSpec{Name: "iam2", Mount: iam2.Mount}
+	}
+	return cloud.MountSpec{Name: "iam", Mount: iam.Mount}
+}
+
 // Wire returns every linked subsystem as a cloud.MountSpec, in mount order. The
 // slice position IS the order: cloud.MountAll iterates it as-given, registering each
 // subsystem's teardown as a zip shutdown hook so teardown runs in reverse (LIFO).
@@ -170,186 +188,204 @@ func init() {
 func Wire() []cloud.MountSpec {
 	return []cloud.MountSpec{
 		// embedded NATS :4222 + JetStream.
-		{Name: "pubsub", Mount: cloud.Typed(pubsub.Mount), Shutdown: pubsub.Shutdown},
+		{Name: "pubsub", Mount: pubsub.Mount, Shutdown: pubsub.Shutdown},
 		// embedded Kafka adaptor :9092.
-		{Name: "kafka", Mount: cloud.Typed(kafka.Mount), Shutdown: kafka.Shutdown},
+		{Name: "kafka", Mount: kafka.Mount, Shutdown: kafka.Shutdown},
 		// /.well-known/agent-skills/* — before IAM's /.well-known/* wildcard (50).
-		{Name: "agentskills", Mount: cloud.Typed(agentskills.Mount)},
+		{Name: "agentskills", Mount: agentskills.Mount},
 		// Insights feature-flag evaluation seam (no routes; a hot value plane).
-		{Name: "featureflags", Mount: cloud.Typed(featureflags.Mount), Shutdown: featureflags.Shutdown, OwnsHealth: true},
+		{Name: "flags", Mount: flags.Mount, Shutdown: flags.Shutdown, OwnsHealth: true},
 		// Embedded KMS secrets plane /v1/kms/*. OwnsHealth: serves its own fail-closed
 		// /v1/kms/health (the generic always-ok route must not shadow it). Fails closed
 		// until the operator injects CLOUD_KMS_MASTER_KEY_REF. (Its in-process client
 		// factory is registered separately via cloud.RegisterKMSClientFactory.)
-		{Name: "kms", Mount: cloud.Typed(kms.Mount), OwnsHealth: true},
+		{Name: "kms", Mount: kms.Mount, OwnsHealth: true},
 		// hanzoai/metrics — native o11y. It declares its OWN narrow metrics.Deps (no
 		// hanzoai/cloud import), so Typed cannot adapt it; mountMetrics builds that Deps
 		// from cloud.Deps and calls metrics.Mount explicitly.
 		{Name: "metrics", Mount: mountMetrics},
 		// Embedded runtime edge (/v1/ingress/*). STAGED — edge listeners stay off unless
 		// the operator names "ingress" in CLOUD_ENABLE.
-		{Name: "ingress", Mount: cloud.Typed(ingress.Mount), Shutdown: ingress.Shutdown},
+		{Name: "ingress", Mount: ingress.Mount, Shutdown: ingress.Shutdown},
 		// SPECIFIC self-service routes (/v1/iam/{keys,onboard}, /v1/csrf, /v1/embed-status,
 		// /v1/commerce/topup/wallet). MUST mount before the IAM /v1/iam/* wildcard (50) so
 		// they win Fiber's first-match scan (framework-guaranteed since zip v1.3.0).
-		{Name: "account", Mount: cloud.Typed(account.MountAccount)},
+		{Name: "account", Mount: account.MountAccount},
 		// Embedded IAM identity plane (/v1/iam/*, /.well-known/*, /login/oauth/*, /_/iam/*,
 		// /cas/*, /scim/*) — the identity authority, mounts before its dependents. STAGED:
 		// the operator adds "iam" to --enable only after IAM config + the fold are verified.
-		{Name: "iam", Mount: cloud.Typed(iam.Mount)},
+		// Which IMPLEMENTATION owns these prefixes is selected by CLOUD_IAM_IMPL
+		// (identitySpec): the clean-room iam2 (zip+orm, beego-free) when =="iam2", else the
+		// legacy beego Casdoor embed — the default (unset = today's behavior, byte-for-byte).
+		// Both register the SAME absolute paths and cannot co-mount, so this is an either/or
+		// switch at this ONE slot, never a shadow prefix.
+		identitySpec(),
 		// Embedded Base app engine + viral waitlist (/v1/waitlist/*). STAGED behind
 		// CLOUD_BASE_EMBED. OwnsHealth: native /v1/base/health.
-		{Name: "base", Mount: cloud.Typed(base.Mount), Shutdown: base.Shutdown, OwnsHealth: true},
-		// In-repo o11y READ plane + the runtime-handler install (o11y.SetHandler). Every
-		// specific /v1/o11y/* route registers INSIDE this one mount, hence BEFORE the
-		// hanzoai/o11y module wildcard (70) — Fiber's in-order match gives them precedence.
-		// OwnsHealth: the module's order-70 co-entry below owns the single /v1/o11y/health.
-		{Name: "o11y", Mount: o11y.MountO11y, Shutdown: o11y.ShutdownO11y, OwnsHealth: true},
-		// hanzoai/o11y module wildcard /v1/o11y/* — co-owner of the ONE o11y concept with
-		// the in-repo entry above (same name), delegated to via o11y.SetHandler.
-		{Name: "o11y", Mount: cloud.Typed(o11ymod.Mount)},
-		{Name: "authz", Mount: cloud.Typed(authz.Mount)},
+		{Name: "base", Mount: base.Mount, Shutdown: base.Shutdown, OwnsHealth: true},
+		// The ONE observability subsystem: the in-repo o11y READ plane + runtime-handler
+		// install (o11y.SetHandler), with the hanzoai/o11y module wildcard /v1/o11y/*
+		// folded in as the TERMINAL sub-mount INSIDE o11y.MountO11y. Every specific
+		// /v1/o11y/* route registers before that wildcard, so Fiber's in-order match gives
+		// them precedence. NOT OwnsHealth: /v1/o11y/health stays the generic always-ok
+		// route (registered before MountAll), exactly as when the former module co-entry —
+		// which also set OwnsHealth=false — triggered it.
+		{Name: "o11y", Mount: o11y.MountO11y, Shutdown: o11y.ShutdownO11y},
+		{Name: "authz", Mount: authz.Mount},
 		// Embedded commerce plane /v1/commerce/*, /_/commerce/* — the hanzoai/commerce
 		// MODULE via the adapter in commerce.go (un-forked; the in-process
 		// CommerceClient is wired directly in pickCommerceClient).
-		{Name: "commerce", Mount: cloud.Typed(mountCommerce)},
-		// hanzoai/licensing. Its Mount is func(any, cloud.Deps) error — a MountFunc
-		// already — so Wire references it DIRECTLY, not through Typed.
+		{Name: "commerce", Mount: mountCommerce},
 		{Name: "licensing", Mount: licensing.Mount},
-		{Name: "plans", Mount: cloud.Typed(plan.Mount), OwnsHealth: true},
-		{Name: "pricing", Mount: cloud.Typed(pricing.Mount), OwnsHealth: true},
+		// clients/plan.Mount. Enable id normalized "plans" -> "plan" to match the
+		// package + generated cmd/plan (one subsystem, one name). Its product routes
+		// stay /v1/plans/* (incl. the OwnsHealth /v1/plans/health probe) — unchanged.
+		{Name: "plan", Mount: plan.Mount, OwnsHealth: true},
+		{Name: "pricing", Mount: pricing.Mount, OwnsHealth: true},
 		// /v1/s3/buckets/* + /v1/s3/health. Mounts BEFORE provisioning (120) so its static
 		// routes win over provisioning's /v1/s3/:name. OwnsHealth (real fail-closed probe).
-		{Name: "storage", Mount: cloud.Typed(storage.Mount), OwnsHealth: true},
+		{Name: "storage", Mount: storage.Mount, OwnsHealth: true},
 		// Provisioning control plane: /v1/sql,/v1/vector,/v1/datastore,/v1/kv,/v1/search,/v1/s3,/v1/docdb.
-		{Name: "provisioning", Mount: cloud.Typed(provisioning.Mount)},
-		{Name: "billing", Mount: cloud.Typed(billing.Mount)},
+		{Name: "provisioning", Mount: provisioning.Mount},
+		{Name: "billing", Mount: billing.Mount},
 		// CATCH-ALL /v1/billing/* + /v1/commerce/* data bridges — AFTER clients/billing
 		// (121) + the commerce embed (100). Same clients/account package as "account" (48).
-		{Name: "account-bridge", Mount: cloud.Typed(account.MountBridge)},
-		{Name: "do", Mount: cloud.Typed(do.Mount)},
-		{Name: "platform", Mount: cloud.Typed(platform.Mount), OwnsHealth: true},
-		{Name: "projects", Mount: cloud.Typed(projects.Mount)},
-		{Name: "prompts", Mount: cloud.Typed(prompts.Mount)},
-		{Name: "agents", Mount: cloud.Typed(agents.Mount), Shutdown: agents.Shutdown},
+		{Name: "account-bridge", Mount: account.MountBridge},
+		{Name: "do", Mount: do.Mount},
+		{Name: "platform", Mount: platform.Mount, OwnsHealth: true},
+		{Name: "projects", Mount: projects.Mount},
+		// The /v1/dns forward head: relays the console DNS dashboard to the DNS
+		// control plane under the caller's own validated bearer (clients/dns).
+		{Name: "dns", Mount: dns.Mount},
+		{Name: "prompts", Mount: prompts.Mount},
+		{Name: "agents", Mount: agents.Mount, Shutdown: agents.Shutdown},
 		// The unified AI login manager registry (/v1/links). Mounts AFTER agents so
 		// a link revoke can stop the affected agent sessions in-process.
-		{Name: "link", Mount: cloud.Typed(link.Mount), Shutdown: link.Shutdown},
-		{Name: "wallets", Mount: cloud.Typed(wallets.Mount), Shutdown: ctxShutdown(wallets.Shutdown)},
+		{Name: "link", Mount: link.Mount, Shutdown: link.Shutdown},
+		{Name: "wallets", Mount: wallets.Mount, Shutdown: ctxShutdown(wallets.Shutdown)},
 		// x402 pay-per-use: settles a signed ERC-3009 authorization to a recipient
 		// wallet through the metering spine. Mounts AFTER wallets (it resolves the
 		// recipient via wallets.ResolvePaymentTarget) and provides the Enforce
 		// middleware a marketplace applies to its priced routes.
-		{Name: "x402", Mount: cloud.Typed(x402.Mount), Shutdown: ctxShutdown(x402.Shutdown)},
-		{Name: "paas", Mount: cloud.Typed(paas.Mount), OwnsHealth: true},
+		{Name: "x402", Mount: x402.Mount, Shutdown: ctxShutdown(x402.Shutdown)},
+		{Name: "paas", Mount: paas.Mount, OwnsHealth: true},
 		// GitOps deploy dashboard /v1/deploy/* (the ArgoCD-grade fleet view over the
 		// operator App CRs). After paas so the release seam paas installs is registered
 		// before a gitops rollback delegates to it; owns its own /v1/deploy/health.
-		{Name: "deploy", Mount: cloud.Typed(deploy.Mount), OwnsHealth: true},
-		{Name: "functions", Mount: cloud.Typed(functions.Mount)},
-		{Name: "tracker", Mount: cloud.Typed(tracker.Mount)},
-		{Name: "templates", Mount: cloud.Typed(templates.Mount)},
-		{Name: "framework", Mount: cloud.Typed(framework.Mount), Shutdown: ctxShutdown(framework.Shutdown)},
-		{Name: "knowledge", Mount: cloud.Typed(knowledge.Mount)},
+		{Name: "deploy", Mount: deploy.Mount, OwnsHealth: true},
+		{Name: "functions", Mount: functions.Mount},
+		{Name: "tracker", Mount: tracker.Mount},
+		{Name: "templates", Mount: templates.Mount},
+		{Name: "framework", Mount: framework.Mount, Shutdown: ctxShutdown(framework.Shutdown)},
+		{Name: "knowledge", Mount: knowledge.Mount},
 		// Marketing content loop /v1/content/* (generate → CMS → transition → publish).
 		// After framework (its DocType store the ops read/write) + knowledge (the sibling
 		// framework lane); before the AI /v1/* catch-all so /v1/content/* resolves here.
 		// CRUD/tenancy/install are framework's; this adds the board, lifecycle transition,
 		// and the generate/publish orchestration over the zen5 + studio + social edges.
-		{Name: "content", Mount: cloud.Typed(content.Mount), Shutdown: ctxShutdown(content.Shutdown)},
+		{Name: "content", Mount: content.Mount, Shutdown: ctxShutdown(content.Shutdown)},
 		// Reverse storefront loop: consume the commerce COMMERCE stream (product.created)
 		// → content.EnsureCatalogAsset (render the new product's ecom asset, design==slug).
 		// After content (whose EnsureCatalogAsset it drives). Inert until CLOUD_COMMERCE_NATS_URL
 		// names the NATS carrying commerce catalog events — the reverse of the forward edge.
-		{Name: "catalogsync", Mount: cloud.Typed(catalogsync.Mount), Shutdown: catalogsync.Shutdown},
-		{Name: "ml", Mount: cloud.Typed(ml.Mount), OwnsHealth: true},
-		{Name: "usage", Mount: cloud.Typed(usage.Mount)},
-		{Name: "crm", Mount: cloud.Typed(crm.Mount)},
+		{Name: "catalogsync", Mount: catalogsync.Mount, Shutdown: catalogsync.Shutdown},
+		{Name: "ml", Mount: ml.Mount, OwnsHealth: true},
+		{Name: "usage", Mount: usage.Mount},
+		{Name: "crm", Mount: crm.Mount},
 		// Native /v1/marketing/* — the in-process fold of github.com/hanzoai/marketing
 		// (per-org campaign store on Base/SQLite), twin of crm. Owns a DB handle, so
 		// its Shutdown closes it cleanly on SIGTERM (ctxShutdown adapts func() error).
-		{Name: "marketing", Mount: cloud.Typed(marketing.Mount), Shutdown: ctxShutdown(marketing.Shutdown)},
+		{Name: "marketing", Mount: marketing.Mount, Shutdown: ctxShutdown(marketing.Shutdown)},
 		// Native /v1/ads/* — the net-new per-org ad-campaign store on Base/SQLite,
 		// twin of crm/marketing. Owns a DB handle, so its Shutdown closes it cleanly
 		// on SIGTERM (ctxShutdown adapts func() error).
-		{Name: "ads", Mount: cloud.Typed(ads.Mount), Shutdown: ctxShutdown(ads.Shutdown)},
+		{Name: "ads", Mount: ads.Mount, Shutdown: ctxShutdown(ads.Shutdown)},
 		// Native /v1/social/* — the in-process fold of the live social stack
 		// (github.com/hanzoai/social: social-backend/frontend/orchestrator, a Postiz-style
 		// scheduler), a per-org accounts+posts store on Base/SQLite, twin of crm. Owns a DB
 		// handle, so its Shutdown closes it cleanly on SIGTERM (ctxShutdown adapts func() error).
-		{Name: "social", Mount: cloud.Typed(social.Mount), Shutdown: ctxShutdown(social.Shutdown)},
-		{Name: "analytics", Mount: cloud.Typed(analytics.Mount), OwnsHealth: true},
-		{Name: "git", Mount: cloud.Typed(git.Mount)},
-		{Name: "visor", Mount: cloud.Typed(visor.Mount)},
+		{Name: "social", Mount: social.Mount, Shutdown: ctxShutdown(social.Shutdown)},
+		{Name: "analytics", Mount: analytics.Mount, OwnsHealth: true},
+		{Name: "git", Mount: git.Mount},
+		// Universal sync (/v1/sync/links + engine). Registers the cloud.SyncEngine the
+		// GitHub/Gitea webhooks enqueue to; git is its first provider. Owns per-org
+		// DB handles, so its Shutdown closes them on SIGTERM.
+		{Name: "sync", Mount: sync.Mount, Shutdown: ctxShutdown(sync.Shutdown)},
+		{Name: "visor", Mount: visor.Mount},
 		// Cap table on Base via goja. STAGED behind CLOUD_ENABLE.
-		{Name: "captable", Mount: cloud.Typed(captable.Mount), Shutdown: captable.Shutdown},
-		{Name: "code", Mount: cloud.Typed(code.Mount), Shutdown: code.Shutdown},
-		{Name: "zero-trust", Mount: cloud.Typed(zt.Mount)},
+		{Name: "captable", Mount: captable.Mount, Shutdown: captable.Shutdown},
+		{Name: "code", Mount: code.Mount, Shutdown: code.Shutdown},
+		{Name: "zero-trust", Mount: zt.Mount},
 		// Data rooms via goja + per-tenant Base. STAGED behind CLOUD_ENABLE. OwnsHealth.
-		{Name: "dataroom", Mount: cloud.Typed(dataroom.Mount), Shutdown: dataroom.Shutdown, OwnsHealth: true},
-		{Name: "graph", Mount: cloud.Typed(graph.Mount)},
-		{Name: "security", Mount: cloud.Typed(security.Mount), Shutdown: ctxShutdown(security.Shutdown), OwnsHealth: true},
-		{Name: "integrations", Mount: cloud.Typed(integrations.Mount), Shutdown: integrations.Shutdown},
-		{Name: "sbom", Mount: cloud.Typed(sbom.Mount), OwnsHealth: true},
-		{Name: "team", Mount: cloud.Typed(team.Mount), Shutdown: ctxShutdown(team.Shutdown)},
-		{Name: "settings", Mount: cloud.Typed(settings.Mount), Shutdown: settings.Shutdown},
-		{Name: "notify", Mount: cloud.Typed(notify.Mount), OwnsHealth: true},
-		{Name: "gateway", Mount: cloud.Typed(gateway.Mount)},
-		{Name: "entitlements", Mount: cloud.Typed(entitlements.Mount), Shutdown: entitlements.Shutdown},
-		{Name: "exec", Mount: cloud.Typed(exec.Mount)},
-		{Name: "websearch", Mount: cloud.Typed(websearch.Mount)},
-		{Name: "world", Mount: cloud.Typed(world.Mount), Shutdown: ctxShutdown(world.Shutdown)},
+		{Name: "dataroom", Mount: dataroom.Mount, Shutdown: dataroom.Shutdown, OwnsHealth: true},
+		{Name: "graph", Mount: graph.Mount},
+		{Name: "security", Mount: security.Mount, Shutdown: ctxShutdown(security.Shutdown), OwnsHealth: true},
+		{Name: "integrations", Mount: integrations.Mount, Shutdown: integrations.Shutdown},
+		// Per-org Cloudflare asset plane /v1/integrations/cloudflare/{pages,workers,r2,kv,d1}/*.
+		// Mounts AFTER integrations because it reads the org's Cloudflare token through
+		// the integrations custody seam (integrations.TokenFor) — one token, one
+		// custody boundary. Stateless: no store, no shutdown.
+		{Name: "cloudflare", Mount: cloudflare.Mount},
+		{Name: "sbom", Mount: sbom.Mount, OwnsHealth: true},
+		{Name: "team", Mount: team.Mount, Shutdown: ctxShutdown(team.Shutdown)},
+		{Name: "settings", Mount: settings.Mount, Shutdown: settings.Shutdown},
+		{Name: "notify", Mount: notify.Mount, OwnsHealth: true},
+		{Name: "gateway", Mount: gateway.Mount},
+		{Name: "entitlements", Mount: entitlements.Mount, Shutdown: entitlements.Shutdown},
+		{Name: "exec", Mount: exec.Mount},
+		{Name: "websearch", Mount: websearch.Mount},
+		{Name: "world", Mount: world.Mount, Shutdown: ctxShutdown(world.Shutdown)},
 		// The bot runtime's ops face (/v1/bot/*). The transport itself is domain-free;
 		// the run control plane is "bots" below.
-		{Name: "runtime", Mount: cloud.Typed(runtime.Mount)},
-		{Name: "authors", Mount: cloud.Typed(authors.Mount), Shutdown: ctxShutdown(authors.Shutdown)},
-		{Name: "bots", Mount: cloud.Typed(bots.Mount)},
-		{Name: "audit", Mount: cloud.Typed(auditlog.Mount)},
-		{Name: "affiliates", Mount: cloud.Typed(affiliates.Mount)},
+		{Name: "runtime", Mount: runtime.Mount},
+		{Name: "authors", Mount: authors.Mount, Shutdown: ctxShutdown(authors.Shutdown)},
+		{Name: "bots", Mount: bots.Mount},
+		{Name: "audit", Mount: auditlog.Mount},
+		{Name: "affiliates", Mount: affiliates.Mount},
 		// Hanzo Sign (e-signature) via goja + per-tenant Base. STAGED behind CLOUD_ENABLE. OwnsHealth.
-		{Name: "sign", Mount: cloud.Typed(sign.Mount), Shutdown: sign.Shutdown, OwnsHealth: true},
-		{Name: "product", Mount: cloud.Typed(product.Mount)},
-		{Name: "evals", Mount: cloud.Typed(eval.Mount)},
-		{Name: "treasury", Mount: cloud.Typed(treasury.Mount), Shutdown: ctxShutdown(treasury.Shutdown)},
-		{Name: "admin", Mount: cloud.Typed(admin.Mount)},
-		// Launch-control plane: per-service waitlist-mode registry (global SQLite)
-		// + admin board/toggle (/v1/admin/services*) + the guard's runtime mode read
-		// (/v1/featuregate/mode). Mounts right after admin so its specific routes bind
-		// ahead of the AI /v1/* catch-all; native Enforce middleware (wired in serve.go)
-		// reads the SAME store in-process. Owns a global store, so Shutdown closes it.
-		{Name: "featuregate", Mount: cloud.Typed(featuregate.Mount), Shutdown: ctxShutdown(featuregate.Shutdown)},
-		{Name: "tasks", Mount: cloud.Typed(tasks.Mount)},
-		// Platform cron: durable schedules on the shared tasks engine replacing
-		// every k8s CronJob — entries are cron.hanzo.ai ConfigMaps (universe git),
-		// runs visible in the Tasks console. Mounts no routes; starts after the
-		// engine is wired.
-		{Name: "cron", Mount: cloud.Typed(cron.Mount)},
-		{Name: "automations", Mount: cloud.Typed(automations.Mount), Shutdown: automations.Shutdown},
-		// Native single-connector execution (HIP-0126): runs an ActivePieces JS
-		// connector action in-process via goja (clients/connectorruntime), retiring
-		// the standalone auto Node engine. Mounts POST /v1/automations/connectors/:id/run,
-		// paired with the automations catalogue above; STAGED like the rest.
-		{Name: "connectorruntime", Mount: cloud.Typed(connectorruntime.Mount)},
+		{Name: "sign", Mount: sign.Mount, Shutdown: sign.Shutdown, OwnsHealth: true},
+		{Name: "product", Mount: product.Mount},
+		{Name: "evals", Mount: eval.Mount},
+		{Name: "treasury", Mount: treasury.Mount, Shutdown: ctxShutdown(treasury.Shutdown)},
+		{Name: "admin", Mount: admin.Mount},
+		// Launch-control gate (per-service waitlist): the COMPLETE feature — host→service
+		// registry + brand seed + the waitlist.<svc> switch registration + the
+		// /v1/flags/waitlist (and /v1/admission/mode compat) mode read + the Enforce
+		// middleware — COMPOSING the flags engine one-way (flags.Bool/Register/
+		// SetPlatformSwitch; flags never imports admission). Mounts AFTER flags so the
+		// engine's platform-switch plane is installed first; the admin board is the
+		// /v1/admin/services lens over it. Owns the registry store handle → Shutdown.
+		{Name: "admission", Mount: admission.Mount, Shutdown: ctxShutdown(admission.Shutdown)},
+		// Tasks: the durable workflow/UI surface AND platform cron (durable schedules
+		// on the same shared engine, replacing every k8s CronJob). cron was a separate
+		// Wire entry; it mounts no routes and only registers schedules, so it is folded
+		// in as a sub-mount of tasks.Mount — ONE tasks subsystem.
+		{Name: "tasks", Mount: tasks.Mount},
+		// Automations: the connector catalogue + flow engine AND native single-connector
+		// execution (POST /v1/automations/connectors/:id/run, HIP-0126). The connector
+		// runner mounts no other routes, so it is folded in as a sub-mount of
+		// automations.Mount (was a separate "connectorruntime" entry) — ONE subsystem.
+		{Name: "automations", Mount: automations.Mount, Shutdown: automations.Shutdown},
 		// Unified tool plane: /v1/tools/* — the ONE registry (connectors, functions,
 		// agents, skills, external MCP servers, full-cloud-control /v1 routes), per-org
 		// activation, and the unified MCP endpoint. Sources register into it from their
 		// own Mounts, so mount position is not load-bearing (List/Dispatch run at request
 		// time); placed after automations, before the zen/ai catch-all so /v1/tools wins.
-		{Name: "tools", Mount: cloud.Typed(tools.Mount), Shutdown: tools.Shutdown},
+		{Name: "tools", Mount: tools.Mount, Shutdown: tools.Shutdown},
 		// Marketplace: /v1/marketplace/* — listing/discovery/install over the tool plane,
 		// with x402-priced monetized listings. Mounts after tools (it fills the price seam).
-		{Name: "marketplace", Mount: cloud.Typed(marketplace.Mount), Shutdown: marketplace.Shutdown},
-		{Name: "referrals", Mount: cloud.Typed(referrals.Mount)},
+		{Name: "marketplace", Mount: marketplace.Mount, Shutdown: marketplace.Shutdown},
+		{Name: "referrals", Mount: referrals.Mount},
 		// Business AI Guide /v1/guide/* — the interactive launch checklist engine +
 		// the agent that executes a step through the per-principal MCP plane. After
 		// automations (whose InvokeTool it drives) and referrals; before the ai
 		// catch-all. Owns per-org SQLite, so its Shutdown closes the stores.
-		{Name: "guide", Mount: cloud.Typed(guide.Mount), Shutdown: ctxShutdown(guide.Shutdown)},
+		{Name: "guide", Mount: guide.Mount, Shutdown: ctxShutdown(guide.Shutdown)},
 		// Hanzo Company — the incorporation + fundraising state machine
 		// (/v1/company/*). Mounts after the seams it composes (integrations for the
 		// google token custody; captable/dataroom facades) and before the /v1/* AI
 		// catch-all so its routes resolve here.
-		{Name: "company", Mount: cloud.Typed(company.Mount), Shutdown: company.Shutdown},
+		{Name: "company", Mount: company.Mount, Shutdown: company.Shutdown},
 		// Chat orchestrator — POST /v1/chat: ONE LLM tool-calling round over the tool
 		// plane. It COMPOSES the ai completion path (in-process, so per-org billing
 		// runs) + the unified tool registry, and splits the model's tool calls into
@@ -357,7 +393,7 @@ func Wire() []cloud.MountSpec {
 		// catch-all so /v1/chat resolves here (Fiber first-match); the ai module's
 		// beego /v1/chat alias behind its /v1/* glob is thereby shadowed, while ai
 		// keeps /v1/chat/completions + /v1/completions.
-		{Name: "chat", Mount: cloud.Typed(chat.Mount)},
+		{Name: "agent", Mount: agent.Mount},
 		// The bare /v1/* AI catch-all — the LAST route position. Every owning subsystem above
 		// wins its own namespace (Fiber first-match); AI is the fallback for the rest of /v1/*.
 		// zen mounts as a /v1-scoped Claim middleware BEFORE ai: it routes zen* models
@@ -366,9 +402,9 @@ func Wire() []cloud.MountSpec {
 		// other model and the /v1/models list. Order is load-bearing — Claim must
 		// run before ai's catch-all. (See hip-00NN.)
 		{Name: "zen", Mount: mountZen},
-		{Name: "ai", Mount: cloud.Typed(ai.Mount)},
+		{Name: "ai", Mount: ai.Mount},
 		// Runtime wasm/proxy plugins — mounts dead last.
-		{Name: "plugins", Mount: cloud.Typed(plugin.Mount)},
+		{Name: "plugins", Mount: plugin.Mount},
 	}
 }
 
@@ -395,11 +431,7 @@ func ServeSingle(name string) error {
 // externals, metrics declares its OWN narrow Deps (Logger, DataDir, Brand) and does
 // not import hanzoai/cloud, so cloud.Typed cannot bridge it: the composition root
 // builds metrics.Deps from cloud.Deps and calls metrics.Mount explicitly here.
-func mountMetrics(app any, deps cloud.Deps) error {
-	a, ok := app.(*zip.App)
-	if !ok {
-		return fmt.Errorf("metrics.Mount: app is %T, want *zip.App", app)
-	}
+func mountMetrics(a *zip.App, deps cloud.Deps) error {
 	return metrics.Mount(a, metrics.Deps{Logger: deps.Logger, DataDir: deps.DataDir, Brand: deps.Brand})
 }
 

apps/commerce.go

@@ -20,15 +20,20 @@ import (
 	"fmt"
 	"net/http"
 	"path/filepath"
+	"strings"
+	"time"
 
 	"github.com/hanzoai/cloud"
 	"github.com/hanzoai/cloud/clients/commerceclient"
 	"github.com/hanzoai/cloud/clients/commerceinproc"
+	financeclient "github.com/hanzoai/cloud/clients/finance"
 	"github.com/hanzoai/commerce"
 	commercebilling "github.com/hanzoai/commerce/api/billing"
 	commercestore "github.com/hanzoai/commerce/api/store"
+	commercedatastore "github.com/hanzoai/commerce/datastore"
 	commercemid "github.com/hanzoai/commerce/middleware"
 	"github.com/hanzoai/commerce/middleware/iammiddleware"
+	commercensctx "github.com/hanzoai/commerce/util/nscontext"
 	log "github.com/luxfi/log"
 	"github.com/zap-proto/zip"
 )
@@ -133,7 +138,7 @@ func mountCommerce(app *zip.App, deps cloud.Deps) error {
 	// mirrors the standalone /v1 bundle: gated request context, host, IAM
 	// resolution; store.Route's own tokenRequired arg gates the CRUD.
 	storeV1 := app.Group("/v1")
-	storeV1.Use(commercemid.AddHost(), commercemid.RequestContext(), commercemid.ErrorHandlerJSON())
+	storeV1.Use(commercemid.AddHost(), commercemid.RequestContext(), commerceErrorScope())
 	// Unconditional, exactly like the standalone bundle: IAMTokenRequired
 	// no-ops gracefully when IAM is not initialized.
 	storeV1.Use(iammiddleware.IAMTokenRequired())
@@ -163,6 +168,19 @@ func mountCommerce(app *zip.App, deps cloud.Deps) error {
 	commerceinproc.SetApp(app)
 	commerceclient.PublishEmbedded(embedded)
 
+	// Usage-cap enforcement on the FINANCE path. The unified binary records usage in
+	// the finance ledger (fin.RecordUsage), NOT commerce's transaction store — which
+	// it leaves empty — so the cap must read spend from, and fire alerts on, the
+	// finance ledger. Two seams, both org-wide (the finance Entry carries no scope;
+	// per-scope caps are a follow-up):
+	//   - SetPeriodSpendReader: AuthorizeSpendCap's scopeSpentCents reads the org's
+	//     finance period spend instead of the empty commerce transaction ledger, so
+	//     a real LLM request increments the cap's `spent` and trips the 402.
+	//   - SetUsageHook: after each finance debit, fire the org's spend-alerts on the
+	//     SAME crossing (the alert half), reading the same finance spend + debouncing.
+	commercebilling.SetPeriodSpendReader(financePeriodSpend)
+	financeclient.SetUsageHook(fireCapAlert)
+
 	lg.Info("commerce embedded natively (hanzoai/commerce module on the shared zip app)",
 		"data_dir", dataDir,
 		"brand", deps.Brand,
@@ -173,6 +191,37 @@ func mountCommerce(app *zip.App, deps cloud.Deps) error {
 
 // mountCommerceFailClosed serves an honest JSON 503 on every commerce prefix when
 // the embed cannot boot, so /v1/commerce/* answers "commerce unavailable" instead
+// commerceErrorScope confines commerce's JSON error envelope to commerce's OWN
+// routes. commercemid.ErrorHandlerJSON is a `/v1` GROUP middleware, and fiber
+// matches group middleware by PREFIX, not by the handle a route registered on —
+// so on the shared `/v1` it wraps every subsystem mounted AFTER commerce and
+// flattens their typed zip.HTTPError (403/400/…) into a blanket 500 (the store
+// envelope always renders 500). Guarded by commercePrefixes, the envelope stays on
+// commerce and every other subsystem renders its own status via zip's default
+// handler — the pre-commerce subsystems (kms, o11y, …) already do; this makes the
+// post-commerce ones (projects, agents, wallets, …) match.
+func commerceErrorScope() zip.Handler {
+	envelope := commercemid.ErrorHandlerJSON()
+	return func(c *zip.Ctx) error {
+		if hasCommercePrefix(c.Path()) {
+			return envelope(c)
+		}
+		return c.Next()
+	}
+}
+
+// hasCommercePrefix reports whether path is a commerce-owned root (an exact prefix
+// or a child of one), the SAME ownership commercePrefixes encodes for the
+// fail-closed mount.
+func hasCommercePrefix(path string) bool {
+	for _, p := range commercePrefixes {
+		if path == p || strings.HasPrefix(path, p+"/") {
+			return true
+		}
+	}
+	return false
+}
+
 // of falling through to another subsystem's catch-all.
 func mountCommerceFailClosed(app *zip.App) {
 	failed := func(c *zip.Ctx) error {
@@ -183,3 +232,41 @@ func mountCommerceFailClosed(app *zip.App) {
 		app.All(p+"/*", failed)
 	}
 }
+
+// financePeriodSpend is the usage-cap's period-spend source (injected into commerce
+// via SetPeriodSpendReader). It returns the org's finance-ledger usage in cents since
+// the start of the CURRENT UTC month — the window the cap resets on (mirrors
+// commerce periodStartUTC). Org-wide: the finance Entry carries no project/service,
+// so scope args are ignored and the org total is returned (what the covering
+// org-wide spend-alert row binds on). A finance impl without the sum capability, or a
+// split deploy (no co-resident finance), reports 0 — the cap can never over-count.
+func financePeriodSpend(ctx context.Context, org string, test bool, _, _ string) (int64, error) {
+	fin := financeclient.Current()
+	if fin == nil {
+		return 0, nil
+	}
+	summer, ok := fin.(interface {
+		SumUsageSince(context.Context, string, bool, int64) (int64, error)
+	})
+	if !ok {
+		return 0, nil
+	}
+	n := time.Now().UTC()
+	since := time.Date(n.Year(), n.Month(), 1, 0, 0, 0, 0, time.UTC).Unix()
+	return summer.SumUsageSince(ctx, org, test, since)
+}
+
+// fireCapAlert fires the org's spend-alerts after a finance usage debit — the alert
+// half of the cap on the finance path (wired via finance.SetUsageHook). It resolves
+// the org's commerce datastore (where the spend-alert rows live) and calls the
+// exported commerce trigger, which reads the org's period spend via financePeriodSpend
+// and stamps/debounces. Detached + best-effort; never blocks the money path. Runs in
+// its own goroutine (the hook is invoked with `go`), so a background context is right.
+func fireCapAlert(org string, test bool, project, service string) {
+	if strings.TrimSpace(org) == "" {
+		return
+	}
+	ctx := commercensctx.WithNamespace(context.Background(), org)
+	db := commercedatastore.New(ctx)
+	commercebilling.FireSpendAlerts(ctx, db, org, test, project, service, nil)
+}

apps/commerce_errorscope_test.go

@@ -0,0 +1,82 @@
+package apps
+
+import (
+	"io"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	commercemid "github.com/hanzoai/commerce/middleware"
+	"github.com/zap-proto/zip"
+)
+
+// TestCommerceErrorScope proves commerceErrorScope() confines commerce's always-500
+// JSON envelope to commerce-owned prefixes: a post-commerce subsystem route
+// (/v1/projects) that returns a typed 403 renders 403 (zip default), while a
+// commerce route (/v1/store/...) still gets commerce's envelope. Mirrors the frozen
+// order: kms (before commerce) → commerce /v1 group chain → projects + store (after).
+// Regression for the release-smoke failure where 14 post-commerce endpoints 500'd.
+func TestCommerceErrorScope(t *testing.T) {
+	app := zip.New(zip.Config{})
+
+	// kms (before commerce) — a clean 403 baseline (never wrapped by commerce).
+	app.Get("/v1/kms/health", func(c *zip.Ctx) error { return zip.ErrForbidden("kms says no") })
+
+	// commerce (position 39): the REAL group chain, now with the scoped envelope.
+	sv1 := app.Group("/v1")
+	sv1.Use(commercemid.AddHost(), commercemid.RequestContext(), commerceErrorScope())
+
+	// projects (after commerce) — typed 403; must NOT be clobbered to 500.
+	app.Get("/v1/projects", func(c *zip.Ctx) error { return zip.ErrForbidden("X-Org-Id required") })
+	// a commerce store route (after commerce) — typed 403; commerce envelope applies.
+	app.Get("/v1/store/current", func(c *zip.Ctx) error { return zip.ErrForbidden("store needs org") })
+
+	probe := func(path string) (int, string) {
+		req := httptest.NewRequest("GET", path, nil)
+		resp, err := app.Fiber().Test(req)
+		if err != nil {
+			t.Fatalf("%s: %v", path, err)
+		}
+		defer resp.Body.Close()
+		b, _ := io.ReadAll(resp.Body)
+		return resp.StatusCode, string(b)
+	}
+
+	// The invariant the scope guards: a typed zip.HTTPError (403) a subsystem returns
+	// is NEVER flattened to a 500 — not before commerce, not after it. (commerce
+	// >=1.48.10 honors the status itself; the scope keeps commerce's error handler off
+	// other subsystems' routes regardless, so a future commerce regression can't
+	// re-clobber them.)
+	for _, tc := range []struct {
+		path     string
+		wantCode int
+	}{
+		{"/v1/kms/health", 403},    // before commerce
+		{"/v1/projects", 403},      // after commerce — must NOT be clobbered to 500
+		{"/v1/store/current", 403}, // commerce's own route — its handler still honors 403
+	} {
+		code, body := probe(tc.path)
+		t.Logf("%-20s -> %d  %s", tc.path, code, body)
+		if code != tc.wantCode {
+			t.Errorf("%s: got %d, want %d (%s)", tc.path, code, tc.wantCode, body)
+		}
+		if strings.Contains(body, "\"status\":5") || code >= 500 {
+			t.Errorf("%s: a typed 403 was flattened to a 5xx (%s)", tc.path, body)
+		}
+	}
+
+	for _, p := range []struct {
+		path string
+		own  bool
+	}{
+		{"/v1/store/current", true},
+		{"/v1/commerce/checkout", true},
+		{"/v1/projects", false},
+		{"/v1/agent/presets", false},
+		{"/v1/agents", false},
+	} {
+		if got := hasCommercePrefix(p.path); got != p.own {
+			t.Errorf("hasCommercePrefix(%q) = %v, want %v", p.path, got, p.own)
+		}
+	}
+}

apps/wire_test.go

@@ -16,100 +16,100 @@ import "testing"
 // origin/main @c504d2b: 68 specs from the live init()-registry, PLUS the two the
 // wave-2 external bumps (ai v1.805.2, o11y v1.5.12) stopped self-registering — ai
 // (@150 catch-all) and the hanzoai/o11y module wildcard (@70), which main currently
-// DROPS and this PR restores at their order-int slots. "o11y" therefore appears
-// twice: the in-repo read plane (order 69) and the module wildcard (70), co-owners.
+// DROPS and this PR restores. The module wildcard (order 70) is now folded in as the
+// terminal sub-mount of the in-repo o11y read plane (order 69), so "o11y" is ONE spec.
 var frozen = []struct {
 	name        string
 	ownsHealth  bool
 	hasShutdown bool
 }{
-	{"pubsub", false, true},            // was order 5
-	{"kafka", false, true},             // was order 6
-	{"agentskills", false, false},      // was order 8
-	{"featureflags", true, true},       // was order 9; native engine: /v1/flags health + store shutdown
-	{"kms", true, false},               // was order 10
-	{"metrics", false, false},          // was order 40
-	{"ingress", false, true},           // was order 42
-	{"account", false, false},          // was order 48
-	{"iam", false, false},              // was order 50
-	{"base", true, true},               // was order 60; per-org embed added Shutdown (#298)
-	{"o11y", true, true},               // was order 69
-	{"o11y", false, false},             // was order 70
-	{"authz", false, false},            // was order 70
-	{"commerce", false, false},         // was order 100
-	{"licensing", false, false},        // was order 110
-	{"plans", true, false},             // was order 111
-	{"pricing", true, false},           // was order 112
-	{"storage", true, false},           // was order 118
-	{"provisioning", false, false},     // was order 120
-	{"billing", false, false},          // was order 121
-	{"account-bridge", false, false},   // was order 122
-	{"do", false, false},               // was order 123
-	{"platform", true, false},          // was order 124
-	{"projects", false, false},         // was order 125
-	{"prompts", false, false},          // was order 126
-	{"agents", false, true},            // was order 127
-	{"link", false, true},              // new: unified AI login manager (/v1/links), after agents
-	{"wallets", false, true},           // was order 127
-	{"x402", false, true},              // new: x402 pay-per-use settlement (after wallets)
-	{"paas", true, false},              // was order 128
-	{"deploy", true, false},            // after paas (release seam), before functions
-	{"functions", false, false},        // was order 128
-	{"tracker", false, false},          // was order 129
-	{"templates", false, false},        // was order 129
-	{"framework", false, true},         // was order 129
-	{"knowledge", false, false},        // was order 130
-	{"content", false, true},           // new: marketing content loop (after knowledge)
-	{"catalogsync", false, true},       // new: reverse loop (product.created → render) after content
-	{"ml", true, false},                // was order 130
-	{"usage", false, false},            // was order 131
-	{"crm", false, false},              // was order 131
-	{"marketing", false, true},         // new: marketing domain fold (after crm)
-	{"ads", false, true},               // new: ads domain fold (after crm)
-	{"social", false, true},            // new: /v1/social fold (after crm)
-	{"analytics", true, false},         // was order 132
-	{"git", false, false},              // was order 132
-	{"visor", false, false},            // was order 133
-	{"captable", false, true},          // was order 133
-	{"code", false, true},              // was order 134
-	{"zero-trust", false, false},       // was order 134
-	{"dataroom", true, true},           // was order 134
-	{"graph", false, false},            // was order 135
-	{"security", true, true},           // was order 136
-	{"integrations", false, true},      // was order 137
-	{"sbom", true, false},              // was order 137
-	{"team", false, true},              // was order 138
-	{"settings", false, true},          // was order 138
-	{"notify", true, false},            // was order 139
-	{"gateway", false, false},          // was order 139
-	{"entitlements", false, true},      // was order 139
-	{"exec", false, false},             // was order 140
-	{"websearch", false, false},        // was order 141
-	{"world", false, true},             // was order 142
-	{"runtime", false, false},          // was order 143; was "bot" until the transport was named for what it is
-	{"authors", false, true},           // was order 143
-	{"bots", false, false},             // was order 143
-	{"audit", false, false},            // was order 144
-	{"affiliates", false, false},       // was order 144
-	{"sign", true, true},               // was order 145
-	{"product", false, false},          // was order 145
-	{"evals", false, false},            // was order 145
-	{"treasury", false, true},          // was order 146
-	{"admin", false, false},            // was order 146
-	{"featuregate", false, true},       // new: launch-control waitlist-mode plane (after admin, was order 147)
-	{"tasks", false, false},            // was order 147
-	{"cron", false, false},             // durable platform cron on the shared engine (post-freeze add)
-	{"automations", false, true},       // was order 148
-	{"connectorruntime", false, false}, // new: native single-connector exec via goja (after automations, HIP-0126)
-	{"tools", false, true},             // new: unified tool plane (after automations)
-	{"marketplace", false, true},       // new: marketplace over the tool plane (after tools)
-	{"referrals", false, false},        // was order 149
-	{"guide", false, true},             // new: Business AI Guide (after referrals, before ai)
-	{"company", false, true},           // new: Hanzo Company formation state machine (after guide)
-	{"chat", false, false},             // new: /v1/chat tool-calling round (before zen/ai catch-all)
-	{"zen", false, false},              // zen* claim middleware before ai's catch-all (hip-00NN)
-	{"ai", false, false},               // was order 150
-	{"plugins", false, false},          // was order 900
+	{"pubsub", false, true},          // was order 5
+	{"kafka", false, true},           // was order 6
+	{"agentskills", false, false},    // was order 8
+	{"flags", true, true},            // was order 9; native engine: /v1/flags health + store shutdown
+	{"kms", true, false},             // was order 10
+	{"metrics", false, false},        // was order 40
+	{"ingress", false, true},         // was order 42
+	{"account", false, false},        // was order 48
+	{"iam", false, false},            // was order 50
+	{"base", true, true},             // was order 60; per-org embed added Shutdown (#298)
+	{"o11y", false, true},            // ONE observability subsystem (was co-owned orders 69+70): read plane + the hanzoai/o11y module wildcard folded in as MountO11y's terminal sub-mount. OwnsHealth=false keeps /v1/o11y/health the generic always-ok route the module co-entry used to trigger.
+	{"authz", false, false},          // was order 70
+	{"commerce", false, false},       // was order 100
+	{"licensing", false, false},      // was order 110
+	{"plan", true, false},            // was order 111; enable id normalized plans->plan (routes stay /v1/plans/*)
+	{"pricing", true, false},         // was order 112
+	{"storage", true, false},         // was order 118
+	{"provisioning", false, false},   // was order 120
+	{"billing", false, false},        // was order 121
+	{"account-bridge", false, false}, // was order 122
+	{"do", false, false},             // was order 123
+	{"platform", true, false},        // was order 124
+	{"projects", false, false},       // was order 125
+	{"dns", false, false},            // new: /v1/dns zone plane (after projects)
+	{"prompts", false, false},        // was order 126
+	{"agents", false, true},          // was order 127
+	{"link", false, true},            // new: unified AI login manager (/v1/links), after agents
+	{"wallets", false, true},         // was order 127
+	{"x402", false, true},            // new: x402 pay-per-use settlement (after wallets)
+	{"paas", true, false},            // was order 128
+	{"deploy", true, false},          // after paas (release seam), before functions
+	{"functions", false, false},      // was order 128
+	{"tracker", false, false},        // was order 129
+	{"templates", false, false},      // was order 129
+	{"framework", false, true},       // was order 129
+	{"knowledge", false, false},      // was order 130
+	{"content", false, true},         // new: marketing content loop (after knowledge)
+	{"catalogsync", false, true},     // new: reverse loop (product.created → render) after content
+	{"ml", true, false},              // was order 130
+	{"usage", false, false},          // was order 131
+	{"crm", false, false},            // was order 131
+	{"marketing", false, true},       // new: marketing domain fold (after crm)
+	{"ads", false, true},             // new: ads domain fold (after crm)
+	{"social", false, true},          // new: /v1/social fold (after crm)
+	{"analytics", true, false},       // was order 132
+	{"git", false, false},            // was order 132
+	{"sync", false, true},            // /v1/sync engine (owns per-org DB handles → Shutdown)
+	{"visor", false, false},          // was order 133
+	{"captable", false, true},        // was order 133
+	{"code", false, true},            // was order 134
+	{"zero-trust", false, false},     // was order 134
+	{"dataroom", true, true},         // was order 134
+	{"graph", false, false},          // was order 135
+	{"security", true, true},         // was order 136
+	{"integrations", false, true},    // was order 137
+	{"cloudflare", false, false},     // new: /v1/cloudflare edge plane (after integrations)
+	{"sbom", true, false},            // was order 137
+	{"team", false, true},            // was order 138
+	{"settings", false, true},        // was order 138
+	{"notify", true, false},          // was order 139
+	{"gateway", false, false},        // was order 139
+	{"entitlements", false, true},    // was order 139
+	{"exec", false, false},           // was order 140
+	{"websearch", false, false},      // was order 141
+	{"world", false, true},           // was order 142
+	{"runtime", false, false},        // was order 143; was "bot" until the transport was named for what it is
+	{"authors", false, true},         // was order 143
+	{"bots", false, false},           // was order 143
+	{"audit", false, false},          // was order 144
+	{"affiliates", false, false},     // was order 144
+	{"sign", true, true},             // was order 145
+	{"product", false, false},        // was order 145
+	{"evals", false, false},          // was order 145
+	{"treasury", false, true},        // was order 146
+	{"admin", false, false},          // was order 146
+	{"admission", false, true},       // launch-control gate: composes flags (registry+seed+mode route+Enforce); Shutdown closes the registry store
+	{"tasks", false, false},          // was order 147; platform cron folded in as a sub-mount of tasks.Mount (was a separate entry)
+	{"automations", false, true},     // was order 148; connectorruntime (POST /v1/automations/connectors/:id/run) folded in as a sub-mount of automations.Mount
+	{"tools", false, true},           // new: unified tool plane (after automations)
+	{"marketplace", false, true},     // new: marketplace over the tool plane (after tools)
+	{"referrals", false, false},      // was order 149
+	{"guide", false, true},           // new: Business AI Guide (after referrals, before ai)
+	{"company", false, true},         // new: Hanzo Company formation state machine (after guide)
+	{"agent", false, false},          // new: /v1/agent tool-calling round (before zen/ai catch-all)
+	{"zen", false, false},            // zen* claim middleware before ai's catch-all (hip-00NN)
+	{"ai", false, false},             // was order 150
+	{"plugins", false, false},        // was order 900
 }
 
 // TestWireOrderMatchesFrozen proves the composition root's mount order is
@@ -136,17 +136,18 @@ func TestWireOrderMatchesFrozen(t *testing.T) {
 	}
 }
 
-// TestWireNoDuplicateEnablement guards the ONE intentional duplicate: only "o11y"
-// may appear twice (the two co-owners of the observability concept). Any other
-// duplicate name is a copy-paste bug — two specs would both mount under one enable id.
+// TestWireNoDuplicateEnablement guards that every subsystem name is unique: a name
+// maps 1:1 to an enable id, so a duplicate would mount two specs under one id. (The
+// former o11y co-ownership was collapsed — the module wildcard is now a sub-mount of
+// the in-repo o11y read plane — so there is no longer any exempt duplicate.)
 func TestWireNoDuplicateEnablement(t *testing.T) {
 	seen := map[string]int{}
 	for _, s := range Wire() {
 		seen[s.Name]++
 	}
 	for name, n := range seen {
-		if n > 1 && name != "o11y" {
-			t.Errorf("subsystem %q wired %d times (only o11y may be a co-owned duplicate)", name, n)
+		if n > 1 {
+			t.Errorf("subsystem %q wired %d times (each enable id must be unique)", name, n)
 		}
 	}
 }

apps/zen.go

@@ -5,9 +5,12 @@ package apps
 import (
 	"context"
 	"fmt"
+	"math/big"
 	"os"
 	"strings"
 
+	aicontrollers "github.com/hanzoai/ai/controllers"
+	aiobject "github.com/hanzoai/ai/object"
 	"github.com/hanzoai/cloud"
 	"github.com/hanzoai/cloud/clients/metering"
 	cloudmoney "github.com/hanzoai/cloud/clients/money"
@@ -53,11 +56,7 @@ import (
 //
 // It is wired BEFORE ai in Wire() so Claim's c.Next() falls through to ai's
 // catch-all. zen's catalog reads its upstream keys from KMS via the Key resolver.
-func mountZen(app any, deps cloud.Deps) error {
-	a, ok := app.(*zip.App)
-	if !ok {
-		return fmt.Errorf("zen.Mount: app is %T, want *zip.App", app)
-	}
+func mountZen(a *zip.App, deps cloud.Deps) error {
 	z, err := zen.New(zen.Config{
 		Logger: deps.Logger,
 		Key:    zenKeyResolver(deps.KMS),
@@ -110,11 +109,17 @@ func commerceGate(m *metering.Client) zen.Gate {
 		if t.BillingOrg == "" {
 			return fmt.Errorf("a billable tenant is required (no anonymous usage)")
 		}
-		// zen's estimate is exact 18-dp atto-USD (hanzoai/money). Fold to whole
-		// cents for the balance check via cloud's typed money.Amount.Cents() — a
-		// sub-cent estimate gates as 0 (any-positive-balance), matching the edge
-		// gate's AmountCents contract. The post-serve Meter debits the exact 18-dp.
-		cents := cloudmoney.FromInt(est.Minor()).Cents()
+		// zen's estimate is an exact 18-dp USD value. Fold it to whole cents for
+		// the balance check via cloud's typed money.Amount.Cents() — a sub-cent
+		// estimate gates as 0 (any-positive-balance), matching the edge gate's
+		// AmountCents contract. The post-serve Meter debits the exact 18-dp.
+		//
+		// The fold is Cents() on the CREDIT amount, never Minor() on the zen one:
+		// Minor() renders money.USD's 2 decimals, so it returned cents that FromInt
+		// then read as atto — every estimate came back 0, and AuthorizeVerdict skips
+		// its `available >= AmountCents` check when AmountCents is 0, admitting a
+		// request of ANY size against any positive balance.
+		cents := credit(est).Cents()
 		v, err := m.AuthorizeVerdict(ctx, metering.AuthInput{
 			User:        t.BillingOrg,
 			Org:         t.BillingOrg,
@@ -158,25 +163,106 @@ func (g commerceMeterImpl) Record(ctx context.Context, u zen.Usage) {
 	if u.Tenant.BillingOrg == "" {
 		return // never debit an unattributable request
 	}
-	usage := metering.Usage{
-		User:            u.Tenant.BillingOrg,
-		Org:             u.Tenant.BillingOrg,
-		Actor:           u.Tenant.User,
-		Model:           u.Model,
-		Provider:        zenProvider,
-		Service:         zenService,
-		Project:         u.Tenant.Project,
-		PromptTokens:    u.PromptTokens,
+	// Beside the commerce debit, land the SAME warehouse row + gen_ai span every
+	// native ai path writes (TraceServedUsage = recordTrace WITHOUT recordUsage —
+	// the debit below is the one billing source, never doubled). zen knows its
+	// EXACT per-tier retail (Charge) and upstream COGS (Cost), so the row carries
+	// true margin (credit → nano). Without this, zen* traffic is
+	// warehouse/o11y-blind exactly where prod runs (the unified binary).
+	aicontrollers.TraceServedUsage(context.Background(), aicontrollers.ServedUsage{
+		Owner:            u.Tenant.BillingOrg,
+		User:             u.Tenant.User,
+		Model:            u.Model,
+		Provider:         zenProvider,
+		RequestID:        u.RequestID,
+		Status:           "success",
+		PromptTokens:     u.PromptTokens,
 		CompletionTokens: u.CompletionTokens,
-		TotalTokens:     u.PromptTokens + u.CompletionTokens,
-		Amount:          cloudmoney.FromInt(u.Cost.Minor()), // exact 18-dp USD, no floor
-		RequestID:       u.RequestID,
-		Currency:        "usd",
-		Status:          "success",
-	}
+		BilledNano:       nano(credit(u.Charge)),
+		CostNano:         nano(credit(u.Cost)),
+	})
 	// Detached: the request context is recycled once the handler returns, so a
 	// background context carries the debit to commerce without racing the reply.
+	usage := meterUsage(u)
 	go func() { _, _ = g.m.Record(context.Background(), usage) }()
+
+	// Enso learning ledger: the embedded zen mount serves the zen catalog in-process
+	// and never reaches ai's pipeToFamily, so ai's family-event writer never runs for
+	// zen traffic. Write the SAME RoutingEvent here (source="family") through the ONE
+	// shared writer, keyed on the client-visible response id (zen.Usage.ResponseID), so
+	// zen* calls land in the same ledger — stats, world, spark retrain, and /v1/feedback
+	// all read these rows. No prompt text; no shadow (zen.Usage carries no request
+	// text — that stays the auto/enso-proxy path's job). Fire-and-forget.
+	owner := u.Tenant.Org
+	if owner == "" {
+		owner = u.Tenant.BillingOrg
+	}
+	go aiobject.RecordFamilyRouting(aiobject.FamilyRoutingInput{
+		Owner:            owner,
+		User:             u.Tenant.User,
+		RequestedModel:   u.Model,
+		RoutedModel:      u.Upstream,
+		ResponseId:       u.ResponseID,
+		PromptTokens:     u.PromptTokens,
+		CompletionTokens: u.CompletionTokens,
+		CostCents:        credit(u.Charge).Cents(),
+		RouterEndpoint:   os.Getenv("ROUTER_ENDPOINT"),
+	})
+}
+
+// meterUsage projects a served zen.Usage onto the commerce debit. It is the ONE
+// place the debit's amount is chosen, and it is pure — no ledger, no warehouse —
+// so the money property is a unit test rather than an integration.
+//
+// The amount is the RETAIL Charge: what the caller pays. Cost is the upstream
+// COGS we pay to serve the call; it is never the debit. It rides only the
+// warehouse row (CostNano), where margin = Charge − Cost stays exact. Debiting
+// Cost would collect our own COGS and book zero margin on every zen call — and
+// because the affiliate and OSS payout bases read this debit, it would fund
+// their shares out of principal. This mirrors ai, whose debit is likewise the
+// customer price (usageBilledCents), never its CostIn/CostOut COGS.
+func meterUsage(u zen.Usage) metering.Usage {
+	return metering.Usage{
+		User:             u.Tenant.BillingOrg,
+		Org:              u.Tenant.BillingOrg,
+		Actor:            u.Tenant.User,
+		Model:            u.Model,
+		Provider:         zenProvider,
+		Service:          zenService,
+		Project:          u.Tenant.Project,
+		PromptTokens:     u.PromptTokens,
+		CompletionTokens: u.CompletionTokens,
+		TotalTokens:      u.PromptTokens + u.CompletionTokens,
+		Amount:           credit(u.Charge), // exact 18-dp USD, no floor
+		RequestID:        u.RequestID,
+		Currency:         "usd",
+		Status:           "success",
+	}
+}
+
+// credit re-denominates a zen price into cloud's credit unit. It is the ONE
+// conversion at this seam — every site below goes through it, so the unit is
+// decided once rather than re-derived per call site.
+//
+// zen prices every SKU as an exact 18-dp value tagged money.USD (meter.go:
+// money.New(<18-dp decimal>, money.USD)), and cloud's credit unit is the SAME USD
+// value at 18-dp storage scale. So the conversion carries the exact decimal across
+// and changes only the minor-unit convention: no rescale, no rounding, no factor.
+// It is right by construction because the decimal is the value — the currency's
+// Decimals is a rendering convention, not part of it.
+//
+// It must NEVER go through Amount.Minor(). money.USD declares 2 decimals, so
+// Minor() rescales zen's 18-dp value to CENTS; feeding cents to the 18-dp
+// FromInt understates the debit by 10^16 (a $17.376 charge debits $0.0000000000000017),
+// and folds every sub-cent charge to a zero the ledger drops entirely.
+func credit(a hmoney.Amount) cloudmoney.Amount { return cloudmoney.FromDecimal(a.Decimal()) }
+
+// nano folds an exact credit Amount to nano-USD (1e-9) for the warehouse margin
+// columns. It takes the typed Amount rather than a bare *big.Int so the unit is
+// carried by the type: a cents integer is not a cloudmoney.Amount and can no
+// longer be passed here. A single request's cost always fits int64 at nano.
+func nano(a cloudmoney.Amount) int64 {
+	return new(big.Int).Div(a.Atto(), big.NewInt(1_000_000_000)).Int64()
 }
 
 // zenService is the commerce service axis zen* spend attributes to. zen serves

apps/zen_test.go

@@ -0,0 +1,305 @@
+// Copyright 2026 Hanzo AI Inc. All Rights Reserved.
+
+package apps
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strconv"
+	"sync/atomic"
+	"testing"
+
+	"github.com/hanzoai/cloud/clients/metering"
+	cloudmoney "github.com/hanzoai/cloud/clients/money"
+	"github.com/hanzoai/decimal"
+	hmoney "github.com/hanzoai/money"
+	"github.com/hanzoai/zen"
+)
+
+// usd builds an exact USD amount from a decimal string, the way zen prices a
+// call (18-dp native, never through float).
+//
+// The decimal here must be hanzoai/decimal, the one hmoney.New takes and the one
+// zen prices with — not shopspring's identically-named type. Money has exactly one
+// decimal; a second one that merely LOOKS like it is how a price silently becomes
+// a different number.
+func usd(t *testing.T, s string) hmoney.Amount {
+	t.Helper()
+	d, err := decimal.Parse(s)
+	if err != nil {
+		t.Fatalf("parse %q: %v", s, err)
+	}
+	return hmoney.New(d, hmoney.USD)
+}
+
+// zen5Usage is one served call at zen5's live rates: 1M in + 1M out, priced at
+// the family's 3× margin (retail = cost × margin).
+//
+//	in : cost 1.392 → retail 4.176
+//	out: cost 4.40  → retail 13.20
+func zen5Usage(t *testing.T) zen.Usage {
+	t.Helper()
+	return zen.Usage{
+		Tenant:           zen.Tenant{BillingOrg: "acme", User: "acme/alice", Project: "p1"},
+		Model:            "zen5",
+		PromptTokens:     1_000_000,
+		CompletionTokens: 1_000_000,
+		Charge:           usd(t, "17.376"), // 4.176 + 13.20 — what the caller pays
+		Cost:             usd(t, "5.792"),  // 1.392 + 4.40  — what we pay upstream
+		RequestID:        "req-1",
+	}
+}
+
+// dollars is the EXPECTED money, built from a plain dollar literal through cloud's
+// OWN ParseUSD — deliberately a DIFFERENT constructor than the code under test uses.
+//
+// That is what pins the UNIT. These tests once asserted meterUsage(u).Amount.Int()
+// against u.Charge.Minor(): both sides re-derived the number through the same
+// conversion, so the assertion only proved an integer round-tripped and was blind
+// to what the integer MEANT. It passed while every zen debit was 10^16 too small.
+// Comparing money to a known dollar amount cannot be blind that way: if the debit
+// is off by any factor, it is not $17.376 and the test fails.
+func dollars(t *testing.T, s string) cloudmoney.Amount {
+	t.Helper()
+	a, err := cloudmoney.ParseUSD(s)
+	if err != nil {
+		t.Fatalf("parse %q: %v", s, err)
+	}
+	return a
+}
+
+// The debit is the RETAIL Charge — what the caller pays — never the upstream
+// COGS. This is the money property: it fails if the Amount is built from
+// u.Cost. The tier is margin-bearing (retail != cost), so the two values are
+// distinguishable and the assertion cannot pass by coincidence.
+func TestMeterUsageDebitsRetailNotCost(t *testing.T) {
+	u := zen5Usage(t)
+
+	if u.Charge.Cmp(u.Cost) == 0 {
+		t.Fatal("fixture is not margin-bearing: retail == cost, so the test could not tell them apart")
+	}
+
+	got := meterUsage(u).Amount
+
+	// The known dollar value the fixture charges — a $17.376 call debits $17.376.
+	if want := dollars(t, "17.376"); got.Cmp(want) != 0 {
+		t.Errorf("debit = $%s, want the retail Charge $%s", got, want)
+	}
+	if cogs := dollars(t, "5.792"); got.Cmp(cogs) == 0 {
+		t.Errorf("debit = $%s, which is the upstream COGS — the caller must be billed retail, not our cost", got)
+	}
+}
+
+// At the family's 3× margin the debit is exactly 3× the COGS: we collect the
+// full retail price, not the wholesale one. Debiting Cost would collect 1/3 —
+// our own COGS — and book zero margin.
+func TestMeterUsageCollectsTheFullMargin(t *testing.T) {
+	u := zen5Usage(t)
+
+	debit := meterUsage(u).Amount
+	// 3x the COGS in the SAME unit as the debit — $5.792 + $5.792 + $5.792.
+	// Summing the credit Amount keeps the comparison in exact dollars; the old
+	// version multiplied Cost.Minor() (cents, 579 after rounding away 5.792's
+	// third decimal) and compared it to a debit that was not cents at all.
+	cogs := dollars(t, "5.792")
+	thriceCOGS := cogs.Add(cogs).Add(cogs)
+
+	if debit.Cmp(thriceCOGS) != 0 {
+		t.Errorf("debit = $%s, want 3x COGS = $%s (retail = cost x margin, margin 3.0)", debit, thriceCOGS)
+	}
+	if want := dollars(t, "17.376"); debit.Cmp(want) != 0 {
+		t.Errorf("debit = $%s, want $%s", debit, want)
+	}
+}
+
+// The debit carries zen's exact 18-dp value with no floor: a sub-cent call must
+// not round to zero on the way to the ledger.
+func TestMeterUsageKeepsExactSubCentCharge(t *testing.T) {
+	u := zen5Usage(t)
+	u.Charge = usd(t, "0.004176") // 1k input tokens at 4.176/MTok — well under a cent
+	u.Cost = usd(t, "0.001392")
+
+	got := meterUsage(u).Amount
+
+	if want := dollars(t, "0.004176"); got.Cmp(want) != 0 {
+		t.Errorf("debit = $%s, want the exact sub-cent charge $%s", got, want)
+	}
+	// metering.Record drops a zero Amount before it ever reaches the ledger
+	// (`if !c.Enabled() || amt.IsZero() ...  return nil`), so a floored sub-cent
+	// charge is not a small debit — it is NO DEBIT ROW, and the call is free.
+	if got.IsZero() {
+		t.Error("sub-cent charge floored to zero — Record drops a zero Amount, so the call is served free with no debit row")
+	}
+}
+
+// The unit trap this seam shipped with, pinned so it cannot come back. zen prices
+// an exact 18-dp value but tags it money.USD, whose Currency declares 2 decimals —
+// so Charge.Minor() renders CENTS, while cloudmoney.FromInt reads its argument as
+// 18-dp. Composing them silently divides every debit by 10^16.
+//
+// The tests above already fail if credit() regresses to that composition; this one
+// names WHY, and proves the two conversions are still distinguishable — an
+// assertion that cannot tell right from wrong is worse than no assertion.
+func TestCreditIsNotTheMinorUnit(t *testing.T) {
+	charge := usd(t, "17.376")
+
+	if got, want := credit(charge), dollars(t, "17.376"); got.Cmp(want) != 0 {
+		t.Fatalf("credit($17.376) = $%s, want $%s", got, want)
+	}
+
+	// The conversion that shipped, spelled out.
+	old := cloudmoney.FromAtto(charge.Minor())
+	if old.Cmp(credit(charge)) == 0 {
+		t.Fatal("FromInt(Minor()) agrees with credit() — the fixture can no longer tell the units apart, so these tests prove nothing")
+	}
+	if old.Cents() != 0 {
+		t.Errorf("FromInt(Minor()).Cents() = %d, want 0 — that this was ALWAYS 0 is what made the spend gate admit every request", old.Cents())
+	}
+}
+
+// The identity fields ride with the debit unchanged: the debit lands on the org
+// that PAYS (BillingOrg), scoped to its project, with the actor for the audit
+// trail.
+func TestMeterUsageAttribution(t *testing.T) {
+	u := zen5Usage(t)
+	m := meterUsage(u)
+
+	for _, c := range []struct{ name, got, want string }{
+		{"User", m.User, "acme"},
+		{"Org", m.Org, "acme"},
+		{"Actor", m.Actor, "acme/alice"},
+		{"Project", m.Project, "p1"},
+		{"Model", m.Model, "zen5"},
+		{"Provider", m.Provider, zenProvider},
+		{"Service", m.Service, zenService},
+		{"Currency", m.Currency, "usd"},
+	} {
+		if c.got != c.want {
+			t.Errorf("%s = %q, want %q", c.name, c.got, c.want)
+		}
+	}
+	if m.TotalTokens != 2_000_000 {
+		t.Errorf("TotalTokens = %d, want 2000000", m.TotalTokens)
+	}
+}
+
+// The spend gate must REFUSE a request whose estimate exceeds the balance.
+//
+// This is the sharpest edge of the unit bug and the reason it is a security
+// finding, not only a revenue one. AuthorizeVerdict gates size like this:
+//
+//	funded := available > 0
+//	if in.AmountCents > 0 { funded = available >= in.AmountCents }
+//
+// The estimate reached it as FromInt(est.Minor()).Cents(), which is ALWAYS 0 —
+// so the size branch was DEAD and every request rode `available > 0`. Any org
+// with a single cent of balance could draw an unbounded call. The debits were
+// dust too, so the balance never fell and the cap could never trip.
+func TestCommerceGateRefusesAnOverCapRequest(t *testing.T) {
+	const availableCents = 500 // the org holds $5.00
+
+	var authorized atomic.Bool
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/v1/billing/balance":
+			_, _ = io.WriteString(w, `{"available":`+strconv.Itoa(availableCents)+`}`)
+		case "/v1/billing/spend-alerts/authorize":
+			authorized.Store(true)
+			_, _ = io.WriteString(w, `{"allow":true}`)
+		default:
+			_, _ = io.WriteString(w, `[]`)
+		}
+	}))
+	defer srv.Close()
+
+	m, err := metering.New(metering.Config{BaseURL: srv.URL, Token: "t", Org: "acme"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	gate := commerceGate(m)
+	if gate == nil {
+		t.Fatal("gate is nil — the client must be enabled for this proof to mean anything")
+	}
+	tenant := zen.Tenant{BillingOrg: "acme", User: "acme/alice", Project: "p1"}
+
+	// $17.376 against a $5.00 balance: over cap, must be refused.
+	if err := gate(context.Background(), tenant, "zen5", usd(t, "17.376")); err == nil {
+		t.Error("gate ADMITTED a $17.376 request against a $5.00 balance — the estimate is reaching AuthorizeVerdict as 0 cents, so the size check never runs")
+	}
+
+	// The same balance must still admit a request it can actually cover, or the
+	// test would pass by refusing everything.
+	if err := gate(context.Background(), tenant, "zen5", usd(t, "1.00")); err != nil {
+		t.Errorf("gate refused an affordable $1.00 request against a $5.00 balance: %v", err)
+	}
+	if !authorized.Load() {
+		t.Error("the affordable request never reached the spend-cap authorize step")
+	}
+}
+
+// The estimate must reach the balance check as the RIGHT number of cents. The
+// refusal test above proves the gate says no; this proves it says no for the
+// right reason — that $17.376 is folded to 1738 cents, not to 0.
+func TestCommerceGateFoldsTheEstimateToRealCents(t *testing.T) {
+	for _, c := range []struct {
+		charge string
+		want   int64
+	}{
+		{"17.376", 1738}, // rounds half-away-from-zero at the cent
+		{"1000.00", 100000},
+		{"1.00", 100},
+		{"0.004176", 0}, // sub-cent gates as "any positive balance"
+	} {
+		if got := credit(usd(t, c.charge)).Cents(); got != c.want {
+			t.Errorf("$%s folds to %d cents, want %d", c.charge, got, c.want)
+		}
+	}
+}
+
+// A zero-value zen price must convert and fold without panicking. zen leaves
+// Charge/Cost as the zero Amount for a free SKU, and that value carries the
+// EMPTY currency code, not "USD" — so this also pins that credit() reads the
+// decimal rather than dispatching on the currency. The old attoToNano guarded a
+// nil *big.Int here; nano() needs no guard because decimal.Coef() returns a real
+// zero big.Int, never nil, but the property is worth holding.
+func TestCreditAndNanoHandleTheZeroPrice(t *testing.T) {
+	var free hmoney.Amount // zero value: no currency, no coefficient
+
+	got := credit(free)
+	if !got.IsZero() {
+		t.Errorf("credit(zero) = $%s, want $0", got)
+	}
+	if n := nano(got); n != 0 {
+		t.Errorf("nano(credit(zero)) = %d, want 0", n)
+	}
+	if c := got.Cents(); c != 0 {
+		t.Errorf("credit(zero).Cents() = %d, want 0", c)
+	}
+	// And a free call books no debit row, which is correct — nothing is owed.
+	u := zen5Usage(t)
+	u.Charge, u.Cost = free, free
+	if amt := meterUsage(u).Amount; !amt.IsZero() {
+		t.Errorf("free call debits $%s, want $0", amt)
+	}
+}
+
+// nano carries real money to the warehouse margin columns. It is the last place
+// the unit could silently collapse: attoToNano(cents) divided a cents integer by
+// 1e9 and produced 0 for every charge under $10,000,000.
+func TestNanoFoldsCreditToRealNano(t *testing.T) {
+	for _, c := range []struct {
+		charge string
+		want   int64
+	}{
+		{"17.376", 17_376_000_000},
+		{"5.792", 5_792_000_000},
+		{"0.004176", 4_176_000},
+		{"1000.00", 1_000_000_000_000},
+	} {
+		if got := nano(credit(usd(t, c.charge))); got != c.want {
+			t.Errorf("nano($%s) = %d, want %d", c.charge, got, c.want)
+		}
+	}
+}

audit_serve.go

@@ -15,6 +15,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"strings"
 	"time"
 
 	"github.com/hanzoai/cloud/audit"
@@ -57,6 +58,18 @@ func buildAuditRecorder(cfg *Config, logger luxlog.Logger) (*audit.Recorder, err
 		return nil, fmt.Errorf("open audit store: %w", err)
 	}
 
+	// PER-SHARD audit under horizontal scale. The trail lives at {DataDir}/audit.db on
+	// THIS pod's own RWO PVC, so under shard routing each pod's chain covers ONLY the
+	// tenants routed to it (its shard) — and org-scoped audit queries route to the
+	// owning shard where those records live. Soundness: the chain is a per-FILE hash
+	// chain whose head is recovered at open; because no two pods share the file, there
+	// is no cross-pod head to fork (the very failure that pinned cloud to replicas:1 was
+	// two pods on ONE audit file). Integrity is preserved WITHIN each partition; a
+	// deployment-wide view is the union of the N per-shard chains. The shard id is
+	// stamped on the AU-9 checkpoint stream below so the external tail-truncation monitor
+	// tracks N heads (one per shard) rather than expecting a single global head.
+	shard := strings.TrimSpace(cfg.ShardSelf) // "" when single-pod — a harmless empty tag
+
 	// AU-9 tail-truncation anchor: emit a periodic head-digest checkpoint to the
 	// append-only observability log (and, when a mirror supports it, an
 	// independent digest store). An external o11y monitor compares consecutive
@@ -67,7 +80,7 @@ func buildAuditRecorder(cfg *Config, logger luxlog.Logger) (*audit.Recorder, err
 	if logger != nil {
 		rec.StartCheckpoints(interval, func(cp audit.Checkpoint) {
 			logger.Info("audit_head_checkpoint",
-				"count", cp.Count, "head", cp.Head, "ts", cp.Time.Format(time.RFC3339Nano))
+				"shard", shard, "count", cp.Count, "head", cp.Head, "ts", cp.Time.Format(time.RFC3339Nano))
 		})
 	} else {
 		rec.StartCheckpoints(interval, nil)
@@ -76,7 +89,7 @@ func buildAuditRecorder(cfg *Config, logger luxlog.Logger) (*audit.Recorder, err
 	if logger != nil {
 		count, head := rec.Head()
 		logger.Info("audit trail ready (tamper-evident, append-only)",
-			"store", dbPath, "records", count, "head", head,
+			"store", dbPath, "shard", shard, "records", count, "head", head,
 			"mirror", mirror != nil, "checkpoint_interval", interval.String())
 	}
 	return rec, nil

auth_apikey.go

@@ -37,31 +37,86 @@ type keyResolver interface {
 // A brief cache keeps the hot auth path off the network; it caches misses too, so a
 // bad key cannot hammer IAM.
 type iamKeys struct {
-	base   string
-	auth   string // client_secret_basic, or "" when unconfigured
-	http   *http.Client
-	cache  cache[string, *idClaims]
+	base  string
+	auth  string // client_secret_basic, or "" when unconfigured
+	http  *http.Client
+	cache cache[string, *idClaims]
 }
 
 // newIAMKeys reads the same IAM env clients/account does. With no confidential
 // credential it returns a resolver that resolves nothing (keys stay anonymous —
 // never a fabricated principal), so a deployment lacking the credential is safe.
 func newIAMKeys() *iamKeys {
-	base := strings.TrimRight(env("IAM_URL", "IAM_INTERNAL_URL"), "/")
-	id := strings.TrimSpace(os.Getenv("IAM_MINT_CLIENT_ID"))
-	secret := strings.TrimSpace(os.Getenv("IAM_MINT_CLIENT_SECRET"))
-	var auth string
-	if id != "" && secret != "" {
-		auth = "Basic " + base64.StdEncoding.EncodeToString([]byte(id+":"+secret))
-	}
 	return &iamKeys{
-		base:  base,
-		auth:  auth,
+		base:  iamHost(),
+		auth:  iamCred(),
 		http:  &http.Client{Timeout: 5 * time.Second},
 		cache: newCache[string, *idClaims](60 * time.Second),
 	}
 }
 
+// sharedKeys memoizes ONE API-key resolver (and its 60s cache) for the whole
+// binary. The identity boundary (SanitizeIdentity, via newIdentityValidator) and
+// any subsystem that must resolve a key OUT-OF-BAND of the Authorization header
+// (analytics capture: a project key posted in the SDK body/query) both go through
+// this ONE seam, so a key resolves to the SAME org either way and IAM sees one
+// warm cache — never a second, drifting resolver.
+var (
+	sharedKeysOnce sync.Once
+	sharedKeysInst *iamKeys
+)
+
+func sharedKeys() *iamKeys {
+	sharedKeysOnce.Do(func() { sharedKeysInst = newIAMKeys() })
+	return sharedKeysInst
+}
+
+// maxKeyOrgLen bounds a resolved org key the same way principal.MaxOrgLen does: the
+// org becomes a warehouse partition key, so an over-long value (malformed / hostile)
+// is refused rather than stored.
+const maxKeyOrgLen = 128
+
+// OrgForKey resolves an opaque Hanzo API key (hk-/sk-/pk-/fw_/hz_) to the org it
+// belongs to — the SAME owner org SanitizeIdentity mints when that key arrives as a
+// bearer — through the ONE IAM key seam (get-user?accessKey). It is the exported
+// door a keyed, bearer-less SDK path uses to attribute a project key to a tenant.
+//
+// FAILS CLOSED: ("", false) for a non-key-shaped string, an unknown/unresolvable
+// key, an unconfigured resolver, or an out-of-bounds org — never a fabricated or
+// default tenant, so a bad key can never be written into another org's partition.
+// The isAPIKey prefix gate keeps garbage strings off the IAM network path.
+func OrgForKey(ctx context.Context, key string) (string, bool) {
+	key = strings.TrimSpace(key)
+	if !isAPIKey(key) {
+		return "", false
+	}
+	claims := sharedKeys().resolve(ctx, key)
+	if claims == nil {
+		return "", false
+	}
+	owner := strings.TrimSpace(claims.Owner)
+	if owner == "" || len(owner) > maxKeyOrgLen {
+		return "", false
+	}
+	return owner, true
+}
+
+// iamHost is the standalone IAM origin cloud talks to; iamCred is the service
+// credential (client_secret_basic) it presents — the ONE IAM identity, shared by
+// the API-key resolver here and the /v1/iam edge (iam_edge.go), so both
+// authenticate to IAM the same way. Empty cred → a deployment lacking the
+// credential stays safe (the caller treats "" as unconfigured).
+func iamHost() string { return strings.TrimRight(env("IAM_URL", "IAM_INTERNAL_URL"), "/") }
+
+func iamCred() string {
+	id := strings.TrimSpace(os.Getenv("IAM_MINT_CLIENT_ID"))
+	secret := strings.TrimSpace(os.Getenv("IAM_MINT_CLIENT_SECRET"))
+	if id == "" || secret == "" {
+		return ""
+	}
+	return "Basic " + base64.StdEncoding.EncodeToString([]byte(id+":"+secret))
+}
+
 func env(names ...string) string {
 	for _, n := range names {
 		if v := strings.TrimSpace(os.Getenv(n)); v != "" {

auth_identity.go

@@ -42,6 +42,7 @@ type idClaims struct {
 
 	Owner             string `json:"owner"`              // org slug (the org)
 	Project           string `json:"project"`            // org SUB-SCOPE within owner (empty ⟹ default project)
+	BillingAccount    string `json:"billing_account"`    // WHO PAYS, stated by IAM (empty ⟹ pre-claim token)
 	Name              string `json:"name"`               // display name (id fallback)
 	PreferredUsername string `json:"preferred_username"` // id fallback
 	Email             string `json:"email"`
@@ -62,6 +63,20 @@ func (c *idClaims) mintedProject() string {
 	return strings.TrimSpace(c.Project)
 }
 
+// mintedBillingAccount returns the funding account to stamp into
+// X-Billing-Account-Id, or "" when the header must be OMITTED (a token minted
+// before IAM shipped the claim, or one IAM could not attribute).
+//
+// WHO PAYS IS NOT A CLIENT'S TO NAME. This rides the validated `billing_account`
+// claim — IAM's signed statement, resolved at the identity boundary from the real
+// grant context — exactly like `owner` and `project`. It mirrors the edge
+// (iamauth.Claims.MintedBillingAccount) byte-for-byte, so the in-binary path binds
+// the same header the gateway would, and ai/object.Payer reads the same payer on
+// both. The raw client copy is deleted on ingress and NEVER restored.
+func (c *idClaims) mintedBillingAccount() string {
+	return strings.TrimSpace(c.BillingAccount)
+}
+
 // userID resolves the canonical user id: sub, then preferred_username, then
 // name. IAM may leave sub empty. This is the STABLE identifier (a UUID when IAM
 // sets sub) stamped as X-User-Id and consumed as the attribution key everywhere.
@@ -123,7 +138,7 @@ func newIdentityValidator(issuer, jwksURL string, audiences []string, ttl time.D
 		issuers:   trustedIssuers(issuer),
 		audiences: audiences,
 		cache:     newJWKSCache(jwksURL, ttl),
-		keys:      newIAMKeys(),
+		keys:      sharedKeys(), // ONE resolver+cache, shared with OrgForKey (analytics capture)
 	}
 }
 
@@ -197,6 +212,16 @@ func (v *identityValidator) validate(raw string) (*idClaims, error) {
 		return nil, err
 	}
 
+	// Fail SECURE on a misconfigured (empty) trust set: an empty issuer OR audience
+	// allowlist must REJECT every token, never silently disable that axis. In
+	// production both are always resolved non-empty (BrandIssuers + the baked
+	// audience defaults, unioned in config.go so they are "never empty"), so this
+	// fires ONLY on an operator misconfiguration (CLOUD_JWT_AUDIENCES="" emptying the
+	// resolved set, or an empty issuer set) — and then it denies, it never admits (I2).
+	if len(v.issuers) == 0 || len(v.audiences) == 0 {
+		return nil, fmt.Errorf("identity validator misconfigured: empty issuer or audience allowlist")
+	}
+
 	// Reject a missing issuer: an empty issuer must never pass the set check.
 	if claims.Issuer == "" {
 		return nil, fmt.Errorf("missing issuer")
@@ -221,14 +246,13 @@ func (v *identityValidator) validate(raw string) (*idClaims, error) {
 	// gates on owner == :org. Without this, a real client_credentials machine token
 	// (aud == its per-org clientId, never in the allowlist) fails here and the
 	// sync silently stays pending — the activation blocker.
-	expected := jwt.Expected{}
-	if len(v.audiences) > 0 {
-		auds := v.audiences
-		if mach := kmsMachineAudience(claims.Owner); mach != "" {
-			auds = append(append(make([]string, 0, len(v.audiences)+1), v.audiences...), mach)
-		}
-		expected.AnyAudience = jwt.Audience(auds)
+	// The audience allowlist is guaranteed non-empty (checked above), so the
+	// audience axis is ALWAYS enforced — never silently skipped.
+	auds := v.audiences
+	if mach := kmsMachineAudience(claims.Owner); mach != "" {
+		auds = append(append(make([]string, 0, len(v.audiences)+1), v.audiences...), mach)
 	}
+	expected := jwt.Expected{AnyAudience: jwt.Audience(auds)}
 	if err := claims.Claims.ValidateWithLeeway(expected, 2*time.Minute); err != nil {
 		return nil, fmt.Errorf("claims: %w", err)
 	}
@@ -425,14 +449,11 @@ func trustedIssuers(primary string) []string {
 	return out
 }
 
-// issuerAllowed reports whether iss is one of the trusted issuers. An empty set
-// (no primary, no brands — never the case in production) skips the check, matching
-// the prior "empty issuer disables the check" behavior; a non-empty set is
-// fail-secure (a token whose iss is not in the set is rejected).
+// issuerAllowed reports whether iss is one of the trusted issuers. It is
+// fail-secure in BOTH directions: an empty trusted set matches NOTHING (deny), so
+// a misconfiguration that empties the issuer allowlist rejects every token instead
+// of silently disabling the check (I2); a non-empty set rejects any iss not in it.
 func issuerAllowed(iss string, trusted []string) bool {
-	if len(trusted) == 0 {
-		return true
-	}
 	for _, t := range trusted {
 		if iss == t {
 			return true

auth_identity_whitelabel_test.go

@@ -1,10 +1,45 @@
 package cloud
 
 import (
+	"crypto/rand"
+	"crypto/rsa"
 	"os"
 	"testing"
+	"time"
 )
 
+// TestValidate_FailSecureOnEmptyTrustSet proves I2: a validator whose resolved
+// issuer OR audience allowlist is empty REJECTS an otherwise-valid, correctly
+// signed token — the axis is never silently disabled. Production always resolves
+// non-empty sets; this guards the misconfiguration path (CLOUD_JWT_AUDIENCES=""
+// or an empty issuer set), which must fail closed, not open.
+func TestValidate_FailSecureOnEmptyTrustSet(t *testing.T) {
+	key, err := rsa.GenerateKey(rand.Reader, 2048)
+	if err != nil {
+		t.Fatalf("genkey: %v", err)
+	}
+	jwks := jwksServer(t, &key.PublicKey)
+	future := time.Now().Add(time.Hour)
+	tok := signWith(t, key, tokenClaims("hanzo-console", "acme", "", false, future))
+
+	// Sanity: a properly configured validator accepts the token.
+	if _, err := newIdentityValidator(testIssuer, jwks.URL, []string{"hanzo-console"}, 0).validate(tok); err != nil {
+		t.Fatalf("baseline valid token must be accepted, got %v", err)
+	}
+
+	// Empty audience set → deny.
+	if _, err := newIdentityValidator(testIssuer, jwks.URL, nil, 0).validate(tok); err == nil {
+		t.Error("empty audience allowlist must REJECT (fail-secure), not accept")
+	}
+
+	// Empty issuer set → deny (construct directly; trustedIssuers never yields empty
+	// with a primary, so bypass it to exercise the guard).
+	vEmptyIss := &identityValidator{issuers: nil, audiences: []string{"hanzo-console"}, cache: newJWKSCache(jwks.URL, 0), keys: newIAMKeys()}
+	if _, err := vEmptyIss.validate(tok); err == nil {
+		t.Error("empty issuer allowlist must REJECT (fail-secure), not accept")
+	}
+}
+
 // TestTrustedIssuers_WhiteLabel proves the in-binary validator's trusted-issuer
 // set is the primary issuer UNIONED with every white-label brand issuer plus the
 // WHITELABEL_ISSUERS override, deduped, primary-first.
@@ -40,7 +75,8 @@ func TestTrustedIssuers_WhiteLabel(t *testing.T) {
 }
 
 // TestIssuerAllowed proves the set membership check: brand issuers pass, an
-// outsider is rejected, and an empty set (never in prod) skips the check.
+// outsider is rejected, and an empty set is fail-secure — it matches NOTHING (I2),
+// so a misconfiguration that empties the allowlist denies every token.
 func TestIssuerAllowed(t *testing.T) {
 	set := []string{"https://hanzo.id", "https://lux.id"}
 	if !issuerAllowed("https://lux.id", set) {
@@ -49,8 +85,8 @@ func TestIssuerAllowed(t *testing.T) {
 	if issuerAllowed("https://attacker.id", set) {
 		t.Error("attacker.id must be rejected")
 	}
-	if !issuerAllowed("anything", nil) {
-		t.Error("empty set must skip the check (matches prior empty-issuer behavior)")
+	if issuerAllowed("anything", nil) {
+		t.Error("empty set must DENY (fail-secure), never skip the check")
 	}
 }
 

build.go

@@ -14,7 +14,7 @@ import (
 
 	"github.com/hanzoai/cloud/clients"
 	"github.com/hanzoai/cloud/clients/finance"
-	"github.com/hanzoai/cloud/clients/gatewaypolicy"
+	"github.com/hanzoai/cloud/clients/gateway/edge"
 	"github.com/hanzoai/cloud/clients/money"
 	"github.com/hanzoai/cloud/clients/s3admin"
 	"github.com/hanzoai/cloud/types"
@@ -90,6 +90,7 @@ func BuildDeps(cfg *Config) Deps {
 	// commerce URL yields a !Enabled() client, so the wrap is a transparent
 	// pass-through and a dev deployment is never blocked.
 	deps.Metering = buildMeteringClient(cfg, logger)
+	wireTierReader(deps.Metering, logger)
 	deps.AI = meteredAIClient(pickAIClient(cfg, logger), deps)
 	wireFinance(cfg, logger)
 	deps.O11y = pick(cfg, logger, "o11y", "O11y", cfg.O11yZAPAddr, clients.O11yRPCAt, clients.DisabledO11y)
@@ -107,7 +108,7 @@ func BuildDeps(cfg *Config) Deps {
 	// working *Store (static-only if the SQLite file can't open), so the edge
 	// middleware is never left without a policy source — a store-open error is
 	// logged, not fatal.
-	gp, err := gatewaypolicy.New(cfg.DataDir, cfg.AdminOrg, staticEdgePolicy(cfg))
+	gp, err := edge.New(cfg.DataDir, cfg.AdminOrg, staticEdgePolicy(cfg))
 	if err != nil {
 		logger.Warn("gateway policy store degraded to static-only", "err", err)
 	}
@@ -117,10 +118,10 @@ func BuildDeps(cfg *Config) Deps {
 }
 
 // staticEdgePolicy projects the static env/flag edge config into the boot-default
-// policy the gatewaypolicy.Store layers runtime overrides on top of. A disabled
+// policy the edge.Store layers runtime overrides on top of. A disabled
 // per-IP limiter (CLOUD_EDGE_RATELIMIT=false) maps to PerIPRPM 0 (a live no-op).
-func staticEdgePolicy(cfg *Config) gatewaypolicy.Policy {
-	p := gatewaypolicy.Policy{
+func staticEdgePolicy(cfg *Config) edge.Policy {
+	p := edge.Policy{
 		CORSOrigins: cfg.CORSOrigins,
 		WindowSec:   cfg.EdgeRateWindowSec,
 	}
@@ -157,9 +158,15 @@ func buildMeteringClient(cfg *Config, log luxlog.Logger) *metering.Client {
 		httpClient = commerceinproc.Client(0) // in-process dispatch; no network timeout
 	}
 	m, err := metering.New(metering.Config{
-		BaseURL:    base,
-		Token:      cfg.CommerceServiceToken,
-		Org:        cfg.Brand, // X-Org-Id default for S2S; per-request org overrides.
+		BaseURL: base,
+		Token:   cfg.CommerceServiceToken,
+		Org:     cfg.Brand, // X-Org-Id default for S2S; per-request org overrides.
+		// Honor the documented METERING_TEST env: when "true", route every debit to
+		// commerce's TEST/sandbox books (fin.RecordUsage in.Test=true) so a staging /
+		// canary deployment records NO real money — and the usage-cap read (org.TestMode
+		// via SQUARE_ENVIRONMENT=sandbox) sees the SAME test books. Unset in prod → live,
+		// unchanged. Without this the flag was silently ignored (always live).
+		Test:       strings.EqualFold(strings.TrimSpace(os.Getenv(metering.EnvTest)), "true"),
 		FailOpen:   cfg.BillingFailOpen,
 		HTTPClient: httpClient, // nil off the co-resident path → metering builds its own
 	})
@@ -169,6 +176,11 @@ func buildMeteringClient(cfg *Config, log luxlog.Logger) *metering.Client {
 		log.Error("billing: invalid commerce URL, gate disabled", "err", err)
 		m, _ = metering.New(metering.Config{})
 	}
+	// Observe every cap-check fail-open (timeout / slow / broken commerce) — a cap that
+	// silently allows must never be silent. The completion still proceeds (fail-open).
+	metering.OnCapError = func(err error) {
+		log.Warn("spend-cap check failed open (allowing completion) — commerce authorize slow/unavailable", "err", err)
+	}
 	if m.Enabled() {
 		log.Info("billing gate enabled", "commerce", boolStr(inProcess, "in-process", "http:"+base), "fail_open", cfg.BillingFailOpen)
 	} else {
@@ -184,6 +196,28 @@ func boolStr(b bool, t, f string) string {
 	return f
 }
 
+// wireTierReader installs the embedded ai module's per-tier SKU gate reader so it
+// resolves the caller's commerce subscription tier through the SAME co-resident
+// commerce client the metering gate bills over — in-process (commerceinproc) when
+// commerce is folded in, S2S HTTP with the service token otherwise — NEVER an authed
+// self-call to the cloud edge. That self-call is the toothless-gate bug: the edge
+// 401/403s a service call to /v1/billing/*, so the ai module's own HTTP lookup always
+// returned "" in-cluster and every tier-gated SKU failed OPEN. This mirrors
+// wireFinance's SetBalanceReader: cloud owns the co-resident read, ai stays
+// transport-agnostic. Fail-safe is preserved — Client.Tier folds a commerce error or
+// an unknown plan to "", which the gate treats as ALLOW, so a commerce blip never
+// locks out a paying caller. No-op when commerce is unreachable (metering !Enabled),
+// leaving ai's standalone HTTP fallback in place.
+func wireTierReader(m *metering.Client, log luxlog.Logger) {
+	if m == nil || !m.Enabled() {
+		return
+	}
+	aiobject.SetTierReader(func(ctx context.Context, subject, namespace string) (string, error) {
+		return m.Tier(ctx, subject, namespace)
+	})
+	log.Info("ai per-tier SKU gate wired to co-resident commerce (in-process tier read, fail-safe)")
+}
+
 // wireFinance constructs the ONE in-process finance ledger (per-org SQLite
 // double-entry prepaid wallet), publishes it for every money consumer to resolve by
 // the narrow finance.Client, and installs the embedded ai router's balance-read +
@@ -656,30 +690,17 @@ func pickVaultClient(cfg *Config, log luxlog.Logger) VaultClient {
 	return clients.DisabledVault()
 }
 
-// MountFunc is a subsystem's mount contract. app is `any`, not *zip.App, and that
-// is load-bearing: some external modules expose Mount as func(any, Deps) error
-// (e.g. hanzoai/licensing), which apps.Wire references DIRECTLY — a
-// func(any,…) value is not assignable to a func(*zip.App,…) parameter, so
-// narrowing the type would break them at compile time. The concrete value is
-// always a *zip.App; strongly-typed Mounts (func(*zip.App, Deps) error, what every
-// in-repo subsystem exports) are adapted by Typed, which recovers it in ONE place.
-type MountFunc func(app any, deps Deps) error
-
-// Typed adapts a strongly-typed subsystem Mount — func(*zip.App, Deps) error,
-// the signature every in-repo subsystem already exports — into the registry's
-// MountFunc. It performs the *zip.App recovery in ONE place, fail-closed with a
-// clear error, so no subsystem repeats the `a, ok := app.(*zip.App)` boilerplate.
-// The concrete value MountAll passes is always a *zip.App, so the assertion is
-// total in practice; it stays as a defensive, self-documenting guard.
-func Typed(mount func(*zip.App, Deps) error) MountFunc {
-	return func(app any, deps Deps) error {
-		a, ok := app.(*zip.App)
-		if !ok {
-			return fmt.Errorf("cloud.Mount: app is %T, want *zip.App", app)
-		}
-		return mount(a, deps)
-	}
-}
+// MountFunc is a subsystem's mount contract: register your routes on app, using
+// deps for everything shared. Every subsystem in the fleet exports exactly this
+// signature, so Wire references each one directly and the compiler checks it.
+//
+// app was once `any`, on the stated grounds that an external module (licensing)
+// exposed func(any, Deps) error and narrowing would break it — while licensing
+// said it used `any` to avoid an import cycle in pkg/cloud. Each cited the other,
+// and the cycle could not exist: this package already imports zip, and zip does
+// not import cloud. The `any` bought nothing and cost every subsystem a Typed()
+// wrapper plus a runtime type assertion whose failure branch was unreachable.
+type MountFunc func(app *zip.App, deps Deps) error
 
 // ShutdownFunc releases a subsystem's process-lifetime resources (background
 // goroutines, open DB handles) on graceful shutdown. It must be idempotent and
@@ -704,8 +725,7 @@ type MountSpec struct {
 
 // MountAll mounts every ENABLED subsystem in specs, in slice order — the order is
 // the composition root's (apps.Wire()); MountAll does NOT sort. app is the
-// concrete *zip.App from Serve; the MountFunc accepts it as `any` and in-repo
-// subsystems recover it via Typed.
+// concrete *zip.App from Serve, handed to each MountFunc as itself.
 //
 // Teardown is wired HERE, at mount time: right after a subsystem mounts, its
 // ShutdownFunc (if any) is registered via app.OnShutdown. zip drains those hooks

build_onshutdown_test.go

@@ -16,7 +16,7 @@ import (
 
 // noopMount mounts nothing: the fake specs below carry the behavior under test in
 // their Shutdown, not their Mount.
-func noopMount(any, cloud.Deps) error { return nil }
+func noopMount(*zip.App, cloud.Deps) error { return nil }
 
 // freeAddr reserves an ephemeral loopback port and hands back its address; the
 // listener is closed so the app under test can bind it.

build_registration_test.go

@@ -1,49 +1,28 @@
 package cloud_test
 
 import (
-	"strings"
 	"testing"
 
 	"github.com/hanzoai/cloud"
 	"github.com/zap-proto/zip"
 )
 
-// TestTyped_RecoversZipApp verifies cloud.Typed adapts a strongly-typed
-// func(*zip.App, Deps) into the registry MountFunc: it hands the concrete
-// *zip.App straight through to the wrapped mount.
-func TestTyped_RecoversZipApp(t *testing.T) {
-	app := zip.New(zip.Config{})
-	var got *zip.App
-	mf := cloud.Typed(func(a *zip.App, _ cloud.Deps) error {
-		got = a
-		return nil
-	})
-	if err := mf(app, cloud.Deps{}); err != nil {
-		t.Fatalf("Typed mount returned error: %v", err)
-	}
-	if got != app {
-		t.Fatalf("Typed did not pass the concrete *zip.App through (got %p, want %p)", got, app)
-	}
-}
-
-// TestTyped_WrongTypeFailsClosed verifies cloud.Typed fails closed with a clear
-// error — never a panic — when the registry passes a value that is not a
-// *zip.App. This is the single, central replacement for the per-subsystem
-// assertion boilerplate.
-func TestTyped_WrongTypeFailsClosed(t *testing.T) {
-	called := false
-	mf := cloud.Typed(func(*zip.App, cloud.Deps) error {
-		called = true
-		return nil
-	})
-	err := mf("not-a-zip-app", cloud.Deps{})
-	if err == nil {
-		t.Fatal("Typed must return an error on a non-*zip.App value")
-	}
-	if called {
-		t.Fatal("Typed must NOT invoke the wrapped mount on a type mismatch")
-	}
-	if !strings.Contains(err.Error(), "*zip.App") {
-		t.Errorf("error should name the wanted type *zip.App, got: %v", err)
-	}
+// TestMountFunc_IsTheSubsystemSignature pins the registry's mount contract: the
+// signature every subsystem exports IS a cloud.MountFunc, checked by the compiler.
+//
+// This file used to test cloud.Typed, the adapter that took a MountFunc's `any`
+// app and asserted it back to *zip.App. Both of its tests went with it, and
+// neither is a loss:
+//
+//   - "Typed recovers the *zip.App" only ever proved the adapter handed through
+//     the value it was given. MountFunc now names *zip.App, so there is no
+//     recovery step left to get wrong.
+//   - "Typed fails closed on a wrong type" can no longer be written: passing
+//     "not-a-zip-app" to a MountFunc is a compile error, so the runtime branch it
+//     exercised does not exist. A test asserting a wrong type is rejected is
+//     precisely what a type already is.
+//
+// What remains is the only claim worth making, and the build enforces it.
+func TestMountFunc_IsTheSubsystemSignature(t *testing.T) {
+	var _ cloud.MountFunc = func(*zip.App, cloud.Deps) error { return nil }
 }

caller_bearer_test.go

@@ -0,0 +1,51 @@
+package cloud
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	luxlog "github.com/luxfi/log"
+	"github.com/zap-proto/zip"
+)
+
+// CallerBearer relays the caller's OWN validated JWT bearer and nothing else: a JWT
+// passes through unchanged, an opaque API key is not relayable, and no credential
+// yields "". This is the token a downstream org-scoped service (the DNS forward
+// head) re-validates to enforce tenant isolation across the hop.
+func TestCallerBearer(t *testing.T) {
+	app := zip.New(zip.Config{Logger: luxlog.New("test")})
+	app.Get("/probe", func(c *zip.Ctx) error { return c.Bytes(200, []byte(CallerBearer(c))) })
+
+	probe := func(setup func(*http.Request)) string {
+		req := httptest.NewRequest(http.MethodGet, "/probe", nil)
+		if setup != nil {
+			setup(req)
+		}
+		res, err := app.Fiber().Test(req)
+		if err != nil {
+			t.Fatal(err)
+		}
+		b, _ := io.ReadAll(res.Body)
+		_ = res.Body.Close()
+		return string(b)
+	}
+
+	cases := []struct {
+		name  string
+		setup func(*http.Request)
+		want  string
+	}{
+		{"jwt bearer relayed unchanged", func(r *http.Request) { r.Header.Set("Authorization", "Bearer jwt.header.sig") }, "jwt.header.sig"},
+		{"X-Authorization fallback", func(r *http.Request) { r.Header.Set("X-Authorization", "Bearer x.y.z") }, "x.y.z"},
+		{"opaque hk- api key is NOT relayable", func(r *http.Request) { r.Header.Set("Authorization", "Bearer hk-secret") }, ""},
+		{"opaque sk- api key is NOT relayable", func(r *http.Request) { r.Header.Set("Authorization", "Bearer sk-secret") }, ""},
+		{"no credential yields empty", nil, ""},
+	}
+	for _, c := range cases {
+		if got := probe(c.setup); got != c.want {
+			t.Errorf("%s: CallerBearer = %q, want %q", c.name, got, c.want)
+		}
+	}
+}

cli/auth.go

@@ -160,24 +160,21 @@ type loginFlags struct {
 }
 
 func runLogin(env *Env, lf *loginFlags, cmd *cobra.Command) error {
-	creds, err := LoadCredentials()
-	if err != nil {
-		return err
-	}
+	var creds *Credentials
 
 	switch {
 	case lf.token != "":
 		// Paste an externally-minted token. Decode claims for identity.
-		tr := &tokenResp{AccessToken: lf.token, TokenType: "Bearer"}
-		creds = credsFromToken(tr)
+		creds = credsFromToken(&tokenResp{AccessToken: lf.token, TokenType: "Bearer"})
 	case lf.username != "" || lf.passwordStdin:
 		// Password grant — kept for automation (--username/--password-stdin).
 		username := lf.username
 		if username == "" {
-			username, err = prompt(cmd, "Email: ")
+			u, err := prompt(cmd, "Email: ")
 			if err != nil {
 				return err
 			}
+			username = u
 		}
 		password, err := readPassword(cmd, lf.passwordStdin)
 		if err != nil {
@@ -192,14 +189,15 @@ func runLogin(env *Env, lf *loginFlags, cmd *cobra.Command) error {
 	default:
 		// The ONE interactive way: RFC 8628 device flow — link + QR + code,
 		// approve from any signed-in browser or phone. Headless-safe.
-		creds, err = runDeviceLogin(cmd, env, lf.scope)
+		c, err := runDeviceLogin(cmd, env, lf.scope)
 		if err != nil {
 			return err
 		}
+		creds = c
 	}
 
 	// Optional machine-to-machine tokens for the platform control plane,
-	// stored alongside the identity so apps/deploy work post-login.
+	// stored with this identity so apps/deploy work post-login.
 	if lf.platformToken != "" {
 		creds.PlatformToken = lf.platformToken
 	}
@@ -207,14 +205,23 @@ func runLogin(env *Env, lf *loginFlags, cmd *cobra.Command) error {
 		creds.BuildToken = lf.buildToken
 	}
 
-	if err := creds.Save(); err != nil {
+	// Persist under this identity's stable key and make it active. A second
+	// login as a different owner (e.g. admin vs hanzo for the same email, via a
+	// different --client-id) is stored beside the first, never over it;
+	// credentials.json mirrors whichever is active for legacy readers.
+	store, err := LoadIdentities()
+	if err != nil {
+		return err
+	}
+	key := store.Put(creds)
+	if err := store.Save(); err != nil {
 		return err
 	}
 	who := firstNonEmpty(creds.Subject, "(unknown)")
 	if creds.Owner != "" {
 		who += " @ " + creds.Owner
 	}
-	fmt.Fprintf(cmd.OutOrStdout(), "Logged in as %s (token expires %s)\n", who, shortTime(creds.Expiry))
+	fmt.Fprintf(cmd.OutOrStdout(), "Logged in as %s [%s] (token expires %s)\n", who, key, shortTime(creds.Expiry))
 	return nil
 }
 
@@ -249,28 +256,92 @@ func bindLoginFlags(cmd *cobra.Command, lf *loginFlags) {
 
 func newLogoutCmd() *cobra.Command {
 	return &cobra.Command{
-		Use:               "logout",
-		Short:             "Remove stored credentials",
-		Args:              cobra.NoArgs,
+		Use:               "logout [<owner>]",
+		Short:             "Remove a stored identity (the active one, or the named owner)",
+		Args:              cobra.MaximumNArgs(1),
 		PersistentPreRunE: func(*cobra.Command, []string) error { return nil },
-		RunE: func(cmd *cobra.Command, _ []string) error {
-			if err := DeleteCredentials(); err != nil {
+		RunE: func(cmd *cobra.Command, args []string) error {
+			store, err := LoadIdentities()
+			if err != nil {
 				return err
 			}
-			fmt.Fprintln(cmd.OutOrStdout(), "Logged out.")
+			if len(store.Identities) == 0 {
+				fmt.Fprintln(cmd.OutOrStdout(), "Not logged in.")
+				return nil
+			}
+			target := store.Active
+			if len(args) == 1 {
+				if target, err = store.resolve(args[0]); err != nil {
+					return err
+				}
+			}
+			store.Remove(target)
+			if err := store.Save(); err != nil {
+				return err
+			}
+			msg := "Logged out of " + target + "."
+			if store.Active != "" {
+				msg += " Active is now " + store.Active + "."
+			}
+			fmt.Fprintln(cmd.OutOrStdout(), msg)
 			return nil
 		},
 	}
 }
 
+// identityRow is the JSON/table projection of one stored identity.
+type identityRow struct {
+	Key     string `json:"key"`
+	Owner   string `json:"owner"`
+	Subject string `json:"subject"`
+	Expiry  int64  `json:"expiry,omitempty"`
+	Active  bool   `json:"active"`
+}
+
+// listIdentities renders every stored identity (active marked with *) — the
+// shared body of `hanzo auth list` and `hanzo whoami --all`.
+func listIdentities(env *Env, _ *cobra.Command) error {
+	store, err := LoadIdentities()
+	if err != nil {
+		return err
+	}
+	rows := make([]identityRow, 0, len(store.Identities))
+	for _, k := range store.keys() {
+		c := store.Identities[k]
+		rows = append(rows, identityRow{
+			Key: k, Owner: c.Owner, Subject: c.Subject, Expiry: c.Expiry,
+			Active: k == store.Active,
+		})
+	}
+	return env.emit(rows, func(w io.Writer) {
+		if len(rows) == 0 {
+			fmt.Fprintln(w, "No stored identities. Run `hanzo login`.")
+			return
+		}
+		tw := newTab(w)
+		fmt.Fprintln(tw, "ACTIVE\tKEY\tOWNER\tSUBJECT\tEXPIRES")
+		for _, r := range rows {
+			active := ""
+			if r.Active {
+				active = "*"
+			}
+			fmt.Fprintf(tw, "%s\t%s\t%s\t%s\t%s\n", active, r.Key, r.Owner, r.Subject, shortTime(r.Expiry))
+		}
+		tw.Flush()
+	})
+}
+
 func newWhoamiCmd(envOf func() *Env) *cobra.Command {
-	var verify bool
+	var verify, all bool
 	cmd := &cobra.Command{
 		Use:   "whoami",
-		Short: "Show the current identity from the stored token",
+		Short: "Show the active identity from the stored token (--all lists every stored identity)",
 		Args:  cobra.NoArgs,
 		RunE: func(cmd *cobra.Command, _ []string) error {
 			env := envOf()
+			if all {
+				return listIdentities(env, cmd)
+			}
 			tok := env.accessToken()
 			if tok == "" {
 				return fmt.Errorf("not logged in: run `hanzo login`")
@@ -300,6 +371,7 @@ func newWhoamiCmd(envOf func() *Env) *cobra.Command {
 		},
 	}
 	cmd.Flags().BoolVar(&verify, "verify", false, "verify the token against the IAM userinfo endpoint")
+	cmd.Flags().BoolVar(&all, "all", false, "list every stored identity (active marked with *)")
 	return cmd
 }
 
@@ -329,11 +401,11 @@ func verifyUserInfo(ctx context.Context, env *Env, token string) error {
 func newAuthCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 	cmd := &cobra.Command{
 		Use:   "auth",
-		Short: "Manage authentication",
+		Short: "Manage authentication and stored identities",
 	}
 	tokenCmd := &cobra.Command{
 		Use:   "token",
-		Short: "Print the stored access token",
+		Short: "Print the active access token",
 		Args:  cobra.NoArgs,
 		RunE: func(cmd *cobra.Command, _ []string) error {
 			tok := envOf().accessToken()
@@ -344,7 +416,41 @@ func newAuthCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 			return nil
 		},
 	}
-	cmd.AddCommand(newLoginCmd(envOf, gf), newLogoutCmd(), newWhoamiCmd(envOf), tokenCmd)
+	listCmd := &cobra.Command{
+		Use:     "list",
+		Aliases: []string{"ls", "identities"},
+		Short:   "List stored identities (active marked with *)",
+		Args:    cobra.NoArgs,
+		RunE:    func(cmd *cobra.Command, _ []string) error { return listIdentities(envOf(), cmd) },
+	}
+	switchCmd := &cobra.Command{
+		Use:     "switch <owner>",
+		Aliases: []string{"use"},
+		Short:   "Make a stored identity active (accepts an owner, or a full owner/name key)",
+		Args:    cobra.ExactArgs(1),
+		RunE: func(cmd *cobra.Command, args []string) error {
+			store, err := LoadIdentities()
+			if err != nil {
+				return err
+			}
+			key, err := store.resolve(args[0])
+			if err != nil {
+				return err
+			}
+			store.Active = key
+			if err := store.Save(); err != nil {
+				return err
+			}
+			c := store.Identities[key]
+			who := firstNonEmpty(c.Subject, "(unknown)")
+			if c.Owner != "" {
+				who += " @ " + c.Owner
+			}
+			fmt.Fprintf(cmd.OutOrStdout(), "Switched to %s [%s] (token expires %s)\n", who, key, shortTime(c.Expiry))
+			return nil
+		},
+	}
+	cmd.AddCommand(newLoginCmd(envOf, gf), newLogoutCmd(), newWhoamiCmd(envOf), tokenCmd, listCmd, switchCmd)
 	return cmd
 }
 

cli/auth_test.go

@@ -208,3 +208,178 @@ func TestAuthTokenCommand(t *testing.T) {
 		t.Fatalf("auth token output: %q", out)
 	}
 }
+
+// TestMultiIdentityLoginSwitch is the full multi-identity story: two logins for
+// the same email under different owners (admin vs hanzo — the privilege-
+// separation case) coexist, `auth list` shows both, `switch` flips the active
+// pointer and rewrites credentials.json, and legacy single-file readers always
+// see the active identity.
+func TestMultiIdentityLoginSwitch(t *testing.T) {
+	sandbox(t)
+	adminTok := makeJWT(map[string]any{"email": "z@hanzo.ai", "owner": "admin", "sub": "u-admin", "exp": float64(2000000000)})
+	hanzoTok := makeJWT(map[string]any{"email": "z@hanzo.ai", "owner": "hanzo", "sub": "u-hanzo", "exp": float64(2000000001)})
+
+	// First login → admin/z is stored and active.
+	out, err := runRoot(t, "", "login", "--token", adminTok)
+	if err != nil {
+		t.Fatalf("login admin: %v", err)
+	}
+	if !strings.Contains(out, "admin/z") {
+		t.Fatalf("login should report the key: %q", out)
+	}
+	if c, _ := LoadCredentials(); c.Owner != "admin" || c.Subject != "z@hanzo.ai" {
+		t.Fatalf("active not admin after first login: %+v", c)
+	}
+
+	// Second login (different owner) → added beside admin/z, becomes active,
+	// does NOT clobber the first.
+	if _, err := runRoot(t, "", "login", "--token", hanzoTok); err != nil {
+		t.Fatalf("login hanzo: %v", err)
+	}
+	store, err := LoadIdentities()
+	if err != nil {
+		t.Fatalf("load identities: %v", err)
+	}
+	if len(store.Identities) != 2 {
+		t.Fatalf("want 2 identities, got %d: %v", len(store.Identities), store.keys())
+	}
+	if store.Identities["admin/z"] == nil || store.Identities["hanzo/z"] == nil {
+		t.Fatalf("both identities must persist, got %v", store.keys())
+	}
+	if store.Active != "hanzo/z" {
+		t.Fatalf("active = %q, want hanzo/z (last login)", store.Active)
+	}
+	// Legacy reader sees the active (hanzo) identity.
+	if c, _ := LoadCredentials(); c.Owner != "hanzo" {
+		t.Fatalf("credentials.json not mirroring active: %+v", c)
+	}
+
+	// auth list shows both, with the active row marked.
+	out, err = runRoot(t, "", "auth", "list")
+	if err != nil {
+		t.Fatalf("auth list: %v", err)
+	}
+	for _, want := range []string{"admin/z", "hanzo/z", "z@hanzo.ai", "*"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("auth list missing %q in:\n%s", want, out)
+		}
+	}
+
+	// switch admin → active flips + credentials.json is rewritten to admin.
+	if _, err := runRoot(t, "", "auth", "switch", "admin"); err != nil {
+		t.Fatalf("auth switch admin: %v", err)
+	}
+	if st, _ := LoadIdentities(); st.Active != "admin/z" {
+		t.Fatalf("active after switch = %q, want admin/z", st.Active)
+	}
+	if c, _ := LoadCredentials(); c.Owner != "admin" || c.Subject != "z@hanzo.ai" {
+		t.Fatalf("switch did not rewrite credentials.json: %+v", c)
+	}
+
+	// whoami (top-level, reads the active token) reflects admin.
+	out, err = runRoot(t, "", "whoami")
+	if err != nil {
+		t.Fatalf("whoami: %v", err)
+	}
+	if !strings.Contains(out, "admin") || !strings.Contains(out, "z@hanzo.ai") {
+		t.Fatalf("whoami not reflecting the switched-to identity: %q", out)
+	}
+
+	// switch by the full owner/name key works too.
+	if _, err := runRoot(t, "", "auth", "switch", "hanzo/z"); err != nil {
+		t.Fatalf("auth switch hanzo/z: %v", err)
+	}
+	if c, _ := LoadCredentials(); c.Owner != "hanzo" {
+		t.Fatalf("switch by full key failed: %+v", c)
+	}
+}
+
+// TestAuthListJSON checks the machine-readable projection.
+func TestAuthListJSON(t *testing.T) {
+	sandbox(t)
+	tok := makeJWT(map[string]any{"email": "z@hanzo.ai", "owner": "admin", "sub": "a"})
+	if _, err := runRoot(t, "", "login", "--token", tok); err != nil {
+		t.Fatalf("login: %v", err)
+	}
+	out, err := runRoot(t, "", "auth", "list", "-o", "json")
+	if err != nil {
+		t.Fatalf("auth list json: %v", err)
+	}
+	var rows []identityRow
+	if err := json.Unmarshal([]byte(out), &rows); err != nil {
+		t.Fatalf("json unmarshal: %v\n%s", err, out)
+	}
+	if len(rows) != 1 || rows[0].Key != "admin/z" || rows[0].Owner != "admin" || !rows[0].Active {
+		t.Fatalf("json rows wrong: %+v", rows)
+	}
+}
+
+// TestLogoutOneOfMany removes a single identity and, only when the last one is
+// gone, clears the store entirely.
+func TestLogoutOneOfMany(t *testing.T) {
+	sandbox(t)
+	admin := makeJWT(map[string]any{"email": "z@hanzo.ai", "owner": "admin", "sub": "a"})
+	hanzo := makeJWT(map[string]any{"email": "z@hanzo.ai", "owner": "hanzo", "sub": "h"})
+	if _, err := runRoot(t, "", "login", "--token", admin); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := runRoot(t, "", "login", "--token", hanzo); err != nil { // active = hanzo/z
+		t.Fatal(err)
+	}
+
+	// logout of the named owner (admin) leaves hanzo/z active.
+	if _, err := runRoot(t, "", "logout", "admin"); err != nil {
+		t.Fatalf("logout admin: %v", err)
+	}
+	store, _ := LoadIdentities()
+	if store.Identities["admin/z"] != nil {
+		t.Fatalf("admin/z not removed: %v", store.keys())
+	}
+	if store.Active != "hanzo/z" {
+		t.Fatalf("active = %q, want hanzo/z", store.Active)
+	}
+	if c, _ := LoadCredentials(); c.Owner != "hanzo" {
+		t.Fatalf("credentials.json not mirroring survivor: %+v", c)
+	}
+
+	// logout of the active (no arg) removes the last identity → both files gone.
+	if _, err := runRoot(t, "", "logout"); err != nil {
+		t.Fatalf("logout active: %v", err)
+	}
+	if c, _ := LoadCredentials(); c.AccessToken != "" {
+		t.Fatalf("credentials.json not cleared: %+v", c)
+	}
+	if st, _ := LoadIdentities(); len(st.Identities) != 0 {
+		t.Fatalf("identity store not cleared: %v", st.keys())
+	}
+}
+
+// TestMigrateLegacyCredentials proves a pre-multi-identity credentials.json is
+// adopted into the store and preserved when a new identity is added.
+func TestMigrateLegacyCredentials(t *testing.T) {
+	sandbox(t)
+	// Simulate an old single-file login: only credentials.json exists.
+	legacy := credsFromToken(&tokenResp{AccessToken: makeJWT(map[string]any{
+		"email": "z@hanzo.ai", "owner": "hanzo", "sub": "h",
+	})})
+	if err := legacy.Save(); err != nil {
+		t.Fatal(err)
+	}
+	store, err := LoadIdentities()
+	if err != nil {
+		t.Fatalf("load identities: %v", err)
+	}
+	if store.Identities["hanzo/z"] == nil || store.Active != "hanzo/z" {
+		t.Fatalf("legacy credentials not migrated: active=%q keys=%v", store.Active, store.keys())
+	}
+	// A fresh login as a different owner preserves the migrated identity.
+	if _, err := runRoot(t, "", "login", "--token", makeJWT(map[string]any{
+		"email": "z@hanzo.ai", "owner": "admin", "sub": "a",
+	})); err != nil {
+		t.Fatal(err)
+	}
+	st2, _ := LoadIdentities()
+	if len(st2.Identities) != 2 || st2.Identities["hanzo/z"] == nil {
+		t.Fatalf("migrated identity lost after new login: %v", st2.keys())
+	}
+}

cli/cli.go

@@ -55,7 +55,7 @@ var controlCommands = map[string]string{
 	"login":    "authenticate against Hanzo IAM (hanzo.id) and store a token",
 	"logout":   "remove stored credentials",
 	"whoami":   "show the current identity from the stored token",
-	"auth":     "manage authentication (login, logout, whoami, token)",
+	"auth":     "manage authentication + stored identities (login, logout, whoami, list, switch, token)",
 	"apps":     "list/get the platform apps board (declared/running/drift)",
 	"deploy":   "drive a platform redeploy (rolling restart, zero-downtime)",
 	"clusters": "provision/list/select dedicated DOKS clusters",
@@ -236,6 +236,170 @@ func DeleteCredentials() error {
 	return nil
 }
 
+// ---------------------------------------------------------------------------
+// Identity store — ~/.hanzo/identities.json. Holds EVERY logged-in identity
+// keyed by its stable "<owner>/<name>" key, with an Active pointer. On every
+// write the active identity is mirrored into credentials.json (above), so every
+// legacy single-file reader keeps seeing the current identity unchanged. This
+// is the ONE credential store; credentials.json is its active-view mirror.
+// ---------------------------------------------------------------------------
+
+// IdentityStore is the on-disk shape of ~/.hanzo/identities.json.
+type IdentityStore struct {
+	Active     string                  `json:"active,omitempty"`
+	Identities map[string]*Credentials `json:"identities,omitempty"`
+}
+
+// key is the stable per-identity store key "<owner>/<name>", where name is the
+// email local-part (else the raw subject). The same identity yields the same
+// key every login, so re-login updates in place; the same email under a
+// different org (privilege separation) yields a distinct key (admin/z vs
+// hanzo/z) and is stored side by side rather than clobbering.
+func (c *Credentials) key() string {
+	name := c.Subject
+	if i := strings.IndexByte(name, '@'); i > 0 {
+		name = name[:i]
+	}
+	return firstNonEmpty(c.Owner, "-") + "/" + firstNonEmpty(name, "-")
+}
+
+func identitiesPath() (string, error) {
+	dir, err := hanzoDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(dir, "identities.json"), nil
+}
+
+// LoadIdentities reads the store. A pre-existing single-file credentials.json
+// with no store yet is migrated in (read-only) as the sole, active identity, so
+// upgrades are seamless — the first write persists it into the store.
+func LoadIdentities() (*IdentityStore, error) {
+	p, err := identitiesPath()
+	if err != nil {
+		return nil, err
+	}
+	s := &IdentityStore{Identities: map[string]*Credentials{}}
+	if err := loadJSON(p, s); err != nil {
+		return nil, err
+	}
+	if s.Identities == nil {
+		s.Identities = map[string]*Credentials{}
+	}
+	if len(s.Identities) == 0 {
+		if c, err := LoadCredentials(); err == nil && c.AccessToken != "" {
+			k := c.key()
+			s.Identities[k] = c
+			s.Active = k
+		}
+	}
+	return s, nil
+}
+
+// keys returns the identity keys, sorted, for deterministic output.
+func (s *IdentityStore) keys() []string {
+	ks := make([]string, 0, len(s.Identities))
+	for k := range s.Identities {
+		ks = append(ks, k)
+	}
+	sort.Strings(ks)
+	return ks
+}
+
+// Put stores c under its key and makes it active, returning the key.
+func (s *IdentityStore) Put(c *Credentials) string {
+	if s.Identities == nil {
+		s.Identities = map[string]*Credentials{}
+	}
+	k := c.key()
+	s.Identities[k] = c
+	s.Active = k
+	return k
+}
+
+// Remove deletes an identity; Save re-points Active if it was the one removed.
+func (s *IdentityStore) Remove(key string) { delete(s.Identities, key) }
+
+// resolve turns a user selector into a stored key: an exact key wins; otherwise
+// a bare owner matches iff exactly one identity carries it.
+func (s *IdentityStore) resolve(sel string) (string, error) {
+	if _, ok := s.Identities[sel]; ok {
+		return sel, nil
+	}
+	var match []string
+	for _, k := range s.keys() {
+		if s.Identities[k].Owner == sel {
+			match = append(match, k)
+		}
+	}
+	switch len(match) {
+	case 1:
+		return match[0], nil
+	case 0:
+		return "", fmt.Errorf("no stored identity for %q (see `hanzo auth list`)", sel)
+	default:
+		return "", fmt.Errorf("%q is ambiguous across %s — pass the full owner/name key", sel, strings.Join(match, ", "))
+	}
+}
+
+// Save persists the store (0600) and mirrors the active identity into
+// credentials.json for legacy single-file readers. When the store is empty it
+// removes both files. Active is normalized to a real key first.
+func (s *IdentityStore) Save() error {
+	if _, ok := s.Identities[s.Active]; !ok {
+		s.Active = ""
+		if ks := s.keys(); len(ks) > 0 {
+			s.Active = ks[0]
+		}
+	}
+	if len(s.Identities) == 0 {
+		return clearCredentialStore()
+	}
+	p, err := identitiesPath()
+	if err != nil {
+		return err
+	}
+	if err := writeJSON(p, s, 0o600); err != nil {
+		return err
+	}
+	return s.Identities[s.Active].Save() // mirror active → credentials.json (0600)
+}
+
+// SaveActive writes c back as the active identity (store + mirror), keeping the
+// two consistent after an in-place token refresh. With no store yet it falls
+// back to the single-file write.
+func SaveActive(c *Credentials) error {
+	s, err := LoadIdentities()
+	if err != nil {
+		return err
+	}
+	if len(s.Identities) == 0 {
+		return c.Save()
+	}
+	k := s.Active
+	if k == "" || s.Identities[k] == nil {
+		k = c.key()
+	}
+	s.Identities[k] = c
+	s.Active = k
+	return s.Save()
+}
+
+// clearCredentialStore removes the identity store and its credentials.json
+// mirror (used by logout when the last identity is removed).
+func clearCredentialStore() error {
+	for _, pathOf := range []func() (string, error){credentialsPath, identitiesPath} {
+		p, err := pathOf()
+		if err != nil {
+			return err
+		}
+		if err := os.Remove(p); err != nil && !os.IsNotExist(err) {
+			return err
+		}
+	}
+	return nil
+}
+
 // ---------------------------------------------------------------------------
 // Env — the effective, resolved settings a command operates with.
 // ---------------------------------------------------------------------------
@@ -310,11 +474,14 @@ func (e *Env) freshAccessToken() string {
 	return tok
 }
 
-// platformToken resolves the platform control-plane service token. The
-// platform REST surface is machine-to-machine (it cannot validate IAM user
-// tokens), so apps/clusters/redeploy authenticate with this, sourced from
-// (in precedence) the bound --platform-token flag, the environment, then the
-// credential store. Never hardcoded.
+// platformToken resolves the bearer the platform control plane authenticates
+// apps/clusters/redeploy with. ONE identity authorizes everything: after a plain
+// `hanzo login` the IAM access token is the FINAL fallback, so no separate
+// --platform-token is needed — the platform verifies the IAM JWT (signature,
+// issuer, expiry) and org-scopes the caller. A dedicated service token still
+// wins when present (flag > env > credential store > IAM login), so purpose-minted
+// machine tokens keep their precedence and internal automation is unchanged.
+// Never hardcoded.
 func (e *Env) platformToken(flagVal string) string {
 	return firstNonEmpty(
 		flagVal,
@@ -322,17 +489,22 @@ func (e *Env) platformToken(flagVal string) string {
 		os.Getenv("PLATFORM_SERVICE_TOKEN"),
 		os.Getenv("PAAS_SERVICE_TOKEN"),
 		e.creds.PlatformToken,
+		e.accessToken(), // IAM login is the one identity that authorizes control-plane ops
 	)
 }
 
-// buildToken resolves the platform build-enqueue token (a distinct credential
-// from the service token — see /v1/runner).
+// buildToken resolves the bearer `hanzo build` sends to the platform build
+// enqueue (/v1/runner). Same unify-infra contract as platformToken: a dedicated
+// build token wins when present, but a plain IAM login is the FINAL fallback, so
+// `hanzo build` works off the one identity with no separate --build-token — the
+// platform verifies the IAM JWT and authorizes the build by org + role.
 func (e *Env) buildToken(flagVal string) string {
 	return firstNonEmpty(
 		flagVal,
 		os.Getenv("HANZO_BUILD_TOKEN"),
 		os.Getenv("PLATFORM_BUILD_CALLBACK_TOKEN"),
 		e.creds.BuildToken,
+		e.accessToken(), // IAM login is the one identity that authorizes builds
 	)
 }
 
@@ -424,7 +596,6 @@ func newRootCmd() *cobra.Command {
 		newDeployCmd(envOf, &f),
 		newClustersCmd(envOf, &f),
 		newBuildCmd(envOf, &f),
-		newK8sCmd(envOf, &f),
 		newConfigCmd(),
 		newSecurityCmd(envOf),
 		newGPUCmd(envOf, &f),

cli/cli_test.go

@@ -141,6 +141,29 @@ func TestPlatformTokenPrecedence(t *testing.T) {
 	}
 }
 
+// TestPlatformTokenFallsBackToIAM is the UNIFY-INFRA contract for the control
+// plane: after a plain `hanzo login`, the IAM access token is the FINAL fallback
+// so `hanzo apps`/`hanzo deploy` authorize off the one identity. An explicit
+// platform service token (creds/env/flag) still wins.
+func TestPlatformTokenFallsBackToIAM(t *testing.T) {
+	sandbox(t)
+	// Only an IAM login: no platform token anywhere ⇒ the IAM access token is sent.
+	e := resolve(&Config{}, &Credentials{AccessToken: "iam-jwt"}, globalFlags{})
+	if got := e.platformToken(""); got != "iam-jwt" {
+		t.Fatalf("IAM access token should be the final platform-token fallback: %q", got)
+	}
+	// A dedicated platform service token still beats the IAM token.
+	e = resolve(&Config{}, &Credentials{AccessToken: "iam-jwt", PlatformToken: "svc"}, globalFlags{})
+	if got := e.platformToken(""); got != "svc" {
+		t.Fatalf("dedicated platform token must beat the IAM fallback: %q", got)
+	}
+	// No login at all ⇒ empty (caller surfaces "run `hanzo login`").
+	e = resolve(&Config{}, &Credentials{}, globalFlags{})
+	if got := e.platformToken(""); got != "" {
+		t.Fatalf("no token and no login should resolve empty: %q", got)
+	}
+}
+
 func TestBuildTokenPrecedence(t *testing.T) {
 	sandbox(t)
 	e := resolve(&Config{}, &Credentials{BuildToken: "creds"}, globalFlags{})
@@ -156,6 +179,37 @@ func TestBuildTokenPrecedence(t *testing.T) {
 	}
 }
 
+// TestBuildTokenFallsBackToIAM is the UNIFY-INFRA contract: after a plain
+// `hanzo login` (no --build-token), the IAM access token is the FINAL fallback,
+// so `hanzo build` authorizes off the one identity. An explicit build token
+// (creds/env/flag) still wins — the IAM token is the LAST resort, never an
+// override of a purpose-minted machine token.
+func TestBuildTokenFallsBackToIAM(t *testing.T) {
+	sandbox(t)
+	// Only an IAM login: no build token anywhere ⇒ the IAM access token is sent.
+	e := resolve(&Config{}, &Credentials{AccessToken: "iam-jwt"}, globalFlags{})
+	if got := e.buildToken(""); got != "iam-jwt" {
+		t.Fatalf("IAM access token should be the final build-token fallback: %q", got)
+	}
+	// A dedicated build token still beats the IAM token (precedence preserved).
+	e = resolve(&Config{}, &Credentials{AccessToken: "iam-jwt", BuildToken: "creds"}, globalFlags{})
+	if got := e.buildToken(""); got != "creds" {
+		t.Fatalf("dedicated build token must beat the IAM fallback: %q", got)
+	}
+	// HANZO_TOKEN (the env form of the IAM token) is also honored via accessToken().
+	e = resolve(&Config{}, &Credentials{}, globalFlags{})
+	t.Setenv("HANZO_TOKEN", "iam-env")
+	if got := e.buildToken(""); got != "iam-env" {
+		t.Fatalf("HANZO_TOKEN should back the build-token fallback: %q", got)
+	}
+	// No login at all ⇒ empty, so the caller can surface "run `hanzo login`".
+	t.Setenv("HANZO_TOKEN", "")
+	e = resolve(&Config{}, &Credentials{}, globalFlags{})
+	if got := e.buildToken(""); got != "" {
+		t.Fatalf("no token and no login should resolve empty: %q", got)
+	}
+}
+
 func TestAccessTokenFromEnvOverCreds(t *testing.T) {
 	sandbox(t)
 	e := resolve(&Config{}, &Credentials{AccessToken: "creds"}, globalFlags{})

cli/code.go

@@ -147,6 +147,7 @@ type codeAgent struct {
 	bin          string                     // executable to exec
 	wire         wire                       // how it finds the cloud
 	fullAuto     []string                   // flags that bypass approval prompts
+	continueArgs []string                   // harness-native form of Hanzo -c/--continue
 	modelArg     []string                   // how the model is passed on argv (empty: via env)
 	carrier      func(model string) string  // maps the resolved model to a client-recognized id (claude: zen→carrier); nil = pass through
 	provider     func(base string) []string // agents that need the endpoint declared, not just env'd
@@ -163,10 +164,11 @@ type codeAgent struct {
 // declared, so declare Hanzo as the provider and select it.
 func codexLike(bin, install string) codeAgent {
 	return codeAgent{
-		bin:      bin,
-		wire:     openaiWire,
-		fullAuto: []string{"--dangerously-bypass-approvals-and-sandbox"},
-		modelArg: []string{"-m"},
+		bin:          bin,
+		wire:         openaiWire,
+		fullAuto:     []string{"--dangerously-bypass-approvals-and-sandbox"},
+		continueArgs: []string{"resume", "--last"},
+		modelArg:     []string{"-m"},
 		provider: func(base string) []string {
 			return []string{
 				"-c", "model_provider=hanzo",
@@ -174,6 +176,12 @@ func codexLike(bin, install string) codeAgent {
 				"-c", fmt.Sprintf(`model_providers.hanzo.base_url="%s/v1"`, strings.TrimSuffix(base, "/")),
 				"-c", `model_providers.hanzo.env_key="OPENAI_API_KEY"`,
 				"-c", `model_providers.hanzo.wire_api="responses"`,
+				// api.hanzo.ai exposes the standard OpenAI /v1/models shape,
+				// not Codex's private remote model-catalog schema. Skip that
+				// optional refresh and supply the coding model's metadata here.
+				"-c", `features.remote_models=false`,
+				"-c", `model_context_window=262144`,
+				"-c", `model_auto_compact_token_limit=235929`,
 			}
 		},
 		install: install,
@@ -190,9 +198,10 @@ const zenIdentityPrompt = "You are running through the Hanzo AI cloud as a Hanzo
 
 var codeAgents = map[string]codeAgent{
 	"claude": {
-		bin:      "claude",
-		wire:     anthropicWire,
-		fullAuto: []string{"--dangerously-skip-permissions"},
+		bin:          "claude",
+		wire:         anthropicWire,
+		fullAuto:     []string{"--dangerously-skip-permissions"},
+		continueArgs: []string{"--continue"},
 		// --model forces the session model on argv. Claude Code persists the
 		// user's last /model selection (e.g. the reserved word "best"), and that
 		// persisted choice OVERRIDES ANTHROPIC_MODEL — so the env var alone cannot
@@ -246,9 +255,12 @@ func newCodeCmd(envOf func() *Env, _ *globalFlags) *cobra.Command {
 		Long: "Run @hanzo/dev, Claude Code, or Codex against api.hanzo.ai with the endpoint,\n" +
 			"credential and model injected — no env vars to remember. `hanzo code` alone runs\n" +
 			"dev (the Hanzo agent); name an agent to pick another. Model ids resolve fuzzily\n" +
-			"(glm5.2 -> glm-5.2) and agents run full-auto unless you pass --safe.",
+			"(glm5.2 -> glm-5.2), -c resumes either harness, and agents run full-auto unless\n" +
+			"you pass --safe. Unknown options pass through; -- forces verbatim passthrough.",
 		Example: "  hanzo code                 # dev, the default agent\n" +
 			"  hanzo code claude\n" +
+			"  hanzo code claude -c\n" +
+			"  hanzo code codex -c\n" +
 			"  hanzo code codex deepseek-v4-pro\n" +
 			"  hanzo code dev glm5.2 -- --resume\n" +
 			"  hanzo code ls",
@@ -320,19 +332,10 @@ func runCode(env *Env, agent codeAgent, args []string) error {
 	}
 	base := strings.TrimSuffix(firstNonEmpty(env.CloudURL, "https://api.hanzo.ai"), "/")
 
-	// First non-flag arg is the model; --safe is ours; the rest is the agent's.
-	model, safe, rest := "", false, make([]string, 0, len(args))
-	for _, a := range args {
-		switch {
-		case a == "--": // the separator is ours; the agent must not see it
-		case a == "--safe" || a == "--ask":
-			safe = true
-		case model == "" && !strings.HasPrefix(a, "-") && len(rest) == 0:
-			model = a
-		default:
-			rest = append(rest, a)
-		}
-	}
+	// First non-flag arg before -- is the model; --safe and --continue are ours.
+	// Unknown options pass through unchanged. Everything after -- belongs to the
+	// agent, including positional subcommands and raw Codex -c config overrides.
+	model, safe, continueLast, rest := splitCodeArgs(args)
 	if model == "" {
 		model = defaultCodeModel
 	}
@@ -382,7 +385,7 @@ func runCode(env *Env, agent codeAgent, args []string) error {
 		}
 	}
 
-	argv := codeArgv(agent, base, model, safe, rest)
+	argv := codeArgv(agent, base, model, safe, codeAgentRest(agent, continueLast, rest))
 
 	for k, v := range agent.wire(base, token, model) {
 		if err := os.Setenv(k, v); err != nil {
@@ -397,6 +400,45 @@ func runCode(env *Env, agent codeAgent, args []string) error {
 	return execEngine(bin, argv) // exec: signals + exit code flow straight through
 }
 
+// splitCodeArgs pulls the launcher-owned tokens (the model, --safe, -c/--continue)
+// out of the raw args; everything else is the agent's. The `--` separator is ours
+// and switches on verbatim passthrough — every token after it goes to the agent
+// untouched, including positional subcommands (`codex exec`) and raw Codex -c
+// config overrides that would otherwise look like our --continue.
+func splitCodeArgs(args []string) (model string, safe, continueLast bool, rest []string) {
+	rest = make([]string, 0, len(args))
+	passthrough := false
+	for _, a := range args {
+		switch {
+		case passthrough:
+			rest = append(rest, a)
+		case a == "--": // the separator is ours; the agent must not see it
+			passthrough = true
+		case a == "--safe" || a == "--ask":
+			safe = true
+		case a == "-c" || a == "--continue":
+			continueLast = true
+		case model == "" && !strings.HasPrefix(a, "-") && len(rest) == 0:
+			model = a
+		default:
+			rest = append(rest, a)
+		}
+	}
+	return model, safe, continueLast, rest
+}
+
+// codeAgentRest prepends the agent's harness-native resume tokens when -c/--continue
+// was given, so one Hanzo flag resumes the last session on either harness (`--continue`
+// for Claude Code, `resume --last` for Codex/dev).
+func codeAgentRest(agent codeAgent, continueLast bool, rest []string) []string {
+	if !continueLast {
+		return rest
+	}
+	args := make([]string, 0, len(agent.continueArgs)+len(rest))
+	args = append(args, agent.continueArgs...)
+	return append(args, rest...)
+}
+
 // codeArgv builds the final agent command line. Permission bypass is the
 // launcher default for every agent; --safe is the single explicit opt-out.
 func codeArgv(agent codeAgent, base, model string, safe bool, rest []string) []string {

cli/code_test.go

@@ -18,6 +18,7 @@ import (
 	"encoding/json"
 	"os"
 	"path/filepath"
+	"reflect"
 	"slices"
 	"testing"
 	"time"
@@ -48,6 +49,72 @@ func TestCodeAgentsBypassPermissionsByDefault(t *testing.T) {
 	}
 }
 
+func TestCodeArgsSeparatorPreservesAgentSubcommand(t *testing.T) {
+	model, safe, continueLast, rest := splitCodeArgs([]string{"--safe", "--", "exec", "--ephemeral", "do it"})
+	if model != "" || !safe || continueLast {
+		t.Fatalf("model=%q safe=%v continue=%v, want default model and safe mode", model, safe, continueLast)
+	}
+	if want := []string{"exec", "--ephemeral", "do it"}; !reflect.DeepEqual(rest, want) {
+		t.Fatalf("agent args = %q, want %q", rest, want)
+	}
+}
+
+func TestCodeArgsExplicitModelBeforeSeparator(t *testing.T) {
+	model, safe, continueLast, rest := splitCodeArgs([]string{"zen5-max", "--", "exec"})
+	if model != "zen5-max" || safe || continueLast || !reflect.DeepEqual(rest, []string{"exec"}) {
+		t.Fatalf("model=%q safe=%v continue=%v rest=%q", model, safe, continueLast, rest)
+	}
+}
+
+func TestCodeContinueIsNormalizedForBothHarnesses(t *testing.T) {
+	for _, tt := range []struct {
+		name string
+		want []string
+	}{
+		{name: "claude", want: []string{"--continue"}},
+		{name: "codex", want: []string{"resume", "--last"}},
+	} {
+		t.Run(tt.name, func(t *testing.T) {
+			model, safe, continueLast, rest := splitCodeArgs([]string{"-c"})
+			if model != "" || safe || !continueLast || len(rest) != 0 {
+				t.Fatalf("model=%q safe=%v continue=%v rest=%q", model, safe, continueLast, rest)
+			}
+			if got := codeAgentRest(codeAgents[tt.name], continueLast, rest); !reflect.DeepEqual(got, tt.want) {
+				t.Fatalf("normalized continue args = %q, want %q", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestCodeUnknownOptionsAndPostSeparatorArgsPassThrough(t *testing.T) {
+	unknown := []string{"--mystery", "value", "--other=1"}
+	model, safe, continueLast, rest := splitCodeArgs(unknown)
+	if model != "" || safe || continueLast || !reflect.DeepEqual(rest, unknown) {
+		t.Fatalf("unknown options changed: model=%q safe=%v continue=%v rest=%q", model, safe, continueLast, rest)
+	}
+
+	_, _, continueLast, rest = splitCodeArgs([]string{"--", "-c", "model=x"})
+	if continueLast || !reflect.DeepEqual(rest, []string{"-c", "model=x"}) {
+		t.Fatalf("post-separator Codex config must pass verbatim: continue=%v rest=%q", continueLast, rest)
+	}
+}
+
+func TestCodexProviderUsesNativeResponsesMetadata(t *testing.T) {
+	argv := codeArgv(codeAgents["codex"], "https://api.hanzo.ai", defaultCodeModel, false, nil)
+	for _, want := range []string{
+		`model_provider=hanzo`,
+		`model_providers.hanzo.base_url="https://api.hanzo.ai/v1"`,
+		`model_providers.hanzo.wire_api="responses"`,
+		`features.remote_models=false`,
+		`model_context_window=262144`,
+		`model_auto_compact_token_limit=235929`,
+	} {
+		if !slices.Contains(argv, want) {
+			t.Errorf("Codex argv %q does not contain %q", argv, want)
+		}
+	}
+}
+
 // TestCodeTokenPrecedence locks in the 402 unblock: a fresh `hanzo login` JWT
 // (which carries owner/project/sub on EVERY deployment) beats the hk- API key
 // (which only mints a billing principal where the server has IAM_MINT_CLIENT_*).
@@ -60,10 +127,10 @@ func TestCodeTokenPrecedence(t *testing.T) {
 	freshExpiry := time.Now().Add(1 * time.Hour).Unix()
 
 	cases := []struct {
-		name    string
-		envKey  string // HANZO_API_KEY override
-		creds   Credentials
-		want    string
+		name   string
+		envKey string // HANZO_API_KEY override
+		creds  Credentials
+		want   string
 	}{
 		{
 			name:  "fresh JWT beats hk- key",
@@ -81,16 +148,16 @@ func TestCodeTokenPrecedence(t *testing.T) {
 			want:  "hk-stored",
 		},
 		{
-			name:  "HANZO_API_KEY overrides everything (deliberate operator override)",
+			name:   "HANZO_API_KEY overrides everything (deliberate operator override)",
 			envKey: "hk-explicit",
-			creds: Credentials{AccessToken: "jwt-live", Expiry: freshExpiry},
-			want:  "hk-explicit",
+			creds:  Credentials{AccessToken: "jwt-live", Expiry: freshExpiry},
+			want:   "hk-explicit",
 		},
 		{
-			name:  "HANZO_API_KEY overrides even an expired JWT",
+			name:   "HANZO_API_KEY overrides even an expired JWT",
 			envKey: "hk-explicit",
-			creds: Credentials{AccessToken: "jwt-dead", Expiry: time.Now().Add(-1 * time.Hour).Unix()},
-			want:  "hk-explicit",
+			creds:  Credentials{AccessToken: "jwt-dead", Expiry: time.Now().Add(-1 * time.Hour).Unix()},
+			want:   "hk-explicit",
 		},
 	}
 

cli/commands.go

@@ -3,6 +3,7 @@ package cli
 import (
 	"fmt"
 	"io"
+	"strings"
 	"text/tabwriter"
 
 	"github.com/spf13/cobra"
@@ -15,12 +16,12 @@ func (e *Env) platform(gf *globalFlags) *Platform {
 	return newPlatform(e.PlatformURL, e.platformToken(gf.platformToken))
 }
 
-// deref renders a *string for a table cell, "-" when nil/empty.
-func deref(p *string) string {
-	if p == nil || *p == "" {
+// dashIfEmpty renders a string cell, "-" when empty.
+func dashIfEmpty(s string) string {
+	if s == "" {
 		return "-"
 	}
-	return *p
+	return s
 }
 
 // yesno renders a bool for a table cell.
@@ -37,7 +38,8 @@ func newTab(w io.Writer) *tabwriter.Writer {
 }
 
 // ---------------------------------------------------------------------------
-// apps — the observe surface.
+// apps — the fleet drift board (GET /v1/paas/apps). Org-confined server-side by
+// the IAM identity: a superadmin sees the whole fleet, an org-admin only its own.
 // ---------------------------------------------------------------------------
 
 func newAppsCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
@@ -56,7 +58,6 @@ func newAppsCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 		RunE: func(cmd *cobra.Command, _ []string) error {
 			e := envOf()
 			res, err := e.platform(gf).Apps(cmd.Context(), AppsQuery{
-				Org:    e.Org, // empty == all (single-tenant default)
 				Env:    envFilter,
 				Health: healthFilter,
 				Drift:  driftOnly,
@@ -69,8 +70,8 @@ func newAppsCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 				fmt.Fprintln(tw, "ORG\tAPP\tENV\tDECLARED\tRUNNING\tHEALTH\tDRIFT")
 				for _, a := range res.Apps {
 					fmt.Fprintf(tw, "%s\t%s\t%s\t%s\t%s\t%s\t%s\n",
-						a.Org, a.App, a.Env, deref(a.DeclaredTag), deref(a.RunningTag),
-						deref(a.Health), driftSeverity(a.Drift))
+						a.Org, a.App, a.Env, dashIfEmpty(a.DeclaredTag), dashIfEmpty(a.RunningTag),
+						dashIfEmpty(a.Health), driftSeverity(a.Drift))
 				}
 				tw.Flush()
 				fmt.Fprintf(w, "\n%d apps  (ok=%d yellow=%d red=%d)\n",
@@ -79,17 +80,17 @@ func newAppsCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 			})
 		},
 	}
-	list.Flags().StringVar(&envFilter, "env", "", "filter by env: dev|test|main")
+	list.Flags().StringVar(&envFilter, "env", "", "filter by env: main|test|dev")
 	list.Flags().StringVar(&healthFilter, "health", "", "filter by health: green|yellow|red")
 	list.Flags().BoolVar(&driftOnly, "drift", false, "only rows that are drifting")
 
 	get := &cobra.Command{
-		Use:   "get <org/app/env>",
-		Short: "Get one app row by its <org>/<app>/<env> id",
+		Use:   "get <app>",
+		Short: "Get one app row by its CR name (production by default)",
 		Args:  cobra.ExactArgs(1),
 		RunE: func(cmd *cobra.Command, args []string) error {
 			e := envOf()
-			a, err := e.platform(gf).App(cmd.Context(), args[0], e.Org)
+			a, err := e.platform(gf).App(cmd.Context(), args[0])
 			if err != nil {
 				return err
 			}
@@ -101,109 +102,93 @@ func newAppsCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 				fmt.Fprintf(tw, "env:\t%s\n", a.Env)
 				fmt.Fprintf(tw, "repo:\t%s\n", a.Repo)
 				fmt.Fprintf(tw, "registry:\t%s\n", a.Registry)
-				fmt.Fprintf(tw, "declared:\t%s\n", deref(a.DeclaredTag))
-				fmt.Fprintf(tw, "running:\t%s\n", deref(a.RunningTag))
-				fmt.Fprintf(tw, "latest:\t%s\n", deref(a.LatestTag))
-				fmt.Fprintf(tw, "health:\t%s\n", deref(a.Health))
+				fmt.Fprintf(tw, "declared:\t%s\n", dashIfEmpty(a.DeclaredTag))
+				fmt.Fprintf(tw, "running:\t%s\n", dashIfEmpty(a.RunningTag))
+				fmt.Fprintf(tw, "health:\t%s\n", dashIfEmpty(a.Health))
+				fmt.Fprintf(tw, "phase:\t%s\n", dashIfEmpty(a.Phase))
 				fmt.Fprintf(tw, "drift:\t%s\n", driftSeverity(a.Drift))
-				fmt.Fprintf(tw, "cluster:\t%s\n", deref(a.Cluster))
-				fmt.Fprintf(tw, "namespace:\t%s\n", deref(a.Namespace))
-				fmt.Fprintf(tw, "updated:\t%s\n", a.UpdatedAt)
+				fmt.Fprintf(tw, "cluster:\t%s\n", dashIfEmpty(a.Cluster))
+				fmt.Fprintf(tw, "namespace:\t%s\n", dashIfEmpty(a.Namespace))
+				if len(a.Endpoints) > 0 {
+					fmt.Fprintf(tw, "endpoints:\t%s\n", strings.Join(a.Endpoints, ", "))
+				}
 				tw.Flush()
 			})
 		},
 	}
 
-	sync := &cobra.Command{
-		Use:   "sync",
-		Short: "Trigger an inventory refresh of the apps board",
-		Args:  cobra.NoArgs,
-		RunE: func(cmd *cobra.Command, _ []string) error {
-			e := envOf()
-			if err := e.platform(gf).SyncApps(cmd.Context()); err != nil {
-				return err
-			}
-			fmt.Fprintln(cmd.OutOrStdout(), "apps sync triggered")
-			return nil
-		},
-	}
-
-	cmd.AddCommand(list, get, sync)
+	cmd.AddCommand(list, get)
 	return cmd
 }
 
 // ---------------------------------------------------------------------------
-// deploy — the drive surface (rolling restart, zero-downtime).
+// deploy — POST /v1/paas/apps/{app}/deploy: a zero-downtime rolling restart.
 // ---------------------------------------------------------------------------
 
 func newDeployCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
-	var project, environment string
+	var environment string
 	cmd := &cobra.Command{
-		Use:   "deploy <container>",
-		Short: "Redeploy a container (rolling restart, zero-downtime)",
-		Long: "Drive a platform redeploy: a rolling restart of the container's k8s\n" +
-			"Deployment (re-pulls the image, recreates pods, zero downtime). Coordinates\n" +
-			"are exact — org (--org/config), project (--project), env (--env) and the\n" +
-			"container id (positional). This is the canonical PaaS-driven deploy.",
+		Use:   "deploy <app>",
+		Short: "Redeploy an app (rolling restart, zero-downtime) — requires --env",
+		Long: "Drive a platform redeploy: a rolling restart of the app's k8s Deployment\n" +
+			"(re-pulls the declared image, recreates pods, zero downtime). The app is the\n" +
+			"operator App CR name; the org comes from your IAM identity. --env is REQUIRED\n" +
+			"(main|test|dev) — deploy never silently targets production. Restarting a shared\n" +
+			"platform service is a platform-operator action, so this needs a superadmin\n" +
+			"identity. A TAG change is still a git commit — this restarts what is declared.",
 		Args: cobra.ExactArgs(1),
 		RunE: func(cmd *cobra.Command, args []string) error {
+			if strings.TrimSpace(environment) == "" {
+				return fmt.Errorf("--env is required (main|test|dev) — deploy will not default to production")
+			}
 			e := envOf()
-			org, err := e.requireOrg()
+			res, err := e.platform(gf).Redeploy(cmd.Context(), args[0], environment)
 			if err != nil {
 				return err
 			}
-			if project == "" || environment == "" {
-				return fmt.Errorf("--project and --env are required (the container's project/environment ids)")
-			}
-			container := args[0]
-			if err := e.platform(gf).Redeploy(cmd.Context(), org, project, environment, container); err != nil {
-				return err
-			}
-			fmt.Fprintf(cmd.OutOrStdout(), "redeployed %s (org=%s project=%s env=%s)\n", container, org, project, environment)
-			return nil
+			return e.emit(res, func(w io.Writer) {
+				fmt.Fprintf(w, "restarted %s (namespace=%s env=%s at %s)\n",
+					res.App, res.Namespace, dashIfEmpty(res.Env), res.RestartedAt)
+			})
 		},
 	}
-	cmd.Flags().StringVar(&project, "project", "", "project id")
-	cmd.Flags().StringVar(&environment, "env", "", "environment id")
+	cmd.Flags().StringVar(&environment, "env", "", "lifecycle env: main|test|dev (REQUIRED)")
 	return cmd
 }
 
 // ---------------------------------------------------------------------------
-// clusters — dedicated DOKS cluster lifecycle.
+// clusters — GET /v1/clusters: the org's compute fleet (Visor-managed + BYO),
+// tenant-scoped server-side by the IAM identity.
 // ---------------------------------------------------------------------------
 
 func newClustersCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 	cmd := &cobra.Command{
 		Use:     "clusters",
 		Aliases: []string{"cluster"},
-		Short:   "Provision/list/select dedicated DOKS clusters",
+		Short:   "List the org's clusters (managed + BYO)",
 	}
 
 	list := &cobra.Command{
 		Use:   "list",
-		Short: "List the org's dedicated clusters",
+		Short: "List the org's clusters",
 		Args:  cobra.NoArgs,
 		RunE: func(cmd *cobra.Command, _ []string) error {
 			e := envOf()
-			org, err := e.requireOrg()
-			if err != nil {
-				return err
-			}
-			cs, err := e.platform(gf).Clusters(cmd.Context(), org)
+			cs, err := e.platform(gf).Clusters(cmd.Context())
 			if err != nil {
 				return err
 			}
 			return e.emit(cs, func(w io.Writer) {
 				tw := newTab(w)
-				fmt.Fprintln(tw, "NAME\tID\tREGION\tSTATUS\tPHASE\tACTIVE\tOPERATOR\tBASELINE")
+				fmt.Fprintln(tw, "NAME\tID\tREGION\tSTATUS\tKIND\tNODES\tSIZE\tGPUS")
 				for _, c := range cs {
-					fmt.Fprintf(tw, "%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\n",
-						c.Name, c.DoksClusterID, c.Region, c.Status, c.Phase,
-						yesno(c.Active), yesno(c.OperatorInstalled), yesno(c.BaselineInstalled))
+					fmt.Fprintf(tw, "%s\t%s\t%s\t%s\t%s\t%d\t%s\t%s\n",
+						c.Name, dashIfEmpty(c.ID()), dashIfEmpty(c.Region), dashIfEmpty(c.Status),
+						dashIfEmpty(c.Kind), c.NodeCount, dashIfEmpty(c.NodeSize), gpuCell(c))
 				}
 				tw.Flush()
 				if len(cs) == 0 {
-					fmt.Fprintln(w, "(no dedicated clusters)")
+					fmt.Fprintln(w, "(no clusters)")
 				}
 			})
 		},
@@ -215,151 +200,51 @@ func newClustersCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 		Args:  cobra.ExactArgs(1),
 		RunE: func(cmd *cobra.Command, args []string) error {
 			e := envOf()
-			org, err := e.requireOrg()
-			if err != nil {
-				return err
-			}
-			cs, err := e.platform(gf).Clusters(cmd.Context(), org)
+			cs, err := e.platform(gf).Clusters(cmd.Context())
 			if err != nil {
 				return err
 			}
 			for _, c := range cs {
-				if c.DoksClusterID == args[0] || c.Name == args[0] {
+				if c.ID() == args[0] || c.Name == args[0] {
 					return e.emit(c, func(w io.Writer) { printCluster(w, c) })
 				}
 			}
-			return fmt.Errorf("cluster %q not found in org %s", args[0], org)
+			return fmt.Errorf("cluster %q not found", args[0])
 		},
 	}
 
-	var region, nodeSize string
-	var ha bool
-	var nodeCount int
-	create := &cobra.Command{
-		Use:   "create",
-		Short: "Provision a new dedicated DOKS cluster for the org",
-		Args:  cobra.NoArgs,
-		RunE: func(cmd *cobra.Command, _ []string) error {
-			e := envOf()
-			org, err := e.requireOrg()
-			if err != nil {
-				return err
-			}
-			c, err := e.platform(gf).ProvisionCluster(cmd.Context(), org, ProvisionReq{
-				Region: region, HA: ha, NodeSize: nodeSize, NodeCount: nodeCount,
-			})
-			if err != nil {
-				return err
-			}
-			return e.emit(c, func(w io.Writer) {
-				fmt.Fprintf(w, "provisioning cluster %s (%s)\n", c.Name, c.DoksClusterID)
-				printCluster(w, *c)
-			})
-		},
-	}
-	create.Flags().StringVar(&region, "region", "", "DO region (default sfo3)")
-	create.Flags().BoolVar(&ha, "ha", false, "highly-available control plane")
-	create.Flags().StringVar(&nodeSize, "node-size", "", "node size slug (e.g. s-2vcpu-4gb)")
-	create.Flags().IntVar(&nodeCount, "node-count", 0, "node count")
-
-	var shared bool
-	selectCmd := &cobra.Command{
-		Use:   "select <cluster-id>",
-		Short: "Set the org's active deploy target (or --shared to revert)",
-		Args:  cobra.MaximumNArgs(1),
-		RunE: func(cmd *cobra.Command, args []string) error {
-			e := envOf()
-			org, err := e.requireOrg()
-			if err != nil {
-				return err
-			}
-			var clusterID *string
-			switch {
-			case shared:
-				clusterID = nil
-			case len(args) == 1:
-				clusterID = &args[0]
-			default:
-				return fmt.Errorf("give a cluster id, or --shared to revert to the shared cluster")
-			}
-			t, err := e.platform(gf).SelectTarget(cmd.Context(), org, clusterID)
-			if err != nil {
-				return err
-			}
-			return e.emit(t, func(w io.Writer) { printTarget(w, t) })
-		},
-	}
-	selectCmd.Flags().BoolVar(&shared, "shared", false, "revert to the shared cluster")
-
-	installBaseline := &cobra.Command{
-		Use:   "install-baseline <cluster-id>",
-		Short: "Install the hanzo-operator + per-tenant baseline on a cluster",
-		Args:  cobra.ExactArgs(1),
-		RunE: func(cmd *cobra.Command, args []string) error {
-			e := envOf()
-			org, err := e.requireOrg()
-			if err != nil {
-				return err
-			}
-			if err := e.platform(gf).InstallBaseline(cmd.Context(), org, args[0]); err != nil {
-				return err
-			}
-			fmt.Fprintf(cmd.OutOrStdout(), "baseline install requested for %s\n", args[0])
-			return nil
-		},
-	}
-
-	target := &cobra.Command{
-		Use:   "target",
-		Short: "Show the org's current resolved deploy target",
-		Args:  cobra.NoArgs,
-		RunE: func(cmd *cobra.Command, _ []string) error {
-			e := envOf()
-			org, err := e.requireOrg()
-			if err != nil {
-				return err
-			}
-			t, err := e.platform(gf).Target(cmd.Context(), org)
-			if err != nil {
-				return err
-			}
-			return e.emit(t, func(w io.Writer) { printTarget(w, t) })
-		},
-	}
-
-	cmd.AddCommand(list, get, create, selectCmd, installBaseline, target)
+	cmd.AddCommand(list, get)
 	return cmd
 }
 
+// gpuCell renders the live GPU inventory of a cluster ("-" when none).
+func gpuCell(c Cluster) string {
+	var parts []string
+	if c.NvidiaGPU > 0 {
+		parts = append(parts, fmt.Sprintf("%d nvidia", c.NvidiaGPU))
+	}
+	if c.AmdGPU > 0 {
+		parts = append(parts, fmt.Sprintf("%d amd", c.AmdGPU))
+	}
+	if len(parts) == 0 {
+		return "-"
+	}
+	return strings.Join(parts, "+")
+}
+
 func printCluster(w io.Writer, c Cluster) {
 	tw := newTab(w)
-	fmt.Fprintf(tw, "id:\t%s\n", c.DoksClusterID)
+	fmt.Fprintf(tw, "id:\t%s\n", dashIfEmpty(c.ID()))
 	fmt.Fprintf(tw, "name:\t%s\n", c.Name)
-	fmt.Fprintf(tw, "region:\t%s\n", c.Region)
-	fmt.Fprintf(tw, "status:\t%s\n", c.Status)
-	fmt.Fprintf(tw, "phase:\t%s\n", c.Phase)
-	fmt.Fprintf(tw, "active:\t%s\n", yesno(c.Active))
-	fmt.Fprintf(tw, "operatorInstalled:\t%s\n", yesno(c.OperatorInstalled))
-	fmt.Fprintf(tw, "baselineInstalled:\t%s\n", yesno(c.BaselineInstalled))
-	fmt.Fprintf(tw, "endpoint:\t%s\n", deref(c.Endpoint))
-	fmt.Fprintf(tw, "k8sVersion:\t%s\n", deref(c.K8sVersion))
-	fmt.Fprintf(tw, "created:\t%s\n", c.CreatedAt)
-	if c.BaselineError != nil && *c.BaselineError != "" {
-		fmt.Fprintf(tw, "baselineError:\t%s\n", *c.BaselineError)
-	}
-	tw.Flush()
-}
-
-func printTarget(w io.Writer, t *Target) {
-	tw := newTab(w)
-	kind := "shared"
-	if t.Dedicated {
-		kind = "dedicated"
-	}
-	fmt.Fprintf(tw, "cluster:\t%s\n", t.Cluster)
-	fmt.Fprintf(tw, "kind:\t%s\n", kind)
-	for ns, env := range t.Namespaces {
-		fmt.Fprintf(tw, "namespace:\t%s -> %s\n", ns, env)
+	fmt.Fprintf(tw, "region:\t%s\n", dashIfEmpty(c.Region))
+	fmt.Fprintf(tw, "status:\t%s\n", dashIfEmpty(c.Status))
+	fmt.Fprintf(tw, "kind:\t%s\n", dashIfEmpty(c.Kind))
+	fmt.Fprintf(tw, "nodeCount:\t%d\n", c.NodeCount)
+	fmt.Fprintf(tw, "nodeSize:\t%s\n", dashIfEmpty(c.NodeSize))
+	fmt.Fprintf(tw, "gpus:\t%s\n", gpuCell(c))
+	fmt.Fprintf(tw, "created:\t%s\n", dashIfEmpty(c.CreatedAt))
+	for _, np := range c.NodePools {
+		fmt.Fprintf(tw, "pool:\t%s (%s x%d, autoscale=%s)\n", np.Name, np.Size, np.Count, yesno(np.AutoScale))
 	}
 	tw.Flush()
 }
@@ -387,6 +272,10 @@ func newBuildCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 			if br.Repo == "" || br.SHA == "" || br.Image == "" {
 				return fmt.Errorf("--repo (or positional), --sha and --image are required")
 			}
+			// The platform build muscle clones an https git URL; accept the
+			// idiomatic `owner/name` shorthand and expand it to GitHub (the host
+			// for every hanzoai/luxfi/zooai repo). A full URL passes through.
+			br.Repo = normalizeRepoURL(br.Repo)
 			if br.OrganizationID == "" {
 				br.OrganizationID = e.Org // optional; server defaults to DEFAULT_BUILD_ORG_ID
 			}
@@ -419,32 +308,24 @@ func newBuildCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
 	return cmd
 }
 
-// ---------------------------------------------------------------------------
-// k8s — deploy-target helpers.
-// ---------------------------------------------------------------------------
-
-func newK8sCmd(envOf func() *Env, gf *globalFlags) *cobra.Command {
-	cmd := &cobra.Command{
-		Use:   "k8s",
-		Short: "Kubernetes deploy-target helpers",
+// normalizeRepoURL expands the idiomatic `owner/name` shorthand to a full GitHub
+// https URL (the platform build muscle clones https), and leaves an explicit URL
+// (http/https/git/ssh scheme, or a scp-style git@host:owner/name) untouched. Only
+// a bare single-segment `owner/name` — two path parts, no scheme, no host — is
+// expanded; anything else is the caller's explicit choice and passes through.
+func normalizeRepoURL(repo string) string {
+	r := strings.TrimSpace(repo)
+	if r == "" {
+		return r
 	}
-	target := &cobra.Command{
-		Use:   "target",
-		Short: "Show the org's current resolved deploy target (cluster + namespaces)",
-		Args:  cobra.NoArgs,
-		RunE: func(cmd *cobra.Command, _ []string) error {
-			e := envOf()
-			org, err := e.requireOrg()
-			if err != nil {
-				return err
-			}
-			t, err := e.platform(gf).Target(cmd.Context(), org)
-			if err != nil {
-				return err
-			}
-			return e.emit(t, func(w io.Writer) { printTarget(w, t) })
-		},
+	// Already a URL or scp-style remote → leave as-is.
+	if strings.Contains(r, "://") || strings.Contains(r, "@") {
+		return r
 	}
-	cmd.AddCommand(target)
-	return cmd
+	// Bare owner/name (exactly two non-empty segments, no host dot in the first).
+	parts := strings.Split(strings.Trim(r, "/"), "/")
+	if len(parts) == 2 && parts[0] != "" && parts[1] != "" && !strings.Contains(parts[0], ".") {
+		return "https://github.com/" + parts[0] + "/" + parts[1]
+	}
+	return r
 }

cli/commands_test.go

@@ -8,6 +8,24 @@ import (
 	"testing"
 )
 
+func TestNormalizeRepoURL(t *testing.T) {
+	cases := map[string]string{
+		"luxfi/wallet":                    "https://github.com/luxfi/wallet",
+		"hanzoai/cloud":                   "https://github.com/hanzoai/cloud",
+		"https://github.com/luxfi/wallet": "https://github.com/luxfi/wallet", // full URL untouched
+		"git@github.com:luxfi/wallet.git": "git@github.com:luxfi/wallet.git", // scp-style untouched
+		"https://gitlab.com/org/repo":     "https://gitlab.com/org/repo",     // non-github URL untouched
+		"owner/name/extra":                "owner/name/extra",                // not a bare owner/name
+		"single":                          "single",                          // not two segments
+		"":                                "",                                // empty
+	}
+	for in, want := range cases {
+		if got := normalizeRepoURL(in); got != want {
+			t.Errorf("normalizeRepoURL(%q) = %q, want %q", in, got, want)
+		}
+	}
+}
+
 // withPlatform points the CLI at an httptest platform via env (HANZO_PLATFORM_URL
 // + HANZO_PLATFORM_TOKEN), the same resolution path the real binary uses.
 func withPlatform(t *testing.T, h http.HandlerFunc) string {
@@ -20,11 +38,15 @@ func withPlatform(t *testing.T, h http.HandlerFunc) string {
 	return srv.URL
 }
 
+// apps list hits the LIVE board path /v1/paas/apps and renders the fleet table.
 func TestAppsListCommandTable(t *testing.T) {
-	withPlatform(t, func(w http.ResponseWriter, _ *http.Request) {
+	withPlatform(t, func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/paas/apps" {
+			t.Errorf("apps path = %s, want /v1/paas/apps", r.URL.Path)
+		}
 		_ = json.NewEncoder(w).Encode(AppsList{
 			Apps: []AppView{
-				{Org: "hanzoai", App: "iam", Env: "main", DeclaredTag: strptr("v1.2.3"), RunningTag: strptr("v1.2.3"), Health: strptr("green"), Drift: json.RawMessage(`{"severity":"ok"}`)},
+				{Org: "hanzoai", App: "iam", Env: "main", DeclaredTag: "v1.2.3", RunningTag: "v1.2.3", Health: "green", Drift: json.RawMessage(`{"severity":"ok"}`)},
 			},
 			Summary: struct {
 				Total   int            `json:"total"`
@@ -43,6 +65,20 @@ func TestAppsListCommandTable(t *testing.T) {
 	}
 }
 
+// apps list honors --env/--health/--drift as server query params (the board filters).
+func TestAppsListCommandFilters(t *testing.T) {
+	withPlatform(t, func(w http.ResponseWriter, r *http.Request) {
+		q := r.URL.Query()
+		if q.Get("env") != "main" || q.Get("health") != "red" || q.Get("drift") != "1" {
+			t.Errorf("filters not forwarded: %s", r.URL.RawQuery)
+		}
+		_ = json.NewEncoder(w).Encode(AppsList{})
+	})
+	if _, err := runRoot(t, "", "apps", "list", "--env", "main", "--health", "red", "--drift"); err != nil {
+		t.Fatalf("apps list filters: %v", err)
+	}
+}
+
 func TestAppsListCommandJSON(t *testing.T) {
 	withPlatform(t, func(w http.ResponseWriter, _ *http.Request) {
 		_ = json.NewEncoder(w).Encode(AppsList{Apps: []AppView{{Org: "hanzoai", App: "iam", Env: "main"}}})
@@ -60,69 +96,112 @@ func TestAppsListCommandJSON(t *testing.T) {
 	}
 }
 
+// apps get hits /v1/paas/apps/{app}.
+func TestAppsGetCommand(t *testing.T) {
+	withPlatform(t, func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/paas/apps/iam" {
+			t.Errorf("path = %s, want /v1/paas/apps/iam", r.URL.Path)
+		}
+		_ = json.NewEncoder(w).Encode(AppView{ID: "hanzoai/iam/main", Org: "hanzoai", App: "iam", Env: "main", DeclaredTag: "v1.2.3", Health: "green", Phase: "Running"})
+	})
+	out, err := runRoot(t, "", "apps", "get", "iam")
+	if err != nil {
+		t.Fatalf("apps get: %v", err)
+	}
+	for _, want := range []string{"hanzoai/iam/main", "Running", "v1.2.3"} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("apps get missing %q in:\n%s", want, out)
+		}
+	}
+}
+
+// deploy hits /v1/paas/apps/{app}/deploy — a rolling restart, org from identity.
 func TestDeployCommand(t *testing.T) {
 	withPlatform(t, func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/v1/org/acme/project/p1/env/e1/container/app-x/redeploy" {
-			t.Errorf("redeploy path = %s", r.URL.Path)
+		if r.URL.Path != "/v1/paas/apps/app-x/deploy" || r.URL.Query().Get("env") != "main" {
+			t.Errorf("redeploy request = %s?%s", r.URL.Path, r.URL.RawQuery)
 		}
-		_ = json.NewEncoder(w).Encode(map[string]bool{"ok": true})
+		w.WriteHeader(202)
+		_ = json.NewEncoder(w).Encode(DeployResult{OK: true, App: "app-x", Namespace: "hanzo", Env: "main", RestartedAt: "2026-07-18T12:00:00Z"})
 	})
-	out, err := runRoot(t, "", "deploy", "app-x", "--org", "acme", "--project", "p1", "--env", "e1")
+	out, err := runRoot(t, "", "deploy", "app-x", "--env", "main")
 	if err != nil {
 		t.Fatalf("deploy: %v", err)
 	}
-	if !strings.Contains(out, "redeployed app-x") {
+	if !strings.Contains(out, "restarted app-x") || !strings.Contains(out, "namespace=hanzo") {
 		t.Fatalf("deploy output: %q", out)
 	}
 }
 
-func TestDeployRequiresProjectEnv(t *testing.T) {
-	withPlatform(t, func(w http.ResponseWriter, _ *http.Request) { w.WriteHeader(200) })
-	if _, err := runRoot(t, "", "deploy", "app-x", "--org", "acme"); err == nil {
-		t.Fatalf("deploy must require --project/--env")
+// deploy REQUIRES --env — a bare deploy errors CLI-side, never silently prod.
+func TestDeployRequiresEnv(t *testing.T) {
+	withPlatform(t, func(w http.ResponseWriter, _ *http.Request) {
+		t.Error("deploy without --env must not reach the server")
+		w.WriteHeader(202)
+	})
+	if _, err := runRoot(t, "", "deploy", "app-x"); err == nil {
+		t.Fatalf("deploy must require --env")
 	}
 }
 
-func TestDeployRequiresOrg(t *testing.T) {
-	withPlatform(t, func(w http.ResponseWriter, _ *http.Request) { w.WriteHeader(200) })
-	if _, err := runRoot(t, "", "deploy", "app-x", "--project", "p1", "--env", "e1"); err == nil {
-		t.Fatalf("deploy must require an org")
+// deploy --env selects the lifecycle namespace via the ?env query param.
+func TestDeployCommandEnv(t *testing.T) {
+	withPlatform(t, func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/paas/apps/chat/deploy" || r.URL.Query().Get("env") != "test" {
+			t.Errorf("deploy env request = %s?%s", r.URL.Path, r.URL.RawQuery)
+		}
+		w.WriteHeader(202)
+		_ = json.NewEncoder(w).Encode(DeployResult{OK: true, App: "chat", Namespace: "hanzo-testnet", Env: "test", RestartedAt: "2026-07-18T12:00:00Z"})
+	})
+	if _, err := runRoot(t, "", "deploy", "chat", "--env", "test"); err != nil {
+		t.Fatalf("deploy --env: %v", err)
 	}
 }
 
+// A non-ok deploy response is surfaced as an error.
+func TestDeployNotOK(t *testing.T) {
+	withPlatform(t, func(w http.ResponseWriter, _ *http.Request) {
+		_ = json.NewEncoder(w).Encode(DeployResult{OK: false})
+	})
+	if _, err := runRoot(t, "", "deploy", "app-x", "--env", "main"); err == nil {
+		t.Fatalf("deploy must error when the server does not report ok")
+	}
+}
+
+// clusters list hits the LIVE /v1/clusters (org from identity, not the path).
 func TestClustersListCommand(t *testing.T) {
 	withPlatform(t, func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/v1/org/acme/cluster" {
-			t.Errorf("path = %s", r.URL.Path)
+		if r.URL.Path != "/v1/clusters" {
+			t.Errorf("path = %s, want /v1/clusters", r.URL.Path)
 		}
 		_ = json.NewEncoder(w).Encode(map[string]any{"clusters": []Cluster{
-			{DoksClusterID: "c1", Name: "hanzo-acme", Region: "sfo3", Status: "running", Phase: "ready", Active: true, OperatorInstalled: true, BaselineInstalled: true},
+			{DoksClusterID: "c1", Name: "hanzo-acme", Region: "sfo3", Status: "running", Kind: "managed", NodeCount: 3, NodeSize: "s-2vcpu-4gb", NvidiaGPU: 2},
 		}})
 	})
-	out, err := runRoot(t, "", "clusters", "list", "--org", "acme")
+	out, err := runRoot(t, "", "clusters", "list")
 	if err != nil {
 		t.Fatalf("clusters list: %v", err)
 	}
-	for _, want := range []string{"NAME", "hanzo-acme", "c1", "ready", "yes"} {
+	for _, want := range []string{"NAME", "hanzo-acme", "c1", "managed", "2 nvidia"} {
 		if !strings.Contains(out, want) {
 			t.Fatalf("clusters list missing %q in:\n%s", want, out)
 		}
 	}
 }
 
-func TestK8sTargetCommand(t *testing.T) {
-	withPlatform(t, func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/v1/org/acme/cluster/select" {
-			t.Errorf("path = %s", r.URL.Path)
-		}
-		_ = json.NewEncoder(w).Encode(map[string]any{"target": Target{Cluster: "hanzo-k8s", Dedicated: false, Namespaces: map[string]string{"hanzo": "main"}}})
+// clusters get filters the live list client-side by id or name.
+func TestClustersGetCommand(t *testing.T) {
+	withPlatform(t, func(w http.ResponseWriter, _ *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{"clusters": []Cluster{
+			{DoksClusterID: "c1", Name: "hanzo-acme", Region: "sfo3", Status: "running", Kind: "byo", NodeCount: 1},
+		}})
 	})
-	out, err := runRoot(t, "", "k8s", "target", "--org", "acme")
+	out, err := runRoot(t, "", "clusters", "get", "c1")
 	if err != nil {
-		t.Fatalf("k8s target: %v", err)
+		t.Fatalf("clusters get: %v", err)
 	}
-	if !strings.Contains(out, "hanzo-k8s") || !strings.Contains(out, "shared") {
-		t.Fatalf("k8s target output: %q", out)
+	if !strings.Contains(out, "hanzo-acme") || !strings.Contains(out, "byo") {
+		t.Fatalf("clusters get output: %q", out)
 	}
 }
 
@@ -167,5 +246,3 @@ func TestConfigSetGetCommand(t *testing.T) {
 		t.Fatalf("config get = %q", out)
 	}
 }
-
-func strptr(s string) *string { return &s }

cli/gpu.go

@@ -54,6 +54,10 @@ const (
 	heartbeatEvery = 30 * time.Second
 	claimPoll      = 2 * time.Second
 	claimLeaseSecs = 120
+	// renderWindow matches the dispatch cap (studio gpu_dispatch sets
+	// startToCloseTimeout 14400s). The old 10m local poll undercut it and
+	// marked live renders failed while they kept sampling (observed 8-70m).
+	renderWindow = 4 * time.Hour
 	// localComfyUI is the studio render backend the studio.render handler drives.
 	localComfyUI = "http://127.0.0.1:8188"
 	// defaultStudioUploadURL is where finished render outputs are POSTed so they
@@ -95,6 +99,7 @@ func newGPUCmd(envOf func() *Env, _ *globalFlags) *cobra.Command {
 	var engineEndpoint string
 	var registerProvider bool
 	var studioDir string
+	var studioURL string
 	connect := &cobra.Command{
 		Use:   "connect",
 		Short: "Register this GPU and run the outbound worker loop",
@@ -107,6 +112,7 @@ func newGPUCmd(envOf func() *Env, _ *globalFlags) *cobra.Command {
 				engineEndpoint:   engineEndpoint,
 				registerProvider: registerProvider,
 				studioDir:        studioDir,
+				studioURL:        studioURL,
 			}
 			if daemon {
 				return installDaemon(cmd, opts)
@@ -121,6 +127,7 @@ func newGPUCmd(envOf func() *Env, _ *globalFlags) *cobra.Command {
 	connect.Flags().StringVar(&engineEndpoint, "engine-endpoint", "", "public URL to advertise for gateway routing (defaults to --engine-url; a BYO node needs a reachable URL/tunnel)")
 	connect.Flags().BoolVar(&registerProvider, "register-provider", false, "auto-register the engine endpoint as an org model provider (POST /v1/add-provider)")
 	connect.Flags().StringVar(&studioDir, "studio-dir", os.Getenv("HANZO_STUDIO_DIR"), "local Hanzo Studio checkout; when set, connect launches and supervises the render backend on 127.0.0.1:8188")
+	connect.Flags().StringVar(&studioURL, "studio-url", firstNonEmpty(os.Getenv("HANZO_STUDIO_UPLOAD_URL"), defaultStudioUploadURL), "studio base URL the render mirror uploads finished images to (POST /v1/library/upload)")
 
 	status := &cobra.Command{
 		Use:   "status",
@@ -370,6 +377,7 @@ type connectOpts struct {
 	engineEndpoint   string // public URL to advertise (defaults to engineURL)
 	registerProvider bool   // auto POST /v1/add-provider for the engine
 	studioDir        string // local Studio checkout to launch + supervise on :8188
+	studioURL        string // studio base the render mirror uploads finished images to
 }
 
 func runConnect(cmd *cobra.Command, env *Env, opts connectOpts) error {
@@ -422,6 +430,25 @@ func runConnect(cmd *cobra.Command, env *Env, opts connectOpts) error {
 	poll := time.NewTicker(claimPoll)
 	defer poll.Stop()
 
+	// Render mirror — independent of claims by design. It scans the local studio
+	// output tree every heartbeatEvery and uploads every image to the org's library
+	// (POST /v1/library/upload), so EVERY render lands in studio.hanzo.ai even when
+	// it was produced outside the job path — a graph hand-run on this node, or a
+	// render that finished after its activity was reaped (the stranded-late-render
+	// class). Active only when a studio checkout is named (there is local output to
+	// mirror); a nil channel case never fires when it is not.
+	w.studioUploadURL = firstNonEmpty(opts.studioURL, w.studioUploadURL)
+	mirrorBase := w.studioUploadURL
+	mirrorDir := ""
+	seen := map[string]int64{}
+	var mirC <-chan time.Time
+	if opts.studioDir != "" {
+		mirrorDir = filepath.Join(opts.studioDir, "output")
+		mir := time.NewTicker(heartbeatEvery)
+		defer mir.Stop()
+		mirC = mir.C
+	}
+
 	// Heartbeat once immediately so the machine reports online without waiting a
 	// full interval.
 	_ = w.heartbeat(ctx)
@@ -449,6 +476,8 @@ func runConnect(cmd *cobra.Command, env *Env, opts connectOpts) error {
 			if err := w.claimAndRun(ctx, out); err != nil {
 				fmt.Fprintf(cmd.ErrOrStderr(), "claim: %v\n", err)
 			}
+		case <-mirC:
+			w.mirrorRenders(ctx, out, mirrorDir, mirrorBase, seen)
 		}
 	}
 }
@@ -732,7 +761,7 @@ func (e *Env) ensureToken(ctx context.Context) (string, error) {
 				nc.RefreshToken = e.creds.RefreshToken
 			}
 			*e.creds = *nc
-			_ = e.creds.Save()
+			_ = SaveActive(e.creds) // refresh the active identity in the store + mirror
 		}
 		// On refresh failure fall through: the current token may still be valid
 		// (clock skew) and the server is the authority.
@@ -795,6 +824,14 @@ func (w *worker) studioRenderHandler(ctx context.Context, input json.RawMessage)
 		return nil, fmt.Errorf("studio.render: input needs a `prompt` graph")
 	}
 	cl := &http.Client{Timeout: 60 * time.Second}
+	// The claim-to-submit window: hold off the supervisor's recycle, and wait out
+	// one if it is already mid-flight — a claimed job must never die on staging
+	// because the engine happened to be restarting.
+	staging.Add(1)
+	defer staging.Add(-1)
+	if err := waitEngine(ctx, cl); err != nil {
+		return nil, fmt.Errorf("studio.render: %w", err)
+	}
 	// Materialize any uploaded inputs (they live in orgs/{org}/input on the cloud
 	// pod, which this worker cannot read) into the LOCAL studio input dir via its
 	// own /upload/image, so LoadImage resolves them before we render.
@@ -823,7 +860,7 @@ func (w *worker) studioRenderHandler(ctx context.Context, input json.RawMessage)
 		return nil, fmt.Errorf("studio.render: no prompt_id in /prompt response")
 	}
 	// Poll history until the prompt shows up (completed).
-	deadline := time.Now().Add(10 * time.Minute)
+	deadline := time.Now().Add(renderWindow)
 	for time.Now().Before(deadline) {
 		select {
 		case <-ctx.Done():
@@ -847,6 +884,11 @@ func (w *worker) studioRenderHandler(ctx context.Context, input json.RawMessage)
 			if uerr != nil {
 				return nil, fmt.Errorf("studio.render: prompt %s rendered but gallery upload failed: %w", pr.PromptID, uerr)
 			}
+			// The engine leaks ~58GB per render; recycling after each completed
+			// render caps it at one render's worth. Boot (~20s) is noise next to
+			// 8-70m renders. Never recycle on the timeout path — the engine may
+			// still be sampling and the mirror rescues late finishes.
+			requestStudioRecycle()
 			return map[string]any{"promptId": pr.PromptID, "outputs": outputs, "gallery": gallery}, nil
 		}
 	}
@@ -961,6 +1003,120 @@ func (w *worker) postGalleryOutput(ctx context.Context, base, tok, org, name, su
 	return filepath.Join(out.Subfolder, out.Name), nil
 }
 
+// isImageFile reports whether name carries a render image extension the library accepts.
+func isImageFile(name string) bool {
+	switch strings.ToLower(filepath.Ext(name)) {
+	case ".png", ".jpg", ".jpeg", ".webp":
+		return true
+	}
+	return false
+}
+
+// mirrorRenders scans dir (the local studio output tree) for image files new or
+// changed since the last scan and POSTs each to base/v1/library/upload with the
+// worker's bearer, tagged with this node's identity, so EVERY render lands in the
+// org's studio library — including ones produced OUTSIDE the job path. seen (rel
+// path -> size) skips unchanged files; the endpoint dedupes, so a re-scan after a
+// restart is cheap and harmless. One log line per newly stored file; upload
+// failures are summarized once per scan and retried next tick (no 5xx log spam).
+func (w *worker) mirrorRenders(ctx context.Context, out io.Writer, dir, base string, seen map[string]int64) {
+	tok, err := w.env.ensureToken(ctx)
+	if err != nil {
+		return
+	}
+	base = strings.TrimRight(base, "/")
+	failed := 0
+	var firstErr error
+	_ = filepath.Walk(dir, func(p string, info os.FileInfo, werr error) error {
+		if werr != nil || info == nil || info.IsDir() || !isImageFile(p) {
+			return nil
+		}
+		// Hidden files and AppleDouble forks (`._*`, `.DS_Store`) ride along with
+		// mac scp and are not renders — `._foo.png` passes the extension check
+		// but is a 4KB resource fork that poisons the library.
+		if strings.HasPrefix(filepath.Base(p), ".") {
+			return nil
+		}
+		rel, rerr := filepath.Rel(dir, p)
+		if rerr != nil {
+			return nil
+		}
+		rel = filepath.ToSlash(rel)
+		if seen[rel] == info.Size() {
+			return nil
+		}
+		data, derr := os.ReadFile(p)
+		if derr != nil || len(data) == 0 {
+			return nil
+		}
+		sub, name := "", rel
+		if i := strings.LastIndex(rel, "/"); i >= 0 {
+			sub, name = rel[:i], rel[i+1:]
+		}
+		existed, perr := w.postLibraryUpload(ctx, base, tok, sub, name, data)
+		if perr != nil {
+			failed++
+			if firstErr == nil {
+				firstErr = perr
+			}
+			return nil
+		}
+		seen[rel] = info.Size()
+		if !existed {
+			fmt.Fprintf(out, "mirrored %s (%d bytes) -> %s\n", rel, len(data), base)
+		}
+		return nil
+	})
+	if failed > 0 {
+		fmt.Fprintf(out, "mirror: %d file(s) failed to upload, will retry: %v\n", failed, firstErr)
+	}
+}
+
+// postLibraryUpload multipart-POSTs one image to base/v1/library/upload with the
+// worker's IAM bearer, landing it in the org's library (orgs/{org}/output). The
+// file's subfolder rides as ?subpath and this node's identity as ?node so the
+// render is filterable by its source in Queue & History. Returns whether the
+// endpoint already had a byte-identical copy (dedup).
+func (w *worker) postLibraryUpload(ctx context.Context, base, tok, sub, name string, data []byte) (bool, error) {
+	var buf bytes.Buffer
+	mw := multipart.NewWriter(&buf)
+	part, err := mw.CreateFormFile("image", name)
+	if err != nil {
+		return false, err
+	}
+	if _, err := part.Write(data); err != nil {
+		return false, err
+	}
+	if err := mw.Close(); err != nil {
+		return false, err
+	}
+	q := url.Values{"node": {w.identity}}
+	if sub != "" {
+		q.Set("subpath", sub)
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, base+"/v1/library/upload?"+q.Encode(), &buf)
+	if err != nil {
+		return false, err
+	}
+	req.Header.Set("Authorization", "Bearer "+tok)
+	req.Header.Set("Content-Type", mw.FormDataContentType())
+	req.Header.Set("Accept", "application/json")
+	resp, err := w.http.Do(req)
+	if err != nil {
+		return false, err
+	}
+	defer resp.Body.Close()
+	raw, _ := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
+	if resp.StatusCode/100 != 2 {
+		return false, fmt.Errorf("POST /v1/library/upload HTTP %d: %s", resp.StatusCode, serverMessage(raw))
+	}
+	var out struct {
+		Existed bool `json:"existed"`
+	}
+	_ = json.Unmarshal(raw, &out)
+	return out.Existed, nil
+}
+
 // inputImage is one uploaded input shipped with the job: a base64 blob plus the
 // input-dir-relative location it must occupy on this worker so LoadImage finds it.
 type inputImage struct {
@@ -1015,14 +1171,47 @@ func (w *worker) materializeInputs(ctx context.Context, cl *http.Client, inputs
 	return nil
 }
 
-// collectOutputs pulls the output image/file names out of a ComfyUI history entry.
+// waitEngine blocks until the local engine answers its /queue — up to 90s, which
+// outlasts any supervisor recycle (engine restart is seconds, model reload longer).
+func waitEngine(ctx context.Context, cl *http.Client) error {
+	deadline := time.Now().Add(90 * time.Second)
+	for {
+		req, err := http.NewRequestWithContext(ctx, http.MethodGet, localComfyUI+"/queue", nil)
+		if err != nil {
+			return err
+		}
+		resp, err := cl.Do(req)
+		if err == nil {
+			_, _ = io.Copy(io.Discard, io.LimitReader(resp.Body, 1<<10))
+			resp.Body.Close()
+			if resp.StatusCode/100 == 2 {
+				return nil
+			}
+		}
+		if time.Now().After(deadline) {
+			return fmt.Errorf("engine not up: %v", err)
+		}
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-time.After(3 * time.Second):
+		}
+	}
+}
+
+// collectOutputs pulls the output file names out of a ComfyUI history entry. Savers
+// publish under different keys — SaveImage/SaveVideo under "images", SaveGLB under
+// "3d" — so every saver's outputs are gathered, not just images (a 3D mesh would
+// otherwise never travel back to the library).
 func collectOutputs(entry json.RawMessage) []string {
+	type namedFile struct {
+		Filename  string `json:"filename"`
+		Subfolder string `json:"subfolder"`
+	}
 	var e struct {
 		Outputs map[string]struct {
-			Images []struct {
-				Filename  string `json:"filename"`
-				Subfolder string `json:"subfolder"`
-			} `json:"images"`
+			Images []namedFile `json:"images"`
+			ThreeD []namedFile `json:"3d"`
 		} `json:"outputs"`
 	}
 	if err := json.Unmarshal(entry, &e); err != nil {
@@ -1030,8 +1219,8 @@ func collectOutputs(entry json.RawMessage) []string {
 	}
 	var files []string
 	for _, node := range e.Outputs {
-		for _, img := range node.Images {
-			files = append(files, filepath.Join(img.Subfolder, img.Filename))
+		for _, f := range append(append([]namedFile{}, node.Images...), node.ThreeD...) {
+			files = append(files, filepath.Join(f.Subfolder, f.Filename))
 		}
 	}
 	return files

cli/gpu_mirror_test.go

@@ -0,0 +1,92 @@
+package cli
+
+import (
+	"bytes"
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+// TestMirrorRenders verifies the render mirror: it scans the local studio output
+// tree and POSTs every image (new or changed) to /v1/library/upload with the node's
+// identity + subfolder + bearer, skips unchanged files across scans, and re-uploads
+// a changed file. This is the path that lands EVERY render in studio.hanzo.ai even
+// when it was produced outside the job/claim path.
+func TestMirrorRenders(t *testing.T) {
+	t.Setenv("HANZO_TOKEN", "test-bearer")
+
+	dir := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(dir, "renders"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	write := func(rel, body string) {
+		if err := os.WriteFile(filepath.Join(dir, filepath.FromSlash(rel)), []byte(body), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("renders/a.png", "\x89PNG-a")
+	write("top.jpg", "jpg-top")
+	write("notes.txt", "not an image")
+
+	type up struct{ name, node, subpath, auth string }
+	var got []up
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/library/upload" {
+			http.Error(w, "not found", http.StatusNotFound)
+			return
+		}
+		_ = r.ParseMultipartForm(1 << 20)
+		name := ""
+		if r.MultipartForm != nil {
+			for _, fh := range r.MultipartForm.File["image"] {
+				name = fh.Filename
+			}
+		}
+		got = append(got, up{
+			name:    name,
+			node:    r.URL.Query().Get("node"),
+			subpath: r.URL.Query().Get("subpath"),
+			auth:    r.Header.Get("Authorization"),
+		})
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"ok":true,"path":"x"}`))
+	}))
+	defer srv.Close()
+
+	w := &worker{env: &Env{}, http: &http.Client{Timeout: 10 * time.Second}, identity: "spark"}
+	seen := map[string]int64{}
+	var buf bytes.Buffer
+
+	w.mirrorRenders(context.Background(), &buf, dir, srv.URL, seen)
+	if len(got) != 2 {
+		t.Fatalf("uploaded %d files, want 2 images (the .txt is skipped): %+v", len(got), got)
+	}
+	var a *up
+	for i := range got {
+		if got[i].name == "a.png" {
+			a = &got[i]
+		}
+	}
+	if a == nil || a.subpath != "renders" || a.node != "spark" || a.auth != "Bearer test-bearer" {
+		t.Fatalf("renders/a.png upload = %+v, want subpath=renders node=spark bearer set", a)
+	}
+
+	// A second scan re-uploads nothing (seen matches every size).
+	got = nil
+	w.mirrorRenders(context.Background(), &buf, dir, srv.URL, seen)
+	if len(got) != 0 {
+		t.Fatalf("second scan uploaded %d files, want 0 (all unchanged): %+v", len(got), got)
+	}
+
+	// A changed file is re-uploaded on the next scan.
+	write("renders/a.png", "\x89PNG-a-grew")
+	got = nil
+	w.mirrorRenders(context.Background(), &buf, dir, srv.URL, seen)
+	if len(got) != 1 || got[0].name != "a.png" {
+		t.Fatalf("after change, uploaded %+v, want just renders/a.png", got)
+	}
+}

cli/platform.go

@@ -12,11 +12,13 @@ import (
 	"time"
 )
 
-// Platform is a thin client over the platform.hanzo.ai /v1 control plane. That
-// surface is machine-to-machine (service-token, "No OIDC" — it cannot validate
-// IAM user tokens), so the token here is the platform service token, resolved
-// from flag/env/credential store by the caller; the build endpoint takes its
-// own token per call.
+// Platform is a thin client over the LIVE Hanzo Cloud control plane
+// (platform.hanzo.ai / api.hanzo.ai → svc `cloud`, the Go binary). Every route it
+// calls is served by that one binary and authorized off ONE IAM identity: after
+// `hanzo login` the CLI sends the IAM access token as the bearer, and the cloud's
+// identity middleware (SanitizeIdentity) validates the JWT and org-scopes the
+// caller — no separate platform/service token. A purpose-minted machine token
+// still works (flag > env > credential store > IAM login) for automation.
 type Platform struct {
 	baseURL string
 	token   string
@@ -45,8 +47,8 @@ func (e *apiError) Error() string {
 		msg = http.StatusText(e.status)
 	}
 	hint := ""
-	if e.status == http.StatusUnauthorized {
-		hint = " (set the platform service token: --platform-token, HANZO_PLATFORM_TOKEN, or `hanzo login --platform-token`)"
+	if e.status == http.StatusUnauthorized || e.status == http.StatusForbidden {
+		hint = " (run `hanzo login` — your IAM identity authorizes the platform; admin ops need an org-admin or superadmin identity)"
 	}
 	return fmt.Sprintf("platform %s: HTTP %d: %s%s", e.path, e.status, msg, hint)
 }
@@ -55,7 +57,7 @@ func (e *apiError) Error() string {
 // into out (when non-nil) and mapping a non-2xx into an *apiError.
 func (p *Platform) do(ctx context.Context, method, path, token string, body, out any) error {
 	if token == "" {
-		return fmt.Errorf("no platform token: pass --platform-token, set HANZO_PLATFORM_TOKEN, or run `hanzo login --platform-token <tok>`")
+		return fmt.Errorf("not authenticated: run `hanzo login` (an IAM login now authorizes the platform; a --platform-token / HANZO_PLATFORM_TOKEN still works for machine automation)")
 	}
 	var rdr io.Reader
 	if body != nil {
@@ -94,8 +96,8 @@ func (p *Platform) do(ctx context.Context, method, path, token string, body, out
 	return nil
 }
 
-// serverMessage pulls the `{ "message": … }` field platform errors use, falling
-// back to the raw (truncated) body.
+// serverMessage pulls the `{ "message": … }` / `{ "error": … }` field the cloud's
+// errors use, falling back to the raw (truncated) body.
 func serverMessage(raw []byte) string {
 	var e struct {
 		Message string `json:"message"`
@@ -117,33 +119,35 @@ func serverMessage(raw []byte) string {
 }
 
 // ---------------------------------------------------------------------------
-// Apps board — GET /v1/apps, GET /v1/apps/{id}, POST /v1/apps/sync.
+// Apps board — GET /v1/paas/apps, GET /v1/paas/apps/{app}. The live Go cloud's
+// fleet drift board (clients/paas): the operator App CRs across the platform
+// namespaces, declared/running/latest tags + health + the drift verdict. It is
+// org-confined server-side (a SuperAdmin sees the fleet; an OrgAdmin only its own
+// org), so the CLI sends NO org filter — identity scopes the view.
 // ---------------------------------------------------------------------------
 
-// AppView mirrors the platform apps-lifecycle DTO. Nullable columns are *string
-// so JSON null round-trips; Drift is kept raw so --json is byte-faithful and
-// the drift schema can evolve without a client bump.
+// AppView mirrors clients/paas.AppView (the LIVE board DTO). Tags/health are plain
+// strings ("" == unknown, rendered "-"); Drift is kept raw so --json is
+// byte-faithful and the drift schema can evolve without a client bump.
 type AppView struct {
-	ID            string          `json:"id"`
-	Org           string          `json:"org"`
-	App           string          `json:"app"`
-	Env           string          `json:"env"`
-	Repo          string          `json:"repo"`
-	Registry      string          `json:"registry"`
-	DeclaredTag   *string         `json:"declaredTag"`
-	RunningTag    *string         `json:"runningTag"`
-	LatestTag     *string         `json:"latestTag"`
-	ReleaseURL    *string         `json:"releaseUrl"`
-	ReleaseAssets int             `json:"releaseAssets"`
-	Health        *string         `json:"health"`
-	Cluster       *string         `json:"cluster"`
-	Namespace     *string         `json:"namespace"`
-	LastObserved  *string         `json:"lastObserved"`
-	UpdatedAt     string          `json:"updatedAt"`
-	Drift         json.RawMessage `json:"drift"`
+	ID          string          `json:"id"`  // <org>/<app>/<env>, e.g. hanzoai/iam/main
+	Org         string          `json:"org"` // image namespace, e.g. hanzoai
+	App         string          `json:"app"`
+	Env         string          `json:"env"`
+	Repo        string          `json:"repo"`
+	Registry    string          `json:"registry"`
+	DeclaredTag string          `json:"declaredTag"`
+	RunningTag  string          `json:"runningTag"`
+	LatestTag   string          `json:"latestTag"`
+	Health      string          `json:"health"`
+	Phase       string          `json:"phase"`
+	Cluster     string          `json:"cluster"`
+	Namespace   string          `json:"namespace"`
+	Endpoints   []string        `json:"endpoints"`
+	Drift       json.RawMessage `json:"drift"`
 }
 
-// AppsList is the /v1/apps envelope: ordered rows + a drift summary.
+// AppsList is the /v1/paas/apps envelope: ordered rows + a drift summary.
 type AppsList struct {
 	Apps    []AppView `json:"apps"`
 	Summary struct {
@@ -152,9 +156,10 @@ type AppsList struct {
 	} `json:"summary"`
 }
 
-// AppsQuery are the optional /v1/apps filters.
+// AppsQuery are the optional /v1/paas/apps filters (server-honored). Env/Health/
+// Drift narrow the board; there is deliberately no org filter — the board is
+// confined to the caller's org by the validated identity, never a client value.
 type AppsQuery struct {
-	Org    string
 	Env    string
 	Health string
 	Drift  bool
@@ -162,9 +167,6 @@ type AppsQuery struct {
 
 func (p *Platform) Apps(ctx context.Context, q AppsQuery) (*AppsList, error) {
 	v := url.Values{}
-	if q.Org != "" {
-		v.Set("org", q.Org)
-	}
 	if q.Env != "" {
 		v.Set("env", q.Env)
 	}
@@ -174,7 +176,7 @@ func (p *Platform) Apps(ctx context.Context, q AppsQuery) (*AppsList, error) {
 	if q.Drift {
 		v.Set("drift", "1")
 	}
-	path := "/v1/apps"
+	path := "/v1/paas/apps"
 	if len(v) > 0 {
 		path += "?" + v.Encode()
 	}
@@ -182,17 +184,11 @@ func (p *Platform) Apps(ctx context.Context, q AppsQuery) (*AppsList, error) {
 	return out, p.do(ctx, http.MethodGet, path, p.token, nil, out)
 }
 
-func (p *Platform) App(ctx context.Context, id, org string) (*AppView, error) {
-	path := "/v1/apps/" + id
-	if org != "" {
-		path += "?org=" + url.QueryEscape(org)
-	}
+// App gets one app row by its <app> CR name (production by default; the server
+// scans the caller's authorized namespaces main→test→dev).
+func (p *Platform) App(ctx context.Context, app string) (*AppView, error) {
 	out := &AppView{}
-	return out, p.do(ctx, http.MethodGet, path, p.token, nil, out)
-}
-
-func (p *Platform) SyncApps(ctx context.Context) error {
-	return p.do(ctx, http.MethodPost, "/v1/apps/sync", p.token, nil, nil)
+	return out, p.do(ctx, http.MethodGet, "/v1/paas/apps/"+url.PathEscape(app), p.token, nil, out)
 }
 
 // driftSeverity extracts the severity string from the raw drift object.
@@ -207,110 +203,92 @@ func driftSeverity(raw json.RawMessage) string {
 }
 
 // ---------------------------------------------------------------------------
-// Dedicated clusters — /v1/org/{org}/cluster[ /select | /{id}/install-baseline ].
+// Clusters — GET /v1/clusters. The live Go cloud's compute fleet (clients/visor):
+// Visor-managed node pools + the org's BYO clusters, tenant-scoped server-side by
+// the validated org (?owner is the caller's IAM org). No org in the path.
 // ---------------------------------------------------------------------------
 
-// Cluster mirrors a doks_cluster record. `status` is DigitalOcean state; `phase`
-// is the platform provisioning lifecycle — orthogonal (a DO-running cluster is
-// not a usable target until phase=ready).
+// NodePool mirrors clients/visor.nodePoolView.
+type NodePool struct {
+	PoolID    string `json:"poolId"`
+	Name      string `json:"name"`
+	Size      string `json:"size"`
+	Count     int    `json:"count"`
+	MinNodes  int    `json:"minNodes"`
+	MaxNodes  int    `json:"maxNodes"`
+	AutoScale bool   `json:"autoScale"`
+}
+
+// Cluster mirrors clients/visor.clusterView — the LIVE cluster DTO. `kind` is
+// "managed" (Visor-provisioned) or "byo" (attached kubeconfig).
 type Cluster struct {
-	DoksClusterID     string          `json:"doksClusterId"`
-	Name              string          `json:"name"`
-	DoClusterID       *string         `json:"doClusterId"`
-	Region            string          `json:"region"`
-	Status            string          `json:"status"`
-	Endpoint          *string         `json:"endpoint"`
-	K8sVersion        *string         `json:"k8sVersion"`
-	HA                bool            `json:"ha"`
-	Phase             string          `json:"phase"`
-	OperatorInstalled bool            `json:"operatorInstalled"`
-	BaselineInstalled bool            `json:"baselineInstalled"`
-	Active            bool            `json:"active"`
-	BaselineError     *string         `json:"baselineError"`
-	OrganizationID    string          `json:"organizationId"`
-	CreatedAt         string          `json:"createdAt"`
-	Tags              []string        `json:"tags"`
-	MaintenancePolicy json.RawMessage `json:"maintenancePolicy,omitempty"`
+	DoksClusterID string     `json:"doksClusterId"`
+	DoClusterID   string     `json:"doClusterId"`
+	Name          string     `json:"name"`
+	Region        string     `json:"region"`
+	Status        string     `json:"status"`
+	NodePools     []NodePool `json:"nodePools"`
+	NodeSize      string     `json:"nodeSize"`
+	NodeCount     int        `json:"nodeCount"`
+	CreatedAt     string     `json:"createdAt"`
+	Kind          string     `json:"kind"`
+	NvidiaGPU     int        `json:"nvidiaGpu"`
+	AmdGPU        int        `json:"amdGpu"`
 }
 
-// ProvisionReq is the dedicated-cluster provisioning body (org forced by path).
-type ProvisionReq struct {
-	Region    string `json:"region,omitempty"`
-	HA        bool   `json:"ha,omitempty"`
-	NodeSize  string `json:"nodeSize,omitempty"`
-	NodeCount int    `json:"nodeCount,omitempty"`
+// ID is the stable cluster identifier for display/lookup: the DOKS id when managed,
+// else the name (a BYO cluster keys on its attached name).
+func (c Cluster) ID() string {
+	if c.DoksClusterID != "" {
+		return c.DoksClusterID
+	}
+	return c.Name
 }
 
-// Target is the redacted ClusterTargetView — the kubeconfig is never present.
-type Target struct {
-	Cluster    string            `json:"cluster"`
-	Namespaces map[string]string `json:"namespaces"`
-	Dedicated  bool              `json:"dedicated"`
-}
-
-func (p *Platform) Clusters(ctx context.Context, org string) ([]Cluster, error) {
+func (p *Platform) Clusters(ctx context.Context) ([]Cluster, error) {
 	var out struct {
 		Clusters []Cluster `json:"clusters"`
 	}
-	err := p.do(ctx, http.MethodGet, "/v1/org/"+url.PathEscape(org)+"/cluster", p.token, nil, &out)
+	err := p.do(ctx, http.MethodGet, "/v1/clusters", p.token, nil, &out)
 	return out.Clusters, err
 }
 
-func (p *Platform) ProvisionCluster(ctx context.Context, org string, req ProvisionReq) (*Cluster, error) {
-	var out struct {
-		Cluster Cluster `json:"cluster"`
-	}
-	err := p.do(ctx, http.MethodPost, "/v1/org/"+url.PathEscape(org)+"/cluster", p.token, req, &out)
-	return &out.Cluster, err
-}
-
-func (p *Platform) Target(ctx context.Context, org string) (*Target, error) {
-	var out struct {
-		Target Target `json:"target"`
-	}
-	err := p.do(ctx, http.MethodGet, "/v1/org/"+url.PathEscape(org)+"/cluster/select", p.token, nil, &out)
-	return &out.Target, err
-}
-
-// SelectTarget activates a dedicated cluster as the org's deploy target, or
-// reverts to the shared cluster when clusterID is nil.
-func (p *Platform) SelectTarget(ctx context.Context, org string, clusterID *string) (*Target, error) {
-	var out struct {
-		Target Target `json:"target"`
-	}
-	body := map[string]any{"doksClusterId": clusterID}
-	err := p.do(ctx, http.MethodPost, "/v1/org/"+url.PathEscape(org)+"/cluster/select", p.token, body, &out)
-	return &out.Target, err
-}
-
-func (p *Platform) InstallBaseline(ctx context.Context, org, clusterID string) error {
-	path := "/v1/org/" + url.PathEscape(org) + "/cluster/" + url.PathEscape(clusterID) + "/install-baseline"
-	return p.do(ctx, http.MethodPost, path, p.token, nil, nil)
-}
-
 // ---------------------------------------------------------------------------
-// Deploy — POST …/container/{id}/redeploy (rolling restart, zero-downtime).
+// Deploy — POST /v1/paas/apps/{app}/deploy: a zero-downtime ROLLING RESTART of the
+// app's Deployment (re-pulls the declared image, recreates pods). Org-confined
+// server-side; an optional env selects the lifecycle namespace (main|test|dev).
 // ---------------------------------------------------------------------------
 
-// Redeploy triggers a rolling restart of the container's k8s Deployment. The
-// coordinates are exact (the platform validates org+project+env+container scope).
-func (p *Platform) Redeploy(ctx context.Context, org, project, env, container string) error {
-	path := fmt.Sprintf("/v1/org/%s/project/%s/env/%s/container/%s/redeploy",
-		url.PathEscape(org), url.PathEscape(project), url.PathEscape(env), url.PathEscape(container))
-	var out struct {
-		OK bool `json:"ok"`
+// Redeploy triggers a rolling restart of the named app. env is optional
+// (main|test|dev); empty targets production (the first match, main→test→dev).
+func (p *Platform) Redeploy(ctx context.Context, app, env string) (*DeployResult, error) {
+	path := "/v1/paas/apps/" + url.PathEscape(app) + "/deploy"
+	if env != "" {
+		path += "?env=" + url.QueryEscape(env)
 	}
-	if err := p.do(ctx, http.MethodPost, path, p.token, nil, &out); err != nil {
-		return err
+	out := &DeployResult{}
+	if err := p.do(ctx, http.MethodPost, path, p.token, nil, out); err != nil {
+		return nil, err
 	}
 	if !out.OK {
-		return fmt.Errorf("redeploy did not report ok")
+		return nil, fmt.Errorf("redeploy did not report ok")
 	}
-	return nil
+	return out, nil
+}
+
+// DeployResult is the /deploy acceptance (202): the restarted app + its namespace.
+type DeployResult struct {
+	OK          bool   `json:"ok"`
+	App         string `json:"app"`
+	Namespace   string `json:"namespace"`
+	Env         string `json:"env"`
+	RestartedAt string `json:"restartedAt"`
 }
 
 // ---------------------------------------------------------------------------
-// Build — POST /v1/runner (platform-native CI, no GitHub builders).
+// Build — POST /v1/runner (platform-native CI, no GitHub builders). Authorized off
+// the IAM login exactly like the surfaces above (or a dedicated build token for
+// machine automation). Unchanged wire contract.
 // ---------------------------------------------------------------------------
 
 // BuildReq is the direct-enqueue body. Repo/SHA/Image are required.
@@ -337,11 +315,11 @@ type BuildJob struct {
 	Target     string `json:"target"`
 }
 
-// EnqueueBuild enqueues a native build. It authenticates with the dedicated
-// build-callback token, not the service token.
+// EnqueueBuild enqueues a native build. buildToken is resolved by the caller (IAM
+// login is the final fallback; a dedicated build token wins when present).
 func (p *Platform) EnqueueBuild(ctx context.Context, req BuildReq, buildToken string) (*BuildJob, error) {
 	if buildToken == "" {
-		return nil, fmt.Errorf("no build token: set HANZO_BUILD_TOKEN / PLATFORM_BUILD_CALLBACK_TOKEN or `hanzo login --build-token <tok>`")
+		return nil, fmt.Errorf("not authenticated: run `hanzo login` (an IAM login now authorizes builds; HANZO_BUILD_TOKEN / --build-token still works for machine automation)")
 	}
 	out := &BuildJob{}
 	return out, p.do(ctx, http.MethodPost, "/v1/runner", buildToken, req, out)

cli/platform_test.go

@@ -18,17 +18,22 @@ func platformStub(t *testing.T, token string, h http.HandlerFunc) (*Platform, fu
 	return newPlatform(srv.URL, token), srv.Close
 }
 
+// Apps hits the LIVE board /v1/paas/apps with the IAM bearer; it sends NO org
+// filter (the board is org-confined server-side by the validated identity).
 func TestPlatformAuthHeaderAndApps(t *testing.T) {
 	p, done := platformStub(t, "svc-tok", func(w http.ResponseWriter, r *http.Request) {
 		if got := r.Header.Get("Authorization"); got != "Bearer svc-tok" {
 			t.Errorf("auth header = %q", got)
 		}
-		if r.URL.Path != "/v1/apps" {
-			t.Errorf("path = %s", r.URL.Path)
+		if r.URL.Path != "/v1/paas/apps" {
+			t.Errorf("path = %s, want /v1/paas/apps", r.URL.Path)
 		}
 		if r.URL.Query().Get("env") != "main" || r.URL.Query().Get("drift") != "1" {
 			t.Errorf("query = %s", r.URL.RawQuery)
 		}
+		if r.URL.Query().Has("org") {
+			t.Errorf("client must NOT send an org filter (identity confines the board): %s", r.URL.RawQuery)
+		}
 		_ = json.NewEncoder(w).Encode(AppsList{
 			Apps: []AppView{{ID: "hanzoai/iam/main", Org: "hanzoai", App: "iam", Env: "main", Drift: json.RawMessage(`{"severity":"red"}`)}},
 		})
@@ -47,126 +52,75 @@ func TestPlatformAuthHeaderAndApps(t *testing.T) {
 	}
 }
 
+// App hits /v1/paas/apps/{app}; no org query (identity scopes it).
 func TestPlatformApp(t *testing.T) {
 	p, done := platformStub(t, "t", func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/v1/apps/hanzoai/iam/main" {
-			t.Errorf("path = %s", r.URL.Path)
+		if r.URL.Path != "/v1/paas/apps/iam" {
+			t.Errorf("path = %s, want /v1/paas/apps/iam", r.URL.Path)
 		}
-		if r.URL.Query().Get("org") != "hanzoai" {
-			t.Errorf("org query = %s", r.URL.RawQuery)
+		if r.URL.RawQuery != "" {
+			t.Errorf("app get must carry no query, got %s", r.URL.RawQuery)
 		}
-		_ = json.NewEncoder(w).Encode(AppView{ID: "hanzoai/iam/main", App: "iam"})
+		_ = json.NewEncoder(w).Encode(AppView{ID: "hanzoai/iam/main", App: "iam", Phase: "Running"})
 	})
 	defer done()
-	a, err := p.App(context.Background(), "hanzoai/iam/main", "hanzoai")
+	a, err := p.App(context.Background(), "iam")
 	if err != nil || a.App != "iam" {
 		t.Fatalf("App: %v %+v", err, a)
 	}
 }
 
-func TestPlatformSyncApps(t *testing.T) {
+// Clusters hits the LIVE /v1/clusters (org from identity, not the path).
+func TestPlatformClusters(t *testing.T) {
 	p, done := platformStub(t, "t", func(w http.ResponseWriter, r *http.Request) {
-		if r.Method != http.MethodPost || r.URL.Path != "/v1/apps/sync" {
-			t.Errorf("sync = %s %s", r.Method, r.URL.Path)
-		}
-		w.WriteHeader(200)
-	})
-	defer done()
-	if err := p.SyncApps(context.Background()); err != nil {
-		t.Fatalf("SyncApps: %v", err)
-	}
-}
-
-func TestPlatformClustersAndProvision(t *testing.T) {
-	p, done := platformStub(t, "t", func(w http.ResponseWriter, r *http.Request) {
-		switch {
-		case r.Method == http.MethodGet && r.URL.Path == "/v1/org/acme/cluster":
-			_ = json.NewEncoder(w).Encode(map[string]any{"clusters": []Cluster{{DoksClusterID: "c1", Name: "hanzo-acme", Region: "sfo3", Status: "running", Phase: "ready", Active: true}}})
-		case r.Method == http.MethodPost && r.URL.Path == "/v1/org/acme/cluster":
-			body, _ := io.ReadAll(r.Body)
-			var req ProvisionReq
-			_ = json.Unmarshal(body, &req)
-			if req.Region != "sfo3" || !req.HA {
-				t.Errorf("provision body = %+v", req)
-			}
-			w.WriteHeader(201)
-			_ = json.NewEncoder(w).Encode(map[string]any{"cluster": Cluster{DoksClusterID: "c2", Name: "new", Phase: "requested"}})
-		default:
-			t.Errorf("unexpected %s %s", r.Method, r.URL.Path)
+		if r.Method != http.MethodGet || r.URL.Path != "/v1/clusters" {
+			t.Errorf("clusters = %s %s, want GET /v1/clusters", r.Method, r.URL.Path)
 		}
+		_ = json.NewEncoder(w).Encode(map[string]any{"clusters": []Cluster{
+			{DoksClusterID: "c1", Name: "hanzo-acme", Region: "sfo3", Status: "running", Kind: "managed", NodeCount: 3},
+		}})
 	})
 	defer done()
 
-	cs, err := p.Clusters(context.Background(), "acme")
-	if err != nil || len(cs) != 1 || cs[0].DoksClusterID != "c1" {
+	cs, err := p.Clusters(context.Background())
+	if err != nil || len(cs) != 1 || cs[0].ID() != "c1" || cs[0].Kind != "managed" {
 		t.Fatalf("Clusters: %v %+v", err, cs)
 	}
-	c, err := p.ProvisionCluster(context.Background(), "acme", ProvisionReq{Region: "sfo3", HA: true})
-	if err != nil || c.DoksClusterID != "c2" {
-		t.Fatalf("ProvisionCluster: %v %+v", err, c)
-	}
 }
 
-func TestPlatformTargetAndSelect(t *testing.T) {
-	p, done := platformStub(t, "t", func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/v1/org/acme/cluster/select" {
-			t.Errorf("path = %s", r.URL.Path)
-		}
-		if r.Method == http.MethodPost {
-			body, _ := io.ReadAll(r.Body)
-			var m map[string]any
-			_ = json.Unmarshal(body, &m)
-			if m["doksClusterId"] != "c1" {
-				t.Errorf("select body = %v", m)
-			}
-		}
-		_ = json.NewEncoder(w).Encode(map[string]any{"target": Target{Cluster: "hanzo-acme", Dedicated: true, Namespaces: map[string]string{"acme": "main"}}})
-	})
-	defer done()
-
-	tg, err := p.Target(context.Background(), "acme")
-	if err != nil || tg.Cluster != "hanzo-acme" || !tg.Dedicated {
-		t.Fatalf("Target: %v %+v", err, tg)
-	}
-	id := "c1"
-	if _, err := p.SelectTarget(context.Background(), "acme", &id); err != nil {
-		t.Fatalf("SelectTarget: %v", err)
-	}
-}
-
-func TestPlatformInstallBaseline(t *testing.T) {
-	p, done := platformStub(t, "t", func(w http.ResponseWriter, r *http.Request) {
-		if r.Method != http.MethodPost || r.URL.Path != "/v1/org/acme/cluster/c1/install-baseline" {
-			t.Errorf("install-baseline = %s %s", r.Method, r.URL.Path)
-		}
-		w.WriteHeader(200)
-	})
-	defer done()
-	if err := p.InstallBaseline(context.Background(), "acme", "c1"); err != nil {
-		t.Fatalf("InstallBaseline: %v", err)
+// A BYO cluster with no DOKS id keys on its name via ID().
+func TestClusterIDFallsBackToName(t *testing.T) {
+	c := Cluster{Name: "byo-1", Kind: "byo"}
+	if c.ID() != "byo-1" {
+		t.Fatalf("ID() = %q, want byo-1", c.ID())
 	}
 }
 
+// Redeploy hits /v1/paas/apps/{app}/deploy (rolling restart), org from identity.
 func TestPlatformRedeploy(t *testing.T) {
 	p, done := platformStub(t, "t", func(w http.ResponseWriter, r *http.Request) {
-		want := "/v1/org/acme/project/p1/env/e1/container/app-x/redeploy"
-		if r.Method != http.MethodPost || r.URL.Path != want {
-			t.Errorf("redeploy path = %s %s", r.Method, r.URL.Path)
+		if r.Method != http.MethodPost || r.URL.Path != "/v1/paas/apps/app-x/deploy" {
+			t.Errorf("redeploy = %s %s, want POST /v1/paas/apps/app-x/deploy", r.Method, r.URL.Path)
 		}
-		_ = json.NewEncoder(w).Encode(map[string]bool{"ok": true})
+		if r.URL.Query().Get("env") != "test" {
+			t.Errorf("env query = %s", r.URL.RawQuery)
+		}
+		w.WriteHeader(202)
+		_ = json.NewEncoder(w).Encode(DeployResult{OK: true, App: "app-x", Namespace: "hanzo-testnet", Env: "test", RestartedAt: "2026-07-18T00:00:00Z"})
 	})
 	defer done()
-	if err := p.Redeploy(context.Background(), "acme", "p1", "e1", "app-x"); err != nil {
-		t.Fatalf("Redeploy: %v", err)
+	res, err := p.Redeploy(context.Background(), "app-x", "test")
+	if err != nil || res.Namespace != "hanzo-testnet" {
+		t.Fatalf("Redeploy: %v %+v", err, res)
 	}
 }
 
 func TestPlatformRedeployNotOK(t *testing.T) {
 	p, done := platformStub(t, "t", func(w http.ResponseWriter, _ *http.Request) {
-		_ = json.NewEncoder(w).Encode(map[string]bool{"ok": false})
+		_ = json.NewEncoder(w).Encode(DeployResult{OK: false})
 	})
 	defer done()
-	if err := p.Redeploy(context.Background(), "o", "p", "e", "c"); err == nil {
+	if _, err := p.Redeploy(context.Background(), "c", ""); err == nil {
 		t.Fatalf("expected error when ok=false")
 	}
 }
@@ -207,14 +161,16 @@ func TestPlatformError401Hint(t *testing.T) {
 	})
 	defer done()
 	_, err := p.Apps(context.Background(), AppsQuery{})
-	if err == nil || !strings.Contains(err.Error(), "HTTP 401") || !strings.Contains(err.Error(), "platform service token") {
-		t.Fatalf("401 error should carry a token hint, got %v", err)
+	if err == nil || !strings.Contains(err.Error(), "HTTP 401") || !strings.Contains(err.Error(), "hanzo login") {
+		t.Fatalf("401 error should point at `hanzo login`, got %v", err)
 	}
 }
 
 func TestPlatformNoTokenError(t *testing.T) {
 	p := newPlatform("https://platform.hanzo.ai", "")
-	if _, err := p.Apps(context.Background(), AppsQuery{}); err == nil || !strings.Contains(err.Error(), "no platform token") {
-		t.Fatalf("expected no-token error, got %v", err)
+	// After unify-infra, the "no credential" error points at `hanzo login` — the one
+	// identity that authorizes the platform — not a separate platform token.
+	if _, err := p.Apps(context.Background(), AppsQuery{}); err == nil || !strings.Contains(err.Error(), "hanzo login") {
+		t.Fatalf("expected a `hanzo login` hint, got %v", err)
 	}
 }

cli/studio.go

@@ -12,6 +12,7 @@ package cli
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"io"
 	"net"
@@ -20,6 +21,7 @@ import (
 	"os/exec"
 	"path/filepath"
 	"strconv"
+	"sync/atomic"
 	"syscall"
 	"time"
 )
@@ -107,6 +109,48 @@ func stopStudio(cmd *exec.Cmd) {
 	}
 }
 
+// studioRecycle carries at most one pending recycle request; the render
+// handler signals it after each completed render (see gpu.go).
+var studioRecycle = make(chan struct{}, 1)
+
+// staging guards the claim-to-submit window: a claimed job is real work the
+// engine queue cannot see yet, so the supervisor must never recycle over it
+// (observed: jobs claimed during a recycle failed staging on a dead engine
+// and were consumed).
+var staging atomic.Int32
+
+func requestStudioRecycle() {
+	select {
+	case studioRecycle <- struct{}{}:
+	default:
+	}
+}
+
+// studioBusy reports whether the engine holds queued or running prompts.
+// A generous timeout: a saturated GB10 answers slowly mid-render — slow is
+// alive, and killing a live render costs 8-70 minutes of GPU work.
+func studioBusy(ctx context.Context) (busy, ok bool) {
+	ctx, cancel := context.WithTimeout(ctx, 20*time.Second)
+	defer cancel()
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, "http://"+studioAddr+"/queue", nil)
+	if err != nil {
+		return false, false
+	}
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return false, false
+	}
+	defer resp.Body.Close()
+	var q struct {
+		Running []json.RawMessage `json:"queue_running"`
+		Pending []json.RawMessage `json:"queue_pending"`
+	}
+	if err := json.NewDecoder(io.LimitReader(resp.Body, 32<<20)).Decode(&q); err != nil {
+		return false, false
+	}
+	return len(q.Running)+len(q.Pending) > 0, true
+}
+
 // superviseStudio keeps the local render backend on :8188 alive until ctx
 // ends. Quiet by design: one line per restart event, not a probe firehose.
 func superviseStudio(ctx context.Context, dir string, out io.Writer) {
@@ -138,6 +182,12 @@ func superviseStudio(ctx context.Context, dir string, out io.Writer) {
 
 	tick := time.NewTicker(studioProbeEvery)
 	defer tick.Stop()
+	// recyclePending defers the post-render recycle until the queue is EMPTY:
+	// short jobs complete while a long render is mid-sample, and recycling on
+	// their completion killed the live render (observed: every direct render
+	// died within ~6 minutes while probe jobs cycled).
+	recyclePending := false
+	unhealthy := 0
 	for {
 		select {
 		case <-ctx.Done():
@@ -145,19 +195,32 @@ func superviseStudio(ctx context.Context, dir string, out io.Writer) {
 				_ = syscall.Kill(-cmd.Process.Pid, syscall.SIGTERM)
 			}
 			return
+		case <-studioRecycle:
+			recyclePending = true
 		case <-tick.C:
+			busy, ok := studioBusy(ctx)
+			if recyclePending && ok && !busy && staging.Load() == 0 {
+				recyclePending = false
+				unhealthy = 0
+				restart("recycle")
+				continue
+			}
 			if studioHealthy(ctx) {
+				unhealthy = 0
 				continue
 			}
-			// Grace re-check: it may be momentarily busy mid-render.
-			select {
-			case <-ctx.Done():
+			if ok && busy {
+				// Alive-busy: slow health under render load is not death.
+				unhealthy = 0
 				continue
-			case <-time.After(studioGraceWait):
 			}
-			if !studioHealthy(ctx) {
-				restart("unresponsive")
+			// Sustained silence with an idle or unreadable queue = actually dead.
+			unhealthy++
+			if unhealthy < 3 {
+				continue
 			}
+			unhealthy = 0
+			restart("unresponsive")
 		}
 	}
 }

clients/account/account.go

@@ -161,6 +161,10 @@ func routesBridge(s *cloud.Service[state], app *zip.App) {
 	// console calls, forwarded to commerce with the admin service token and SCOPED to the
 	// validated caller's own subject (billing.go). Registered AFTER clients/billing's
 	// specific routes (121 < 122) so those win and this catches the rest. GET+POST only.
+	// The wildcard is what the ROUTER matches; it is NOT the forwardable set — billing.go's
+	// billingForwardable allowlist decides that, per method, and 404s everything else
+	// BEFORE the admin service token is attached. Widening this pattern grants nothing on
+	// its own; adding a line to that table is the only way to expose an endpoint.
 	app.Get("/v1/billing/*", cloud.Handle(s, billingData))
 	app.Post("/v1/billing/*", requireCSRF(s, cloud.Handle(s, billingData)))
 	// Per-tenant STORE DATA bridge — the canonical /v1/commerce/* the console calls,

clients/account/billing.go

@@ -7,6 +7,14 @@
 // read/act on its OWN ledger (balance / usage / invoices / subscriptions /
 // payment-methods / spend-alerts / …), never another's.
 //
+// TWO INDEPENDENT BOUNDS, because the token makes this a privileged forwarder:
+//  1. WHICH ENDPOINT — billingForwardable, the per-method allowlist below. It is the
+//     authorization gate: an unlisted path is 404'd before the token is ever attached, so
+//     no money-MINT route (deposit/credit/refund/…) can be reached through this bridge.
+//  2. WHOSE DATA — the subject-pinning below. It aims a permitted call at the caller's own
+//     ledger. It is an IDOR control and NOT an authority control: on a mint route it would
+//     have pinned the CREDIT to the attacker's own account. (1) is what stops that.
+//
 // WHY A SERVER HANDLER (not a same-origin passthrough). Commerce's billing surface is
 // service-token-gated and filters DIFFERENT endpoints on DIFFERENT subject params —
 // subscriptions on ?userId, payment-methods on ?customerId, usage on ?user. Pinning
@@ -22,16 +30,124 @@ package account
 
 import (
 	"bytes"
+	"crypto/subtle"
 	"encoding/json"
 	"net/http"
 	"net/url"
 	"strings"
 	"unicode"
 
+	"github.com/hanzoai/account"
+
 	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/principal"
 	"github.com/zap-proto/zip"
 )
 
+// billingForwardable — THE allowlist of billing endpoints this bridge may forward, keyed
+// by method. It is the whole authorization story of the bridge, because forwarding IS
+// authorization here: every forwarded request carries the admin COMMERCE_SERVICE_TOKEN,
+// and commerce's money gate is MayMintMoney(c) = IsServiceToken(c) || IsSuperAdmin(c)
+// (middleware/platformonly.go). The token satisfies IsServiceToken, so ANY subpath that
+// reaches commerce is executed with PLATFORM authority — not the caller's. Commerce 403s
+// an org admin who calls POST /v1/billing/deposit directly; without this table the bridge
+// handed that same person the platform's own credential and minted it for them, scoped —
+// by the subject-pinning below — to their OWN account. That is the escalation, and
+// subject-pinning is what AIMS it, not what stops it. Only a path gate stops it.
+//
+// It is an ALLOWLIST, never a denylist: a denylist must enumerate every mint route
+// (deposit/credit/refund/credit-grants/payouts/husd/allotment…) and stays correct only
+// until commerce adds the next one — a route this file has never heard of is then
+// forwarded by default. Here the default is REFUSE, so a new commerce mint route is
+// unreachable the day it lands, with no change on this side. One table, one place; a path
+// not in it cannot reach commerce, by construction.
+//
+// GET and POST are SEPARATE sets because a read bridge and a write bridge are different
+// concerns: `payouts` is a legitimate read and a money-MINT write (api/billing/handlers.go
+// `api.Get("/payouts", ListPayouts)` vs `api.Post("/payouts", mintRequired, CreatePayout)`),
+// so one method-blind set would hand the mint to every reader. The POST set is therefore
+// deliberately tiny and holds NOTHING that creates spendable balance from a client-named
+// amount: cancel/reactivate a subscription, vault a card, create a budget, and a top-up
+// that CHARGES a real card (money in, not minted). Every entry is a call the console
+// actually makes; `{}` matches exactly one opaque id segment.
+//
+// EVIDENCE — each entry is a live console call (repo hanzoai/console):
+//
+//	GET  balance             src/lib/api/billing.ts:397   sidebar wallet + billing overview
+//	GET  usage               src/lib/api/billing.ts:415   cost reports / AI metrics
+//	GET  invoices            src/lib/api/billing.ts:419   invoice history table
+//	GET  invoices/{}/pdf     src/components/products/billing/BillingInvoices.tsx:31
+//	GET  subscriptions       src/lib/api/billing.ts:423   subscriptions list
+//	GET  payment-methods     src/lib/api/billing.ts:450   saved cards (masked)
+//	GET  spend-alerts        src/lib/api/billing.ts:482   budgets / spend caps
+//	GET  payment-config      src/lib/api/billing.ts:552   public Square app/location id
+//	GET  plans               src/lib/api/plans.ts:126     published tiers
+//	GET  payouts             src/components/products/SettlementModule.tsx:61  settlement view
+//	POST subscriptions/{}/cancel      src/lib/api/billing.ts:434
+//	POST subscriptions/{}/reactivate  src/lib/api/billing.ts:444
+//	POST payment-methods              src/lib/api/billing.ts:461  vault a Square nonce (no PAN)
+//	POST spend-alerts                 src/lib/api/billing.ts:500  create a budget
+//	POST topup/token                  src/lib/api/billing.ts:565  charge a card → credit
+//
+// balance/usage/payment-methods are ALSO served natively by clients/billing (order 121),
+// which wins over this catch-all (122), so those entries are reached only on a deploy
+// where that subsystem is disabled. They are listed because they are legitimate reads of
+// the caller's own ledger, not because this bridge is their primary route.
+//
+// NOT LISTED, deliberately: `me/welcome` and `grant-starter` (console calls the first at
+// billing.ts:407 and the second server-side at src/lib/server/billing-grant.ts:35) exist
+// in NEITHER the pinned commerce (v1.48.5) route table — both 404 today whether or not
+// this bridge forwards them, and grant-starter is mint-gated and browser-unreachable by
+// design. The console's PATCH/DELETE calls (spend-alerts/{}, payment-methods/{}) are absent
+// because routesBridge mounts GET+POST only, so they never reached this handler.
+var billingForwardable = map[string][]string{
+	http.MethodGet: {
+		"balance",
+		"usage",
+		"invoices",
+		"invoices/{}/pdf",
+		"subscriptions",
+		"payment-methods",
+		"spend-alerts",
+		"spend-alerts/authorize", // the S2S cap-verdict read (metering gate); 2 segments need their own entry
+		"payment-config",
+		"plans",
+		"payouts",
+	},
+	http.MethodPost: {
+		"subscriptions/{}/cancel",
+		"subscriptions/{}/reactivate",
+		"payment-methods",
+		"spend-alerts",
+		"topup/token",
+	},
+}
+
+// isForwardableBilling reports whether method+sub is in billingForwardable. sub has
+// already passed isSafeSegment, so no segment can contain a slash, a percent-escape, or a
+// traversal — a pattern segment therefore matches exactly one real segment and `{}` cannot
+// swallow a path. Fail-closed: an unknown method or an unlisted path is false.
+func isForwardableBilling(method, sub string) bool {
+	got := strings.Split(sub, "/")
+	for _, pattern := range billingForwardable[method] {
+		want := strings.Split(pattern, "/")
+		if len(want) != len(got) {
+			continue
+		}
+		match := true
+		for i, seg := range want {
+			if seg != "{}" && seg != got[i] {
+				match = false
+				break
+			}
+		}
+		if match {
+			return true
+		}
+	}
+	return false
+}
+
 // billingSubjectKeys — every query/body param through which a commerce billing endpoint
 // identifies its subject. Kept identical to commerce's edge-auth billingSubjectKeys
 // {user,userId,customerId} AND console's billing-scope.ts BILLING_SUBJECT_KEYS. Change
@@ -48,17 +164,6 @@ func isSubjectKey(k string) bool {
 	return false
 }
 
-// billingSubject — the commerce billing subject for an org+user: ALWAYS the org
-// (`org`), lowercased. Every member of an org reads/scopes to the ONE org billing
-// account — the same subject the gateway gate reads and debits. `name` is recorded
-// for metrics, never for the billing key. This is the ONE rule; the former
-// PERSONAL_BILLING_ORGS / ORG_BILLING_ORGS allowlists are gone. Keep in lockstep
-// with ai/object.BillingSubject so the console view and the gate never disagree.
-func billingSubject(org, name string) string {
-	_ = name
-	return strings.ToLower(strings.TrimSpace(org))
-}
-
 // scopedBillingSearch — pin every billingSubjectKey to subject (OVERWRITING any client
 // value — the browser cannot widen scope) and DROP org. Every OTHER param (currency,
 // status, date range) passes through untouched. Mirrors billing-scope.ts.
@@ -139,9 +244,26 @@ func commerceCreds() (base, token string) {
 func billingData(s *cloud.Service[state], c *zip.Ctx) error {
 	// IDOR boundary: the subject is the VALIDATED caller's own org/user, never a client
 	// value. requireOwner=true — billing is always org-scoped (a zero-org user has none).
+	// Auth. A browser caller is the VALIDATED principal (customer path — subject-pinned
+	// below). An IN-PROC S2S caller carries the verified COMMERCE_SERVICE_TOKEN (the
+	// metering cap-gate's authorize + the SuperAdmin cap-oversight Forward). The gateway
+	// 401s a public Bearer that is not an IAM JWT / hk-|pk-|sk- key (the 64-hex service
+	// token fails JWT parse at the edge), so an EXTERNAL client can NEVER present it here —
+	// an unauthenticated caller still hits the 403 below. On the S2S path the caller
+	// legitimately names its own subject, so its query is forwarded as-is (no pin), scoped
+	// only by the EdgeAuth-controlled X-Org-Id.
 	cr, ok := resolveCaller(c, true)
+	s2s := false
+	owner := cr.owner
 	if !ok {
-		return zip.ErrForbidden("sign in to view billing")
+		if !s2sBillingCall(c) {
+			return zip.ErrForbidden("sign in to view billing")
+		}
+		owner = strings.TrimSpace(c.Org()) // trusted X-Org-Id (never a client value on a public call)
+		if owner == "" {
+			return zip.ErrForbidden("sign in to view billing")
+		}
+		s2s = true
 	}
 
 	method := c.Method()
@@ -165,19 +287,42 @@ func billingData(s *cloud.Service[state], c *zip.Ctx) error {
 			return zip.ErrBadRequest("invalid billing path")
 		}
 	}
-
-	// Scope EVERY request to the caller's OWN subject — query AND write body — so
-	// commerce's per-tenant isolation can never be crossed from the browser.
-	subject := billingSubject(cr.owner, cr.name)
-	inQuery, _ := url.ParseQuery(string(c.Fiber().Request().URI().QueryString()))
-	q := scopedBillingSearch(inQuery, subject)
-
-	var body []byte
-	if method == http.MethodPost {
-		body = scopedBillingBody(c.Body(), subject)
+	// THE authorization gate. Forwarding is authorization: the request below carries the
+	// admin service token, which satisfies commerce's MayMintMoney. So refuse anything the
+	// console does not actually call — BEFORE the token is attached. Fail closed (404, the
+	// same answer an unrouted path gives, so this leaks no map of the money surface).
+	if !isForwardableBilling(method, sub) {
+		return zip.Errorf(http.StatusNotFound, "not a forwardable billing endpoint")
 	}
 
-	raw, status, err := commerceDo(c.Context(), base, token, method, "/v1/billing/"+sub, q, cr.owner, body)
+	// Scope EVERY request to the caller's OWN subject — query AND write body — so
+	// commerce's per-tenant isolation can never be crossed from the browser. The
+	// subject comes from the ONE rule (ai/object.Payer), fed the account the
+	// credential NAMES (the validated `billing_account` claim) — the same claim the
+	// ai gate reads, so a top-up credits the SAME account the gate debits. Feeding
+	// Payer a different credential here than the gate gets is the modern shape of
+	// the old split: money landing in an account the gate never reads.
+	inQuery, _ := url.ParseQuery(string(c.Fiber().Request().URI().QueryString()))
+	var q url.Values
+	var body []byte
+	if s2s {
+		// Trusted S2S caller: forward its query/body VERBATIM — it legitimately names the
+		// subject (e.g. the metering gate's ?user=<org>&amount=). Scoped by X-Org-Id.
+		q = inQuery
+		if method == http.MethodPost {
+			body = c.Body()
+		}
+	} else {
+		// Browser customer: pin EVERY subject key to the caller's OWN account so commerce's
+		// per-tenant isolation can never be crossed from the client.
+		subject := account.Payer(account.Credential{Owner: cr.owner, Name: cr.username, Account: principal.BillingAccount(c)}).Subject()
+		q = scopedBillingSearch(inQuery, subject)
+		if method == http.MethodPost {
+			body = scopedBillingBody(c.Body(), subject)
+		}
+	}
+
+	raw, status, err := commerceDo(c.Context(), base, token, method, "/v1/billing/"+sub, q, owner, body)
 	if err != nil {
 		return zip.Errorf(http.StatusBadGateway, "billing upstream unreachable: %v", err)
 	}
@@ -187,3 +332,21 @@ func billingData(s *cloud.Service[state], c *zip.Ctx) error {
 	c.SetHeader("Cache-Control", "no-store, must-revalidate")
 	return c.Bytes(status, raw)
 }
+
+// s2sBillingCall reports whether the request carries the verified COMMERCE_SERVICE_TOKEN
+// as its Bearer — a trusted IN-PROC service-to-service caller (the metering cap-gate's
+// authorize, the SuperAdmin cap-oversight Forward). It is the SAME secret this bridge
+// already forwards WITH, so admitting a caller who already holds it grants no authority it
+// could not otherwise wield. Safety rests on the edge: the gateway 401s a public Bearer
+// that is not an IAM JWT / hk-|pk-|sk- API key (the 64-hex service token is a JWT
+// candidate that fails to parse), so an EXTERNAL client can never reach this handler
+// holding it — only in-proc commerceinproc dispatch does. Constant-time compare; the token
+// is never logged.
+func s2sBillingCall(c *zip.Ctx) bool {
+	_, token := commerceCreds()
+	if token == "" {
+		return false
+	}
+	bearer := strings.TrimSpace(strings.TrimPrefix(c.Header("Authorization"), "Bearer "))
+	return bearer != "" && subtle.ConstantTimeCompare([]byte(bearer), []byte(token)) == 1
+}

clients/account/billing_test.go

@@ -2,6 +2,7 @@ package account
 
 import (
 	"encoding/json"
+	"github.com/hanzoai/account"
 	"io"
 	"net/http"
 	"net/http/httptest"
@@ -16,37 +17,44 @@ import (
 
 // ── pure scoping ─────────────────────────────────────────────────────────────
 
+// TestBillingSubject proves the top-up subject is resolved through the ONE rule
+// (ai/object.Payer) — so a top-up credits the SAME account the ai gate debits and
+// the console reads. The signup org bills per-person (matching the gate), which is
+// the whole fix: money and gate land on one account.
 func TestBillingSubject(t *testing.T) {
 	cases := []struct{ org, name, want string }{
-		{"acme", "alice", "acme"},  // any member bills the ONE org account
-		{"hanzo", "Dave", "hanzo"}, // no per-user wallet; org, lowercased
-		{"hanzo", "z", "hanzo"},    // another member — same org account
-		{"hanzo", "", "hanzo"},     // no name → org
-		{"Hanzo", "z", "hanzo"},    // lowercased
-		{"", "x", ""},              // no org → empty subject
+		{"acme", "alice", "acme"},       // real org: any member bills the ONE org account
+		{"hanzo", "Dave", "hanzo/dave"}, // signup org: each person bills their OWN account
+		{"hanzo", "z", "hanzo/z"},       // another signup person — their own account
+		{"hanzo", "", "hanzo"},          // no name (org-owned principal) → org pool
+		{"Hanzo", "Z", "hanzo/z"},       // folded
+		{"", "x", ""},                   // no org → empty subject (cannot bill)
 	}
 	for _, c := range cases {
-		if got := billingSubject(c.org, c.name); got != c.want {
-			t.Fatalf("billingSubject(%q,%q): want %q, got %q", c.org, c.name, c.want, got)
+		got := account.Payer(account.Credential{Owner: c.org, Name: c.name}).Subject()
+		if got != c.want {
+			t.Fatalf("Payer(%q,%q).Subject(): want %q, got %q", c.org, c.name, c.want, got)
 		}
 	}
 }
 
 // TestBillingSubject_IgnoresLegacyEnv locks that the killed allowlist envs have NO
-// effect: the subject is ALWAYS the org, whether or not the old PERSONAL_BILLING_ORGS
-// / ORG_BILLING_ORGS knobs are set. This mirrors ai/object.BillingSubject (one rule,
-// no config) so the console view and the gateway gate can never disagree.
+// effect: nothing reads them. Set to values that WOULD have flipped every
+// resolution — the subject is unchanged. This is the console/top-up half of the
+// same proof ai carries (one rule, no config), so the view and the gate can never
+// disagree, and the deleted CR env is a genuine no-op.
 func TestBillingSubject_IgnoresLegacyEnv(t *testing.T) {
-	t.Setenv("PERSONAL_BILLING_ORGS", "hanzo,acme")
-	t.Setenv("ORG_BILLING_ORGS", "hanzo")
+	t.Setenv("PERSONAL_BILLING_ORGS", "hanzo,acme") // would have split acme per-user
+	t.Setenv("ORG_BILLING_ORGS", "hanzo")           // would have pooled the signup org
 	cases := []struct{ org, name, want string }{
-		{"hanzo", "z", "hanzo"},
-		{"acme", "alice", "acme"},
-		{"maxpower", "dave", "maxpower"},
+		{"hanzo", "z", "hanzo/z"},        // env cannot pool the signup org
+		{"acme", "alice", "acme"},        // env cannot split a real org per-user
+		{"maxpower", "dave", "maxpower"}, // untouched
 	}
 	for _, c := range cases {
-		if got := billingSubject(c.org, c.name); got != c.want {
-			t.Fatalf("legacy env must be ignored: billingSubject(%q,%q) want %q, got %q", c.org, c.name, c.want, got)
+		got := account.Payer(account.Credential{Owner: c.org, Name: c.name}).Subject()
+		if got != c.want {
+			t.Fatalf("legacy env must be ignored: Payer(%q,%q).Subject() want %q, got %q", c.org, c.name, c.want, got)
 		}
 	}
 }
@@ -224,3 +232,65 @@ func TestBilling_RejectsTraversalSegment(t *testing.T) {
 		t.Fatalf("a traversal must never reach commerce, but upstream saw %q", f.path)
 	}
 }
+
+// ── S2S service-token admission (the auth fix; 4 security invariants) ─────────
+
+// Invariant #4 — THE SECURITY GATE: a public/unauthenticated caller (no validated
+// principal AND not the service token) STILL gets 403 on the spend-alert routes, incl.
+// a WRONG bearer. The fix must NEVER open billing to the world.
+func TestBilling_S2S_PublicStill403(t *testing.T) {
+	t.Setenv("COMMERCE_SERVICE_TOKEN", "svc-tok")
+	app := mountApp(t, "http://iam.invalid", "", "")
+	for _, path := range []string{
+		"/v1/billing/spend-alerts",
+		"/v1/billing/spend-alerts/authorize?user=acme&amount=1",
+	} {
+		// forged X-Org-Id, no validated principal, no service token
+		if code, body := callH(t, app, http.MethodGet, path, map[string]string{"X-Org-Id": "victim"}, ""); code != http.StatusForbidden {
+			t.Fatalf("public caller to %s: want 403, got %d (%s)", path, code, body)
+		}
+	}
+	// a WRONG bearer is still just a public caller → 403
+	if code, _ := callH(t, app, http.MethodGet, "/v1/billing/spend-alerts/authorize?user=acme&amount=1",
+		map[string]string{"X-Org-Id": "acme", "Authorization": "Bearer not-the-token"}, ""); code != http.StatusForbidden {
+		t.Fatalf("wrong bearer: want 403")
+	}
+}
+
+// The trusted in-proc S2S caller (verified COMMERCE_SERVICE_TOKEN + X-Org-Id) is admitted
+// and its authorize query is forwarded to commerce VERBATIM (a trusted caller names its
+// own subject), scoped by X-Org-Id — this is what lets the cap gate reach AuthorizeSpendCap.
+func TestBilling_S2S_ServiceTokenForwardsVerbatim(t *testing.T) {
+	f := &fakeBilling{}
+	t.Setenv("COMMERCE_URL", f.server(t).URL)
+	t.Setenv("COMMERCE_SERVICE_TOKEN", "svc-tok")
+	app := mountApp(t, "http://iam.invalid", "", "")
+
+	code, body := callH(t, app, http.MethodGet,
+		"/v1/billing/spend-alerts/authorize?user=acme&amount=100&project=P",
+		map[string]string{"Authorization": "Bearer svc-tok", "X-Org-Id": "acme"}, "")
+	if code != http.StatusOK {
+		t.Fatalf("S2S authorize: want 200, got %d (%s)", code, body)
+	}
+	if f.path != "/v1/billing/spend-alerts/authorize" {
+		t.Fatalf("forwarded path = %q", f.path)
+	}
+	// VERBATIM: the S2S caller's ?user/?amount/?project reach commerce un-pinned.
+	if f.query.Get("user") != "acme" || f.query.Get("amount") != "100" || f.query.Get("project") != "P" {
+		t.Fatalf("S2S query must forward verbatim, got %v", f.query)
+	}
+	if f.org != "acme" || f.auth != "Bearer svc-tok" {
+		t.Fatalf("S2S must send X-Org-Id=acme + service token, got org=%q auth=%q", f.org, f.auth)
+	}
+}
+
+// S2S with the verified token but NO X-Org-Id → 403 (no org to scope the privileged
+// forward to; never fall back to a client value).
+func TestBilling_S2S_NoOrg403(t *testing.T) {
+	t.Setenv("COMMERCE_SERVICE_TOKEN", "svc-tok")
+	app := mountApp(t, "http://iam.invalid", "", "")
+	if code, _ := callH(t, app, http.MethodGet, "/v1/billing/spend-alerts/authorize?user=acme&amount=1",
+		map[string]string{"Authorization": "Bearer svc-tok"}, ""); code != http.StatusForbidden {
+		t.Fatalf("S2S without X-Org-Id: want 403")
+	}
+}

clients/account/bridge_mint_test.go

@@ -0,0 +1,257 @@
+package account
+
+import (
+	"net/http"
+	"strings"
+	"testing"
+
+	commercebilling "github.com/hanzoai/commerce/api/billing"
+	commercemid "github.com/hanzoai/commerce/middleware"
+	"github.com/zap-proto/zip"
+)
+
+// bridge_mint_test.go — the privilege-escalation boundary of the /v1/billing/*
+// bridge: an ordinary signed-in ORG user must never reach commerce's money-MINT
+// surface.
+//
+// THE ESCALATION THIS LOCKS OUT. The bridge forwards to commerce with the admin
+// COMMERCE_SERVICE_TOKEN. Commerce gates every mint on
+// MayMintMoney(c) = IsServiceToken(c) || IsSuperAdmin(c) (middleware/platformonly.go)
+// — and the bridge's service token satisfies IsServiceToken. So ANY subpath the
+// bridge forwards is executed by commerce as the PLATFORM, not as the caller.
+// billingData scopes the SUBJECT to the caller's own account, which is exactly the
+// attack rather than a defense: an org user mints to THEMSELVES. Commerce's own
+// gate comment names this: "let ANY org owner self-credit unlimited balance (POST
+// /v1/billing/deposit &c.) → unlimited free inference."
+//
+// Commerce 403s that same org admin when they call it DIRECTLY
+// (TestC1_OrgAdminDeniedOnEveryMintRoute) and mints 201 for the service token
+// (TestC1_ServiceTokenMintsDeposit). The bridge is what converts the former into
+// the latter. The gate therefore has to live HERE, at the point that hands out the
+// token: forwardable subpaths are an ALLOWLIST, and a mint path is not on it.
+//
+// alice is an ordinary org user — X-Org-Id "acme", owner != "admin", NOT a
+// SuperAdmin — i.e. precisely the principal commerce refuses at the front door.
+
+// TestBridge_OrgUserCannotReachMint is the reproduction. Each of these commerce
+// subpaths is PlatformOnly-gated (api/billing/handlers.go: `mintRequired`), meaning
+// possession of the service token IS authority to create spendable balance. None
+// may leave cloud. A request that never reaches commerce cannot mint, so the
+// assertion is twofold: the caller is refused AND upstream saw nothing.
+// mintSurface asks COMMERCE which routes it gates, rather than keeping a copy.
+//
+// The list used to live here by hand under "kept in lockstep with
+// api/billing/handlers.go" — and it had already drifted: 10 paths here against
+// 16 commerce actually gates. A comment cannot hold two lists together. Now
+// commerce DECLARES its gated surface (middleware.Mint records what it gates)
+// and we read that declaration, so a mint route added there is covered here with
+// nobody remembering to do anything.
+//
+// Registration is what populates the registry, so register first, then read.
+func mintSurface(t *testing.T) []commercemid.MintRoute {
+	t.Helper()
+	commercebilling.Route(zip.New(zip.Config{DisableStartupMessage: true}).Group("/v1"))
+
+	var out []commercemid.MintRoute
+	for _, r := range commercemid.MintRoutes() {
+		// Only what THIS bridge can address: it forwards /v1/billing/* alone.
+		if !strings.HasPrefix(r.Path, "/v1/billing/") {
+			continue
+		}
+		// A wildcard segment needs some concrete value to be requestable; which
+		// one is irrelevant, since a refused call never reaches an id.
+		parts := strings.Split(r.Path, "/")
+		for i, seg := range parts {
+			if strings.HasPrefix(seg, ":") || seg == "{}" {
+				parts[i] = "probe"
+			}
+		}
+		r.Path = strings.Join(parts, "/")
+		out = append(out, r)
+	}
+	if len(out) == 0 {
+		t.Fatal("commerce declared no /v1/billing mint routes — the registry is not being populated")
+	}
+	return out
+}
+
+func TestBridge_OrgUserCannotReachMint(t *testing.T) {
+	mintPaths := mintSurface(t)
+	t.Logf("commerce declares %d gated /v1/billing mint routes", len(mintPaths))
+
+	for _, m := range mintPaths {
+		t.Run(m.Method+" "+m.Path, func(t *testing.T) {
+			f := &fakeBilling{}
+			t.Setenv("COMMERCE_URL", f.server(t).URL)
+			t.Setenv("COMMERCE_SERVICE_TOKEN", "svc-tok")
+			app := mountApp(t, "http://iam.invalid", "", "")
+
+			code, body := callH(t, app, m.Method, m.Path, alice, `{}`)
+
+			// The mint request must NEVER reach commerce: arriving there at all means
+			// it arrived bearing the admin service token, which IS the authority to
+			// mint (MayMintMoney → AuthorizeMint → the ledger write).
+			if f.path != "" {
+				t.Fatalf("ESCALATION: an ordinary org user's %s %s reached commerce at %q "+
+					"carrying %q — the service token that satisfies MayMintMoney. "+
+					"Minted subject=%v amount=%v in org=%q.",
+					m.Method, m.Path, f.path, f.auth, f.body["user"], f.body["amount"], f.org)
+			}
+			if code != http.StatusNotFound {
+				t.Fatalf("%s %s: want 404 (not a forwardable billing endpoint), got %d (%s)",
+					m.Method, m.Path, code, body)
+			}
+		})
+	}
+}
+
+// TestBridge_ConsoleCallsStillForward is the other half of the allowlist: the calls the
+// console ACTUALLY makes must still reach commerce. An allowlist that blocks the product
+// is not a fix, so each entry here is a live console call (cited in billing.go), and this
+// test fails if a future edit narrows the table below the console's real needs.
+func TestBridge_ConsoleCallsStillForward(t *testing.T) {
+	calls := []struct{ method, path, want string }{
+		{http.MethodGet, "/v1/billing/invoices", "/v1/billing/invoices"},
+		{http.MethodGet, "/v1/billing/invoices/inv_123/pdf", "/v1/billing/invoices/inv_123/pdf"},
+		{http.MethodGet, "/v1/billing/subscriptions", "/v1/billing/subscriptions"},
+		{http.MethodGet, "/v1/billing/spend-alerts", "/v1/billing/spend-alerts"},
+		{http.MethodGet, "/v1/billing/payment-config", "/v1/billing/payment-config"},
+		{http.MethodGet, "/v1/billing/plans", "/v1/billing/plans"},
+		{http.MethodGet, "/v1/billing/payouts", "/v1/billing/payouts"},
+		{http.MethodPost, "/v1/billing/subscriptions/sub_1/cancel", "/v1/billing/subscriptions/sub_1/cancel"},
+		{http.MethodPost, "/v1/billing/subscriptions/sub_1/reactivate", "/v1/billing/subscriptions/sub_1/reactivate"},
+		{http.MethodPost, "/v1/billing/payment-methods", "/v1/billing/payment-methods"},
+		{http.MethodPost, "/v1/billing/spend-alerts", "/v1/billing/spend-alerts"},
+		{http.MethodPost, "/v1/billing/topup/token", "/v1/billing/topup/token"},
+	}
+	for _, call := range calls {
+		t.Run(call.method+" "+call.path, func(t *testing.T) {
+			f := &fakeBilling{}
+			t.Setenv("COMMERCE_URL", f.server(t).URL)
+			t.Setenv("COMMERCE_SERVICE_TOKEN", "svc-tok")
+			app := mountApp(t, "http://iam.invalid", "", "")
+
+			code, body := callH(t, app, call.method, call.path, alice, "{}")
+			if code != http.StatusOK {
+				t.Fatalf("%s %s: want 200 (the console needs this), got %d (%s)",
+					call.method, call.path, code, body)
+			}
+			if f.path != call.want {
+				t.Fatalf("%s %s must forward to %q, got %q", call.method, call.path, call.want, f.path)
+			}
+		})
+	}
+}
+
+// TestBridge_UnlistedPathsAreRefused covers the rest of the money surface — routes that
+// are NOT mint-gated but that the console never calls. The bridge is not a general
+// commerce proxy; least privilege means "only what the product needs", so these 404
+// even though commerce would have served them to a service token.
+func TestBridge_UnlistedPathsAreRefused(t *testing.T) {
+	unlisted := []struct{ method, path string }{
+		{http.MethodPost, "/v1/billing/invoices"},          // CreateInvoice (admin group)
+		{http.MethodPost, "/v1/billing/invoices/i1/pay"},   // PayInvoice
+		{http.MethodPost, "/v1/billing/invoices/i1/void"},  // VoidInvoice
+		{http.MethodPost, "/v1/billing/meters"},            // CreateMeter
+		{http.MethodPost, "/v1/billing/pricing-rules"},     // CreatePricingRule
+		{http.MethodPost, "/v1/billing/withdraw"},          // money OUT
+		{http.MethodPost, "/v1/billing/usage"},             // RecordUsage — the meter itself
+		{http.MethodGet, "/v1/billing/balance/all"},        // every subject's balance
+		{http.MethodGet, "/v1/billing/sbom"},               // OSS payout surface
+		{http.MethodGet, "/v1/billing/oss-payout/summary"}, // OSS payout rollup
+		{http.MethodPost, "/v1/billing/subscriptions"},     // CreateBillingSubscription
+	}
+	for _, u := range unlisted {
+		t.Run(u.method+" "+u.path, func(t *testing.T) {
+			f := &fakeBilling{}
+			t.Setenv("COMMERCE_URL", f.server(t).URL)
+			t.Setenv("COMMERCE_SERVICE_TOKEN", "svc-tok")
+			app := mountApp(t, "http://iam.invalid", "", "")
+
+			code, _ := callH(t, app, u.method, u.path, alice, "{}")
+			if f.path != "" {
+				t.Fatalf("%s %s is not a console call and must not reach commerce, but upstream saw %q",
+					u.method, u.path, f.path)
+			}
+			if code != http.StatusNotFound {
+				t.Fatalf("%s %s: want 404, got %d", u.method, u.path, code)
+			}
+		})
+	}
+}
+
+// TestBridge_ReadAllowlistIsNotAWriteAllowlist pins the method split. `payouts` is the
+// proof that one method-blind set would be a hole: GET /payouts is a plain read, POST
+// /payouts is `mintRequired` (api/billing/handlers.go). The same string must resolve
+// differently by method, or reading the settlement view would grant minting a payout.
+func TestBridge_ReadAllowlistIsNotAWriteAllowlist(t *testing.T) {
+	if !isForwardableBilling(http.MethodGet, "payouts") {
+		t.Fatal("GET payouts is a live console read and must be forwardable")
+	}
+	if isForwardableBilling(http.MethodPost, "payouts") {
+		t.Fatal("POST payouts is mint-gated in commerce and must NEVER be forwardable")
+	}
+	// A GET-only entry must not leak into POST, and vice-versa.
+	if isForwardableBilling(http.MethodPost, "invoices") {
+		t.Fatal("POST invoices must not inherit the GET entry")
+	}
+	if isForwardableBilling(http.MethodGet, "topup/token") {
+		t.Fatal("GET topup/token must not inherit the POST entry")
+	}
+	// An unknown method fails closed (the router mounts GET+POST only; defense in depth).
+	for _, m := range []string{http.MethodPut, http.MethodPatch, http.MethodDelete, ""} {
+		if isForwardableBilling(m, "balance") {
+			t.Fatalf("method %q must fail closed", m)
+		}
+	}
+	// `{}` matches exactly ONE segment — it can never swallow a path into a mint route.
+	if isForwardableBilling(http.MethodPost, "subscriptions/a/b/cancel") {
+		t.Fatal("{} must match exactly one segment")
+	}
+}
+
+// TestBridge_StoreBridgeCannotReachBilling is the sibling lock. /v1/commerce/* carries the
+// SAME admin token with FULL CRUD, and its own allowlist (commerceStoreHeads) is what keeps
+// it a store proxy. Prove it cannot tunnel into the money surface — a store head that
+// resolved to `billing` would reopen this hole from the other bridge.
+func TestBridge_StoreBridgeCannotReachBilling(t *testing.T) {
+	for _, p := range []string{
+		"/v1/commerce/billing/deposit",
+		"/v1/commerce/billing",
+		"/v1/commerce/checkout",
+		"/v1/commerce/_/commerce/tenants",
+	} {
+		f := &fakeBilling{}
+		t.Setenv("COMMERCE_URL", f.server(t).URL)
+		t.Setenv("COMMERCE_SERVICE_TOKEN", "svc-tok")
+		app := mountApp(t, "http://iam.invalid", "", "")
+
+		code, _ := callH(t, app, http.MethodPost, p, alice, `{"amount":100000000}`)
+		if f.path != "" {
+			t.Fatalf("store bridge %q must never reach commerce, but upstream saw %q", p, f.path)
+		}
+		if code != http.StatusNotFound {
+			t.Fatalf("store bridge %q: want 404, got %d", p, code)
+		}
+	}
+}
+
+// TestBridge_MintIsRefusedEvenWithForgedSubject proves the refusal does not depend
+// on the subject-pinning. Pinning is an IDOR control, not an authority control: it
+// makes the mint land on the CALLER's own account, which is the attack, not a
+// defense. The path gate must refuse before any of that logic runs.
+func TestBridge_MintIsRefusedEvenWithForgedSubject(t *testing.T) {
+	f := &fakeBilling{}
+	t.Setenv("COMMERCE_URL", f.server(t).URL)
+	t.Setenv("COMMERCE_SERVICE_TOKEN", "svc-tok")
+	app := mountApp(t, "http://iam.invalid", "", "")
+
+	code, _ := callH(t, app, http.MethodPost, "/v1/billing/deposit", alice,
+		`{"user":"victim","userId":"victim","amount":100000000}`)
+	if f.path != "" {
+		t.Fatalf("ESCALATION: deposit reached commerce at %q with the service token", f.path)
+	}
+	if code != http.StatusNotFound {
+		t.Fatalf("forged-subject deposit: want 404, got %d", code)
+	}
+}

clients/admin/admin.go

@@ -38,7 +38,10 @@ import (
 	"github.com/hanzoai/cloud/clients/admin/finance"
 	"github.com/hanzoai/cloud/clients/admin/health"
 	"github.com/hanzoai/cloud/clients/admin/iam"
+	"github.com/hanzoai/cloud/clients/admin/invoices"
+	"github.com/hanzoai/cloud/clients/admin/metrics"
 	"github.com/hanzoai/cloud/clients/admin/revenue"
+	"github.com/hanzoai/cloud/clients/admin/subscriptions"
 	"github.com/hanzoai/cloud/clients/commerceinproc"
 	"github.com/hanzoai/cloud/clients/principal"
 	"github.com/zap-proto/zip"
@@ -100,6 +103,7 @@ func routes(app *zip.App, s *cloud.Service[core.State]) {
 	app.Get("/v1/admin/products", core.Guard(s, products))
 	app.Get("/v1/admin/compute", core.Guard(s, compute))
 	app.Get("/v1/admin/o11y", core.Guard(s, o11y))
+	app.Get("/v1/admin/aimetrics", core.Guard(s, aimetrics))
 	app.Post("/v1/admin/sync", core.Guard(s, syncNow))
 
 	// Product analytics — org-scoped (SuperAdmin: all-orgs; org admin: their own org).
@@ -108,16 +112,28 @@ func routes(app *zip.App, s *cloud.Service[core.State]) {
 	app.Get("/v1/admin/bases", core.GuardScoped(s, bases))
 
 	// ── Platform control plane — SuperAdmin ONLY (launch/release/flags + access). ──
-	app.Get("/v1/admin/flags", core.Guard(s, flags))
+	app.Get("/v1/admin/flags", core.Guard(s, flagsBoard))
 	app.Put("/v1/admin/flags/:key", core.Guard(s, setFlag))
+	// Launch-control services board — the waitlist-mode lens on the flag engine (twin
+	// of /v1/admin/flags), reading the registry + decide the admission gate owns.
+	app.Get("/v1/admin/services", core.Guard(s, services))
+	app.Post("/v1/admin/services", core.Guard(s, upsertService))
+	app.Post("/v1/admin/services/:service/mode", core.Guard(s, setServiceMode))
 	app.Get("/v1/admin/waitlist", core.Guard(s, waitlist))
 	app.Post("/v1/admin/waitlist/boost", core.Guard(s, waitlistBoost))
 
-	// ── Carved-out domains own their routes (audit/customer/revenue/finance). ──
+	// Usage-cap + promo control plane (promos platform-only; spend-caps org-scoped).
+	limitRoutes(app, s)
+
+	// ── Carved-out domains own their routes (audit/customer/revenue/finance +
+	// the billing fleet views metrics/invoices/subscriptions). ──
 	audit.Routes(app, s)
 	customer.Routes(app, s)
 	revenue.Routes(app, s)
 	finance.Routes(app, s)
+	metrics.Routes(app, s)
+	invoices.Routes(app, s)
+	subscriptions.Routes(app, s)
 }
 
 // ── /v1/admin/me — operator identity (AdminMe) ───────────────────────────────

clients/admin/aimetrics.go

@@ -0,0 +1,392 @@
+// Copyright 2023-2026 Hanzo AI Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package admin
+
+// aimetrics — GET /v1/admin/aimetrics, the GLOBAL fleet-wide AI / training / eval
+// read that powers the operator's AI-metrics board on admin.hanzo.ai. It is the
+// AI-and-eval-focused companion to o11y (o11y.go): where o11y answers "how is the
+// FLEET behaving" (RED metrics, logs, usage), this answers "how are the MODELS and
+// EVALS doing" — LLM generations, per-model spend, and eval-run quality/progress —
+// over the SAME ONE datastore (Datastore), the SAME shared client
+// (aiobject.DatastoreQuery), no second connection.
+//
+// Signals, each from its canonical table in the one datastore:
+//   - LLM generations → langfuse.observations : generations, cost (USD), latency
+//                                                (fleet-wide; honest-empty until the
+//                                                Langfuse ingest lands rows)
+//   - Per-model usage → hanzo.cloud_usage      : requests, tokens, cost per model
+//                                                (the live usage ledger the ai gateway
+//                                                writes — populated today)
+//   - Eval runs       → hanzo.eval_traces      : traces, runs, datasets, models under
+//                                                test, per-trace latency
+//   - Eval progress   → hanzo.eval_scores      : score count, avg score, per-score-name
+//                                                distribution, recent-run averages, and
+//                                                the avg-score-over-time TREND — the
+//                                                training/eval progress signal
+//
+// The eval_traces / eval_scores tables are OWNED and written by the eval telemetry
+// store (clients/eval/telemetry.go) — the SAME warehouse, same db ("hanzo"), same
+// shared aiobject client. admin only READS them here. There is deliberately no
+// "training_progress" table: the router's per-request training events live in the ai
+// OLTP Postgres (object.RoutingEvent), NOT the OLAP warehouse, so the honest
+// warehouse-side progress signal is the eval-score trend, not a routing table.
+//
+// SUPERADMIN ONLY (the core.Guard wrap in admin.go), all-orgs, no org filter — the
+// one place a fleet operator crosses tenants for AI/eval metrics; a non-admin bearer
+// is refused 403 before a single row is read. Fail-closed.
+//
+// Honest by construction, exactly like o11y/compute: no datastore connected → the
+// real empty aggregate, never a fabricated fleet; and every signal degrades
+// INDEPENDENTLY — a table that is absent or a column that differs contributes its
+// zero-value (the enclosing `if err == nil`), never a failure, so the board always
+// renders what the datastore actually holds. admin READS only; it owns and creates
+// NO table. Money from cloud_usage is USD cents, from langfuse is USD; latency is
+// milliseconds; time bounds are POSITIONAL parameters (never interpolated), and the
+// bucket interval is a server-side constant — injection-safe.
+
+import (
+	"strconv"
+	"time"
+
+	aiobject "github.com/hanzoai/ai/object"
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/admin/core"
+	"github.com/zap-proto/zip"
+)
+
+// Fully-qualified datastore tables. admin only READS these — the ai gateway owns
+// hanzo.cloud_usage, Langfuse owns langfuse.observations, and the eval telemetry
+// store (clients/eval) owns hanzo.eval_traces / hanzo.eval_scores.
+const (
+	aimUsageTable  = "hanzo.cloud_usage"
+	aimLangfuseObs = "langfuse.observations"
+	aimEvalTraces  = "hanzo.eval_traces"
+	aimEvalScores  = "hanzo.eval_scores"
+	aimTopN        = 12
+)
+
+// aiMetrics is the whole AI-metrics board payload.
+type aiMetrics struct {
+	Range          string           `json:"range"`
+	Start          string           `json:"start"`
+	End            string           `json:"end"`
+	Langfuse       aimLangfuse      `json:"langfuse"`
+	Usage          aimUsage         `json:"usage"`
+	Evals          aimEvals         `json:"evals"`
+	TopModels      []aimModelStat   `json:"topModels"`      // cloud_usage per-model (populated today)
+	LangfuseModels []aimLfModelStat `json:"langfuseModels"` // langfuse per-model (honest-empty today)
+	ScoreNames     []aimScoreStat   `json:"scoreNames"`     // eval_scores per score-name
+	EvalRuns       []aimRunStat     `json:"evalRuns"`       // recent eval runs (progress)
+	ScoreSeries    []aimScorePoint  `json:"scoreSeries"`    // avg eval score over time (progress trend)
+}
+
+// aimLangfuse is the fleet-wide Langfuse generation rollup (honest-empty today).
+// Cost is USD (Langfuse's native unit); latency is milliseconds (end_time-start_time).
+type aimLangfuse struct {
+	Generations  int64   `json:"generations"`
+	CostUsd      float64 `json:"costUsd"`
+	LatencyMsAvg float64 `json:"latencyMsAvg"`
+	LatencyMsP95 float64 `json:"latencyMsP95"`
+}
+
+// aimUsage is the fleet LLM-usage KPI band from the live cloud_usage ledger.
+type aimUsage struct {
+	Requests         int64 `json:"requests"`
+	Tokens           int64 `json:"tokens"`
+	PromptTokens     int64 `json:"promptTokens"`
+	CompletionTokens int64 `json:"completionTokens"`
+	CostCents        int64 `json:"costCents"`
+	Models           int64 `json:"models"`
+}
+
+// aimEvals is the fleet eval KPI band: the trace half (eval_traces) and the score
+// half (eval_scores). LatencyMsAvg is the mean model-under-test call window.
+type aimEvals struct {
+	Runs         int64   `json:"runs"`
+	Traces       int64   `json:"traces"`
+	Datasets     int64   `json:"datasets"`
+	Models       int64   `json:"models"`
+	LatencyMsAvg float64 `json:"latencyMsAvg"`
+	Scores       int64   `json:"scores"`
+	ScoreNames   int64   `json:"scoreNames"`
+	AvgScore     float64 `json:"avgScore"`
+}
+
+// aimModelStat is one row of the per-model usage leaderboard (cloud_usage).
+type aimModelStat struct {
+	Model     string `json:"model"`
+	Requests  int64  `json:"requests"`
+	Tokens    int64  `json:"tokens"`
+	CostCents int64  `json:"costCents"`
+}
+
+// aimLfModelStat is one row of the per-model Langfuse leaderboard (honest-empty today).
+type aimLfModelStat struct {
+	Model       string  `json:"model"`
+	Generations int64   `json:"generations"`
+	CostUsd     float64 `json:"costUsd"`
+}
+
+// aimScoreStat is one row of the per-score-name eval leaderboard (eval_scores).
+type aimScoreStat struct {
+	Name     string  `json:"name"`
+	Count    int64   `json:"count"`
+	AvgValue float64 `json:"avgValue"`
+	MinValue float64 `json:"minValue"`
+	MaxValue float64 `json:"maxValue"`
+}
+
+// aimRunStat is one recent eval run: its dataset, how many scores it recorded, its
+// mean score, and when it last ran — the run-level eval-progress row.
+type aimRunStat struct {
+	RunName  string  `json:"runName"`
+	Dataset  string  `json:"dataset"`
+	Scores   int64   `json:"scores"`
+	AvgValue float64 `json:"avgValue"`
+	LastTs   string  `json:"lastTs"`
+}
+
+// aimScorePoint is one bucket of the avg-eval-score-over-time trend.
+type aimScorePoint struct {
+	Ts       string  `json:"ts"`
+	AvgValue float64 `json:"avgValue"`
+	Count    int64   `json:"count"`
+}
+
+// aimetrics answers GET /v1/admin/aimetrics. ?range=24h|7d|30d bounds the window
+// (default 30d). SUPERADMIN ONLY (core.Guard). Every signal degrades independently:
+// a table that is absent or errors contributes its zero-value, never a failure — the
+// board always renders what the datastore actually holds.
+func aimetrics(s *cloud.Service[core.State], c *zip.Ctx) error {
+	ctx := c.Context()
+	rangeLabel := o11yRange(c.Query("range"))
+	since := computeSince(rangeLabel)
+	payload := aiMetrics{
+		Range:          rangeLabel,
+		Start:          since.Format(time.RFC3339),
+		End:            time.Now().UTC().Format(time.RFC3339),
+		TopModels:      []aimModelStat{},
+		LangfuseModels: []aimLfModelStat{},
+		ScoreNames:     []aimScoreStat{},
+		EvalRuns:       []aimRunStat{},
+		ScoreSeries:    []aimScorePoint{},
+	}
+
+	// Honest-empty when the warehouse is not connected: the board renders its zero
+	// state, never a fabricated fleet.
+	if !aiobject.DatastoreEnabled() {
+		return core.OK(c, payload)
+	}
+
+	sinceTS := chTS(since) // DateTime literal — cloud_usage.timestamp, langfuse.start_time, eval_*.ts
+	interval := o11yBucket(rangeLabel)
+
+	// ── Langfuse generations (fleet) — honest-empty until ingest lands rows ──
+	if rows, err := aiobject.DatastoreQuery(ctx, aimLangfuseTotalsSQL(), sinceTS); err == nil {
+		r := firstRowOr(rows)
+		payload.Langfuse.Generations = chInt64(r["gens"])
+		payload.Langfuse.CostUsd = chFloat64(r["cost"])
+	}
+	// Langfuse latency (separate query so a Nullable end_time / column mismatch never
+	// zeroes the proven generations+cost number above).
+	if rows, err := aiobject.DatastoreQuery(ctx, aimLangfuseLatencySQL(), sinceTS); err == nil {
+		r := firstRowOr(rows)
+		payload.Langfuse.LatencyMsAvg = chFloat64(r["lat_avg"])
+		payload.Langfuse.LatencyMsP95 = chFloat64(r["lat_p95"])
+	}
+	// Langfuse per-model.
+	if rows, err := aiobject.DatastoreQuery(ctx, aimLangfuseModelsSQL(), sinceTS); err == nil {
+		payload.LangfuseModels = lfModelsFromRows(rows)
+	}
+
+	// ── Per-model usage (fleet) from the live cloud_usage ledger ──
+	if rows, err := aiobject.DatastoreQuery(ctx, aimUsageTotalsSQL(), sinceTS); err == nil {
+		fillAimUsage(&payload.Usage, firstRowOr(rows))
+	}
+	if rows, err := aiobject.DatastoreQuery(ctx, aimTopModelsSQL(), sinceTS); err == nil {
+		payload.TopModels = aimModelsFromRows(rows)
+	}
+
+	// ── Evals (fleet): traces + scores + progress ──
+	if rows, err := aiobject.DatastoreQuery(ctx, aimEvalTracesSQL(), sinceTS); err == nil {
+		fillAimEvalTraces(&payload.Evals, firstRowOr(rows))
+	}
+	if rows, err := aiobject.DatastoreQuery(ctx, aimEvalScoresSQL(), sinceTS); err == nil {
+		fillAimEvalScores(&payload.Evals, firstRowOr(rows))
+	}
+	if rows, err := aiobject.DatastoreQuery(ctx, aimScoreNamesSQL(), sinceTS); err == nil {
+		payload.ScoreNames = scoreNamesFromRows(rows)
+	}
+	if rows, err := aiobject.DatastoreQuery(ctx, aimEvalRunsSQL(), sinceTS); err == nil {
+		payload.EvalRuns = evalRunsFromRows(rows)
+	}
+	if rows, err := aiobject.DatastoreQuery(ctx, aimScoreSeriesSQL(interval), sinceTS); err == nil {
+		payload.ScoreSeries = scoreSeriesFromRows(rows)
+	}
+
+	return core.OK(c, payload)
+}
+
+// ── pure SQL builders (static SQL + one positional time bound; unit-tested) ──
+
+func aimLangfuseTotalsSQL() string {
+	return "SELECT count() AS gens, toFloat64(sum(total_cost)) AS cost FROM " + aimLangfuseObs +
+		" WHERE type = 'GENERATION' AND start_time >= ?"
+}
+
+func aimLangfuseLatencySQL() string {
+	lat := "(toUnixTimestamp64Milli(end_time) - toUnixTimestamp64Milli(start_time))"
+	return "SELECT round(avg(" + lat + "), 2) AS lat_avg, round(quantile(0.95)(" + lat + "), 2) AS lat_p95 " +
+		"FROM " + aimLangfuseObs + " WHERE type = 'GENERATION' AND start_time >= ? AND end_time > start_time"
+}
+
+func aimLangfuseModelsSQL() string {
+	return "SELECT provided_model_name AS model, count() AS gens, toFloat64(sum(total_cost)) AS cost " +
+		"FROM " + aimLangfuseObs + " WHERE type = 'GENERATION' AND start_time >= ? AND provided_model_name != '' " +
+		"GROUP BY model ORDER BY gens DESC LIMIT " + strconv.Itoa(aimTopN)
+}
+
+func aimUsageTotalsSQL() string {
+	return "SELECT count() AS requests, sum(total_tokens) AS tokens, " +
+		"sum(prompt_tokens) AS prompt_tokens, sum(completion_tokens) AS completion_tokens, " +
+		"sum(cost_cents) AS cost_cents, uniqExact(model) AS models " +
+		"FROM " + aimUsageTable + " WHERE timestamp >= ?"
+}
+
+func aimTopModelsSQL() string {
+	return "SELECT model, count() AS requests, sum(total_tokens) AS tokens, " +
+		"sum(cost_cents) AS cost_cents FROM " + aimUsageTable +
+		" WHERE timestamp >= ? AND model != '' GROUP BY model ORDER BY requests DESC LIMIT " + strconv.Itoa(aimTopN)
+}
+
+func aimEvalTracesSQL() string {
+	lat := "(toUnixTimestamp64Milli(end_time) - toUnixTimestamp64Milli(start_time))"
+	return "SELECT count() AS traces, uniqExact(run_name) AS runs, uniqExact(dataset) AS datasets, " +
+		"uniqExact(model) AS models, round(avgIf(" + lat + ", end_time > start_time), 2) AS lat_avg " +
+		"FROM " + aimEvalTraces + " WHERE ts >= ?"
+}
+
+func aimEvalScoresSQL() string {
+	return "SELECT count() AS scores, round(avg(value), 4) AS avg_value, uniqExact(name) AS score_names " +
+		"FROM " + aimEvalScores + " WHERE ts >= ?"
+}
+
+func aimScoreNamesSQL() string {
+	return "SELECT name, count() AS n, round(avg(value), 4) AS avg_value, " +
+		"round(min(value), 4) AS min_value, round(max(value), 4) AS max_value " +
+		"FROM " + aimEvalScores + " WHERE ts >= ? AND name != '' GROUP BY name ORDER BY n DESC LIMIT " + strconv.Itoa(aimTopN)
+}
+
+func aimEvalRunsSQL() string {
+	return "SELECT run_name, any(dataset) AS dataset, count() AS scores, round(avg(value), 4) AS avg_value, " +
+		"max(ts) AS last_ts FROM " + aimEvalScores + " WHERE ts >= ? AND run_name != '' " +
+		"GROUP BY run_name ORDER BY last_ts DESC LIMIT " + strconv.Itoa(aimTopN)
+}
+
+func aimScoreSeriesSQL(interval string) string {
+	return "SELECT toStartOfInterval(ts, INTERVAL " + interval + ") AS ts, " +
+		"round(avg(value), 4) AS avg_value, count() AS n FROM " + aimEvalScores +
+		" WHERE ts >= ? GROUP BY ts ORDER BY ts"
+}
+
+// ── pure row parsers (unit-tested) ──
+
+func fillAimUsage(u *aimUsage, r map[string]any) {
+	u.Requests = chInt64(r["requests"])
+	u.Tokens = chInt64(r["tokens"])
+	u.PromptTokens = chInt64(r["prompt_tokens"])
+	u.CompletionTokens = chInt64(r["completion_tokens"])
+	u.CostCents = chInt64(r["cost_cents"])
+	u.Models = chInt64(r["models"])
+}
+
+func fillAimEvalTraces(e *aimEvals, r map[string]any) {
+	e.Traces = chInt64(r["traces"])
+	e.Runs = chInt64(r["runs"])
+	e.Datasets = chInt64(r["datasets"])
+	e.Models = chInt64(r["models"])
+	e.LatencyMsAvg = chFloat64(r["lat_avg"])
+}
+
+func fillAimEvalScores(e *aimEvals, r map[string]any) {
+	e.Scores = chInt64(r["scores"])
+	e.AvgScore = chFloat64(r["avg_value"])
+	e.ScoreNames = chInt64(r["score_names"])
+}
+
+func aimModelsFromRows(rows []map[string]any) []aimModelStat {
+	out := make([]aimModelStat, 0, len(rows))
+	for _, r := range rows {
+		out = append(out, aimModelStat{
+			Model:     chStr(r["model"]),
+			Requests:  chInt64(r["requests"]),
+			Tokens:    chInt64(r["tokens"]),
+			CostCents: chInt64(r["cost_cents"]),
+		})
+	}
+	return out
+}
+
+func lfModelsFromRows(rows []map[string]any) []aimLfModelStat {
+	out := make([]aimLfModelStat, 0, len(rows))
+	for _, r := range rows {
+		out = append(out, aimLfModelStat{
+			Model:       chStr(r["model"]),
+			Generations: chInt64(r["gens"]),
+			CostUsd:     chFloat64(r["cost"]),
+		})
+	}
+	return out
+}
+
+func scoreNamesFromRows(rows []map[string]any) []aimScoreStat {
+	out := make([]aimScoreStat, 0, len(rows))
+	for _, r := range rows {
+		out = append(out, aimScoreStat{
+			Name:     chStr(r["name"]),
+			Count:    chInt64(r["n"]),
+			AvgValue: chFloat64(r["avg_value"]),
+			MinValue: chFloat64(r["min_value"]),
+			MaxValue: chFloat64(r["max_value"]),
+		})
+	}
+	return out
+}
+
+func evalRunsFromRows(rows []map[string]any) []aimRunStat {
+	out := make([]aimRunStat, 0, len(rows))
+	for _, r := range rows {
+		out = append(out, aimRunStat{
+			RunName:  chStr(r["run_name"]),
+			Dataset:  chStr(r["dataset"]),
+			Scores:   chInt64(r["scores"]),
+			AvgValue: chFloat64(r["avg_value"]),
+			LastTs:   chTime(r["last_ts"]),
+		})
+	}
+	return out
+}
+
+func scoreSeriesFromRows(rows []map[string]any) []aimScorePoint {
+	out := make([]aimScorePoint, 0, len(rows))
+	for _, r := range rows {
+		out = append(out, aimScorePoint{
+			Ts:       chTime(r["ts"]),
+			AvgValue: chFloat64(r["avg_value"]),
+			Count:    chInt64(r["n"]),
+		})
+	}
+	return out
+}

clients/admin/aimetrics_test.go

@@ -0,0 +1,171 @@
+// Copyright 2023-2026 Hanzo AI Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package admin
+
+import (
+	"strings"
+	"testing"
+)
+
+// TestAimSQL_ReadsCanonicalTables proves every AI-metrics query reads the ONE
+// datastore's canonical table, binds the time bound as a POSITIONAL param (one
+// `?`), and never interpolates user input. The bucket interval is the only rendered
+// value in the series query and it is a server-side constant.
+func TestAimSQL_ReadsCanonicalTables(t *testing.T) {
+	cases := []struct {
+		name, sql, table string
+		wantQMarks       int
+	}{
+		{"langfuseTotals", aimLangfuseTotalsSQL(), "langfuse.observations", 1},
+		{"langfuseLatency", aimLangfuseLatencySQL(), "langfuse.observations", 1},
+		{"langfuseModels", aimLangfuseModelsSQL(), "langfuse.observations", 1},
+		{"usageTotals", aimUsageTotalsSQL(), "hanzo.cloud_usage", 1},
+		{"topModels", aimTopModelsSQL(), "hanzo.cloud_usage", 1},
+		{"evalTraces", aimEvalTracesSQL(), "hanzo.eval_traces", 1},
+		{"evalScores", aimEvalScoresSQL(), "hanzo.eval_scores", 1},
+		{"scoreNames", aimScoreNamesSQL(), "hanzo.eval_scores", 1},
+		{"evalRuns", aimEvalRunsSQL(), "hanzo.eval_scores", 1},
+		{"scoreSeries", aimScoreSeriesSQL("1 DAY"), "hanzo.eval_scores", 1},
+	}
+	for _, c := range cases {
+		if !strings.Contains(c.sql, "FROM "+c.table) {
+			t.Errorf("%s must read %s; got %q", c.name, c.table, c.sql)
+		}
+		if n := strings.Count(c.sql, "?"); n != c.wantQMarks {
+			t.Errorf("%s: %d bind params, want %d (time bound only) — no interpolation; got %q", c.name, n, c.wantQMarks, c.sql)
+		}
+	}
+}
+
+// TestAimLangfuseScopedToGeneration proves the Langfuse lens is scoped to
+// generations only (not spans/events), matching the o11y LLM lens.
+func TestAimLangfuseScopedToGeneration(t *testing.T) {
+	for _, sql := range []string{aimLangfuseTotalsSQL(), aimLangfuseLatencySQL(), aimLangfuseModelsSQL()} {
+		if !strings.Contains(sql, "type = 'GENERATION'") {
+			t.Errorf("langfuse lens must scope to GENERATION observations; got %q", sql)
+		}
+	}
+}
+
+// TestAimTop_LimitAndOrder proves the leaderboards bound + order the result.
+func TestAimTop_LimitAndOrder(t *testing.T) {
+	if !strings.Contains(aimTopModelsSQL(), "ORDER BY requests DESC LIMIT 12") {
+		t.Errorf("topModels must order by requests desc, limit %d", aimTopN)
+	}
+	if !strings.Contains(aimScoreNamesSQL(), "GROUP BY name ORDER BY n DESC LIMIT 12") {
+		t.Errorf("scoreNames must group+order+limit %d", aimTopN)
+	}
+	if !strings.Contains(aimEvalRunsSQL(), "ORDER BY last_ts DESC LIMIT 12") {
+		t.Errorf("evalRuns must order by last_ts desc, limit %d", aimTopN)
+	}
+}
+
+// TestAimScoreSeries_IntervalBound proves the (constant) bucket interval is
+// rendered into the score-trend series query and grouped/ordered by the bucket.
+func TestAimScoreSeries_IntervalBound(t *testing.T) {
+	for _, iv := range []string{"1 HOUR", "6 HOUR", "1 DAY"} {
+		s := aimScoreSeriesSQL(iv)
+		if !strings.Contains(s, "INTERVAL "+iv) || !strings.Contains(s, "GROUP BY ts ORDER BY ts") {
+			t.Errorf("score series must bucket by INTERVAL %s; got %q", iv, s)
+		}
+	}
+}
+
+// TestAimEvalLatencyGuarded proves the latency expressions guard end_time>start_time
+// so a zero/default end_time never contributes a garbage (negative) latency.
+func TestAimEvalLatencyGuarded(t *testing.T) {
+	if !strings.Contains(aimEvalTracesSQL(), "end_time > start_time") {
+		t.Errorf("eval traces latency must guard end_time>start_time; got %q", aimEvalTracesSQL())
+	}
+	if !strings.Contains(aimLangfuseLatencySQL(), "end_time > start_time") {
+		t.Errorf("langfuse latency must guard end_time>start_time; got %q", aimLangfuseLatencySQL())
+	}
+}
+
+// TestFillAimUsage reads a cloud_usage row into the KPI band across the numeric
+// variants the driver returns (uint64/int64/float64), honest zeros on an empty row.
+func TestFillAimUsage(t *testing.T) {
+	var empty aimUsage
+	fillAimUsage(&empty, map[string]any{})
+	if empty.Requests != 0 || empty.Tokens != 0 || empty.Models != 0 {
+		t.Fatalf("empty row must yield honest zeros; got %+v", empty)
+	}
+	var got aimUsage
+	fillAimUsage(&got, map[string]any{
+		"requests": uint64(274), "tokens": uint64(102597), "prompt_tokens": uint64(60000),
+		"completion_tokens": uint64(42597), "cost_cents": uint64(216), "models": uint64(42),
+	})
+	if got.Requests != 274 || got.Tokens != 102597 || got.CostCents != 216 || got.Models != 42 {
+		t.Fatalf("usage totals mis-parsed: %+v", got)
+	}
+}
+
+// TestFillAimEvals maps both eval halves (traces + scores) into the KPI band,
+// including the float latency/score columns (round()/avg() land as float64; a
+// Decimal-as-string is parsed).
+func TestFillAimEvals(t *testing.T) {
+	var e aimEvals
+	fillAimEvalTraces(&e, map[string]any{
+		"traces": uint64(1280), "runs": uint64(16), "datasets": uint64(4),
+		"models": uint64(6), "lat_avg": float64(842.5),
+	})
+	fillAimEvalScores(&e, map[string]any{
+		"scores": uint64(1280), "avg_value": "0.8125", "score_names": uint64(3),
+	})
+	if e.Traces != 1280 || e.Runs != 16 || e.Datasets != 4 || e.Models != 6 || e.LatencyMsAvg != 842.5 {
+		t.Fatalf("eval traces mis-parsed: %+v", e)
+	}
+	if e.Scores != 1280 || e.ScoreNames != 3 || e.AvgScore != 0.8125 { // string→float64 path
+		t.Fatalf("eval scores mis-parsed: %+v", e)
+	}
+}
+
+// TestAimParsers map datastore rows into the view-models and preserve order (the
+// SQL already ORDER BYs; a parser must not reorder or drop rows), with empty input
+// yielding an empty (non-nil) slice rather than a panic.
+func TestAimParsers(t *testing.T) {
+	models := aimModelsFromRows([]map[string]any{
+		{"model": "glm-5.2", "requests": uint64(154), "tokens": uint64(38966), "cost_cents": uint64(114)},
+		{"model": "deepseek-v4-flash", "requests": uint64(118), "tokens": uint64(61550), "cost_cents": uint64(101)},
+	})
+	if len(models) != 2 || models[0].Model != "glm-5.2" || models[1].Model != "deepseek-v4-flash" || models[0].Requests != 154 {
+		t.Fatalf("top models mis-parsed/reordered: %+v", models)
+	}
+	lf := lfModelsFromRows([]map[string]any{
+		{"model": "gpt-4o", "gens": uint64(42), "cost": float64(1.25)},
+	})
+	if len(lf) != 1 || lf[0].Model != "gpt-4o" || lf[0].Generations != 42 || lf[0].CostUsd != 1.25 {
+		t.Fatalf("langfuse models mis-parsed: %+v", lf)
+	}
+	names := scoreNamesFromRows([]map[string]any{
+		{"name": "accuracy", "n": uint64(320), "avg_value": float64(0.82), "min_value": float64(0), "max_value": float64(1)},
+	})
+	if len(names) != 1 || names[0].Name != "accuracy" || names[0].Count != 320 || names[0].AvgValue != 0.82 || names[0].MaxValue != 1 {
+		t.Fatalf("score names mis-parsed: %+v", names)
+	}
+	runs := evalRunsFromRows([]map[string]any{
+		{"run_name": "nightly-2026-07", "dataset": "gsm8k", "scores": uint64(200), "avg_value": float64(0.9), "last_ts": nil},
+	})
+	if len(runs) != 1 || runs[0].RunName != "nightly-2026-07" || runs[0].Dataset != "gsm8k" || runs[0].Scores != 200 || runs[0].AvgValue != 0.9 {
+		t.Fatalf("eval runs mis-parsed: %+v", runs)
+	}
+	// Empty input → empty (non-nil) slices, never a panic.
+	if got := scoreSeriesFromRows(nil); got == nil || len(got) != 0 {
+		t.Errorf("nil rows must yield empty slice, got %v", got)
+	}
+	if got := aimModelsFromRows(nil); got == nil || len(got) != 0 {
+		t.Errorf("nil rows must yield empty slice, got %v", got)
+	}
+}

clients/admin/commerce/commerce.go

@@ -295,6 +295,237 @@ func (c *Client) Deposit(ctx context.Context, subject string, amount money.Cents
 	return out, nil
 }
 
+// ── SaaS-metrics god-view (fleet-wide, org-independent) ──────────────────────
+
+// SaaSMetrics mirrors commerce's GET /v1/metrics/saas snapshot — the whole-business
+// SaaS-operations aggregate (MRR/ARR, new/churn, plan mix, top customers, recent
+// movements) computed IN commerce across every org namespace. It is org-INDEPENDENT
+// (like Costs) so the reader sends NO subject. Only the fields the admin god-view
+// renders are modeled; commerce fields we don't consume (upgrades/downgrades,
+// untagged-request counts) are simply ignored by the decoder.
+type SaaSMetrics struct {
+	AsOf      string         `json:"asOf"`
+	Currency  string         `json:"currency"`
+	Window    string         `json:"window"`
+	Revenue   SaaSRevenue    `json:"revenue"`
+	Subs      SaaSSubs       `json:"subscriptions"`
+	Usage     SaaSUsage      `json:"usage"`
+	Customers []SaaSCustomer `json:"customers"`
+	Orgs      int            `json:"orgs"`
+	Gaps      []string       `json:"gaps"`
+}
+
+// SaaSRevenue is the recurring-revenue headline (run-rate MRR/ARR + windowed movement).
+type SaaSRevenue struct {
+	MRRCents            money.Cents    `json:"mrrCents"`
+	ARRCents            money.Cents    `json:"arrCents"`
+	ActiveSubscriptions int            `json:"activeSubscriptions"`
+	PayingCustomers     int            `json:"payingCustomers"`
+	Trials              int            `json:"trials"`
+	NewMRRCents         money.Cents    `json:"newMrrCents"`
+	ChurnedMRRCents     money.Cents    `json:"churnedMrrCents"`
+	NetNewMRRCents      money.Cents    `json:"netNewMrrCents"`
+	ByCategory          []SaaSCategory `json:"byCategory"`
+}
+
+// SaaSCategory is one plan-category bucket of run-rate MRR (the plan mix).
+type SaaSCategory struct {
+	Category      string      `json:"category"`
+	MRRCents      money.Cents `json:"mrrCents"`
+	Subscriptions int         `json:"subscriptions"`
+}
+
+// SaaSSubs is the subscription-operations panel (per-plan mix, trials, new/canceled,
+// recent movements).
+type SaaSSubs struct {
+	ByPlan       []SaaSPlan  `json:"byPlan"`
+	TrialsActive int         `json:"trialsActive"`
+	New          int         `json:"new"`
+	Canceled     int         `json:"canceled"`
+	Recent       []SaaSEvent `json:"recent"`
+}
+
+// SaaSPlan is one plan's active/trialing counts, seats, and MRR contribution.
+type SaaSPlan struct {
+	Plan     string      `json:"plan"`
+	Name     string      `json:"name"`
+	Category string      `json:"category"`
+	Active   int         `json:"active"`
+	Trialing int         `json:"trialing"`
+	Seats    int         `json:"seats"`
+	MRRCents money.Cents `json:"mrrCents"`
+}
+
+// SaaSEvent is one recent subscription movement ("created" or "canceled").
+type SaaSEvent struct {
+	At            string      `json:"at"`
+	Org           string      `json:"org"`
+	Type          string      `json:"type"`
+	Plan          string      `json:"plan"`
+	Category      string      `json:"category"`
+	MRRDeltaCents money.Cents `json:"mrrDeltaCents"`
+}
+
+// SaaSUsage is the metered / pay-as-you-go revenue headline for the window.
+type SaaSUsage struct {
+	Instrumented     bool        `json:"instrumented"`
+	WindowUsageCents money.Cents `json:"windowUsageCents"`
+	Requests         int64       `json:"requests"`
+}
+
+// SaaSCustomer is one top customer by MRR + windowed usage.
+type SaaSCustomer struct {
+	Org        string      `json:"org"`
+	Plan       string      `json:"plan"`
+	Category   string      `json:"category"`
+	Status     string      `json:"status"`
+	MRRCents   money.Cents `json:"mrrCents"`
+	UsageCents money.Cents `json:"usageCents"`
+	Seats      int         `json:"seats"`
+	Since      string      `json:"since,omitempty"`
+}
+
+// Metrics reads the fleet SaaS-operations god-view (GET /v1/metrics/saas). Like Costs it
+// is org-INDEPENDENT — the engine walks every org namespace itself — so it authenticates
+// with the admin S2S service token and sends NO subject. Empty (not an error) when
+// commerce is unwired, so a partial deploy degrades to an honest empty snapshot.
+func (c *Client) Metrics(ctx context.Context, window string, limit int) (SaaSMetrics, error) {
+	var out SaaSMetrics
+	if !c.Ready() {
+		return out, nil
+	}
+	q := url.Values{}
+	if window != "" {
+		q.Set("window", window)
+	}
+	if limit > 0 {
+		q.Set("limit", fmt.Sprintf("%d", limit))
+	}
+	body, err := c.get(ctx, "/v1/metrics/saas", q, "")
+	if err != nil {
+		return out, err
+	}
+	if err := json.Unmarshal(body, &out); err != nil {
+		return out, fmt.Errorf("commerce metrics decode: %w", err)
+	}
+	return out, nil
+}
+
+// ── billing invoices + subscriptions (per-subject fleet rows) ────────────────
+
+// Invoice is one issued invoice as the fleet god-view renders it: the id (for a future
+// /v1/billing/invoices/:id detail fetch), the human number, status, amount due,
+// currency, and the issue/due dates. Sourced from GET /v1/billing/invoices
+// (invoiceResponse); all timestamps are RFC3339 strings.
+type Invoice struct {
+	ID        string      `json:"id"`
+	Number    string      `json:"numberStr"`
+	Status    string      `json:"status"`
+	AmountDue money.Cents `json:"amountDue"`
+	Currency  string      `json:"currency"`
+	Issued    string      `json:"createdAt"`
+	Due       string      `json:"dueDate"`
+}
+
+// Invoices lists a subject's invoices (GET /v1/billing/invoices), optionally filtered by
+// status. The subject selects the org's billing namespace via X-Org-Id (trusted only
+// after the service-token bearer verifies). Empty (not an error) when commerce is unwired.
+func (c *Client) Invoices(ctx context.Context, subject, status string) ([]Invoice, error) {
+	if !c.Ready() {
+		return nil, nil
+	}
+	q := url.Values{}
+	if status != "" {
+		q.Set("status", status)
+	}
+	body, err := c.get(ctx, "/v1/billing/invoices", q, subject)
+	if err != nil {
+		return nil, err
+	}
+	var wrap struct {
+		Invoices []Invoice `json:"invoices"`
+	}
+	if err := json.Unmarshal(body, &wrap); err != nil {
+		return nil, fmt.Errorf("commerce invoices decode: %w", err)
+	}
+	return wrap.Invoices, nil
+}
+
+// Subscription is one subscription row the fleet god-view renders: the id, the buyer
+// (userId), plan tier, status, monthly-normalized MRR, and the current-period
+// start/end (started/renews). MRR reuses monthlyNormalized so a yearly plan is
+// comparable to a monthly one in the fleet total.
+type Subscription struct {
+	ID      string      `json:"id"`
+	User    string      `json:"user"`
+	Plan    string      `json:"plan"`
+	Status  string      `json:"status"`
+	MRR     money.Cents `json:"mrrCents"`
+	Started string      `json:"started"`
+	Renews  string      `json:"renews"`
+}
+
+// subscriptionRowWire is the /v1/billing/subscriptions row shape the fleet view folds —
+// richer than subscriptionsWire (which Plan() uses for the MRR sum alone).
+type subscriptionRowWire struct {
+	ID          string `json:"id"`
+	UserID      string `json:"userId"`
+	PlanID      string `json:"planId"`
+	Status      string `json:"status"`
+	Created     string `json:"createdAt"`
+	PeriodStart string `json:"currentPeriodStart"`
+	PeriodEnd   string `json:"currentPeriodEnd"`
+	Plan        struct {
+		Name     string      `json:"name"`
+		Price    money.Cents `json:"price"`
+		Interval string      `json:"interval"`
+	} `json:"plan"`
+}
+
+// Subscriptions lists a subject's subscriptions (GET /v1/billing/subscriptions),
+// optionally filtered by status, as fleet rows with a monthly-normalized MRR. Empty (not
+// an error) when commerce is unwired.
+func (c *Client) Subscriptions(ctx context.Context, subject, status string) ([]Subscription, error) {
+	if !c.Ready() {
+		return nil, nil
+	}
+	q := url.Values{}
+	if status != "" {
+		q.Set("status", status)
+	}
+	body, err := c.get(ctx, "/v1/billing/subscriptions", q, subject)
+	if err != nil {
+		return nil, err
+	}
+	var wrap struct {
+		Subscriptions []subscriptionRowWire `json:"subscriptions"`
+	}
+	if err := json.Unmarshal(body, &wrap); err != nil {
+		return nil, fmt.Errorf("commerce subscriptions decode: %w", err)
+	}
+	out := make([]Subscription, 0, len(wrap.Subscriptions))
+	for _, s := range wrap.Subscriptions {
+		name := strings.TrimSpace(s.Plan.Name)
+		if name == "" {
+			name = strings.TrimSpace(s.PlanID)
+		}
+		started := strings.TrimSpace(s.Created)
+		if started == "" {
+			started = s.PeriodStart
+		}
+		out = append(out, Subscription{
+			ID:      s.ID,
+			User:    s.UserID,
+			Plan:    name,
+			Status:  s.Status,
+			MRR:     monthlyNormalized(s.Plan.Price, s.Plan.Interval),
+			Started: started,
+			Renews:  s.PeriodEnd,
+		})
+	}
+	return out, nil
+}
+
 // post performs one admin-authenticated commerce POST (JSON body) and returns the
 // raw response. The admin S2S service token is the bearer and X-Org-Id=<subject>
 // the per-org namespace selector commerce's EdgeAuth trusts only after verifying
@@ -332,6 +563,48 @@ func (c *Client) post(ctx context.Context, path, subject string, body []byte, id
 	return respBody, nil
 }
 
+// Forward proxies an admin-authenticated request to commerce VERBATIM and returns
+// the raw body + status. It is the ONE seam a SuperAdmin surface drives commerce's
+// own endpoints through — the platform plan-promo config (/v1/platform/promo) and a
+// per-org spend-alert override (/v1/billing/spend-alerts) — without a typed method
+// per shape. subject is the X-Org-Id namespace selector (the target org for a cap
+// override, or the admin org for platform config); body is nil for GET/DELETE. The
+// status is returned so the caller surfaces commerce's OWN verdict (400 validation,
+// 403, 404) instead of flattening every non-2xx into one code.
+func (c *Client) Forward(ctx context.Context, method, path, subject string, body []byte) ([]byte, int, error) {
+	if !c.Ready() {
+		return nil, 0, errUnconfigured
+	}
+	var rdr io.Reader
+	if body != nil {
+		rdr = bytes.NewReader(body)
+	}
+	req, err := http.NewRequestWithContext(ctx, method, c.base+path, rdr)
+	if err != nil {
+		return nil, 0, err
+	}
+	req.Header.Set("Accept", "application/json")
+	if body != nil {
+		req.Header.Set("Content-Type", "application/json")
+	}
+	if c.token != "" {
+		req.Header.Set("Authorization", "Bearer "+c.token)
+	}
+	if subject != "" {
+		req.Header.Set("X-Org-Id", subject)
+	}
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return nil, 0, fmt.Errorf("commerce unreachable: %w", err)
+	}
+	defer resp.Body.Close()
+	raw, err := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
+	if err != nil {
+		return nil, resp.StatusCode, err
+	}
+	return raw, resp.StatusCode, nil
+}
+
 // get performs one admin-authenticated commerce GET and returns the raw body.
 func (c *Client) get(ctx context.Context, path string, q url.Values, subject string) ([]byte, error) {
 	u := c.base + path

clients/admin/core/grant.go

@@ -209,7 +209,7 @@ func grantDeposit(s *cloud.Service[State], c *zip.Ctx, org, currency, notes, tag
 		after = balCents
 		if fin := finance.Current(); fin != nil {
 			if bal, berr := fin.Balance(ctx, org, org, currency, false); berr == nil {
-				afterExact = bal.IntString() // afterExact = the EXACT balance (sub-cent visible)
+				afterExact = bal.AttoString() // afterExact = the EXACT balance (sub-cent visible)
 			}
 		}
 		return before, id, after, afterExact, nil

clients/admin/flags.go

@@ -3,11 +3,11 @@ package admin
 // The PLATFORM CONTROL PLANE board (/v1/admin/flags) — every runtime LAUNCH / RELEASE
 // switch (waitlist, public signup, subsystem activation, gateway limits, network ids)
 // with its LIVE value, evaluated through the embedded native flag engine
-// (clients/featureflags → native/flags, SQLite-per-project definitions + Rust FFI
+// (clients/flags → native/flags, SQLite-per-project definitions + Rust FFI
 // evaluation). SuperAdmin only (mounted behind core.Guard, like every /v1/admin/*).
 //
 // ONE flag engine, TWO verbs. GET reads the board; PUT writes a switch's definition
-// through featureflags.SetPlatformSwitch — the ONE write path, audited in the store's
+// through flags.SetPlatformSwitch — the ONE write path, audited in the store's
 // activity log. A flip is hot: this pod applies immediately, peers converge within one
 // evaluation TTL (default 15s), no redeploy. Org/project product flags are managed on
 // /v1/flags (org-scoped); this surface is the platform's own switchboard.
@@ -18,13 +18,13 @@ import (
 
 	"github.com/hanzoai/cloud"
 	"github.com/hanzoai/cloud/clients/admin/core"
-	"github.com/hanzoai/cloud/clients/featureflags"
+	"github.com/hanzoai/cloud/clients/flags"
 	"github.com/zap-proto/zip"
 )
 
-// flags answers GET /v1/admin/flags — the platform control-plane read board.
-func flags(s *cloud.Service[core.State], c *zip.Ctx) error {
-	return core.OK(c, featureflags.Board())
+// flagsBoard answers GET /v1/admin/flags — the platform control-plane read board.
+func flagsBoard(s *cloud.Service[core.State], c *zip.Ctx) error {
+	return core.OK(c, flags.Board())
 }
 
 // setFlag answers PUT /v1/admin/flags/:key — store/overwrite one platform switch's
@@ -42,8 +42,8 @@ func setFlag(s *cloud.Service[core.State], c *zip.Ctx) error {
 	if len(body) == 0 || !json.Valid(body) {
 		return zip.ErrBadRequest("body must be the flag definition JSON")
 	}
-	if err := featureflags.SetPlatformSwitch(key, json.RawMessage(body), c.UserEmail()); err != nil {
+	if err := flags.SetPlatformSwitch(key, json.RawMessage(body), c.UserEmail()); err != nil {
 		return zip.ErrBadRequest(err.Error())
 	}
-	return core.OK(c, featureflags.Board())
+	return core.OK(c, flags.Board())
 }

clients/admin/invoices/invoices.go

@@ -0,0 +1,137 @@
+// Package invoices is the fleet INVOICE view (/v1/admin/invoices) — every issued
+// invoice across every tenant: number, org, amount, status, issue + due date, plus the
+// id a future detail view fetches /v1/billing/invoices/:id with. SuperAdmin only
+// (core.Guard).
+//
+// Commerce billing is per-tenant (an invoice lives in its org's own datastore
+// namespace), so — like revenue — this fans out the org directory concurrently and
+// reads each org's invoices via the admin S2S seam, tagging every row with its owning
+// org. Best-effort per org: an org whose invoice read fails contributes NO rows rather
+// than failing the fleet view (the SAME honest-degradation contract the customer list
+// uses; an unreachable commerce yields an empty list, never fabricated rows). Optional
+// ?org= scopes to one tenant, ?status= filters, ?limit= caps the merged list.
+package invoices
+
+import (
+	"context"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/admin/core"
+	"github.com/hanzoai/cloud/clients/admin/iam"
+	"github.com/zap-proto/zip"
+)
+
+// defaultLimit caps the merged fleet invoice list when the caller sends none.
+const defaultLimit = 500
+
+// InvoiceRow is one row of GET /v1/admin/invoices — an issued invoice at a glance,
+// tagged with its owning org. Money is USD cents; timestamps are RFC3339 strings.
+type InvoiceRow struct {
+	ID          string `json:"id"`
+	Number      string `json:"number"`
+	Org         string `json:"org"`
+	Display     string `json:"display"`
+	Status      string `json:"status"`
+	AmountCents int64  `json:"amountCents"`
+	Currency    string `json:"currency"`
+	Issued      string `json:"issued"`
+	Due         string `json:"due"`
+}
+
+// Invoices answers GET /v1/admin/invoices.
+//
+//	GET /v1/admin/invoices?org=&status=&limit=
+func Invoices(s *cloud.Service[core.State], c *zip.Ctx) error {
+	ctx := c.Context()
+	cr := core.CallerCreds(c)
+	status := strings.TrimSpace(c.Query("status"))
+	wantOrg := strings.TrimSpace(c.Query("org"))
+	limit := parseLimit(c.Query("limit"))
+
+	orgs, err := core.ListOrgs(s, ctx, cr)
+	if err != nil {
+		return core.Fail(c, err.Error())
+	}
+	if wantOrg != "" {
+		orgs = filterOrg(orgs, wantOrg)
+	}
+
+	// Per-org invoices, fanned out concurrently (best-effort per org).
+	perOrg := make([][]InvoiceRow, len(orgs))
+	sem := make(chan struct{}, core.MaxCustomerConcurrency)
+	var wg sync.WaitGroup
+	for i, o := range orgs {
+		wg.Add(1)
+		sem <- struct{}{}
+		go func(i int, o iam.Org) {
+			defer wg.Done()
+			defer func() { <-sem }()
+			perOrg[i] = invoicesOf(s, ctx, o, status)
+		}(i, o)
+	}
+	wg.Wait()
+
+	rows := make([]InvoiceRow, 0)
+	for _, r := range perOrg {
+		rows = append(rows, r...)
+	}
+	// Newest issued first; cap to the merged limit (total reports the full pre-cap count).
+	sort.Slice(rows, func(i, j int) bool { return rows[i].Issued > rows[j].Issued })
+	total := len(rows)
+	if len(rows) > limit {
+		rows = rows[:limit]
+	}
+	return core.OKList(c, rows, total)
+}
+
+// invoicesOf reads one org's invoices into fleet rows, tagged with the org. Best-effort:
+// a failed read yields no rows so the fleet view degrades honestly, never fabricating.
+func invoicesOf(s *cloud.Service[core.State], ctx context.Context, o iam.Org, status string) []InvoiceRow {
+	entries, err := s.State.Commerce.Invoices(ctx, o.Name, status)
+	if err != nil {
+		return nil
+	}
+	display := core.Display(o.DisplayName, o.Name)
+	rows := make([]InvoiceRow, 0, len(entries))
+	for _, inv := range entries {
+		rows = append(rows, InvoiceRow{
+			ID:          inv.ID,
+			Number:      inv.Number,
+			Org:         o.Name,
+			Display:     display,
+			Status:      inv.Status,
+			AmountCents: int64(inv.AmountDue),
+			Currency:    inv.Currency,
+			Issued:      inv.Issued,
+			Due:         inv.Due,
+		})
+	}
+	return rows
+}
+
+// filterOrg narrows the directory to the one requested org (empty when it does not
+// exist — an honest empty list, never a fabricated tenant).
+func filterOrg(orgs []iam.Org, want string) []iam.Org {
+	for _, o := range orgs {
+		if o.Name == want {
+			return []iam.Org{o}
+		}
+	}
+	return nil
+}
+
+// parseLimit clamps the merged-list cap to [1,5000], defaulting to defaultLimit.
+func parseLimit(s string) int {
+	n, err := strconv.Atoi(strings.TrimSpace(s))
+	if err != nil || n <= 0 {
+		return defaultLimit
+	}
+	if n > 5000 {
+		return 5000
+	}
+	return n
+}

clients/admin/invoices/routes.go

@@ -0,0 +1,12 @@
+package invoices
+
+import (
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/admin/core"
+	"github.com/zap-proto/zip"
+)
+
+// Routes registers the fleet invoice view (SuperAdmin only, cross-tenant).
+func Routes(app *zip.App, s *cloud.Service[core.State]) {
+	app.Get("/v1/admin/invoices", core.Guard(s, Invoices))
+}

clients/admin/limits.go

@@ -0,0 +1,143 @@
+package admin
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/url"
+	"strings"
+
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/admin/core"
+	"github.com/zap-proto/zip"
+)
+
+// The SuperAdmin usage-cap + promo control plane, twinning /v1/admin/flags. It owns
+// no store: it FORWARDS to commerce (the billing source of truth) over the ONE
+// service-token seam —
+//
+//	promos      → commerce /v1/platform/promo   (the admin-configured plan promo)
+//	spend-caps  → commerce /v1/billing/spend-alerts (a per-org usage cap override)
+//
+// so admin.hanzo.ai configures the 50%-off promo and oversees/overrides any org's
+// caps without a parallel model. Promo routes are platform-only (core.Guard); cap
+// routes are org-scoped (core.GuardScoped) so a SuperAdmin targets any org via ?org=
+// while a lesser admin is hard-pinned to their own.
+
+// limitRoutes registers the promo + cap control plane. Called from routes().
+func limitRoutes(app *zip.App, s *cloud.Service[core.State]) {
+	// Platform plan promo — SuperAdmin only.
+	app.Get("/v1/admin/promos", core.Guard(s, getPromo))
+	app.Put("/v1/admin/promos", core.Guard(s, putPromo))
+
+	// Per-org usage-cap oversight/override — SuperAdmin (any org via ?org=) or an org
+	// admin (own org only). Reuses the customer's OWN self-service spend-alert CRUD,
+	// so a platform override and a customer edit are the same rows.
+	app.Get("/v1/admin/spend-caps", core.GuardScoped(s, listSpendCaps))
+	app.Post("/v1/admin/spend-caps", core.GuardScoped(s, createSpendCap))
+	app.Patch("/v1/admin/spend-caps/:id", core.GuardScoped(s, updateSpendCap))
+	app.Delete("/v1/admin/spend-caps/:id", core.GuardScoped(s, deleteSpendCap))
+}
+
+// getPromo returns the current platform plan promo. X-Org-Id is the admin org —
+// commerce stores the singleton in the reserved platform namespace regardless, and
+// the service token is what passes commerce's RequirePlatformAdmin.
+func getPromo(s *cloud.Service[core.State], c *zip.Ctx) error {
+	raw, status, err := s.State.Commerce.Forward(c.Context(), http.MethodGet, "/v1/platform/promo", s.State.AdminOrg, nil)
+	return relay(c, raw, status, err)
+}
+
+// putPromo upserts the platform plan promo from the SuperAdmin's {percentOff,start,
+// end,plans,active} body — the ONE place the 50%-off offer is configured.
+func putPromo(s *cloud.Service[core.State], c *zip.Ctx) error {
+	raw, status, err := s.State.Commerce.Forward(c.Context(), http.MethodPut, "/v1/platform/promo", s.State.AdminOrg, c.Body())
+	return relay(c, raw, status, err)
+}
+
+// listSpendCaps returns a target org's usage caps (spend-alerts + derived period
+// spend/over/warn/resetsAt). The org is the SuperAdmin's ?org= or, for a scoped
+// admin, their own — never a client-widened scope.
+func listSpendCaps(s *cloud.Service[core.State], c *zip.Ctx) error {
+	org, ok := targetOrg(s, c)
+	if !ok {
+		return core.Fail(c, "org required")
+	}
+	raw, status, err := s.State.Commerce.Forward(c.Context(), http.MethodGet, "/v1/billing/spend-alerts", org, nil)
+	return relay(c, raw, status, err)
+}
+
+// createSpendCap sets a cap on a target org (platform override of a customer budget).
+func createSpendCap(s *cloud.Service[core.State], c *zip.Ctx) error {
+	org, ok := targetOrg(s, c)
+	if !ok {
+		return core.Fail(c, "org required")
+	}
+	raw, status, err := s.State.Commerce.Forward(c.Context(), http.MethodPost, "/v1/billing/spend-alerts", org, c.Body())
+	return relay(c, raw, status, err)
+}
+
+// updateSpendCap edits a target org's cap by id (raise/lower the ceiling, flip enforce).
+func updateSpendCap(s *cloud.Service[core.State], c *zip.Ctx) error {
+	org, ok := targetOrg(s, c)
+	if !ok {
+		return core.Fail(c, "org required")
+	}
+	id := strings.TrimSpace(c.Param("id"))
+	if id == "" {
+		return core.Fail(c, "cap id required")
+	}
+	raw, status, err := s.State.Commerce.Forward(c.Context(), http.MethodPatch, "/v1/billing/spend-alerts/"+url.PathEscape(id), org, c.Body())
+	return relay(c, raw, status, err)
+}
+
+// deleteSpendCap removes a target org's cap by id.
+func deleteSpendCap(s *cloud.Service[core.State], c *zip.Ctx) error {
+	org, ok := targetOrg(s, c)
+	if !ok {
+		return core.Fail(c, "org required")
+	}
+	id := strings.TrimSpace(c.Param("id"))
+	if id == "" {
+		return core.Fail(c, "cap id required")
+	}
+	raw, status, err := s.State.Commerce.Forward(c.Context(), http.MethodDelete, "/v1/billing/spend-alerts/"+url.PathEscape(id), org, nil)
+	return relay(c, raw, status, err)
+}
+
+// targetOrg resolves which org a cap operation acts on: a SuperAdmin names it with
+// ?org=; a scoped admin is hard-pinned to their own subtree (?org= ignored). Empty
+// (false) when unresolvable, so the handler fails closed rather than acting on a
+// guessed tenant.
+func targetOrg(s *cloud.Service[core.State], c *zip.Ctx) (string, bool) {
+	sc := core.ResolveScope(s, c)
+	if sc.Super {
+		if org := strings.TrimSpace(c.Query("org")); org != "" {
+			return org, true
+		}
+		return "", false
+	}
+	if len(sc.Orgs) > 0 && strings.TrimSpace(sc.Orgs[0]) != "" {
+		return sc.Orgs[0], true
+	}
+	return "", false
+}
+
+// relay surfaces commerce's OWN verdict in the /v1 envelope: a 2xx passes the raw
+// JSON through as data (so the console decodes the exact SpendAlert/Promo shape), a
+// non-2xx becomes an honest failure carrying commerce's status + message rather than
+// masking a 400 validation as success.
+func relay(c *zip.Ctx, raw []byte, status int, err error) error {
+	if err != nil {
+		return core.Fail(c, err.Error())
+	}
+	if status < 200 || status >= 300 {
+		msg := strings.TrimSpace(string(raw))
+		if msg == "" {
+			msg = http.StatusText(status)
+		}
+		return core.Fail(c, msg)
+	}
+	if len(raw) == 0 {
+		return core.OK(c, map[string]any{"ok": true})
+	}
+	return core.OKRaw(c, json.RawMessage(raw), 0)
+}

clients/admin/limits_test.go

@@ -0,0 +1,120 @@
+package admin
+
+import (
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"testing"
+)
+
+// recCommerce records the X-Org-Id + method + path of the last forwarded request so a
+// test can prove the /v1/admin control plane targets the RIGHT tenant namespace, and
+// serves the promo + spend-alert shapes verbatim.
+type recCommerce struct {
+	server     *httptest.Server
+	mu         sync.Mutex
+	lastOrg    string
+	lastMethod string
+	lastPath   string
+}
+
+func newRecCommerce() *recCommerce {
+	f := &recCommerce{}
+	f.server = httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		f.mu.Lock()
+		f.lastOrg = r.Header.Get("X-Org-Id")
+		f.lastMethod = r.Method
+		f.lastPath = r.URL.Path
+		f.mu.Unlock()
+		w.Header().Set("Content-Type", "application/json")
+		switch {
+		case strings.HasSuffix(r.URL.Path, "/platform/promo"):
+			io.WriteString(w, `{"percentOff":50,"plans":["pro"],"active":true}`)
+		case strings.HasSuffix(r.URL.Path, "/spend-alerts"):
+			io.WriteString(w, `[{"id":"a1","threshold":10000,"enforce":true,"period":"2026-07","resetsAt":"2026-08-01T00:00:00Z"}]`)
+		default:
+			io.WriteString(w, `{}`)
+		}
+	}))
+	return f
+}
+
+func (f *recCommerce) seen() (string, string, string) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.lastMethod, f.lastPath, f.lastOrg
+}
+
+func envStatus(t *testing.T, body []byte) string {
+	t.Helper()
+	var e struct {
+		Status string `json:"status"`
+	}
+	_ = json.Unmarshal(body, &e)
+	return e.Status
+}
+
+// The promo control plane is SuperAdmin-only (core.Guard) and forwards to commerce's
+// platform-promo endpoint.
+func TestLimits_Promo_SuperOnly(t *testing.T) {
+	iam := newScopeIAM()
+	defer iam.server.Close()
+	com := newRecCommerce()
+	defer com.server.Close()
+	do := mount(t, iam.server.URL, com.server.URL, "")
+
+	// SuperAdmin GET → 200 ok, forwarded to /v1/platform/promo.
+	resp, body := do("GET", "/v1/admin/promos", superHdr)
+	if resp.StatusCode != http.StatusOK || envStatus(t, body) != "ok" {
+		t.Fatalf("super GET promos = %d %s", resp.StatusCode, body)
+	}
+	if m, p, _ := com.seen(); m != "GET" || !strings.HasSuffix(p, "/platform/promo") {
+		t.Fatalf("forwarded %s %s, want GET .../platform/promo", m, p)
+	}
+
+	// SuperAdmin PUT → forwarded as PUT.
+	if resp, _ := do("PUT", "/v1/admin/promos", superHdr); resp.StatusCode != http.StatusOK {
+		t.Fatalf("super PUT promos = %d", resp.StatusCode)
+	}
+	if m, _, _ := com.seen(); m != "PUT" {
+		t.Fatalf("promo PUT forwarded as %s, want PUT", m)
+	}
+
+	// A non-super org admin is REFUSED at the platform gate (403), never reaching commerce.
+	if resp, _ := do("GET", "/v1/admin/promos", orgAdminHdr); resp.StatusCode != http.StatusForbidden {
+		t.Fatalf("org-admin GET promos = %d, want 403 (platform-only)", resp.StatusCode)
+	}
+}
+
+// Cap oversight is org-scoped: a SuperAdmin targets any org via ?org=; a scoped admin
+// is hard-pinned to their OWN org (a client ?org= is ignored — the escalation line).
+func TestLimits_SpendCaps_OrgScoped(t *testing.T) {
+	iam := newScopeIAM()
+	defer iam.server.Close()
+	com := newRecCommerce()
+	defer com.server.Close()
+	do := mount(t, iam.server.URL, com.server.URL, "")
+
+	// SuperAdmin with ?org=maxpower → forwards X-Org-Id=maxpower.
+	resp, body := do("GET", "/v1/admin/spend-caps?org=maxpower", superHdr)
+	if resp.StatusCode != http.StatusOK || envStatus(t, body) != "ok" {
+		t.Fatalf("super spend-caps = %d %s", resp.StatusCode, body)
+	}
+	if _, p, org := com.seen(); org != "maxpower" || !strings.HasSuffix(p, "/spend-alerts") {
+		t.Fatalf("forwarded org=%q path=%q, want maxpower .../spend-alerts", org, p)
+	}
+
+	// SuperAdmin WITHOUT ?org → org required (honest error, no guessed tenant).
+	if _, body := do("GET", "/v1/admin/spend-caps", superHdr); envStatus(t, body) != "error" {
+		t.Fatalf("super spend-caps without org must be an error envelope, got %s", body)
+	}
+
+	// A scoped org admin naming a FOREIGN ?org=hanzo is hard-pinned to their OWN org.
+	do("GET", "/v1/admin/spend-caps?org=hanzo", orgAdminHdr)
+	if _, _, org := com.seen(); org != "maxpower" {
+		t.Fatalf("scoped admin forwarded org=%q, want maxpower (client ?org= must be ignored)", org)
+	}
+}

clients/admin/metrics/metrics.go

@@ -0,0 +1,110 @@
+// Package metrics is the fleet SaaS-operations god-view (/v1/admin/metrics) — the
+// operator's business dashboard: MRR/ARR, net-new vs churned MRR, the plan/category
+// mix, the top customers, and the recent subscription movements. SuperAdmin only
+// (core.Guard).
+//
+// It OWNS no aggregation. The whole snapshot is computed IN commerce (the system of
+// record for subscriptions + the usage ledger) by its cross-org SaaS-metrics engine
+// (GET /v1/metrics/saas), which admin PROXIES with the SAME admin-scoped S2S service
+// token finance uses for COGS. The engine is ALREADY fleet-wide — it walks every org
+// namespace itself — so this is a SINGLE upstream read, no per-org fan-out, exactly as
+// finance consumes commerce Costs. An unwired or unreachable commerce degrades to an
+// honest empty snapshot (real zeros, `[]` not null) with a not-ok source, never a
+// fabricated number.
+package metrics
+
+import (
+	"errors"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/admin/commerce"
+	"github.com/hanzoai/cloud/clients/admin/core"
+	"github.com/zap-proto/zip"
+)
+
+// errUnconfigured marks commerce not wired on this deployment — core.SrcOf reports it as
+// a not-ok source so the console renders the honest not-configured state.
+var errUnconfigured = errors.New("commerce metrics not configured")
+
+// defaultLimit caps the top-customers list when the caller sends none (mirrors the
+// commerce engine's own default so the proxy never asks for more than it returns).
+const defaultLimit = 20
+
+// MetricsData is the GET /v1/admin/metrics payload: the commerce SaaS snapshot, flat,
+// plus the admin read time and the upstream freshness strip every god-view carries.
+type MetricsData struct {
+	commerce.SaaSMetrics
+	GeneratedAt string              `json:"generatedAt"`
+	Sources     []core.SourceStatus `json:"sources"`
+}
+
+// Metrics answers GET /v1/admin/metrics by proxying the commerce SaaS-metrics engine
+// (already a fleet-wide cross-org aggregate). SuperAdmin only.
+//
+//	GET /v1/admin/metrics?window=30d&limit=20
+func Metrics(s *cloud.Service[core.State], c *zip.Ctx) error {
+	ctx := c.Context()
+	now := time.Now().UTC().Format(time.RFC3339)
+	window := strings.TrimSpace(c.Query("window"))
+	limit := parseLimit(c.Query("limit"))
+
+	if !s.State.Commerce.Ready() {
+		return core.OK(c, empty(now, window, core.SrcOf("commerce-metrics", errUnconfigured, 0, now)))
+	}
+	m, err := s.State.Commerce.Metrics(ctx, window, limit)
+	if err != nil {
+		return core.OK(c, empty(now, window, core.SrcOf("commerce-metrics", err, 0, now)))
+	}
+	return core.OK(c, MetricsData{
+		SaaSMetrics: normalize(m),
+		GeneratedAt: now,
+		Sources:     []core.SourceStatus{core.SrcOf("commerce-metrics", nil, m.Orgs, now)},
+	})
+}
+
+// empty is the honest not-configured/unreachable snapshot: real zeros + empty slices
+// (never null, never fabricated) plus the not-ok source.
+func empty(now, window string, src core.SourceStatus) MetricsData {
+	return MetricsData{
+		SaaSMetrics: normalize(commerce.SaaSMetrics{AsOf: now, Currency: "usd", Window: window}),
+		GeneratedAt: now,
+		Sources:     []core.SourceStatus{src},
+	}
+}
+
+// normalize replaces nil slices with empty ones so the JSON is honest arrays (`[]`, not
+// null) and the console never has to guard a missing collection.
+func normalize(m commerce.SaaSMetrics) commerce.SaaSMetrics {
+	if m.Revenue.ByCategory == nil {
+		m.Revenue.ByCategory = []commerce.SaaSCategory{}
+	}
+	if m.Subs.ByPlan == nil {
+		m.Subs.ByPlan = []commerce.SaaSPlan{}
+	}
+	if m.Subs.Recent == nil {
+		m.Subs.Recent = []commerce.SaaSEvent{}
+	}
+	if m.Customers == nil {
+		m.Customers = []commerce.SaaSCustomer{}
+	}
+	if m.Gaps == nil {
+		m.Gaps = []string{}
+	}
+	return m
+}
+
+// parseLimit clamps the top-N cap to [1,200], defaulting to defaultLimit — mirrors the
+// commerce engine's clamp exactly.
+func parseLimit(s string) int {
+	n, err := strconv.Atoi(strings.TrimSpace(s))
+	if err != nil || n <= 0 {
+		return defaultLimit
+	}
+	if n > 200 {
+		return 200
+	}
+	return n
+}

clients/admin/metrics/routes.go

@@ -0,0 +1,13 @@
+package metrics
+
+import (
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/admin/core"
+	"github.com/zap-proto/zip"
+)
+
+// Routes registers the SaaS-metrics god-view (SuperAdmin only, cross-tenant business
+// aggregate).
+func Routes(app *zip.App, s *cloud.Service[core.State]) {
+	app.Get("/v1/admin/metrics", core.Guard(s, Metrics))
+}

clients/admin/services.go

@@ -0,0 +1,73 @@
+package admin
+
+// The /v1/admin/services board — the launch-control LENS over the waitlist gate, twin
+// of /v1/admin/flags. Every hosted service (studio/chat/console/app/api/team + runtime
+// onboards) with its LIVE waitlist mode — the switch waitlist.<svc>, evaluated through
+// clients/admission (which composes the flag engine one-way). This is the "remove the
+// waitlist one service at a time" toggle. SuperAdmin only (core.Guard), like every
+// platform /v1/admin/*.
+//
+// The registry + mode decide + these admin control funcs live in clients/admission,
+// the complete launch-gate feature; flags is the pure engine underneath. Per-user
+// approval (the second, orthogonal axis) stays IAM's, reached via the existing admin IAM
+// proxy — not re-served here.
+
+import (
+	"errors"
+	"net/http"
+	"strings"
+
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/admin/core"
+	"github.com/hanzoai/cloud/clients/admission"
+	"github.com/zap-proto/zip"
+)
+
+// services answers GET /v1/admin/services — the launch board (every service + live mode).
+func services(s *cloud.Service[core.State], c *zip.Ctx) error {
+	rows, err := admission.ListWaitlistServices(c.Context())
+	if err != nil {
+		return zip.Errorf(http.StatusInternalServerError, "list services: %v", err)
+	}
+	return core.OK(c, map[string]any{"services": rows})
+}
+
+// upsertService answers POST /v1/admin/services — onboard or edit a hosted service so a
+// new host is governed WITHOUT a redeploy. A re-register PRESERVES the live switch.
+func upsertService(s *cloud.Service[core.State], c *zip.Ctx) error {
+	var in admission.ServiceInput
+	if err := c.Bind(&in); err != nil {
+		return err
+	}
+	if strings.TrimSpace(in.Service) == "" {
+		return zip.ErrBadRequest("service slug is required")
+	}
+	view, err := admission.UpsertWaitlistService(c.Context(), in, c.UserEmail())
+	if err != nil {
+		return zip.ErrBadRequest(err.Error())
+	}
+	return core.OK(c, map[string]any{"service": view})
+}
+
+// setServiceMode answers POST /v1/admin/services/:service/mode — flip one service's
+// waitlist switch {waitlistMode:bool}. The launch lever; hot, no redeploy.
+func setServiceMode(s *cloud.Service[core.State], c *zip.Ctx) error {
+	service := strings.TrimSpace(c.Param("service"))
+	if service == "" {
+		return zip.ErrBadRequest("service is required")
+	}
+	var body struct {
+		WaitlistMode bool `json:"waitlistMode"`
+	}
+	if err := c.Bind(&body); err != nil {
+		return err
+	}
+	view, err := admission.SetWaitlistMode(c.Context(), service, body.WaitlistMode, c.UserEmail())
+	if err != nil {
+		if errors.Is(err, admission.ErrServiceNotFound) {
+			return zip.ErrNotFound("service not found: " + service)
+		}
+		return zip.Errorf(http.StatusInternalServerError, "set mode: %v", err)
+	}
+	return core.OK(c, map[string]any{"service": view})
+}

clients/admin/subscriptions/routes.go

@@ -0,0 +1,12 @@
+package subscriptions
+
+import (
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/admin/core"
+	"github.com/zap-proto/zip"
+)
+
+// Routes registers the fleet subscription view (SuperAdmin only, cross-tenant).
+func Routes(app *zip.App, s *cloud.Service[core.State]) {
+	app.Get("/v1/admin/subscriptions", core.Guard(s, Subscriptions))
+}

clients/admin/subscriptions/subscriptions.go

@@ -0,0 +1,139 @@
+// Package subscriptions is the fleet SUBSCRIPTION view (/v1/admin/subscriptions) —
+// every tenant's plan subscription: customer/org, plan, status, monthly-normalized MRR,
+// and the current-period start/renews. SuperAdmin only (core.Guard).
+//
+// Like invoices (and revenue) it fans out the org directory concurrently and reads each
+// org's subscriptions via the admin S2S seam, tagging every row with its owning org. The
+// MRR is monthly-normalized in the commerce reader so a yearly plan is comparable to a
+// monthly one. Best-effort per org (a failed read contributes no rows, never fabricated
+// ones); optional ?org= scopes to one tenant, ?status= filters, ?limit= caps.
+package subscriptions
+
+import (
+	"context"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/admin/core"
+	"github.com/hanzoai/cloud/clients/admin/iam"
+	"github.com/zap-proto/zip"
+)
+
+// defaultLimit caps the merged fleet subscription list when the caller sends none.
+const defaultLimit = 500
+
+// SubscriptionRow is one row of GET /v1/admin/subscriptions — a tenant's subscription at
+// a glance, tagged with its owning org. MRR is USD cents; timestamps are RFC3339 strings.
+type SubscriptionRow struct {
+	ID       string `json:"id"`
+	Org      string `json:"org"`
+	Display  string `json:"display"`
+	User     string `json:"user"`
+	Plan     string `json:"plan"`
+	Status   string `json:"status"`
+	MRRCents int64  `json:"mrrCents"`
+	Started  string `json:"started"`
+	Renews   string `json:"renews"`
+}
+
+// Subscriptions answers GET /v1/admin/subscriptions.
+//
+//	GET /v1/admin/subscriptions?org=&status=&limit=
+func Subscriptions(s *cloud.Service[core.State], c *zip.Ctx) error {
+	ctx := c.Context()
+	cr := core.CallerCreds(c)
+	status := strings.TrimSpace(c.Query("status"))
+	wantOrg := strings.TrimSpace(c.Query("org"))
+	limit := parseLimit(c.Query("limit"))
+
+	orgs, err := core.ListOrgs(s, ctx, cr)
+	if err != nil {
+		return core.Fail(c, err.Error())
+	}
+	if wantOrg != "" {
+		orgs = filterOrg(orgs, wantOrg)
+	}
+
+	// Per-org subscriptions, fanned out concurrently (best-effort per org).
+	perOrg := make([][]SubscriptionRow, len(orgs))
+	sem := make(chan struct{}, core.MaxCustomerConcurrency)
+	var wg sync.WaitGroup
+	for i, o := range orgs {
+		wg.Add(1)
+		sem <- struct{}{}
+		go func(i int, o iam.Org) {
+			defer wg.Done()
+			defer func() { <-sem }()
+			perOrg[i] = subscriptionsOf(s, ctx, o, status)
+		}(i, o)
+	}
+	wg.Wait()
+
+	rows := make([]SubscriptionRow, 0)
+	for _, r := range perOrg {
+		rows = append(rows, r...)
+	}
+	// Highest-MRR first (ties broken by most-recent start); cap to the merged limit.
+	sort.Slice(rows, func(i, j int) bool {
+		if rows[i].MRRCents != rows[j].MRRCents {
+			return rows[i].MRRCents > rows[j].MRRCents
+		}
+		return rows[i].Started > rows[j].Started
+	})
+	total := len(rows)
+	if len(rows) > limit {
+		rows = rows[:limit]
+	}
+	return core.OKList(c, rows, total)
+}
+
+// subscriptionsOf reads one org's subscriptions into fleet rows, tagged with the org.
+// Best-effort: a failed read yields no rows so the fleet view degrades honestly.
+func subscriptionsOf(s *cloud.Service[core.State], ctx context.Context, o iam.Org, status string) []SubscriptionRow {
+	entries, err := s.State.Commerce.Subscriptions(ctx, o.Name, status)
+	if err != nil {
+		return nil
+	}
+	display := core.Display(o.DisplayName, o.Name)
+	rows := make([]SubscriptionRow, 0, len(entries))
+	for _, sub := range entries {
+		rows = append(rows, SubscriptionRow{
+			ID:       sub.ID,
+			Org:      o.Name,
+			Display:  display,
+			User:     sub.User,
+			Plan:     sub.Plan,
+			Status:   sub.Status,
+			MRRCents: int64(sub.MRR),
+			Started:  sub.Started,
+			Renews:   sub.Renews,
+		})
+	}
+	return rows
+}
+
+// filterOrg narrows the directory to the one requested org (empty when it does not
+// exist — an honest empty list, never a fabricated tenant).
+func filterOrg(orgs []iam.Org, want string) []iam.Org {
+	for _, o := range orgs {
+		if o.Name == want {
+			return []iam.Org{o}
+		}
+	}
+	return nil
+}
+
+// parseLimit clamps the merged-list cap to [1,5000], defaulting to defaultLimit.
+func parseLimit(s string) int {
+	n, err := strconv.Atoi(strings.TrimSpace(s))
+	if err != nil || n <= 0 {
+		return defaultLimit
+	}
+	if n > 5000 {
+		return 5000
+	}
+	return n
+}

clients/admission/approval.go

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-package featuregate
+package admission
 
 import (
 	"context"
@@ -31,7 +31,7 @@ import (
 // approval is FAIL-OPEN: a user is approved unless properties.approvalStatus is
 // EXACTLY "pending" (absent / "approved" / "rejected" all read approved via
 // IsApproved). Only "pending" holds a user on the waitlist. Keeping the literal
-// here (not importing IAM) keeps featuregate self-contained.
+// here (not importing IAM) keeps admission self-contained.
 const approvalStatusPending = "pending"
 
 // approvedHeader is the FORWARD-PERFECT path: once IAM carries approvalStatus in

clients/admission/approval_test.go

@@ -1,7 +1,7 @@
 // Copyright 2023-2026 Hanzo AI Inc. All Rights Reserved.
 // Licensed under the Apache License, Version 2.0.
 
-package featuregate
+package admission
 
 import (
 	"context"

clients/admission/middleware.go

@@ -12,9 +12,35 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-package featuregate
+// Package admission is the launch-control GATE for Hanzo's hosted services — the
+// COMPLETE waitlist feature, COMPOSING the ONE flag engine (clients/flags) one-way. It
+// owns:
+//
+//   - the host→service registry (registry.go) + the brand seed (waitlist.go),
+//   - the per-service MODE decide WaitlistModeForHost — a service's mode IS the switch
+//     waitlist.<svc>, evaluated through the flag engine (flags.Bool),
+//   - the admin control funcs (List/Set/Upsert) the /v1/admin/services board calls,
+//   - the guard's public mode read /v1/flags/waitlist, Mount,
+//   - the native enforcement middleware (Enforce, this file),
+//   - the per-user approval predicate (Approvals, reused from IAM — approval.go).
+//
+// flags NEVER imports admission; admission imports flags. The engine is the pure
+// (Principal, context) -> verdict primitive; this package is its first composed tenant.
+// Enforcement is decomplected into two orthogonal axes:
+//
+//   - PER-SERVICE  waitlist mode on|off  — the switch waitlist.<svc>, resolved for a
+//     request host via WaitlistModeForHost (the decide, waitlist.go).
+//   - PER-USER     approvalStatus pending|approved — owned by IAM (approval.go), REUSED.
+//
+// THE RULE, applied at ONE native enforcement point (Enforce):
+//
+//	if waitlistMode[host] AND NOT user.approved → bounce to the waitlist
+//	if approved OR mode=off                     → allow
+//	unauthenticated                             → login first
+package admission
 
 import (
+	"context"
 	"net/http"
 	"strings"
 
@@ -40,14 +66,15 @@ import (
 // INTEGRATION POINT — wire in serve.go RIGHT AFTER SanitizeIdentity:
 //
 //	app.Use(IdentityMiddleware(cfg))          // establishes the validated principal
-//	app.Use(featuregate.Enforce(featuregate.EnforceConfig{ WaitlistURL: … }))  // ← here
+//	app.Use(admission.Enforce(admission.EnforceConfig{ WaitlistURL: … }))  // ← here
 //
 // It reads the sanitized X-User-Id / X-User-IsAdmin / X-User-Approved that
 // IdentityMiddleware minted, so it MUST run after it and (like BillingGate) before
 // the subsystem handlers. It is deliberately NOT wired here — the unified-binary
-// agent owns serve.go's boot chain; this package exposes Enforce + the store so the
-// one-line app.Use lands without a merge collision. The store is resolved lazily
-// (moduleStore()) so Enforce can be constructed before Mount runs.
+// agent owns serve.go's boot chain; this package exposes Enforce so the one-line
+// app.Use lands without a merge collision. The decide (WaitlistModeForHost) is
+// resolved PER REQUEST and fail-opens until the flags engine has mounted, so Enforce
+// can be constructed before Mount runs.
 //
 // WHY NATIVE IS CANONICAL (in-cluster-bypass). The @file edge guard only gates
 // traffic arriving THROUGH the ingress — a pod reaching another service's pod
@@ -80,28 +107,39 @@ type EnforceConfig struct {
 	// ExemptPrefixes are request-path prefixes never gated (health/metrics/auth).
 	// A sensible default set is used when empty.
 	ExemptPrefixes []string
+
+	// Gate is THE decide: it resolves whether a request host is in waitlist mode,
+	// via the ONE policy engine. When nil it is WaitlistModeForHost —
+	// host→service→waitlist.<svc>. Injected only in tests. Fail-open by contract:
+	// known=false (unmounted / registry error / un-governed host) → not gated.
+	Gate func(ctx context.Context, host string) (mode bool, service string, known bool)
 }
 
 // defaultExemptPrefixes are the paths enforcement must never touch — HIP-0106
 // health, the auth/OIDC handshake, and the waitlist join API itself (so a gated
 // user can still submit the waitlist form).
 var defaultExemptPrefixes = []string{
-	"/v1/featuregate/", // the mode read + the health route
-	"/v1/iam/",         // auth / OIDC / approval-status / get-account handshake
-	"/v1/waitlist",     // the waitlist join API (a gated user must reach it)
+	"/v1/flags/waitlist", // the guard's public mode read (flags engine)
+	"/v1/iam/",           // auth / OIDC / approval-status / get-account handshake
+	"/v1/waitlist",       // the waitlist join API (a gated user must reach it)
 	"/health",
 	"/healthz",
 	"/__guard/", // the @file guard's own callback surface (defense in depth)
 }
 
-// Enforce builds the native enforcement middleware. It is a no-op passthrough when
-// no registry store is resolved yet (moduleStore() nil, i.e. Mount hasn't run) —
-// so a request before boot completes is never wrongly gated.
+// Enforce builds the native enforcement middleware. It is a no-op passthrough when the
+// decide reports the host is not governed (gate known=false — the flags registry not
+// mounted yet, a store error, or an un-governed host), so a request before boot
+// completes is never wrongly gated.
 func Enforce(cfg EnforceConfig) zip.Handler {
 	approvals := cfg.Approvals
 	if approvals == nil {
 		approvals = NewApprovals(cfg.IAMBase, 0)
 	}
+	gate := cfg.Gate
+	if gate == nil {
+		gate = WaitlistModeForHost // the ONE decide: host→service→waitlist.<svc>
+	}
 	exempt := cfg.ExemptPrefixes
 	if len(exempt) == 0 {
 		exempt = defaultExemptPrefixes
@@ -109,10 +147,6 @@ func Enforce(cfg EnforceConfig) zip.Handler {
 	waitlistURL := strings.TrimRight(strings.TrimSpace(cfg.WaitlistURL), "/")
 
 	return func(c *zip.Ctx) error {
-		store := moduleStore()
-		if store == nil {
-			return c.Next() // registry not mounted yet — never gate pre-boot
-		}
 		path := c.Path()
 		for _, p := range exempt {
 			if strings.HasPrefix(path, p) {
@@ -134,14 +168,13 @@ func Enforce(cfg EnforceConfig) zip.Handler {
 			return c.Next()
 		}
 
-		host := c.Fiber().Hostname()
-		mode, _, known, err := store.ModeForHost(c.Context(), host)
-		if err != nil || !known || !mode {
-			// Un-governed host, mode OFF, or a registry read error → allow. A
-			// governed host is opened by flipping mode OFF; an unknown host is
-			// not ours to gate at the shared cloud edge (fail-open on a read
-			// error keeps the API available — the guard is the belt-and-braces
-			// gate for the hosts that must stay closed).
+		mode, _, known := gate(c.Context(), c.Fiber().Hostname())
+		if !known || !mode {
+			// Un-governed host, mode OFF, or a registry read error (the decide folds
+			// all three into known=false) → allow. A governed host is opened by
+			// flipping its waitlist.<svc> switch OFF; an unknown host is not ours to
+			// gate at the shared cloud edge — the @file guard is the belt-and-braces
+			// gate for the hosts that must stay closed.
 			return c.Next()
 		}
 
@@ -173,7 +206,7 @@ func bounce(c *zip.Ctx, waitlistURL string) error {
 }
 
 // apiKeyPrefixes are the Hanzo API-key prefixes. This MIRRORS cloud
-// auth_identity.go isAPIKey (the ONE authority) — kept local so featuregate stays
+// auth_identity.go isAPIKey (the ONE authority) — kept local so admission stays
 // self-contained (no cloud-internal import) while agreeing on the exact contract:
 // a token with one of these prefixes is a possession-gated API key, not a session
 // principal. If cloud adds a prefix there, add it here.
@@ -231,14 +264,3 @@ func isAPIClient(c *zip.Ctx) bool {
 	}
 	return false
 }
-
-// ── package-level store resolution (lazy, for Enforce constructed before Mount) ──
-
-// moduleStore returns the registry store once Mount has opened it, else nil. The
-// Enforce closure calls it PER REQUEST, so serve.go can construct Enforce before
-// MountAll runs (the store is set during Mount, well before the first request).
-func moduleStore() *Store {
-	moduleMu.RLock()
-	defer moduleMu.RUnlock()
-	return moduleStoreRef
-}

clients/admission/middleware_test.go

@@ -1,7 +1,7 @@
 // Copyright 2023-2026 Hanzo AI Inc. All Rights Reserved.
 // Licensed under the Apache License, Version 2.0.
 
-package featuregate
+package admission
 
 import (
 	"context"
@@ -14,27 +14,31 @@ import (
 	"github.com/zap-proto/zip"
 )
 
-// gateApp mounts Enforce over a seeded registry and a catch-all "ok" handler. The
-// injected approval status decides whether the caller is off the waitlist. Returns
-// a driver that issues one request (host + path + identity + Accept).
+// testGate is the injected decide (the flags engine's WaitlistModeForHost seam):
+// hanzo.chat is gated, api.hanzo.ai is open, everything else is un-governed. This is
+// exactly what WaitlistModeForHost returns for the equivalent registry, without
+// standing up the native flag engine (cgo) in a middleware unit test.
+func testGate(_ context.Context, host string) (mode bool, service string, known bool) {
+	switch host {
+	case "hanzo.chat":
+		return true, "chat", true // gated
+	case "api.hanzo.ai":
+		return false, "api", true // open
+	default:
+		return false, "", false // un-governed
+	}
+}
+
+// gateApp mounts Enforce over the injected decide and a catch-all "ok" handler. The
+// injected approval status decides whether the caller is off the waitlist.
 func gateApp(t *testing.T, approvalStatus string) *zip.App {
 	t.Helper()
-	st := newTestStore(t)
-	if _, err := st.Seed(context.Background(), []SeedService{
-		{Service: "chat", Hosts: []string{"hanzo.chat"}, WaitlistMode: true},   // gated
-		{Service: "api", Hosts: []string{"api.hanzo.ai"}, WaitlistMode: false}, // open
-	}, 100); err != nil {
-		t.Fatalf("seed: %v", err)
-	}
-	setModuleStore(st)
-	t.Cleanup(func() { setModuleStore(nil) })
-
 	approvals := newApprovalsWithLookup(func(context.Context, string, string) (string, bool) {
 		return approvalStatus, true
 	}, time.Minute)
 
 	app := zip.New(zip.Config{Logger: luxlog.New("test")})
-	app.Use(Enforce(EnforceConfig{WaitlistURL: "https://waitlist.hanzo.ai", Approvals: approvals}))
+	app.Use(Enforce(EnforceConfig{WaitlistURL: "https://waitlist.hanzo.ai", Approvals: approvals, Gate: testGate}))
 	app.Get("/*", func(c *zip.Ctx) error { return c.String(200, "ok") })
 	return app
 }
@@ -141,16 +145,17 @@ func TestRule_UnauthenticatedBrowser_BouncedToWaitlist(t *testing.T) {
 }
 
 // MONEY-CRITICAL: a paid inference request with a Hanzo API key MUST flow through
-// Enforce even on the seeded waitlist-ON `api` host — it is possession-gated + billed
-// downstream, never waitlist-gated. Without the exemption THE RULE would 401 it and
-// break inference cluster-wide.
+// Enforce even on a waitlist-ON host — it is possession-gated + billed downstream,
+// never waitlist-gated. Without the exemption THE RULE would 401 it and break
+// inference cluster-wide.
 func TestRule_APIKeyInference_NeverGated(t *testing.T) {
-	app := gateApp(t, "pending") // api.hanzo.ai is seeded waitlist ON
+	app := gateApp(t, "pending")
 	for _, key := range []string{"hk-43f50b6b", "sk-hz-abc", "pk-hz-obs", "fw_live_x", "hz_secret"} {
-		// The exact paid-inference shape: Bearer key, JSON accept, NO session/user.
+		// The exact paid-inference shape: Bearer key, JSON accept, NO session/user, on a
+		// GATED host — the exemption, not mode, must carry it through.
 		for _, p := range []string{"/v1/chat/completions", "/v1/models", "/v1/embeddings"} {
 			code, _ := drive(t, app, greq{
-				host: "api.hanzo.ai", path: p, accept: "application/json",
+				host: "hanzo.chat", path: p, accept: "application/json",
 				authorization: "Bearer " + key,
 			})
 			if code != 200 {
@@ -159,7 +164,7 @@ func TestRule_APIKeyInference_NeverGated(t *testing.T) {
 		}
 	}
 	// The api-key / x-api-key header form is exempt too.
-	code, _ := drive(t, app, greq{host: "api.hanzo.ai", path: "/v1/chat/completions", accept: "application/json", apiKeyHeader: "hk-headerform"})
+	code, _ := drive(t, app, greq{host: "hanzo.chat", path: "/v1/chat/completions", accept: "application/json", apiKeyHeader: "hk-headerform"})
 	if code != 200 {
 		t.Fatalf("api-key header inference = %d, want 200", code)
 	}
@@ -181,7 +186,7 @@ func TestRule_UngovernedHost_PassesThrough(t *testing.T) {
 
 func TestRule_ExemptPaths_NeverGated(t *testing.T) {
 	app := gateApp(t, "pending")
-	for _, p := range []string{"/health", "/v1/iam/get-account", "/v1/waitlist/join", "/v1/featuregate/mode"} {
+	for _, p := range []string{"/health", "/v1/iam/get-account", "/v1/waitlist/join", "/v1/flags/waitlist"} {
 		code, _ := drive(t, app, greq{host: "hanzo.chat", path: p, user: "u", org: "acme", accept: html})
 		if code != 200 {
 			t.Fatalf("exempt path %q = %d, want 200 (never gated)", p, code)
@@ -200,14 +205,16 @@ func TestRule_ForwardHeaderApproved_ThroughWithoutLookup(t *testing.T) {
 	}
 }
 
-func TestEnforce_NoStore_IsPassthrough(t *testing.T) {
-	setModuleStore(nil)
+// The DEFAULT gate (nil Gate → WaitlistModeForHost) fail-opens before the flag
+// engine has mounted: with no engine, WaitlistModeForHost returns known=false for every
+// host, so Enforce never gates pre-boot.
+func TestEnforce_DefaultGate_FailsOpenPreBoot(t *testing.T) {
 	app := zip.New(zip.Config{Logger: luxlog.New("test")})
 	app.Use(Enforce(EnforceConfig{WaitlistURL: "https://waitlist.hanzo.ai",
 		Approvals: newApprovalsWithLookup(func(context.Context, string, string) (string, bool) { return "pending", true }, time.Minute)}))
 	app.Get("/*", func(c *zip.Ctx) error { return c.String(200, "ok") })
 	code, _ := drive(t, app, greq{host: "hanzo.chat", path: "/dashboard", user: "u", org: "acme", accept: html})
 	if code != 200 {
-		t.Fatalf("no registry mounted = %d, want 200 (never gate pre-boot)", code)
+		t.Fatalf("default gate pre-boot = %d, want 200 (never gate before flags mounts)", code)
 	}
 }

clients/admission/registry.go

@@ -0,0 +1,270 @@
+package admission
+
+// The launch-registry — the host→service map + service display metadata. It is
+// deliberately MODE-FREE: a service's waitlist mode is NOT a column here, it is the
+// platform switch waitlist.<svc> evaluated through the ONE flag engine (clients/flags,
+// composed one-way from waitlist.go). This store answers only "which service owns this
+// host, and what is its display metadata" — the config the decide needs, with the
+// decision itself owned by the flag engine.
+//
+// It rides the SAME per-(org,project) OrgDB machinery as the flag defs (opened via
+// cloud.OrgStore, encrypted at rest via cek); the registry is PLATFORM-global, so it
+// lives in the reserved platform/platform tenant — one waitlist.db for the deployment.
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// ErrServiceNotFound is returned when a service slug is not in the registry.
+var ErrServiceNotFound = errors.New("admission: waitlist service not found")
+
+// ServiceRow is one hosted service in the registry (host→service + metadata). The
+// waitlist MODE is intentionally absent — it is the platform switch waitlist.<svc>,
+// read through the engine; ListWaitlistServices composes the two into a ServiceView.
+type ServiceRow struct {
+	Service     string   `json:"service"`
+	DisplayName string   `json:"displayName"`
+	Description string   `json:"description"`
+	Hosts       []string `json:"hosts"`
+	CreatedAt   int64    `json:"createdAt"`
+	UpdatedAt   int64    `json:"updatedAt"`
+	UpdatedBy   string   `json:"updatedBy"`
+}
+
+// waitlistStore is the registry over one OrgDB handle. Two tables, normalized:
+//
+//	wl_services(service PK, display_name, description, …)
+//	wl_hosts(host PK, service FK)   -- host → service, the hot lookup index
+type waitlistStore struct {
+	db *sql.DB
+}
+
+// openWaitlistStore migrates the registry schema over an already-opened (pragma'd,
+// cek-wrapped) OrgDB handle — the same open contract as flags' openStore for flag defs.
+func openWaitlistStore(db *sql.DB) (*waitlistStore, error) {
+	const schema = `
+CREATE TABLE IF NOT EXISTS wl_services (
+    service      TEXT PRIMARY KEY,
+    display_name TEXT NOT NULL DEFAULT '',
+    description  TEXT NOT NULL DEFAULT '',
+    created_at   INTEGER NOT NULL,
+    updated_at   INTEGER NOT NULL,
+    updated_by   TEXT NOT NULL DEFAULT ''
+);
+CREATE TABLE IF NOT EXISTS wl_hosts (
+    host    TEXT PRIMARY KEY,
+    service TEXT NOT NULL,
+    FOREIGN KEY(service) REFERENCES wl_services(service) ON DELETE CASCADE
+);
+CREATE INDEX IF NOT EXISTS ix_wl_hosts_service ON wl_hosts(service);
+`
+	if _, err := db.Exec(schema); err != nil {
+		return nil, fmt.Errorf("admission: waitlist migrate: %w", err)
+	}
+	return &waitlistStore{db: db}, nil
+}
+
+func (s *waitlistStore) Close() error { return s.db.Close() }
+
+// NormalizeHost reduces a request Host to the registry key: lowercased, trimmed,
+// port stripped. ONE canonicalization for the seed, onboard, and every lookup, so
+// "Hanzo.Chat:443" and "hanzo.chat" resolve to the same service.
+func NormalizeHost(host string) string {
+	h := strings.ToLower(strings.TrimSpace(host))
+	if i := strings.IndexByte(h, ':'); i >= 0 {
+		h = h[:i]
+	}
+	return h
+}
+
+// Seed inserts the initial registry idempotently (INSERT OR IGNORE on both tables),
+// so a boot never clobbers a runtime onboard. Returns the number of services created
+// (0 on a warm store).
+func (s *waitlistStore) Seed(ctx context.Context, rows []SeedService, now int64) (int, error) {
+	tx, err := s.db.BeginTx(ctx, nil)
+	if err != nil {
+		return 0, fmt.Errorf("waitlist seed tx: %w", err)
+	}
+	created := 0
+	for _, r := range rows {
+		svc := strings.ToLower(strings.TrimSpace(r.Service))
+		if svc == "" {
+			continue
+		}
+		res, err := tx.ExecContext(ctx,
+			`INSERT OR IGNORE INTO wl_services (service, display_name, description, created_at, updated_at, updated_by)
+			 VALUES (?,?,?,?,?,?)`,
+			svc, r.DisplayName, r.Description, now, now, "seed")
+		if err != nil {
+			_ = tx.Rollback()
+			return 0, fmt.Errorf("waitlist seed service %q: %w", svc, err)
+		}
+		if n, _ := res.RowsAffected(); n > 0 {
+			created++
+		}
+		for _, h := range r.Hosts {
+			host := NormalizeHost(h)
+			if host == "" {
+				continue
+			}
+			if _, err := tx.ExecContext(ctx,
+				`INSERT OR IGNORE INTO wl_hosts (host, service) VALUES (?,?)`, host, svc); err != nil {
+				_ = tx.Rollback()
+				return 0, fmt.Errorf("waitlist seed host %q: %w", host, err)
+			}
+		}
+	}
+	if err := tx.Commit(); err != nil {
+		return 0, fmt.Errorf("waitlist seed commit: %w", err)
+	}
+	return created, nil
+}
+
+// List returns every registered service (with its hosts), sorted by slug.
+func (s *waitlistStore) List(ctx context.Context) ([]ServiceRow, error) {
+	rows, err := s.db.QueryContext(ctx,
+		`SELECT service, display_name, description, created_at, updated_at, updated_by FROM wl_services`)
+	if err != nil {
+		return nil, fmt.Errorf("list waitlist services: %w", err)
+	}
+	defer func() { _ = rows.Close() }()
+
+	byService := map[string]*ServiceRow{}
+	out := make([]ServiceRow, 0, 16)
+	for rows.Next() {
+		var r ServiceRow
+		if err := rows.Scan(&r.Service, &r.DisplayName, &r.Description, &r.CreatedAt, &r.UpdatedAt, &r.UpdatedBy); err != nil {
+			return nil, fmt.Errorf("scan waitlist service: %w", err)
+		}
+		r.Hosts = []string{}
+		out = append(out, r)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	for i := range out {
+		byService[out[i].Service] = &out[i]
+	}
+	hostRows, err := s.db.QueryContext(ctx, `SELECT host, service FROM wl_hosts`)
+	if err != nil {
+		return nil, fmt.Errorf("list waitlist hosts: %w", err)
+	}
+	defer func() { _ = hostRows.Close() }()
+	for hostRows.Next() {
+		var host, svc string
+		if err := hostRows.Scan(&host, &svc); err != nil {
+			return nil, fmt.Errorf("scan waitlist host: %w", err)
+		}
+		if r := byService[svc]; r != nil {
+			r.Hosts = append(r.Hosts, host)
+		}
+	}
+	if err := hostRows.Err(); err != nil {
+		return nil, err
+	}
+	for i := range out {
+		sort.Strings(out[i].Hosts)
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].Service < out[j].Service })
+	return out, nil
+}
+
+// Get returns one service by slug, or ErrServiceNotFound.
+func (s *waitlistStore) Get(ctx context.Context, service string) (ServiceRow, error) {
+	svc := strings.ToLower(strings.TrimSpace(service))
+	row := s.db.QueryRowContext(ctx,
+		`SELECT service, display_name, description, created_at, updated_at, updated_by FROM wl_services WHERE service=?`, svc)
+	var out ServiceRow
+	err := row.Scan(&out.Service, &out.DisplayName, &out.Description, &out.CreatedAt, &out.UpdatedAt, &out.UpdatedBy)
+	if errors.Is(err, sql.ErrNoRows) {
+		return ServiceRow{}, ErrServiceNotFound
+	}
+	if err != nil {
+		return ServiceRow{}, fmt.Errorf("get waitlist service: %w", err)
+	}
+	out.Hosts = []string{}
+	hostRows, err := s.db.QueryContext(ctx, `SELECT host FROM wl_hosts WHERE service=? ORDER BY host`, svc)
+	if err != nil {
+		return ServiceRow{}, fmt.Errorf("get waitlist hosts: %w", err)
+	}
+	defer func() { _ = hostRows.Close() }()
+	for hostRows.Next() {
+		var h string
+		if err := hostRows.Scan(&h); err != nil {
+			return ServiceRow{}, fmt.Errorf("scan waitlist host: %w", err)
+		}
+		out.Hosts = append(out.Hosts, h)
+	}
+	return out, hostRows.Err()
+}
+
+// Upsert creates or updates a service's metadata + REPLACES its host set (a host
+// already claimed by ANOTHER service is skipped — first claim wins). It never
+// touches the mode: the mode is the waitlist.<svc> switch, flipped through
+// SetWaitlistMode. Returns the stored row.
+func (s *waitlistStore) Upsert(ctx context.Context, in ServiceRow, by string, now int64) (ServiceRow, error) {
+	svc := strings.ToLower(strings.TrimSpace(in.Service))
+	if svc == "" {
+		return ServiceRow{}, fmt.Errorf("admission: waitlist service slug required")
+	}
+	tx, err := s.db.BeginTx(ctx, nil)
+	if err != nil {
+		return ServiceRow{}, fmt.Errorf("waitlist upsert tx: %w", err)
+	}
+	if _, err := tx.ExecContext(ctx,
+		`INSERT INTO wl_services (service, display_name, description, created_at, updated_at, updated_by)
+		 VALUES (?,?,?,?,?,?)
+		 ON CONFLICT(service) DO UPDATE SET
+		   display_name=excluded.display_name,
+		   description=excluded.description,
+		   updated_at=excluded.updated_at,
+		   updated_by=excluded.updated_by`,
+		svc, in.DisplayName, in.Description, now, now, strings.TrimSpace(by)); err != nil {
+		_ = tx.Rollback()
+		return ServiceRow{}, fmt.Errorf("upsert waitlist service: %w", err)
+	}
+	if _, err := tx.ExecContext(ctx, `DELETE FROM wl_hosts WHERE service=?`, svc); err != nil {
+		_ = tx.Rollback()
+		return ServiceRow{}, fmt.Errorf("clear waitlist hosts: %w", err)
+	}
+	for _, h := range in.Hosts {
+		host := NormalizeHost(h)
+		if host == "" {
+			continue
+		}
+		if _, err := tx.ExecContext(ctx,
+			`INSERT OR IGNORE INTO wl_hosts (host, service) VALUES (?,?)`, host, svc); err != nil {
+			_ = tx.Rollback()
+			return ServiceRow{}, fmt.Errorf("add waitlist host %q: %w", host, err)
+		}
+	}
+	if err := tx.Commit(); err != nil {
+		return ServiceRow{}, fmt.Errorf("waitlist upsert commit: %w", err)
+	}
+	return s.Get(ctx, svc)
+}
+
+// ServiceForHost is the HOT lookup the decide calls once per request: it resolves a
+// request host to its owning service. known is false for an un-governed host (the
+// caller treats it as pass-through). host is normalized here so the caller passes
+// the raw Host.
+func (s *waitlistStore) ServiceForHost(ctx context.Context, host string) (service string, known bool, err error) {
+	h := NormalizeHost(host)
+	if h == "" {
+		return "", false, nil
+	}
+	var svc string
+	scanErr := s.db.QueryRowContext(ctx, `SELECT service FROM wl_hosts WHERE host=?`, h).Scan(&svc)
+	if errors.Is(scanErr, sql.ErrNoRows) {
+		return "", false, nil
+	}
+	if scanErr != nil {
+		return "", false, fmt.Errorf("waitlist service for host: %w", scanErr)
+	}
+	return svc, true, nil
+}

clients/admission/registry_test.go

@@ -0,0 +1,100 @@
+package admission
+
+// Registry coverage for the host→service store. It drives the store over a raw sqlite
+// handle (the same driver OrgDB uses), so it exercises the registry WITHOUT the cek
+// at-rest layer — runnable under CGO=0. The MODE is out of scope here by design (it is
+// the waitlist.<svc> switch, evaluated by the flag engine, covered separately).
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"path/filepath"
+	"testing"
+
+	_ "github.com/hanzoai/sqlite" // registers "sqlite" under both build tags
+)
+
+func newWaitlistStore(t *testing.T) *waitlistStore {
+	t.Helper()
+	db, err := sql.Open("sqlite", filepath.Join(t.TempDir(), "waitlist.db"))
+	if err != nil {
+		t.Fatalf("open: %v", err)
+	}
+	st, err := openWaitlistStore(db)
+	if err != nil {
+		t.Fatalf("openWaitlistStore: %v", err)
+	}
+	t.Cleanup(func() { _ = st.Close() })
+	return st
+}
+
+func TestWaitlistStore_SeedIdempotentAndServiceForHost(t *testing.T) {
+	st := newWaitlistStore(t)
+	ctx := context.Background()
+	seed := []SeedService{
+		{Service: "chat", DisplayName: "Chat", Hosts: []string{"hanzo.chat", "chat.hanzo.ai"}},
+		{Service: "api", DisplayName: "API", Hosts: []string{"api.hanzo.ai"}},
+	}
+	created, err := st.Seed(ctx, seed, 100)
+	if err != nil || created != 2 {
+		t.Fatalf("seed = %d, %v; want 2, nil", created, err)
+	}
+	if created, _ := st.Seed(ctx, seed, 200); created != 0 {
+		t.Fatalf("re-seed created = %d, want 0 (idempotent)", created)
+	}
+	// Host resolution is case-insensitive + port-stripped → the same service.
+	for _, h := range []string{"hanzo.chat", "Hanzo.Chat", "hanzo.chat:443", " HANZO.CHAT "} {
+		svc, known, err := st.ServiceForHost(ctx, h)
+		if err != nil || !known || svc != "chat" {
+			t.Fatalf("ServiceForHost(%q) = %q,%v,%v; want chat,true,nil", h, svc, known, err)
+		}
+	}
+	// An un-governed host is honestly unknown (the decide fail-opens on it).
+	if _, known, _ := st.ServiceForHost(ctx, "example.com"); known {
+		t.Fatal("example.com reported known; want unknown")
+	}
+}
+
+func TestWaitlistStore_ListSortedWithHosts(t *testing.T) {
+	st := newWaitlistStore(t)
+	ctx := context.Background()
+	if _, err := st.Seed(ctx, []SeedService{
+		{Service: "chat", Hosts: []string{"chat.hanzo.ai", "hanzo.chat"}},
+		{Service: "api", Hosts: []string{"api.hanzo.ai"}},
+	}, 100); err != nil {
+		t.Fatalf("seed: %v", err)
+	}
+	list, err := st.List(ctx)
+	if err != nil || len(list) != 2 {
+		t.Fatalf("List = %d, %v; want 2", len(list), err)
+	}
+	if list[0].Service != "api" || list[1].Service != "chat" {
+		t.Fatalf("List order = [%s %s], want [api chat]", list[0].Service, list[1].Service)
+	}
+	if len(list[1].Hosts) != 2 || list[1].Hosts[0] != "chat.hanzo.ai" {
+		t.Fatalf("chat hosts = %v, want sorted [chat.hanzo.ai hanzo.chat]", list[1].Hosts)
+	}
+}
+
+func TestWaitlistStore_UpsertOnboardsAndReplacesHosts(t *testing.T) {
+	st := newWaitlistStore(t)
+	ctx := context.Background()
+	row, err := st.Upsert(ctx, ServiceRow{Service: "search", DisplayName: "Search", Hosts: []string{"search.hanzo.ai"}}, "z@hanzo.ai", 100)
+	if err != nil || len(row.Hosts) != 1 || row.Hosts[0] != "search.hanzo.ai" {
+		t.Fatalf("Upsert new = %+v, %v", row, err)
+	}
+	// A metadata edit REPLACES the host set and updates the display name.
+	row, err = st.Upsert(ctx, ServiceRow{Service: "search", DisplayName: "Search v2", Hosts: []string{"search.hanzo.ai", "find.hanzo.ai"}}, "z@hanzo.ai", 200)
+	if err != nil || row.DisplayName != "Search v2" || len(row.Hosts) != 2 {
+		t.Fatalf("Upsert edit = %+v, %v", row, err)
+	}
+	svc, known, _ := st.ServiceForHost(ctx, "find.hanzo.ai")
+	if !known || svc != "search" {
+		t.Fatalf("onboarded host find.hanzo.ai = %q,%v; want search,true", svc, known)
+	}
+	// An unknown slug is ErrServiceNotFound (the admin lens maps it to 404).
+	if _, err := st.Get(ctx, "nope"); !errors.Is(err, ErrServiceNotFound) {
+		t.Fatalf("Get(nope) err = %v, want ErrServiceNotFound", err)
+	}
+}

clients/admission/waitlist.go

@@ -0,0 +1,351 @@
+package admission
+
+// The launch-control gate — the COMPLETE waitlist feature, COMPOSING the ONE flag
+// engine (clients/flags) one-way. Decomplected into the two orthogonal axes it always
+// was, with a single decision plane:
+//
+//   - MODE (per service):  waitlist.<svc> IS a platform switch, evaluated through the
+//     flag engine (flags.Bool / flags.SetPlatformSwitch / flags.Register). There is no
+//     second mode store.
+//   - HOST MAP + metadata:  the registry (registry.go) resolves a request host to the
+//     service whose switch governs it, and carries display metadata.
+//
+// The decide is WaitlistModeForHost(host) → (mode, service, known): resolve host→svc,
+// then read waitlist.<svc>. Enforce (middleware.go) consumes this decide; the admin
+// board (/v1/admin/services) and the guard's runtime mode read (/v1/flags/waitlist,
+// served here) read it too. Per-user approval (pending|approved) is the second,
+// orthogonal axis — IAM's, in approval.go.
+//
+// flags NEVER imports this package; this package imports flags. That one-way arrow is
+// the whole point of the decomplection: the engine is pure, the feature composes it.
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/flags"
+	luxlog "github.com/luxfi/log"
+	"github.com/zap-proto/zip"
+)
+
+// The reserved platform tenant the launch registry rides in — the SAME reserved
+// (org, project) the flag engine uses for its platform switches, so the registry and
+// the waitlist.<svc> switches co-locate. One waitlist.db for the deployment.
+const (
+	platformOrg     = "platform"
+	platformProject = "platform"
+)
+
+// registryState is admission's process-wide launch state: the platform-tenant
+// host→service registry store + the deployment brand it was seeded for. Installed by
+// Mount, torn down by Shutdown.
+type registryState struct {
+	store *cloud.OrgStore[*waitlistStore]
+	brand string
+}
+
+var mounted *registryState
+
+// SeedService is one row of the launch registry (a hosted service + its hosts). Mode
+// is intentionally absent — the launch posture (gated) is waitlistDef's Default "true".
+type SeedService struct {
+	Service     string
+	DisplayName string
+	Description string
+	Hosts       []string
+}
+
+// ServiceInput is the admin onboard/edit payload for /v1/admin/services. WaitlistMode
+// sets the launch switch for a NEW service; a re-register PRESERVES the live switch.
+type ServiceInput struct {
+	Service      string   `json:"service"`
+	DisplayName  string   `json:"displayName"`
+	Description  string   `json:"description"`
+	Hosts        []string `json:"hosts"`
+	WaitlistMode bool     `json:"waitlistMode"`
+}
+
+// ServiceView is one service as the admin board renders it: the registry row plus its
+// LIVE waitlist mode (the waitlist.<svc> switch evaluated through the engine).
+type ServiceView struct {
+	ServiceRow
+	WaitlistMode bool `json:"waitlistMode"`
+}
+
+// waitlistKey is the ONE naming rule: a service's mode is the switch waitlist.<svc>.
+func waitlistKey(svc string) string { return "waitlist." + strings.ToLower(strings.TrimSpace(svc)) }
+
+// waitlistDef is the platform switch for one service's mode. Default "true" = the
+// launch posture (gated until an admin opens it), so a deployment with no stored flag
+// behaves exactly as the old admission seed (waitlistMode ON).
+func waitlistDef(svc, display string) flags.Def {
+	if strings.TrimSpace(display) == "" {
+		display = svc
+	}
+	return flags.Def{
+		Key:      waitlistKey(svc),
+		Category: "Launch",
+		Label:    "Waitlist · " + display,
+		Desc:     "Waitlist mode for " + display + ": ON gates the service to APPROVED users; OFF opens it.",
+		Type:     flags.TypeBool,
+		Default:  "true",
+	}
+}
+
+// ensureWaitlistDef registers a service's switch if it is not already registered
+// (Mount registers the seed set with nicer labels; this covers runtime onboards).
+func ensureWaitlistDef(svc, display string) {
+	key := waitlistKey(svc)
+	for _, d := range flags.Defs() {
+		if d.Key == key {
+			return
+		}
+	}
+	flags.Register(waitlistDef(svc, display))
+}
+
+// boolDef is the minimal PostHog flag definition for a boolean switch value.
+func boolDef(on bool) json.RawMessage {
+	if on {
+		return json.RawMessage(`{"active":true}`)
+	}
+	return json.RawMessage(`{"active":false}`)
+}
+
+// requireRegistry resolves the platform-tenant registry store, or an error when the
+// gate is not mounted (writes need it; the decide fail-opens instead).
+func requireRegistry() (*waitlistStore, error) {
+	if mounted == nil || mounted.store == nil {
+		return nil, fmt.Errorf("admission: waitlist registry not mounted")
+	}
+	return mounted.store.For(platformOrg, platformProject)
+}
+
+// WaitlistModeForHost is THE decide the Enforce consumer, /v1/flags/waitlist, and
+// the admin board call: resolve host→service, then read the waitlist.<svc> switch
+// through the flag engine. FAIL-OPEN by construction — an unmounted registry, a store
+// error, or an un-governed host all return known=false, so a request is NEVER gated
+// pre-boot or on a registry fault (availability over a hard gate, matching the guard).
+func WaitlistModeForHost(ctx context.Context, host string) (mode bool, service string, known bool) {
+	if mounted == nil || mounted.store == nil {
+		return false, "", false
+	}
+	st, err := mounted.store.For(platformOrg, platformProject)
+	if err != nil {
+		return false, "", false
+	}
+	svc, known, err := st.ServiceForHost(ctx, host)
+	if err != nil || !known {
+		return false, "", false
+	}
+	return flags.Bool(waitlistKey(svc)), svc, true
+}
+
+// ListWaitlistServices returns the admin board: every registered service with its LIVE
+// mode (the waitlist.<svc> switch). SuperAdmin surface (the caller gates).
+func ListWaitlistServices(ctx context.Context) ([]ServiceView, error) {
+	st, err := requireRegistry()
+	if err != nil {
+		return nil, err
+	}
+	rows, err := st.List(ctx)
+	if err != nil {
+		return nil, err
+	}
+	out := make([]ServiceView, 0, len(rows))
+	for _, r := range rows {
+		out = append(out, ServiceView{ServiceRow: r, WaitlistMode: flags.Bool(waitlistKey(r.Service))})
+	}
+	return out, nil
+}
+
+// SetWaitlistMode flips one service's waitlist switch — the launch lever — and returns
+// the updated view. It is the ONE write path (through flags.SetPlatformSwitch, audited
+// in the flag activity log); the flip is hot (this pod applies immediately, peers
+// converge within the eval TTL). ErrServiceNotFound when the slug is unknown.
+func SetWaitlistMode(ctx context.Context, service string, mode bool, actor string) (ServiceView, error) {
+	service = strings.ToLower(strings.TrimSpace(service))
+	if service == "" {
+		return ServiceView{}, fmt.Errorf("admission: service is required")
+	}
+	st, err := requireRegistry()
+	if err != nil {
+		return ServiceView{}, err
+	}
+	row, err := st.Get(ctx, service) // ErrServiceNotFound → 404 upstream
+	if err != nil {
+		return ServiceView{}, err
+	}
+	ensureWaitlistDef(service, row.DisplayName)
+	if err := flags.SetPlatformSwitch(waitlistKey(service), boolDef(mode), actor); err != nil {
+		return ServiceView{}, err
+	}
+	return ServiceView{ServiceRow: row, WaitlistMode: flags.Bool(waitlistKey(service))}, nil
+}
+
+// UpsertWaitlistService onboards or edits a hosted service so a new host is governed
+// WITHOUT a redeploy. A NEW service takes in.WaitlistMode as its launch mode; a
+// re-register PRESERVES the live switch (never silently re-gating an opened service).
+func UpsertWaitlistService(ctx context.Context, in ServiceInput, actor string) (ServiceView, error) {
+	svc := strings.ToLower(strings.TrimSpace(in.Service))
+	if svc == "" {
+		return ServiceView{}, fmt.Errorf("admission: service slug is required")
+	}
+	st, err := requireRegistry()
+	if err != nil {
+		return ServiceView{}, err
+	}
+	_, getErr := st.Get(ctx, svc)
+	isNew := errors.Is(getErr, ErrServiceNotFound)
+	if getErr != nil && !isNew {
+		return ServiceView{}, getErr
+	}
+	row, err := st.Upsert(ctx, ServiceRow{
+		Service:     svc,
+		DisplayName: in.DisplayName,
+		Description: in.Description,
+		Hosts:       in.Hosts,
+	}, actor, time.Now().Unix())
+	if err != nil {
+		return ServiceView{}, err
+	}
+	ensureWaitlistDef(svc, row.DisplayName)
+	if isNew {
+		if err := flags.SetPlatformSwitch(waitlistKey(svc), boolDef(in.WaitlistMode), actor); err != nil {
+			return ServiceView{}, err
+		}
+	}
+	return ServiceView{ServiceRow: row, WaitlistMode: flags.Bool(waitlistKey(svc))}, nil
+}
+
+// seedRegistry seeds the registry and registers a waitlist.<svc> switch per known
+// service, COMPOSING the flag engine (flags.Register). Best-effort + fail-safe: a
+// registry error (e.g. cek master key not yet injected) degrades to the in-memory seed
+// switches — the decide then fail-opens, exactly the flag engine's own boot posture.
+// Returns the number of seeded services (for the mount log). Called from Mount.
+func seedRegistry(brand string, log luxlog.Logger) int {
+	seed := seedWaitlist(brand)
+	for _, sv := range seed { // in-memory switches — always succeeds
+		flags.Register(waitlistDef(sv.Service, sv.DisplayName))
+	}
+	st, err := mounted.store.For(platformOrg, platformProject)
+	if err != nil {
+		log.Warn("waitlist registry unavailable — modes degrade to seed defaults", "err", err)
+		return len(seed)
+	}
+	if _, err := st.Seed(context.Background(), seed, time.Now().Unix()); err != nil {
+		log.Warn("waitlist registry seed failed", "err", err)
+		return len(seed)
+	}
+	if rows, err := st.List(context.Background()); err == nil {
+		for _, r := range rows { // register any persisted onboard beyond the seed
+			ensureWaitlistDef(r.Service, r.DisplayName)
+		}
+	}
+	return len(seed)
+}
+
+// waitlistModeRoute answers GET /v1/flags/waitlist?host=<h> — the runtime lookup the
+// @file waitlist-guard caches. Public (in-cluster) read: it returns ONLY the boolean
+// mode for the ONE queried host, never an enumeration.
+func waitlistModeRoute(c *zip.Ctx) error {
+	host := strings.TrimSpace(c.Query("host"))
+	if host == "" {
+		host = c.Fiber().Hostname()
+	}
+	mode, service, known := WaitlistModeForHost(c.Context(), host)
+	return c.JSON(http.StatusOK, map[string]any{
+		"host":         NormalizeHost(host),
+		"service":      service,
+		"waitlistMode": mode,
+		"known":        known,
+	})
+}
+
+// ── lifecycle ────────────────────────────────────────────────────────────────
+
+// Mount installs the launch-control gate: it opens the platform-tenant host→service
+// registry, seeds it for the deployment brand, registers a waitlist.<svc> switch per
+// service in the flag engine (flags.Register), and serves the guard's public mode read
+// at /v1/flags/waitlist. Fail-safe: a registry error (e.g. cek master key not yet
+// injected) degrades to the in-memory seed switches — WaitlistModeForHost then
+// fail-opens. Mounts AFTER flags so the engine's platform-switch plane is installed first.
+func Mount(app *zip.App, deps cloud.Deps) error {
+	if deps.Logger == nil {
+		return fmt.Errorf("admission.Mount: nil deps.Logger")
+	}
+	if deps.DataDir == "" {
+		return fmt.Errorf("admission.Mount: empty deps.DataDir")
+	}
+	log := deps.Logger.New("subsystem", "admission")
+	mounted = &registryState{
+		store: cloud.NewOrgStore[*waitlistStore](deps.DataDir, "waitlist", openWaitlistStore),
+		brand: deps.Brand,
+	}
+	n := seedRegistry(deps.Brand, log)
+	// The guard's public runtime mode read (host→service→waitlist.<svc>), one namespace
+	// under /v1/flags. Exempt from the Enforce gate (see defaultExemptPrefixes) so a
+	// gated user can still resolve mode.
+	app.Get("/v1/flags/waitlist", waitlistModeRoute)
+	log.Info("admission gate ready", "services", n)
+	return nil
+}
+
+// Shutdown closes the launch registry's per-org store handles.
+func Shutdown() error {
+	if mounted == nil || mounted.store == nil {
+		return nil
+	}
+	return mounted.store.CloseAll()
+}
+
+// ── brand seed (moved verbatim from the former flags/waitlist.go) ────────────────
+
+// seedWaitlist returns the launch registry for a brand. White-labeled so a Lux/Zoo/Pars
+// deployment governs its OWN hosts. New hosted services onboard at runtime via
+// POST /v1/admin/services (no redeploy). admin.<brand> is deliberately NOT seeded (it
+// is admin-guarded, not a waitlist surface).
+func seedWaitlist(brand string) []SeedService {
+	d := domainFor(brand)
+	return []SeedService{
+		{Service: "studio", DisplayName: "Studio", Description: "AI app studio", Hosts: []string{"studio." + d}},
+		{Service: "chat", DisplayName: "Chat", Description: "AI chat", Hosts: hostsFor(brand, "chat", "chat."+d)},
+		{Service: "console", DisplayName: "Console", Description: "Cloud console", Hosts: []string{"console." + d}},
+		{Service: "app", DisplayName: "App", Description: "App builder", Hosts: hostsFor(brand, "app", "app."+d)},
+		{Service: "api", DisplayName: "API", Description: "Inference API gateway", Hosts: []string{"api." + d}},
+		{Service: "team", DisplayName: "Team", Description: "Team workspace", Hosts: hostsFor(brand, "team", "team."+d)},
+	}
+}
+
+// domainFor maps a brand to its primary domain. Defaults to hanzo.ai.
+func domainFor(brand string) string {
+	switch strings.ToLower(strings.TrimSpace(brand)) {
+	case "lux":
+		return "lux.network"
+	case "zoo":
+		return "zoo.ngo"
+	case "pars":
+		return "pars.network"
+	default:
+		return "hanzo.ai"
+	}
+}
+
+// hostsFor returns the apex-brand host (hanzo.chat / zoo.chat style) plus the
+// <label>.<domain> alias when the brand ships an apex-label domain; else just the alias.
+func hostsFor(brand, label, alias string) []string {
+	switch strings.ToLower(strings.TrimSpace(brand)) {
+	case "", "hanzo":
+		return []string{"hanzo." + label, alias}
+	case "zoo":
+		return []string{"zoo." + label, alias}
+	default:
+		return []string{alias}
+	}
+}

clients/affiliates/commerce.go

@@ -1,153 +1,44 @@
 package affiliates
 
 import (
-	"bytes"
 	"context"
-	"encoding/json"
-	"errors"
-	"fmt"
-	"io"
-	"net/http"
-	"net/url"
-	"strings"
-	"time"
 
-	"github.com/hanzoai/cloud/clients/commerceinproc"
+	"github.com/hanzoai/cloud/clients/payout"
 )
 
-// commerce is the narrow money seam the affiliate loop needs: read a referred
-// org's metered spend (the accrual base) and grant a promo credit to a wallet (a
-// payout made in credits). It is an INTERFACE so the store/handler logic is
-// testable with a fake ledger — the HTTP impl below is the ONE production binding.
+// commerce is the narrow money seam the affiliate loop needs: read a referred org's
+// metered spend (the commission accrual base) and grant a promo credit to a wallet
+// (a payout made in credits, ledger tag grant:affiliate). It is an INTERFACE so the
+// store/handler logic is testable with a fake ledger; the production binding is
+// clients/payout, reached through the thin adapter below.
 //
-// This mirrors clients/referrals/commerce.go EXACTLY (which itself mirrors
-// clients/admin/commerce.go): the same COMMERCE_SERVICE_TOKEN S2S path, the same
-// X-Org-Id=<org> namespace + bare org `user` subject that admin.grantCredit uses —
-// so an affiliate payout-in-credits lands in precisely the wallet the balance
-// panel reads, indistinguishable from an admin grant except by its ledger tag
-// (grant:affiliate vs grant:referral / grant:admin, all → the commerce Credit/trial
-// bucket per DepositKind's grant:* rule).
+// The S2S impl (COMMERCE_SERVICE_TOKEN path, X-Org-Id=<org> namespace, bare-org
+// `user` subject) was three byte-identical commerce.go copies; it now lives ONCE in
+// clients/payout. An affiliate payout-in-credits still lands in precisely the wallet
+// the balance panel reads, indistinguishable from an admin grant except by its
+// grant:affiliate tag.
 type commerce interface {
 	configured() bool
-	// deposit grants amountCents to org's wallet (Credit/trial bucket via the
-	// grant:affiliate tag) and returns the ledger transaction id.
 	deposit(ctx context.Context, org, user string, amountCents int64, currency, notes, tags string) (txnID string, err error)
-	// spendCents is a referred org's month-to-date metered consumption — the
-	// commission accrual base (spend × the affiliate's rate).
 	spendCents(ctx context.Context, org, user string) (int64, error)
 }
 
-// errUnconfigured is returned by a deposit against an unwired commerce so the
-// caller records an honest failure rather than reporting a phantom payout.
-var errUnconfigured = errors.New("affiliates: commerce endpoint not configured")
+// errUnconfigured is the shared sentinel a deposit against an unwired commerce
+// returns, so the caller records an honest failure rather than a phantom payout.
+var errUnconfigured = payout.ErrUnconfigured
 
-// httpCommerce is the production commerce binding (COMMERCE_SERVICE_TOKEN S2S).
-type httpCommerce struct {
-	base  string
-	token string
-	http  *http.Client
+// commerceSeam adapts the shared payout.Client onto this program's lowercase seam
+// (Go package-scoped interface methods cannot cross packages). Zero logic — pure
+// delegation; the money path lives in clients/payout.
+type commerceSeam struct{ c *payout.Client }
+
+func (s commerceSeam) configured() bool { return s.c.Configured() }
+func (s commerceSeam) deposit(ctx context.Context, org, user string, amountCents int64, currency, notes, tags string) (string, error) {
+	return s.c.Deposit(ctx, org, user, amountCents, currency, notes, tags)
+}
+func (s commerceSeam) spendCents(ctx context.Context, org, user string) (int64, error) {
+	return s.c.SpendCents(ctx, org, user)
 }
 
-func newCommerceClient(base, token string) *httpCommerce {
-	return &httpCommerce{
-		base:  strings.TrimRight(strings.TrimSpace(base), "/"),
-		token: strings.TrimSpace(token),
-		http:  commerceinproc.Client(15 * time.Second),
-	}
-}
-
-func (c *httpCommerce) configured() bool { return c != nil && c.base != "" && c.token != "" }
-
-// deposit posts POST /v1/billing/deposit — the ONE money-in primitive (identical
-// to admin.commerceClient.deposit). Commerce's EdgeAuth pins the body `user` to
-// the X-Org-Id subject, so a payout can never be mis-targeted to another wallet.
-func (c *httpCommerce) deposit(ctx context.Context, org, user string, amountCents int64, currency, notes, tags string) (string, error) {
-	if !c.configured() {
-		return "", errUnconfigured
-	}
-	if currency == "" {
-		currency = "usd"
-	}
-	body, err := json.Marshal(map[string]any{
-		"user":     user,
-		"currency": currency,
-		"amount":   amountCents,
-		"notes":    notes,
-		"tags":     tags,
-	})
-	if err != nil {
-		return "", err
-	}
-	raw, err := c.do(ctx, http.MethodPost, "/v1/billing/deposit", nil, org, body)
-	if err != nil {
-		return "", err
-	}
-	var out struct {
-		TransactionID string `json:"transactionId"`
-	}
-	if err := json.Unmarshal(raw, &out); err != nil {
-		return "", fmt.Errorf("commerce deposit decode: %w", err)
-	}
-	return out.TransactionID, nil
-}
-
-// spendCents reads GET /v1/billing/usage-rollup and returns consumedCents. Zero
-// (not an error) when commerce is unconfigured so a partial deploy degrades to
-// "no spend to accrue yet" rather than a 5xx.
-func (c *httpCommerce) spendCents(ctx context.Context, org, user string) (int64, error) {
-	if !c.configured() {
-		return 0, nil
-	}
-	q := url.Values{"user": {user}}
-	raw, err := c.do(ctx, http.MethodGet, "/v1/billing/usage-rollup", q, org, nil)
-	if err != nil {
-		return 0, err
-	}
-	var out struct {
-		ConsumedCents int64 `json:"consumedCents"`
-	}
-	if err := json.Unmarshal(raw, &out); err != nil {
-		return 0, fmt.Errorf("commerce rollup decode: %w", err)
-	}
-	return out.ConsumedCents, nil
-}
-
-// do performs one admin-S2S commerce request. X-Org-Id=<org> is the per-org
-// namespace selector commerce's EdgeAuth trusts only behind the service token.
-func (c *httpCommerce) do(ctx context.Context, method, path string, q url.Values, org string, body []byte) ([]byte, error) {
-	u := c.base + path
-	if enc := q.Encode(); enc != "" {
-		u += "?" + enc
-	}
-	var r io.Reader
-	if body != nil {
-		r = bytes.NewReader(body)
-	}
-	req, err := http.NewRequestWithContext(ctx, method, u, r)
-	if err != nil {
-		return nil, err
-	}
-	req.Header.Set("Accept", "application/json")
-	if body != nil {
-		req.Header.Set("Content-Type", "application/json")
-	}
-	if c.token != "" {
-		req.Header.Set("Authorization", "Bearer "+c.token)
-	}
-	if org != "" {
-		req.Header.Set("X-Org-Id", org)
-	}
-	resp, err := c.http.Do(req)
-	if err != nil {
-		return nil, fmt.Errorf("commerce unreachable: %w", err)
-	}
-	defer func() { _ = resp.Body.Close() }()
-	out, err := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
-	if err != nil {
-		return nil, err
-	}
-	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		return nil, fmt.Errorf("commerce status %d", resp.StatusCode)
-	}
-	return out, nil
-}
+// newCommerceClient builds the production binding, delegating to clients/payout.
+func newCommerceClient(base, token string) commerce { return commerceSeam{payout.NewClient(base, token)} }

clients/agent/agent.go

@@ -0,0 +1,141 @@
+// Package agent mounts the hanzoai/agent orchestrator into cloud: POST /v1/agent
+// (+ /v1/agent/presets, /v1/agent/conversations). The orchestrator logic and its
+// per-org conversation history live in github.com/hanzoai/agent, which imports
+// NEITHER cloud NOR ai. Cloud is the composition root: it injects the two seams —
+//   - Completer: the ai subsystem's /v1/chat/completions, replayed in-process (the
+//     one path that returns tool_calls AND carries per-org reserve/settle billing);
+//   - ToolPlane: the unified tool registry (tools.Default()), so /v1/agent's
+//     server-executed tools are the org's activated MCP/registry tools.
+// /v1/agent is a DISTINCT path (not /v1/chat, which ai owns as completions), so a
+// specific route wins over ai's /v1/* glob — no collision.
+package agent
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+
+	hz "github.com/hanzoai/agent"
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/tools"
+	openai "github.com/hanzoai/go-openai"
+	fiber "github.com/zap-proto/fiber/v3"
+	"github.com/zap-proto/zip"
+)
+
+// maxCompletionResponse bounds the in-process completion body read so a hostile or
+// broken upstream cannot balloon memory.
+const maxCompletionResponse = 8 << 20
+
+// Mount wires POST /v1/agent (+ reads) into cloud, injecting the ai completion and
+// the tool plane. The caller identity comes from cloud's validated principal.
+func Mount(app *zip.App, deps cloud.Deps) error {
+	if app == nil {
+		return fmt.Errorf("agent.Mount: nil zip.App")
+	}
+	_, err := hz.Mount(app, hz.Deps{
+		Logger:  deps.Logger,
+		DataDir: deps.DataDir,
+		Brand:   deps.Brand,
+		Model:   deps.AIDefaultModel,
+		Principal: func(c *zip.Ctx) (hz.Principal, bool) {
+			p, ok := tools.PrincipalFrom(c)
+			if !ok {
+				return hz.Principal{}, false
+			}
+			return hz.Principal{Org: p.Org, Project: p.Project, User: p.User, Cred: credential(c)}, true
+		},
+	}, aiCompleter{app: app}, toolPlane{})
+	return err
+}
+
+// ── Completer: replay /v1/chat/completions in-process ─────────────────────────────
+
+type aiCompleter struct{ app *zip.App }
+
+// Complete replays the request against the SAME app at /v1/chat/completions, so it
+// flows the whole middleware chain (per-org reserve/settle billing) and returns
+// tool_calls. Non-streaming. Mirrors the tool plane's in-process dispatch contract:
+// the caller's OWN credential headers are replayed; no minted authority header.
+func (a aiCompleter) Complete(ctx context.Context, cred map[string]string, req openai.ChatCompletionRequest) (openai.ChatCompletionResponse, error) {
+	req.Stream = false
+	b, err := json.Marshal(req)
+	if err != nil {
+		return openai.ChatCompletionResponse{}, err
+	}
+	hreq := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", bytes.NewReader(b)).WithContext(ctx)
+	hreq.Header.Set("Content-Type", "application/json")
+	for k, v := range cred {
+		hreq.Header.Set(k, v)
+	}
+	resp, err := a.app.Fiber().Test(hreq, fiber.TestConfig{Timeout: 0})
+	if err != nil {
+		return openai.ChatCompletionResponse{}, err
+	}
+	defer func() { _ = resp.Body.Close() }()
+	raw, _ := io.ReadAll(io.LimitReader(resp.Body, maxCompletionResponse))
+	if resp.StatusCode/100 != 2 {
+		// Carry the completion's OWN status + body so the round can pass a
+		// caller-facing refusal (402 insufficient_balance, 429, 403) straight
+		// through instead of masking it as a gateway 502. hz.UpstreamError is the
+		// agent's typed seam for exactly this.
+		return openai.ChatCompletionResponse{}, &hz.UpstreamError{Status: resp.StatusCode, Body: raw}
+	}
+	var out openai.ChatCompletionResponse
+	if err := json.Unmarshal(raw, &out); err != nil {
+		return openai.ChatCompletionResponse{}, fmt.Errorf("decode completion: %w", err)
+	}
+	return out, nil
+}
+
+// ── ToolPlane: adapter over the unified registry ──────────────────────────────────
+
+type toolPlane struct{}
+
+func (toolPlane) List(ctx context.Context, scope hz.Scope) []hz.Tool {
+	ts := tools.Default().List(ctx, tools.Scope{Org: scope.Org, Project: scope.Project})
+	out := make([]hz.Tool, 0, len(ts))
+	for _, t := range ts {
+		out = append(out, hz.Tool{
+			Name:         t.Name,
+			Description:  t.Description,
+			Schema:       t.Schema,
+			Activated:    t.Activated,
+			Dispatchable: t.Dispatchable,
+		})
+	}
+	return out
+}
+
+func (toolPlane) Exists(ctx context.Context, scope hz.Scope, name string) bool {
+	return tools.Default().Exists(ctx, tools.Scope{Org: scope.Org, Project: scope.Project}, name)
+}
+
+// Dispatch resolves the caller the ONE canonical way — tools.PrincipalFrom(c) — so
+// the tool runs under the SAME validated identity + credential as a direct call.
+// No reconstruction: the credential is only ever read from the live request.
+func (toolPlane) Dispatch(c *zip.Ctx, name string, args map[string]any) (any, error) {
+	p, ok := tools.PrincipalFrom(c)
+	if !ok {
+		return nil, zip.ErrForbidden("a validated principal is required")
+	}
+	return tools.Default().Dispatch(c.Context(), p, name, args)
+}
+
+// ── helpers ───────────────────────────────────────────────────────────────────────
+
+// credential extracts the caller's replayable credential headers (the same set the
+// tool plane replays) so the in-process completion runs as the caller.
+func credential(c *zip.Ctx) map[string]string {
+	cred := map[string]string{}
+	for _, h := range []string{"Authorization", "X-Authorization", "Cookie", "Accept-Language", "X-Forwarded-For"} {
+		if v := c.Header(h); v != "" {
+			cred[h] = v
+		}
+	}
+	return cred
+}

clients/agents/inproc.go

@@ -71,6 +71,44 @@ func OpenSession(ctx context.Context, org, actor, agent, title string) (string,
 	return id, nil
 }
 
+// OpenSessionOn is OpenSession with the run's dispatch TARGET recorded, so
+// mission-control shows a routed run on the machine it was sent to (session.target
+// == the target id) exactly as a locally-linked run shows its host. The target is
+// re-resolved org-scoped and MUST belong to this org — a session can never claim
+// to run on another tenant's machine (the same fail-closed rule sessionContext
+// enforces on the HTTP register path). An empty target falls back to OpenSession.
+func OpenSessionOn(ctx context.Context, org, actor, agent, title, target string) (string, error) {
+	target = strings.TrimSpace(target)
+	if target == "" {
+		return OpenSession(ctx, org, actor, agent, title)
+	}
+	if mounted == nil {
+		return "", fmt.Errorf("agents: not mounted")
+	}
+	org = strings.TrimSpace(org)
+	if org == "" {
+		return "", fmt.Errorf("agents: org required")
+	}
+	if _, err := mounted.State.store.GetTarget(ctx, org, target); err != nil {
+		if err == errTargetNotFound {
+			return "", fmt.Errorf("agents: target not found in this org")
+		}
+		return "", fmt.Errorf("agents: resolve target: %w", err)
+	}
+	id, err := OpenSession(ctx, org, actor, agent, title)
+	if err != nil {
+		return "", err
+	}
+	// Stamp the target onto the freshly-opened row (org-scoped update); a failure
+	// here is non-fatal — the session is live, it simply lacks its machine tag.
+	if x, gerr := mounted.State.store.GetSession(ctx, org, id); gerr == nil {
+		x.Target = target
+		x.UpdatedAt = time.Now().Unix()
+		_ = mounted.State.store.UpdateSession(ctx, x)
+	}
+	return id, nil
+}
+
 // LogSessionEvent appends one ordered event (message|tool-call|spawn|log|status|
 // control) to an org's session and fans it out live. The (org, id) pair is
 // re-resolved so a caller can only write to a session THIS org owns; kind is

clients/agents/inproc_test.go

@@ -114,3 +114,50 @@ func TestInproc_NotMounted_FailsClosed(t *testing.T) {
 		t.Fatal("unmounted OpenSession must fail closed")
 	}
 }
+
+// ResolveTarget turns a human's reference (id or friendly label) into the org's
+// target, org-scoped and fail-closed: an id wins, else an exact case-folded label,
+// and a reference matching neither — or another org's machine — is not found.
+func TestResolveTarget_IdThenLabel_OrgScoped(t *testing.T) {
+	mountInproc(t)
+	ctx := context.Background()
+	now := int64(1000)
+	acme := Target{ID: "tgt_acme1", Org: "acme", Label: "evo", Kind: TargetGPU, Status: TargetOnline, Host: "evo", CreatedAt: now, UpdatedAt: now}
+	evil := Target{ID: "tgt_evil1", Org: "evil", Label: "evo", Kind: TargetGPU, Status: TargetOnline, Host: "evo", CreatedAt: now, UpdatedAt: now}
+	if err := mounted.State.store.CreateTarget(ctx, acme); err != nil {
+		t.Fatal(err)
+	}
+	if err := mounted.State.store.CreateTarget(ctx, evil); err != nil {
+		t.Fatal(err)
+	}
+
+	// By id.
+	if got, err := ResolveTarget(ctx, "acme", "tgt_acme1"); err != nil || got.ID != "tgt_acme1" {
+		t.Fatalf("resolve by id: %+v %v", got, err)
+	}
+	// By label (case-folded), scoped to the caller's org — never evil's same-labelled box.
+	if got, err := ResolveTarget(ctx, "acme", "EVO"); err != nil || got.ID != "tgt_acme1" {
+		t.Fatalf("resolve by label must find acme's own, got %+v %v", got, err)
+	}
+	// Another org's id is not found (no cross-tenant leak).
+	if _, err := ResolveTarget(ctx, "acme", "tgt_evil1"); err != errTargetNotFound {
+		t.Fatalf("cross-org id must be not-found, got %v", err)
+	}
+	// An unknown reference is not found — the caller renders an honest error.
+	if _, err := ResolveTarget(ctx, "acme", "nope"); err != errTargetNotFound {
+		t.Fatalf("unknown ref must be not-found, got %v", err)
+	}
+	// Empty ref is not found (never resolves to "some" machine).
+	if _, err := ResolveTarget(ctx, "acme", ""); err != errTargetNotFound {
+		t.Fatalf("empty ref must be not-found, got %v", err)
+	}
+}
+
+func TestResolveTarget_NotMounted_FailsClosed(t *testing.T) {
+	prev := mounted
+	mounted = nil
+	t.Cleanup(func() { mounted = prev })
+	if _, err := ResolveTarget(context.Background(), "acme", "evo"); err == nil {
+		t.Fatal("unmounted ResolveTarget must fail closed")
+	}
+}

clients/agents/mailbox.go

@@ -0,0 +1,236 @@
+package agents
+
+import (
+	"context"
+	"sync"
+)
+
+// mailbox.go is the LIVE hand-off between a routed run's durable owner (the
+// coding RoutedRunWorkflow, running on the embedded tasks engine) and the
+// external machine that claims and executes it over HTTP. It is the rendezvous
+// ONLY — never the durable queue. The tasks engine is the queue of record: it
+// survives a cloud restart, times a never-claimed run out, and retries. On every
+// (re)start of the delivery activity the run is (re-)Offered here, so a machine
+// that long-polls Claim always finds work the engine still owns; a cloud restart
+// simply re-populates the mailbox from durable history.
+//
+// ISOLATION IS STRUCTURAL. Every offer is filed under the key (org, target), and
+// Claim/Report only ever touch that one key's slot. A run offered for (orgB,
+// targetY) is unreachable from a Claim or Report for (orgA, targetX) — the tenant
+// + machine boundary is a property of the map key, not a check a caller can skip.
+
+// RoutedRun is the NON-SECRET spec of one coding run dispatched to a target. It
+// carries no credential by design: the executing machine authenticates git +
+// model routing with its OWN already-held credentials (the same ones `hanzo code`
+// uses), so no secret ever enters the durable store or crosses to the machine in
+// the claim response. Everything here is safe to persist in the tasks engine.
+type RoutedRun struct {
+	Org            string `json:"org"`
+	TargetID       string `json:"targetId"`
+	SessionID      string `json:"sessionId"` // the live session opened at dispatch; the machine streams into it
+	Repo           string `json:"repo"`
+	Project        string `json:"project,omitempty"`
+	Base           string `json:"base,omitempty"`
+	Branch         string `json:"branch"`
+	Prompt         string `json:"prompt"`
+	CloneURL       string `json:"cloneUrl"`
+	TimeoutSeconds int    `json:"timeoutSeconds,omitempty"`
+}
+
+// RoutedResult is a routed run's terminal outcome, reported by the machine and
+// returned to the durable activity so the workflow completes.
+type RoutedResult struct {
+	OK        bool   `json:"ok"`
+	Changed   bool   `json:"changed"`
+	Branch    string `json:"branch,omitempty"`
+	CommitSha string `json:"commitSha,omitempty"`
+	Diffstat  string `json:"diffstat,omitempty"`
+	Error     string `json:"error,omitempty"`
+}
+
+// offer is one run waiting to be claimed, plus the channel its durable owner
+// blocks on for the terminal result. result is buffered(1) so Report never blocks
+// even if the owner is between selects; closed fires when the offer is finished
+// (reported OR abandoned) so a waiter always unblocks.
+type offer struct {
+	mb     *mailbox
+	key    string // (org,target)
+	rk     string // (org,target,sessionID)
+	run    RoutedRun
+	result chan RoutedResult
+	closed chan struct{}
+	once   sync.Once
+}
+
+// Await blocks until the machine reports this run's result, the offer is
+// abandoned, or ctx (the activity's StartToClose budget) fires. It is the
+// durable owner's half of the rendezvous.
+func (o *offer) Await(ctx context.Context) (RoutedResult, bool) {
+	select {
+	case res := <-o.result:
+		return res, true
+	case <-o.closed:
+		// Abandoned or reported-then-closed: drain a delivered result if one raced in.
+		select {
+		case res := <-o.result:
+			return res, true
+		default:
+			return RoutedResult{}, false
+		}
+	case <-ctx.Done():
+		return RoutedResult{}, false
+	}
+}
+
+// Close removes the offer from the mailbox (if still present) and unblocks any
+// waiter. Idempotent — the durable owner defers it so a timed-out or crashed
+// delivery never leaks a queued or claimed offer.
+func (o *offer) Close() { o.mb.discard(o) }
+
+// mailbox is the process-wide rendezvous. queues holds each key's FIFO of
+// unclaimed offers; byRun indexes every live offer by (org,target,sessionID) for
+// Report + re-offer dedupe; signal is a per-key broadcast channel (closed and
+// recreated on Offer) that Claim waits on.
+type mailbox struct {
+	mu     sync.Mutex
+	queues map[string][]*offer
+	byRun  map[string]*offer
+	signal map[string]chan struct{}
+}
+
+func newMailbox() *mailbox {
+	return &mailbox{
+		queues: map[string][]*offer{},
+		byRun:  map[string]*offer{},
+		signal: map[string]chan struct{}{},
+	}
+}
+
+// routedMailbox is the ONE process-wide rendezvous, shared by the coding
+// delivery activity (Offer/Await) and the machine-facing HTTP surface
+// (Claim/Report). One mailbox, one way.
+var routedMailbox = newMailbox()
+
+func mbKey(org, target string) string      { return org + "\x00" + target }
+func runKey(org, target, sess string) string { return org + "\x00" + target + "\x00" + sess }
+
+// Offer files run for its (org,target) and returns the handle its durable owner
+// awaits. A re-offer of the same (org,target,sessionID) — the workflow retrying
+// or replaying after a restart — supersedes the stale prior offer (removing it
+// from the queue and unblocking its dead waiter) so a machine never claims a run
+// whose owner has already moved on.
+func (m *mailbox) Offer(run RoutedRun) *offer {
+	key := mbKey(run.Org, run.TargetID)
+	rk := runKey(run.Org, run.TargetID, run.SessionID)
+	o := &offer{mb: m, key: key, rk: rk, run: run, result: make(chan RoutedResult, 1), closed: make(chan struct{})}
+	m.mu.Lock()
+	if prev := m.byRun[rk]; prev != nil {
+		m.removeFromQueueLocked(key, prev)
+		prev.finish()
+	}
+	m.byRun[rk] = o
+	m.queues[key] = append(m.queues[key], o)
+	m.broadcastLocked(key)
+	m.mu.Unlock()
+	return o
+}
+
+// Claim blocks until an unclaimed run exists for (org,target) or ctx fires,
+// returning the oldest. The claimed offer leaves the queue but stays in byRun,
+// awaiting Report. Only this key's queue is ever read, so a claim can never
+// surface another tenant's or another machine's run.
+func (m *mailbox) Claim(ctx context.Context, org, target string) (RoutedRun, bool) {
+	key := mbKey(org, target)
+	for {
+		m.mu.Lock()
+		if q := m.queues[key]; len(q) > 0 {
+			o := q[0]
+			m.queues[key] = q[1:]
+			m.mu.Unlock()
+			return o.run, true
+		}
+		sig := m.signalLocked(key)
+		m.mu.Unlock()
+		select {
+		case <-sig:
+			// a new offer (or a superseding one) arrived — re-check
+		case <-ctx.Done():
+			return RoutedRun{}, false
+		}
+	}
+}
+
+// Report delivers a terminal result to the run's durable owner. Scoped to
+// (org,target,sessionID): a report can only ever complete a run that exact key
+// owns, so one machine can never report on behalf of another. Returns false when
+// no live offer matches (already reported, abandoned, or never existed).
+func (m *mailbox) Report(org, target, sess string, res RoutedResult) bool {
+	rk := runKey(org, target, sess)
+	m.mu.Lock()
+	o := m.byRun[rk]
+	if o == nil {
+		m.mu.Unlock()
+		return false
+	}
+	delete(m.byRun, rk)
+	m.removeFromQueueLocked(mbKey(org, target), o)
+	m.mu.Unlock()
+	o.deliver(res)
+	return true
+}
+
+// discard drops an offer the owner is done with (ctx timeout / crash / normal
+// close) so neither the queue nor byRun retains it.
+func (m *mailbox) discard(o *offer) {
+	m.mu.Lock()
+	if m.byRun[o.rk] == o {
+		delete(m.byRun, o.rk)
+	}
+	m.removeFromQueueLocked(o.key, o)
+	m.mu.Unlock()
+	o.finish()
+}
+
+func (m *mailbox) removeFromQueueLocked(key string, o *offer) {
+	q := m.queues[key]
+	for i, e := range q {
+		if e == o {
+			m.queues[key] = append(q[:i:i], q[i+1:]...)
+			return
+		}
+	}
+}
+
+// broadcastLocked wakes every Claim waiting on key by closing its signal channel;
+// a fresh channel replaces it for the next wait.
+func (m *mailbox) broadcastLocked(key string) {
+	if ch, ok := m.signal[key]; ok {
+		close(ch)
+		delete(m.signal, key)
+	}
+}
+
+func (m *mailbox) signalLocked(key string) chan struct{} {
+	ch, ok := m.signal[key]
+	if !ok {
+		ch = make(chan struct{})
+		m.signal[key] = ch
+	}
+	return ch
+}
+
+func (o *offer) deliver(res RoutedResult) {
+	o.once.Do(func() {
+		o.result <- res // buffered(1) — never blocks
+		close(o.closed)
+	})
+}
+
+func (o *offer) finish() {
+	o.once.Do(func() { close(o.closed) })
+}
+
+// OfferRoutedRun is the exported seam the coding delivery activity uses to place
+// a run into the live rendezvous. Kept here (agents owns targets + sessions) so
+// the machine-facing HTTP surface and the durable activity share ONE mailbox.
+func OfferRoutedRun(run RoutedRun) *offer { return routedMailbox.Offer(run) }

clients/agents/mailbox_test.go

@@ -0,0 +1,166 @@
+package agents
+
+import (
+	"context"
+	"sync"
+	"testing"
+	"time"
+)
+
+func mkRun(org, target, sess string) RoutedRun {
+	return RoutedRun{Org: org, TargetID: target, SessionID: sess, Repo: "api", Branch: "agent/" + sess}
+}
+
+// A claimed run comes back to exactly one claimer, then its report reaches the
+// offerer that is awaiting it.
+func TestMailbox_OfferClaimReport(t *testing.T) {
+	m := newMailbox()
+	off := m.Offer(mkRun("acme", "tgt_1", "sess_1"))
+
+	got, ok := m.Claim(context.Background(), "acme", "tgt_1")
+	if !ok || got.SessionID != "sess_1" {
+		t.Fatalf("claim wrong: ok=%v run=%+v", ok, got)
+	}
+
+	done := make(chan RoutedResult, 1)
+	go func() {
+		res, _ := off.Await(context.Background())
+		done <- res
+	}()
+	if !m.Report("acme", "tgt_1", "sess_1", RoutedResult{OK: true, CommitSha: "abc"}) {
+		t.Fatal("report should deliver to the awaiting offer")
+	}
+	select {
+	case res := <-done:
+		if !res.OK || res.CommitSha != "abc" {
+			t.Fatalf("await got wrong result: %+v", res)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("await never received the reported result")
+	}
+}
+
+// THE tenant + machine boundary: a claim for (org,target) can NEVER surface a run
+// offered for a different org OR a different target — it is a property of the key.
+func TestMailbox_CrossTenantAndCrossMachineIsolation(t *testing.T) {
+	m := newMailbox()
+	m.Offer(mkRun("orgB", "tgt_Y", "sess_foreign_org"))
+	m.Offer(mkRun("acme", "tgt_Y", "sess_foreign_machine"))
+	m.Offer(mkRun("acme", "tgt_X", "sess_mine"))
+
+	// A claim for (acme, tgt_X) gets ONLY acme/tgt_X's run.
+	got, ok := m.Claim(context.Background(), "acme", "tgt_X")
+	if !ok || got.SessionID != "sess_mine" {
+		t.Fatalf("claim leaked across a boundary: ok=%v run=%+v", ok, got)
+	}
+	// And that queue is now empty — no foreign run fell through.
+	ctx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	defer cancel()
+	if _, ok := m.Claim(ctx, "acme", "tgt_X"); ok {
+		t.Fatal("a foreign run must never be claimable as acme/tgt_X")
+	}
+
+	// A Report can only complete a run under its exact key: reporting the foreign
+	// machine's session under tgt_X does nothing.
+	if m.Report("acme", "tgt_X", "sess_foreign_machine", RoutedResult{OK: true}) {
+		t.Fatal("report crossed the machine boundary")
+	}
+	if m.Report("acme", "tgt_Y", "sess_foreign_org", RoutedResult{OK: true}) {
+		t.Fatal("report crossed the org boundary")
+	}
+}
+
+// Two racing claimers, one run: exactly one wins.
+func TestMailbox_NoDoubleClaim(t *testing.T) {
+	m := newMailbox()
+	m.Offer(mkRun("acme", "tgt_1", "sess_1"))
+
+	var wins int
+	var mu sync.Mutex
+	var wg sync.WaitGroup
+	for i := 0; i < 8; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			ctx, cancel := context.WithTimeout(context.Background(), 200*time.Millisecond)
+			defer cancel()
+			if _, ok := m.Claim(ctx, "acme", "tgt_1"); ok {
+				mu.Lock()
+				wins++
+				mu.Unlock()
+			}
+		}()
+	}
+	wg.Wait()
+	if wins != 1 {
+		t.Fatalf("exactly one claimer must win, got %d", wins)
+	}
+}
+
+// A claim with no work times out on ctx and reports no run — fail closed, never hang.
+func TestMailbox_ClaimTimesOut(t *testing.T) {
+	m := newMailbox()
+	ctx, cancel := context.WithTimeout(context.Background(), 80*time.Millisecond)
+	defer cancel()
+	if _, ok := m.Claim(ctx, "acme", "tgt_empty"); ok {
+		t.Fatal("an empty mailbox must not yield a run")
+	}
+}
+
+// The durable owner's Await unblocks (fail-closed) when its budget ctx fires with
+// no report — the machine never claimed, or claimed and died.
+func TestMailbox_AwaitFailsClosedOnDeadline(t *testing.T) {
+	m := newMailbox()
+	off := m.Offer(mkRun("acme", "tgt_1", "sess_1"))
+	ctx, cancel := context.WithTimeout(context.Background(), 80*time.Millisecond)
+	defer cancel()
+	if _, ok := off.Await(ctx); ok {
+		t.Fatal("await must fail closed when the deadline fires without a report")
+	}
+}
+
+// A re-offer of the same run (workflow retry / cloud restart) supersedes the stale
+// offer: the old waiter unblocks abandoned, and the fresh run is claimable.
+func TestMailbox_ReOfferSupersedes(t *testing.T) {
+	m := newMailbox()
+	old := m.Offer(mkRun("acme", "tgt_1", "sess_1"))
+
+	// re-offer BEFORE anyone claims the first
+	fresh := m.Offer(mkRun("acme", "tgt_1", "sess_1"))
+
+	// old is abandoned
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+	defer cancel()
+	if _, ok := old.Await(ctx); ok {
+		t.Fatal("the superseded offer must not complete")
+	}
+
+	// exactly one claimable run remains, and reporting reaches the fresh offer
+	got, ok := m.Claim(context.Background(), "acme", "tgt_1")
+	if !ok || got.SessionID != "sess_1" {
+		t.Fatalf("fresh run not claimable: %+v", got)
+	}
+	if _, ok := m.Claim(ctxShort(), "acme", "tgt_1"); ok {
+		t.Fatal("the stale offer must not have left a duplicate in the queue")
+	}
+	done := make(chan struct{})
+	go func() { fresh.Await(context.Background()); close(done) }()
+	if !m.Report("acme", "tgt_1", "sess_1", RoutedResult{OK: true}) {
+		t.Fatal("report must reach the fresh offer")
+	}
+	<-done
+}
+
+// Report for an unknown/already-finished run is a clean false.
+func TestMailbox_ReportUnknownIsNoOp(t *testing.T) {
+	m := newMailbox()
+	if m.Report("acme", "tgt_1", "nope", RoutedResult{OK: true}) {
+		t.Fatal("report for an unknown run must be a no-op")
+	}
+}
+
+func ctxShort() context.Context {
+	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Millisecond)
+	_ = cancel
+	return ctx
+}

clients/agents/routing.go

@@ -0,0 +1,179 @@
+package agents
+
+import (
+	"context"
+	"crypto/rand"
+	"crypto/sha256"
+	"crypto/subtle"
+	"database/sql"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+)
+
+// routing.go is the machine-identity + liveness plane for routed runs (#48 half
+// B). A run dispatched to a target is executed by an EXTERNAL machine (`hanzo
+// code --serve`) that claims it over HTTP. Two properties make that safe:
+//
+//   - MACHINE IDENTITY. A target carries a claim key — a high-entropy capability
+//     minted server-side, returned to the daemon ONCE, and stored only as a
+//     SHA-256 hash (never plaintext, like every other secret). A claim/report
+//     must present the key; cloud verifies it in constant time, scoped to
+//     (org, target). Possession of the key IS being that machine, so one machine
+//     can never claim another's runs even within the same org.
+//
+//   - LIVENESS. Dispatch routes ONLY to a target with a live runner. A serve
+//     daemon proves liveness by polling Claim, which stamps serving_at; the gate
+//     rejects a target whose last poll is older than servingTTL. A run to a dead
+//     or absent runner fails closed at dispatch (never silently runs elsewhere),
+//     and one that dies mid-flight is re-queued/timed-out by the durable owner.
+//
+// The claim key lives in its own table so the live (A) target CRUD is untouched.
+
+const (
+	// servingTTL bounds how stale a target's last claim poll may be and still be
+	// considered "a runner is listening". The serve daemon re-polls right after a
+	// 25s long-poll returns empty, so a healthy runner stamps well inside this.
+	servingTTL = 90 * time.Second
+
+	claimKeyPrefix = "tgtk_"
+	claimKeyBytes  = 32 // 256-bit capability
+	maxClaimKey    = 128
+)
+
+var (
+	errNoClaimKey     = errors.New("agents: target has no claim key")
+	errClaimKeyBad    = errors.New("agents: claim key mismatch")
+	errTargetNotLive  = errors.New("agents: target has no live runner")
+	errTargetNotReady = errors.New("agents: target is not online")
+)
+
+// migrateClaimKeys creates the per-target claim-key + serving-liveness table in
+// the SAME agents.db (one store, one tenancy column). Idempotent.
+func (s *Store) migrateClaimKeys() error {
+	const ddl = `
+CREATE TABLE IF NOT EXISTS agent_target_claim_keys (
+  org        TEXT NOT NULL,
+  target_id  TEXT NOT NULL,
+  key_hash   TEXT NOT NULL,
+  serving_at INTEGER NOT NULL DEFAULT 0,
+  updated_at INTEGER NOT NULL,
+  PRIMARY KEY (org, target_id)
+);`
+	if _, err := s.db.Exec(ddl); err != nil {
+		return fmt.Errorf("migrate claim keys: %w", err)
+	}
+	return nil
+}
+
+// UpsertClaimKeyHash stores (or rotates) a target's claim-key hash. serving_at is
+// reset to 0 on a fresh mint — the daemon proves liveness by its first poll.
+func (s *Store) UpsertClaimKeyHash(ctx context.Context, org, targetID, hash string, now int64) error {
+	_, err := s.db.ExecContext(ctx,
+		`INSERT INTO agent_target_claim_keys (org,target_id,key_hash,serving_at,updated_at)
+		 VALUES (?,?,?,0,?)
+		 ON CONFLICT(org,target_id) DO UPDATE SET key_hash=excluded.key_hash, serving_at=0, updated_at=excluded.updated_at`,
+		org, targetID, hash, now)
+	if err != nil {
+		return fmt.Errorf("upsert claim key: %w", err)
+	}
+	return nil
+}
+
+// ClaimKeyHash returns a target's stored hash + last serving stamp, or
+// errNoClaimKey when none was ever minted.
+func (s *Store) ClaimKeyHash(ctx context.Context, org, targetID string) (hash string, servingAt int64, err error) {
+	row := s.db.QueryRowContext(ctx,
+		`SELECT key_hash, serving_at FROM agent_target_claim_keys WHERE org=? AND target_id=?`, org, targetID)
+	err = row.Scan(&hash, &servingAt)
+	if errors.Is(err, sql.ErrNoRows) {
+		return "", 0, errNoClaimKey
+	}
+	if err != nil {
+		return "", 0, fmt.Errorf("get claim key: %w", err)
+	}
+	return hash, servingAt, nil
+}
+
+// StampServing records that a target's runner polled at now (its liveness
+// heartbeat). Best-effort by the caller; a missing row is a no-op.
+func (s *Store) StampServing(ctx context.Context, org, targetID string, now int64) error {
+	_, err := s.db.ExecContext(ctx,
+		`UPDATE agent_target_claim_keys SET serving_at=? WHERE org=? AND target_id=?`, now, org, targetID)
+	return err
+}
+
+// hashClaimKey is the at-rest form: SHA-256 hex of a high-entropy token. A random
+// 256-bit key needs no password KDF; SHA-256 gives a fixed-size, constant-time-
+// comparable digest and the plaintext is never stored.
+func hashClaimKey(key string) string {
+	sum := sha256.Sum256([]byte(key))
+	return hex.EncodeToString(sum[:])
+}
+
+// newClaimKey mints a fresh capability token.
+func newClaimKey() (string, error) {
+	b := make([]byte, claimKeyBytes)
+	if _, err := rand.Read(b); err != nil {
+		return "", err
+	}
+	return claimKeyPrefix + hex.EncodeToString(b), nil
+}
+
+// verifyClaimKey checks a presented key against the target's stored hash in
+// constant time. Fail-closed: no key on file, or an empty presented key, is a
+// mismatch — never an accidental pass.
+func (s *Store) verifyClaimKey(ctx context.Context, org, targetID, presented string) error {
+	presented = strings.TrimSpace(presented)
+	if presented == "" || len(presented) > maxClaimKey {
+		return errClaimKeyBad
+	}
+	stored, _, err := s.ClaimKeyHash(ctx, org, targetID)
+	if err != nil {
+		return err // errNoClaimKey or a real DB error
+	}
+	if subtle.ConstantTimeCompare([]byte(stored), []byte(hashClaimKey(presented))) != 1 {
+		return errClaimKeyBad
+	}
+	return nil
+}
+
+// TargetDispatchable is the DRY liveness gate, used at dispatch (fail closed
+// before enqueue) AND re-checked at claim. A run is dispatchable only to a target
+// that (a) exists in this org, (b) is online, and (c) has a live runner — a claim
+// poll within servingTTL. Any failure is an explicit error the dispatcher renders
+// honestly; it NEVER falls back to running elsewhere.
+func (s *Store) TargetDispatchable(ctx context.Context, org, targetID string) error {
+	t, err := s.GetTarget(ctx, org, targetID)
+	if err != nil {
+		return err // errTargetNotFound or a real DB error
+	}
+	if t.Status != TargetOnline {
+		return errTargetNotReady
+	}
+	_, servingAt, err := s.ClaimKeyHash(ctx, org, targetID)
+	if err != nil {
+		return errTargetNotLive // no claim key => no runner ever attached
+	}
+	if servingAt <= 0 || time.Now().Unix()-servingAt > int64(servingTTL/time.Second) {
+		return errTargetNotLive
+	}
+	return nil
+}
+
+// TargetDispatchable is the exported gate the coding dispatcher injects (it never
+// imports the store directly). Returns nil when a run may be routed to (org,
+// targetID), else a descriptive error.
+func TargetDispatchable(ctx context.Context, org, targetID string) error {
+	if mounted == nil {
+		return fmt.Errorf("agents: not mounted")
+	}
+	org = strings.TrimSpace(org)
+	targetID = strings.TrimSpace(targetID)
+	if org == "" || targetID == "" {
+		return fmt.Errorf("agents: org and target required")
+	}
+	return mounted.State.store.TargetDispatchable(ctx, org, targetID)
+}

clients/agents/routing_http.go

@@ -0,0 +1,163 @@
+package agents
+
+import (
+	"context"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/hanzoai/cloud"
+	"github.com/zap-proto/zip"
+)
+
+// routing_http.go is the machine-facing surface a `hanzo code --serve` daemon
+// uses to CLAIM and complete routed runs. Every route is BOTH org-scoped (the
+// gateway-minted X-Org-Id, exactly like the rest of the targets plane) AND
+// machine-authenticated (the target claim key in X-Target-Key): a caller must
+// prove it is acting in the target's org and that it holds that specific
+// machine's capability. A run offered to target X is never reachable from a claim
+// for target Y, and a claim for another org's target 404s at the org boundary.
+//
+//	POST /v1/agents/targets/:id/claim-key            mint/rotate this target's claim key -> {claimKey}
+//	POST /v1/agents/targets/:id/claim                long-poll for the next routed run (X-Target-Key)
+//	POST /v1/agents/targets/:id/runs/:runId/report   report a routed run's terminal result (X-Target-Key)
+
+// claimLongPoll bounds one Claim wait; on expiry the daemon gets 204 and re-polls
+// immediately, which also refreshes its serving liveness. A var (not a const) so a
+// test can shrink the empty-poll window without waiting the full window.
+var claimLongPoll = 25 * time.Second
+
+const (
+	// claimKeyHeader carries the machine capability. Distinct from Authorization
+	// (which carries the org bearer): org identity and machine identity are two
+	// independent proofs, both required.
+	claimKeyHeader = "X-Target-Key"
+	maxReportField = 64 << 10
+)
+
+// mountRouting registers the route-work machine surface. Called from mountTargets
+// AFTER the target CRUD routes so the extra-segment paths are unambiguous.
+func mountRouting(s *cloud.Service[state], app *zip.App) {
+	app.Post("/v1/agents/targets/:id/claim-key", cloud.Handle(s, mintClaimKey))
+	app.Post("/v1/agents/targets/:id/claim", cloud.Handle(s, claimRoutedRun))
+	app.Post("/v1/agents/targets/:id/runs/:runId/report", cloud.Handle(s, reportRoutedRun))
+}
+
+// mintClaimKey (re)mints the target's claim key and returns it ONCE. Only the
+// SHA-256 hash is stored. Org-scoped: only a caller in the target's org can mint,
+// and the key is bound to (org, target). Rotating supersedes any prior daemon.
+func mintClaimKey(s *cloud.Service[state], c *zip.Ctx) error {
+	org, ok := tenant(c)
+	if !ok {
+		return zip.ErrForbidden("X-Org-Id required")
+	}
+	id := idParam(c)
+	// The target must exist in this org before it can carry a capability.
+	if _, err := s.State.store.GetTarget(c.Context(), org, id); err == errTargetNotFound {
+		return zip.ErrNotFound("target not found")
+	} else if err != nil {
+		return zip.Errorf(http.StatusInternalServerError, "target: %v", err)
+	}
+	key, err := newClaimKey()
+	if err != nil {
+		return zip.Errorf(http.StatusInternalServerError, "rng: %v", err)
+	}
+	if err := s.State.store.UpsertClaimKeyHash(c.Context(), org, id, hashClaimKey(key), time.Now().Unix()); err != nil {
+		return zip.Errorf(http.StatusInternalServerError, "persist: %v", err)
+	}
+	return c.JSON(http.StatusOK, map[string]any{"targetId": id, "claimKey": key})
+}
+
+// claimRoutedRun authenticates the machine, refreshes its serving liveness, and
+// long-polls the rendezvous for the next run addressed to THIS (org, target).
+// 200 + the run on a claim; 204 when the poll window elapses with no work.
+func claimRoutedRun(s *cloud.Service[state], c *zip.Ctx) error {
+	org, ok := tenant(c)
+	if !ok {
+		return zip.ErrForbidden("X-Org-Id required")
+	}
+	id := idParam(c)
+	if err := s.State.store.verifyClaimKey(c.Context(), org, id, c.Header(claimKeyHeader)); err != nil {
+		return claimAuthError(err)
+	}
+	// The poll itself is the runner's liveness proof — stamp it so the dispatch
+	// gate (TargetDispatchable) sees a live runner. Best-effort.
+	_ = s.State.store.StampServing(c.Context(), org, id, time.Now().Unix())
+
+	ctx, cancel := context.WithTimeout(c.Context(), claimLongPoll)
+	defer cancel()
+	run, got := routedMailbox.Claim(ctx, org, id)
+	if !got {
+		return c.NoContent(http.StatusNoContent)
+	}
+	return c.JSON(http.StatusOK, routedRunView(run))
+}
+
+type reportReq struct {
+	OK        bool   `json:"ok"`
+	Changed   bool   `json:"changed"`
+	Branch    string `json:"branch"`
+	CommitSha string `json:"commitSha"`
+	Diffstat  string `json:"diffstat"`
+	Error     string `json:"error"`
+}
+
+// reportRoutedRun completes a claimed run: it delivers the terminal result to the
+// run's durable owner (the RoutedRunWorkflow activity), which lets the workflow
+// finish. Scoped to (org, target, runId) AND claim-key-authenticated, so a
+// machine can only ever report a run it legitimately holds. Idempotent: a report
+// for an unknown/already-finished run is a clean no-op (the session terminal was
+// already set by the machine's own stream).
+func reportRoutedRun(s *cloud.Service[state], c *zip.Ctx) error {
+	org, ok := tenant(c)
+	if !ok {
+		return zip.ErrForbidden("X-Org-Id required")
+	}
+	id := idParam(c)
+	runID := strings.TrimSpace(c.Param("runId"))
+	if err := s.State.store.verifyClaimKey(c.Context(), org, id, c.Header(claimKeyHeader)); err != nil {
+		return claimAuthError(err)
+	}
+	var body reportReq
+	if err := c.Bind(&body); err != nil {
+		return err
+	}
+	res := RoutedResult{
+		OK: body.OK, Changed: body.Changed,
+		Branch:    clampStr(body.Branch, maxRepo),
+		CommitSha: clampStr(body.CommitSha, 128),
+		Diffstat:  clampStr(body.Diffstat, maxReportField),
+		Error:     clampStr(body.Error, maxReportField),
+	}
+	delivered := routedMailbox.Report(org, id, runID, res)
+	return c.JSON(http.StatusOK, map[string]any{"delivered": delivered})
+}
+
+// claimAuthError maps the claim-key verdict onto a fail-closed HTTP status. A
+// missing target row, a missing/mismatched key, and an unknown org all collapse
+// to 403 so the surface never distinguishes "wrong key" from "no such target" —
+// an unauthorized caller learns nothing about what exists.
+func claimAuthError(err error) error {
+	switch err {
+	case errNoClaimKey, errClaimKeyBad, errTargetNotFound:
+		return zip.ErrForbidden("target claim rejected")
+	default:
+		return zip.Errorf(http.StatusInternalServerError, "claim auth: %v", err)
+	}
+}
+
+// routedRunView is the non-secret run spec handed to the machine. It carries no
+// credential by design — the machine authenticates git + model routing with its
+// own already-held credentials.
+func routedRunView(run RoutedRun) map[string]any {
+	return map[string]any{
+		"sessionId":      run.SessionID,
+		"repo":           run.Repo,
+		"project":        run.Project,
+		"base":           run.Base,
+		"branch":         run.Branch,
+		"prompt":         run.Prompt,
+		"cloneUrl":       run.CloneURL,
+		"timeoutSeconds": run.TimeoutSeconds,
+	}
+}

clients/agents/routing_http_test.go

@@ -0,0 +1,250 @@
+package agents
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"io"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/zap-proto/zip"
+)
+
+// doKey is a keyless-body request with a machine claim key (X-Target-Key) attached.
+func doKey(t *testing.T, app *zip.App, method, path, org, key string) (int, []byte) {
+	t.Helper()
+	req := httptest.NewRequest(method, path, nil)
+	if org != "" {
+		req.Header.Set("X-Org-Id", org)
+		req.Header.Set("X-User-Id", "u-"+org)
+	}
+	if key != "" {
+		req.Header.Set(claimKeyHeader, key)
+	}
+	resp, err := app.Fiber().Test(req)
+	if err != nil {
+		t.Fatalf("Test %s %s: %v", method, path, err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+	b, _ := io.ReadAll(resp.Body)
+	return resp.StatusCode, b
+}
+
+// registerAndMint registers a target for org and mints its claim key, returning
+// (targetID, claimKey).
+func registerAndMint(t *testing.T, app *zip.App, org, host string) (string, string) {
+	t.Helper()
+	code, body := do(t, app, "POST", "/v1/agents/targets", org, map[string]any{"label": host, "host": host})
+	if code != 201 && code != 200 {
+		t.Fatalf("register target: %d %s", code, body)
+	}
+	var tv struct{ ID string `json:"id"` }
+	_ = json.Unmarshal(body, &tv)
+	code, body = doKey(t, app, "POST", "/v1/agents/targets/"+tv.ID+"/claim-key", org, "")
+	if code != 200 {
+		t.Fatalf("mint claim key: %d %s", code, body)
+	}
+	var kv struct{ ClaimKey string `json:"claimKey"` }
+	_ = json.Unmarshal(body, &kv)
+	if kv.ClaimKey == "" {
+		t.Fatal("claim key empty")
+	}
+	return tv.ID, kv.ClaimKey
+}
+
+// A claim without the machine's key, or with the WRONG key, is refused — org
+// membership alone is not enough to claim a machine's runs.
+func TestClaim_RequiresMachineKey(t *testing.T) {
+	app := mountApp(t, &fakeAI{content: "x"})
+	old := claimLongPoll
+	claimLongPoll = 150 * time.Millisecond
+	defer func() { claimLongPoll = old }()
+
+	id, key := registerAndMint(t, app, "acme", "evo")
+
+	// No key => 403.
+	if code, _ := doKey(t, app, "POST", "/v1/agents/targets/"+id+"/claim", "acme", ""); code != 403 {
+		t.Fatalf("claim with no key must be 403, got %d", code)
+	}
+	// Wrong key => 403.
+	if code, _ := doKey(t, app, "POST", "/v1/agents/targets/"+id+"/claim", "acme", "tgtk_wrong"); code != 403 {
+		t.Fatalf("claim with wrong key must be 403, got %d", code)
+	}
+	// Right key, no work => 204 (never 200, never another tenant's run).
+	if code, _ := doKey(t, app, "POST", "/v1/agents/targets/"+id+"/claim", "acme", key); code != 204 {
+		t.Fatalf("claim with right key + no work must be 204, got %d", code)
+	}
+}
+
+// THE machine boundary: a key minted for target A cannot claim target B, and a
+// different org cannot claim at all — even with a real key for its own target.
+func TestClaim_CrossMachineAndCrossOrgDenied(t *testing.T) {
+	app := mountApp(t, &fakeAI{content: "x"})
+	old := claimLongPoll
+	claimLongPoll = 150 * time.Millisecond
+	defer func() { claimLongPoll = old }()
+
+	idA, keyA := registerAndMint(t, app, "acme", "evoA")
+	idB, _ := registerAndMint(t, app, "acme", "evoB")
+
+	// A's key against B => 403 (constant-time mismatch on B's stored hash).
+	if code, _ := doKey(t, app, "POST", "/v1/agents/targets/"+idB+"/claim", "acme", keyA); code != 403 {
+		t.Fatalf("A's key claiming B must be 403, got %d", code)
+	}
+
+	// Offer a run for acme/idA, then a DIFFERENT org cannot claim idA at all (its
+	// org scope resolves no such target => 403), and the run is never handed out.
+	OfferRoutedRun(RoutedRun{Org: "acme", TargetID: idA, SessionID: "sess_a", Repo: "api"})
+	if code, _ := doKey(t, app, "POST", "/v1/agents/targets/"+idA+"/claim", "evil", keyA); code != 403 {
+		t.Fatalf("another org claiming acme's target must be 403, got %d", code)
+	}
+	// acme WITH A's key claims its own run.
+	code, body := doKey(t, app, "POST", "/v1/agents/targets/"+idA+"/claim", "acme", keyA)
+	if code != 200 {
+		t.Fatalf("acme must claim its own run, got %d %s", code, body)
+	}
+	var rv struct{ SessionID string `json:"sessionId"` }
+	_ = json.Unmarshal(body, &rv)
+	if rv.SessionID != "sess_a" {
+		t.Fatalf("claimed wrong run: %s", body)
+	}
+}
+
+// The end-to-end machine round trip: offer -> claim -> report reaches the durable
+// owner awaiting the result.
+func TestClaimReport_RoundTrip(t *testing.T) {
+	app := mountApp(t, &fakeAI{content: "x"})
+	id, key := registerAndMint(t, app, "acme", "evo")
+
+	off := OfferRoutedRun(RoutedRun{Org: "acme", TargetID: id, SessionID: "sess_rt", Repo: "api", Branch: "agent/rt"})
+
+	code, body := doKey(t, app, "POST", "/v1/agents/targets/"+id+"/claim", "acme", key)
+	if code != 200 {
+		t.Fatalf("claim: %d %s", code, body)
+	}
+
+	got := make(chan RoutedResult, 1)
+	go func() { res, _ := off.Await(context.Background()); got <- res }()
+
+	code, _ = doKeyBody(t, app, "POST", "/v1/agents/targets/"+id+"/runs/sess_rt/report", "acme", key,
+		map[string]any{"ok": true, "changed": true, "commitSha": "cafe"})
+	if code != 200 {
+		t.Fatalf("report: %d", code)
+	}
+	select {
+	case res := <-got:
+		if !res.OK || res.CommitSha != "cafe" {
+			t.Fatalf("report did not reach the owner: %+v", res)
+		}
+	case <-time.After(2 * time.Second):
+		t.Fatal("owner never received the report")
+	}
+}
+
+// doKeyBody is doKey with a JSON body.
+func doKeyBody(t *testing.T, app *zip.App, method, path, org, key string, body any) (int, []byte) {
+	t.Helper()
+	b, _ := json.Marshal(body)
+	req := httptest.NewRequest(method, path, bytes.NewReader(b))
+	req.Header.Set("Content-Type", "application/json")
+	if org != "" {
+		req.Header.Set("X-Org-Id", org)
+		req.Header.Set("X-User-Id", "u-"+org)
+	}
+	if key != "" {
+		req.Header.Set(claimKeyHeader, key)
+	}
+	resp, err := app.Fiber().Test(req)
+	if err != nil {
+		t.Fatalf("Test %s %s: %v", method, path, err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+	out, _ := io.ReadAll(resp.Body)
+	return resp.StatusCode, out
+}
+
+// ---- store-level liveness gate ----
+
+// TargetDispatchable is the fail-closed gate: online + a live runner (a recent
+// claim poll). Offline, no key, or a stale poll all reject.
+func TestTargetDispatchable_LivenessGate(t *testing.T) {
+	s := testSessionStore(t)
+	ctx := context.Background()
+	now := time.Now().Unix()
+	tgt := Target{ID: "t1", Org: "acme", Label: "evo", Kind: TargetMachine, Status: TargetOnline, Host: "evo", CreatedAt: now, UpdatedAt: now}
+	if err := s.CreateTarget(ctx, tgt); err != nil {
+		t.Fatal(err)
+	}
+
+	// No claim key yet => not live => not dispatchable.
+	if err := s.TargetDispatchable(ctx, "acme", "t1"); err != errTargetNotLive {
+		t.Fatalf("no runner => not dispatchable, got %v", err)
+	}
+	// Mint + a fresh serving stamp => dispatchable.
+	if err := s.UpsertClaimKeyHash(ctx, "acme", "t1", hashClaimKey("k"), now); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.StampServing(ctx, "acme", "t1", now); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.TargetDispatchable(ctx, "acme", "t1"); err != nil {
+		t.Fatalf("online + fresh runner => dispatchable, got %v", err)
+	}
+	// A stale serving stamp => not dispatchable (dead runner).
+	if err := s.StampServing(ctx, "acme", "t1", now-int64(servingTTL/time.Second)-5); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.TargetDispatchable(ctx, "acme", "t1"); err != errTargetNotLive {
+		t.Fatalf("stale runner => not dispatchable, got %v", err)
+	}
+	// Fresh again but OFFLINE => not dispatchable.
+	_ = s.StampServing(ctx, "acme", "t1", time.Now().Unix())
+	tgt.Status = TargetOffline
+	tgt.UpdatedAt = time.Now().Unix()
+	if err := s.UpdateTarget(ctx, tgt); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.TargetDispatchable(ctx, "acme", "t1"); err != errTargetNotReady {
+		t.Fatalf("offline => not dispatchable, got %v", err)
+	}
+	// Unknown target / cross-org => fail closed.
+	if err := s.TargetDispatchable(ctx, "acme", "nope"); err != errTargetNotFound {
+		t.Fatalf("unknown target => not found, got %v", err)
+	}
+	if err := s.TargetDispatchable(ctx, "evil", "t1"); err != errTargetNotFound {
+		t.Fatalf("cross-org => not found, got %v", err)
+	}
+}
+
+// The claim key is stored ONLY as a hash; verify is constant-time + fail closed.
+func TestClaimKey_HashedAtRestAndVerified(t *testing.T) {
+	s := testSessionStore(t)
+	ctx := context.Background()
+	now := time.Now().Unix()
+	_ = s.CreateTarget(ctx, Target{ID: "t1", Org: "acme", Status: TargetOnline, CreatedAt: now, UpdatedAt: now})
+
+	key, _ := newClaimKey()
+	if err := s.UpsertClaimKeyHash(ctx, "acme", "t1", hashClaimKey(key), now); err != nil {
+		t.Fatal(err)
+	}
+	// The stored value is a hash, never the plaintext.
+	stored, _, _ := s.ClaimKeyHash(ctx, "acme", "t1")
+	if stored == key || stored != hashClaimKey(key) {
+		t.Fatalf("claim key must be stored as a hash, not plaintext")
+	}
+	if err := s.verifyClaimKey(ctx, "acme", "t1", key); err != nil {
+		t.Fatalf("correct key must verify: %v", err)
+	}
+	if err := s.verifyClaimKey(ctx, "acme", "t1", "tgtk_wrong"); err != errClaimKeyBad {
+		t.Fatalf("wrong key must fail: %v", err)
+	}
+	if err := s.verifyClaimKey(ctx, "acme", "t1", ""); err != errClaimKeyBad {
+		t.Fatalf("empty key must fail: %v", err)
+	}
+	// Cross-org verify resolves no key => fail closed.
+	if err := s.verifyClaimKey(ctx, "evil", "t1", key); err != errNoClaimKey {
+		t.Fatalf("cross-org verify must fail closed, got %v", err)
+	}
+}

clients/agents/store.go

@@ -180,6 +180,10 @@ CREATE INDEX IF NOT EXISTS ix_runs_org_agent_created ON agent_runs(org, agent_na
 	if err := s.migrateTargets(); err != nil {
 		return err
 	}
+	// Per-target claim keys + serving liveness (the #48 route-work machine plane).
+	if err := s.migrateClaimKeys(); err != nil {
+		return err
+	}
 	return nil
 }
 

clients/agents/targets.go

@@ -10,6 +10,8 @@ import (
 	"time"
 
 	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/principal"
+	"github.com/hanzoai/cloud/clients/samples"
 	"github.com/zap-proto/zip"
 )
 
@@ -17,10 +19,11 @@ import (
 // box, a GPU host, or a whole cluster. It is the #48 link-a-compute seam over the
 // SAME agents.db (one store, one tenancy column) as sessions/events — NOT a rival
 // device registry. It composes with the compute fleet rather than duplicating it: a
-// session records the target id it runs on (agent_sessions.target), and the mission-
-// control devices view unions these registered targets with the org's BYO workers
-// (GET /v1/fleet/workers) and BYO clusters (GET /v1/clusters) at the view layer — the
-// console's established pattern for folding compute sources.
+// session records the target id it runs on (agent_sessions.target), and the org's
+// unified board (GET /v1/fleet, clients/visor/board.go) unions these registered
+// targets with its BYO workers (GET /v1/fleet/workers), BYO clusters and Visor
+// machines — reading this registry through the in-process seam below rather than
+// copying it.
 //
 //	POST   /v1/agents/targets        register a target -> Target
 //	GET    /v1/agents/targets        list the org's targets (+ live session load)
@@ -30,6 +33,12 @@ import (
 //
 // Every route is org-scoped through principal.Org (tenant), fail-closed — a tenant
 // can never see or mutate another org's targets, exactly like sessions.
+//
+// A write carrying `metrics` IS a heartbeat, and a heartbeat is two facts, not one:
+// the LAST sample (kept on the row, rendered by the views here) and one point in a
+// utilization SERIES (appended to clients/samples). The row answers "is this machine
+// alive and what is it doing now"; the series answers "how hot has it been". The
+// append is best-effort and detached — see recordSample.
 
 // Target kinds — the closed vocabulary of dispatch destinations.
 const (
@@ -233,6 +242,90 @@ func (s *Store) GetTargetByHost(ctx context.Context, org, host string) (Target,
 	return t, nil
 }
 
+// ---- the in-process seam (org-scoped, fail-closed) ----
+//
+// TargetsForOrg / LoadOn are the exported twins of the list + detail reads above:
+// the ONE way another in-process subsystem (the /v1/fleet board in clients/visor)
+// reads this registry WITHOUT an HTTP hop back through the gateway — the same
+// shape ListForOrg gives the agent registry. They are two ORTHOGONAL values on
+// purpose: a target is what the machine IS, its load is what is running on it, and
+// a caller that only needs the inventory does not pay for the rollups.
+//
+// ISOLATION: org is the ONLY tenant key and is threaded verbatim into the
+// org-scoped store methods, so a caller for org A can never enumerate or resolve
+// org B's targets. The caller MUST pass an org it already validated server-side
+// (principal.Org), never a raw client header.
+
+// TargetsForOrg returns the org's registered run-targets from the in-process
+// store, newest first. Fails closed when the subsystem is not mounted or the org
+// is empty/oversized.
+func TargetsForOrg(ctx context.Context, org string) ([]Target, error) {
+	if mounted == nil || mounted.State.store == nil {
+		return nil, fmt.Errorf("agents: not mounted")
+	}
+	org = strings.TrimSpace(org)
+	if org == "" || len(org) > principal.MaxOrgLen {
+		return nil, fmt.Errorf("agents: invalid org")
+	}
+	return mounted.State.store.ListTargets(ctx, org)
+}
+
+// ResolveTarget resolves a human's target REFERENCE — a target id or its friendly
+// label (the hostname the CLI registers) — to the org's target, org-scoped and
+// fail-closed. It is the ONE way a trigger surface (the Slack `code: <repo> on
+// <target>` grammar, a console picker) turns "on evo" into a target id without
+// leaking another tenant's inventory: an id or label that resolves to no target in
+// THIS org returns errTargetNotFound, never another org's machine.
+//
+// Precedence: an exact id match wins (ids are unambiguous), else an exact,
+// case-folded label match (newest first, so a re-registered machine's live row is
+// preferred). A reference that matches neither is not found — the caller renders an
+// honest error and NEVER falls back to a local run.
+func ResolveTarget(ctx context.Context, org, ref string) (Target, error) {
+	if mounted == nil || mounted.State.store == nil {
+		return Target{}, fmt.Errorf("agents: not mounted")
+	}
+	org = strings.TrimSpace(org)
+	ref = strings.TrimSpace(ref)
+	if org == "" || len(org) > principal.MaxOrgLen {
+		return Target{}, fmt.Errorf("agents: invalid org")
+	}
+	if ref == "" || len(ref) > maxTargetID {
+		return Target{}, errTargetNotFound
+	}
+	// An id is exact and unambiguous — try it first.
+	if t, err := mounted.State.store.GetTarget(ctx, org, ref); err == nil {
+		return t, nil
+	} else if err != errTargetNotFound {
+		return Target{}, err
+	}
+	// Else an exact, case-folded label match within this org.
+	rows, err := mounted.State.store.ListTargets(ctx, org)
+	if err != nil {
+		return Target{}, err
+	}
+	for _, t := range rows { // ListTargets is newest-first: the live row wins a label tie
+		if strings.EqualFold(strings.TrimSpace(t.Label), ref) {
+			return t, nil
+		}
+	}
+	return Target{}, errTargetNotFound
+}
+
+// LoadOn returns the live session load on one of the org's targets — the same
+// (target id OR host) mapping the HTTP views use, so the board and /v1/agents/
+// targets can never disagree about what is running where.
+func LoadOn(ctx context.Context, org, id, host string) (TargetLoad, error) {
+	if mounted == nil || mounted.State.store == nil {
+		return TargetLoad{}, fmt.Errorf("agents: not mounted")
+	}
+	org = strings.TrimSpace(org)
+	if org == "" || len(org) > principal.MaxOrgLen {
+		return TargetLoad{}, fmt.Errorf("agents: invalid org")
+	}
+	return mounted.State.store.SessionLoad(ctx, org, id, host)
+}
+
 // DeleteTarget removes an org's target. Sessions keep their recorded target id (a
 // historical fact); a detached target simply stops appearing in the registry.
 func (s *Store) DeleteTarget(ctx context.Context, org, id string) (bool, error) {
@@ -262,6 +355,83 @@ func (s *Store) SessionLoad(ctx context.Context, org, id, host string) (TargetLo
 	return TargetLoad{Sessions: total, Running: running}, nil
 }
 
+// ---- the fleet time series ----
+//
+// A heartbeat is the ONE moment this process learns what a linked machine is
+// doing, so it is also where the fleet's utilization series is fed. The target row
+// keeps the LAST sample (the snapshot the views render, unchanged); clients/samples
+// keeps every sample over time. Two different questions — "is it alive now" and
+// "how hot has it been" — so two homes, one write.
+
+// sampleTimeout bounds the warehouse write. Generous (the insert is one small row
+// in-cluster) but finite, so a wedged datastore can never hold the goroutine open.
+const sampleTimeout = 5 * time.Second
+
+// sampleOf projects a target's server-stamped heartbeat into a fleet sample. PURE
+// (no clock, no I/O, no store) so the whole projection is unit-testable and the
+// caller decides when it runs.
+//
+// cost_cents is 0: an agent run-target is the operator's OWN machine (a laptop, a
+// dialed-in box) — the fleet meters its utilization, it does not resell it. A
+// priced source (visor/cloud) fills that column from its own resale price.
+func sampleOf(t Target) samples.Sample {
+	var model string
+	if len(t.Spec.GPUs) > 0 {
+		// The representative accelerator: the count already rides in GPUs, so the
+		// first card's model names the row. A heterogeneous host is rare enough
+		// that naming its first card beats inventing a summary string here.
+		model = t.Spec.GPUs[0].Model
+		if model == "" {
+			model = t.Spec.GPUs[0].Vendor
+		}
+	}
+	return samples.Sample{
+		Org:    t.Org,
+		Source: samples.SourceAgent,
+		Unit:   t.ID,
+		Host:   t.Host,
+		Kind:   t.Kind,
+		At:     time.Unix(t.MetricsAt, 0).UTC(),
+
+		CPUs:     t.Spec.CPUs,
+		Memory:   t.Spec.Memory,
+		MemUsed:  t.Metrics.MemUsed,
+		MemFree:  t.Metrics.MemFree,
+		Load1:    t.Metrics.Load1,
+		Load5:    t.Metrics.Load5,
+		Load15:   t.Metrics.Load15,
+		GPUUtil:  t.Metrics.GPUUtil,
+		GPUs:     len(t.Spec.GPUs),
+		GPUModel: model,
+	}
+}
+
+// recordSample appends a heartbeat to the fleet series. Best-effort and DETACHED
+// on purpose — the warehouse is never in the heartbeat's critical path:
+//
+//   - it runs on its own bounded context, so neither a slow datastore nor the
+//     client hanging up mid-request can stall or cancel the write;
+//   - it never touches the response, so the /v1/agents/targets contract is
+//     byte-identical whether the warehouse is present, absent or on fire;
+//   - a failure is logged, never surfaced — a dropped sample must not cost a
+//     machine its heartbeat.
+//
+// This is the shape the billing warehouse write already uses (`go zapWriteUsage`):
+// the seam is synchronous, the CALLER owns the concurrency.
+func recordSample(s *cloud.Service[state], t Target) {
+	if t.MetricsAt == 0 {
+		return // no heartbeat in this write — nothing to append
+	}
+	sample := sampleOf(t) // project on the caller's goroutine: t must not escape mutably
+	go func() {
+		ctx, cancel := context.WithTimeout(context.Background(), sampleTimeout)
+		defer cancel()
+		if err := samples.Record(ctx, sample); err != nil {
+			s.Log.Warn("fleet sample write failed", "org", sample.Org, "unit", sample.Unit, "err", err)
+		}
+	}()
+}
+
 // ---- HTTP shapes (the published contract) ----
 
 type targetView struct {
@@ -311,6 +481,10 @@ func mountTargets(s *cloud.Service[state], app *zip.App) {
 	app.Get("/v1/agents/targets/:id", cloud.Handle(s, getTarget))
 	app.Patch("/v1/agents/targets/:id", cloud.Handle(s, patchTarget))
 	app.Delete("/v1/agents/targets/:id", cloud.Handle(s, deleteTarget))
+	// The #48 route-work machine surface (claim-key, claim long-poll, report)
+	// lives on the same target routes; register after the CRUD so the
+	// extra-segment paths are unambiguous.
+	mountRouting(s, app)
 }
 
 // ---- register ----
@@ -386,6 +560,7 @@ func registerTarget(s *cloud.Service[state], c *zip.Ctx) error {
 			if err := s.State.store.UpdateTarget(c.Context(), existing); err != nil {
 				return zip.Errorf(http.StatusInternalServerError, "persist: %v", err)
 			}
+			recordSample(s, existing) // a re-link carrying metrics IS a heartbeat
 			load, _ := s.State.store.SessionLoad(c.Context(), org, existing.ID, existing.Host)
 			return c.JSON(http.StatusOK, toTargetView(existing, load))
 		}
@@ -403,6 +578,7 @@ func registerTarget(s *cloud.Service[state], c *zip.Ctx) error {
 	if err := s.State.store.CreateTarget(c.Context(), t); err != nil {
 		return zip.Errorf(http.StatusInternalServerError, "persist: %v", err)
 	}
+	recordSample(s, t) // a registration carrying metrics is the target's first sample
 	return c.JSON(http.StatusCreated, toTargetView(t, TargetLoad{}))
 }
 
@@ -538,6 +714,9 @@ func patchTarget(s *cloud.Service[state], c *zip.Ctx) error {
 		}
 		return zip.Errorf(http.StatusInternalServerError, "update: %v", err)
 	}
+	if body.Metrics != nil {
+		recordSample(s, t) // THE heartbeat: append it to the fleet series too
+	}
 	load, _ := s.State.store.SessionLoad(c.Context(), org, t.ID, t.Host)
 	return c.JSON(http.StatusOK, toTargetView(t, load))
 }

clients/agents/targetsample_test.go

@@ -0,0 +1,227 @@
+package agents
+
+import (
+	"encoding/json"
+	"net/http"
+	"testing"
+	"time"
+
+	"github.com/hanzoai/cloud/clients/samples"
+)
+
+// targetsample_test.go covers the FIRST emitter: a run-target heartbeat also
+// appends to the fleet series (clients/samples).
+//
+// The datastore is absent under test, so samples.Record is a proven no-op (its own
+// package tests that). What MUST be proven here is everything this side owns:
+// the projection is faithful, the vocabularies agree, and the HTTP contract is
+// untouched whether or not the warehouse exists.
+
+// ---- the projection (pure) ----
+
+// A heartbeat projects onto a sample with no loss and no invention.
+func TestSampleOfProjectsTheHeartbeat(t *testing.T) {
+	at := time.Now().Unix()
+	tg := Target{
+		ID: "tgt-1", Org: "acme", Kind: TargetGPU, Host: "box.local", Label: "Box",
+		Spec: Spec{OS: "linux", Arch: "arm64", CPUs: 20, Memory: 128 << 30,
+			GPUs: []GPU{{Vendor: "nvidia", Model: "GB10", Memory: 96 << 30}}},
+		Metrics: Metrics{Load1: 2.5, Load5: 2, Load15: 1.5,
+			MemUsed: 64 << 30, MemFree: 64 << 30, GPUUtil: 0.75},
+		MetricsAt: at,
+	}
+	s := sampleOf(tg)
+
+	if s.Org != "acme" || s.Unit != "tgt-1" || s.Host != "box.local" {
+		t.Fatalf("identity did not project: %+v", s)
+	}
+	if s.Source != samples.SourceAgent {
+		t.Fatalf("source want %q, got %q", samples.SourceAgent, s.Source)
+	}
+	if s.Kind != TargetGPU {
+		t.Fatalf("kind want %q, got %q", TargetGPU, s.Kind)
+	}
+	if !s.At.Equal(time.Unix(at, 0).UTC()) {
+		t.Fatalf("at must be the SERVER-stamped heartbeat clock, got %v", s.At)
+	}
+	if s.CPUs != 20 || s.Memory != 128<<30 {
+		t.Fatalf("spec did not project: %+v", s)
+	}
+	if s.MemUsed != 64<<30 || s.MemFree != 64<<30 || s.Load1 != 2.5 || s.Load5 != 2 || s.Load15 != 1.5 {
+		t.Fatalf("metrics did not project: %+v", s)
+	}
+	if s.GPUUtil != 0.75 || s.GPUs != 1 || s.GPUModel != "GB10" {
+		t.Fatalf("gpu did not project: %+v", s)
+	}
+	// An agent's own machine is metered, never resold.
+	if s.CostCents != 0 {
+		t.Fatalf("an agent sample must be unpriced, got %d", s.CostCents)
+	}
+	// The projection must be acceptable to the plane it feeds.
+	if err := samples.Record(t.Context(), s); err != nil {
+		t.Fatalf("a projected sample must be recordable: %v", err)
+	}
+}
+
+// The accelerator count comes from the spec, and the row is named by the first
+// card's model — falling back to its vendor when the model is unknown.
+func TestSampleOfGPUSummary(t *testing.T) {
+	cases := []struct {
+		name      string
+		gpus      []GPU
+		wantN     int
+		wantModel string
+	}{
+		{"none", nil, 0, ""},
+		{"model", []GPU{{Vendor: "nvidia", Model: "GB10"}}, 1, "GB10"},
+		{"vendor fallback", []GPU{{Vendor: "amd"}}, 1, "amd"},
+		{"multi is counted, first names it", []GPU{{Model: "GB10"}, {Model: "GB10"}}, 2, "GB10"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			s := sampleOf(Target{ID: "t", Org: "o", Kind: TargetGPU, Spec: Spec{GPUs: tc.gpus}, MetricsAt: 1})
+			if s.GPUs != tc.wantN || s.GPUModel != tc.wantModel {
+				t.Fatalf("want (%d, %q), got (%d, %q)", tc.wantN, tc.wantModel, s.GPUs, s.GPUModel)
+			}
+		})
+	}
+}
+
+// THE cross-package contract: every kind a target can be must be a kind the fleet
+// series accepts, or heartbeats would silently stop being recorded. This fails the
+// day someone adds a target kind without teaching the series about it.
+func TestEveryTargetKindIsAFleetKind(t *testing.T) {
+	fleet := map[string]bool{
+		samples.KindLaptop: true, samples.KindCloud: true, samples.KindGPU: true,
+		samples.KindCluster: true, samples.KindMachine: true, samples.KindWorker: true,
+	}
+	for _, k := range []string{TargetLaptop, TargetCloud, TargetGPU, TargetCluster, TargetMachine} {
+		if !fleet[k] {
+			t.Fatalf("target kind %q is not a fleet sample kind — its heartbeats would be dropped", k)
+		}
+		// Proven end to end: a sample carrying this kind validates.
+		s := sampleOf(Target{ID: "t", Org: "o", Kind: k, MetricsAt: 1})
+		if err := samples.Record(t.Context(), s); err != nil {
+			t.Fatalf("kind %q must be recordable: %v", k, err)
+		}
+	}
+}
+
+// A write with no heartbeat in it appends nothing — recordSample is a no-op when
+// the server never stamped a metrics clock.
+func TestRecordSampleSkipsWhenNoHeartbeat(t *testing.T) {
+	mountApp(t, nil) // sets the `mounted` singleton recordSample logs through
+	// No panic, no goroutine, no write: MetricsAt == 0 means "no sample here".
+	recordSample(mounted, Target{ID: "tgt-1", Org: "acme", Kind: TargetGPU, MetricsAt: 0})
+}
+
+// ---- (c) the HTTP contract is untouched by the series ----
+
+// The heartbeat still 200s with no warehouse, and still returns the snapshot on
+// the row exactly as before — the series is strictly additive.
+func TestHeartbeatStill200sWithoutDatastore(t *testing.T) {
+	app := mountApp(t, nil)
+
+	code, body := do(t, app, http.MethodPost, "/v1/agents/targets", "acme", map[string]any{
+		"label": "Box", "kind": TargetGPU, "host": "box.local",
+		"spec":    map[string]any{"os": "linux", "cpus": 20, "gpus": []map[string]any{{"vendor": "nvidia", "model": "GB10"}}},
+		"metrics": map[string]any{"load1": 2.5, "gpuUtil": 0.75, "memUsed": 100},
+	})
+	if code != http.StatusCreated {
+		t.Fatalf("register want 201 without a datastore, got %d (%s)", code, body)
+	}
+	var created targetView
+	if err := json.Unmarshal(body, &created); err != nil {
+		t.Fatalf("shape: %v (%s)", err, body)
+	}
+	if created.Metrics == nil || created.Metrics.GPUUtil != 0.75 {
+		t.Fatalf("the snapshot on the row must be unchanged: %+v", created.Metrics)
+	}
+	if created.MetricsAt == "" {
+		t.Fatal("the server must still stamp the heartbeat clock")
+	}
+
+	// The heartbeat itself.
+	code, body = do(t, app, http.MethodPatch, "/v1/agents/targets/"+created.ID, "acme", map[string]any{
+		"metrics": map[string]any{"load1": 4, "gpuUtil": 0.9, "memUsed": 200},
+	})
+	if code != http.StatusOK {
+		t.Fatalf("heartbeat want 200 without a datastore, got %d (%s)", code, body)
+	}
+	var beat targetView
+	if err := json.Unmarshal(body, &beat); err != nil {
+		t.Fatalf("shape: %v (%s)", err, body)
+	}
+	if beat.Metrics == nil || beat.Metrics.GPUUtil != 0.9 || beat.Metrics.Load1 != 4 {
+		t.Fatalf("the heartbeat must still refresh the row snapshot: %+v", beat.Metrics)
+	}
+
+	// A re-link (same org+host) is idempotent and still carries a heartbeat.
+	code, body = do(t, app, http.MethodPost, "/v1/agents/targets", "acme", map[string]any{
+		"label": "Box", "kind": TargetGPU, "host": "box.local",
+		"metrics": map[string]any{"load1": 1},
+	})
+	if code != http.StatusOK {
+		t.Fatalf("re-link want 200 (idempotent), got %d (%s)", code, body)
+	}
+	var relinked targetView
+	if err := json.Unmarshal(body, &relinked); err != nil {
+		t.Fatalf("shape: %v (%s)", err, body)
+	}
+	if relinked.ID != created.ID {
+		t.Fatalf("a re-link must refresh the SAME target: %s != %s", relinked.ID, created.ID)
+	}
+}
+
+// ---- the in-process seam ----
+
+// TargetsForOrg / LoadOn are org-keyed and fail closed — the board reads through
+// them, so a cross-tenant id must never resolve.
+func TestInProcessSeamIsOrgScopedAndFailsClosed(t *testing.T) {
+	app := mountApp(t, nil)
+	code, body := do(t, app, http.MethodPost, "/v1/agents/targets", "acme", map[string]any{
+		"label": "Secret", "kind": TargetGPU, "host": "secret.local",
+	})
+	if code != http.StatusCreated {
+		t.Fatalf("register: %d (%s)", code, body)
+	}
+	var created targetView
+	_ = json.Unmarshal(body, &created)
+
+	// The owner sees it.
+	own, err := TargetsForOrg(t.Context(), "acme")
+	if err != nil {
+		t.Fatalf("TargetsForOrg(acme): %v", err)
+	}
+	if len(own) != 1 || own[0].ID != created.ID {
+		t.Fatalf("the owner must see its target, got %+v", own)
+	}
+
+	// Another tenant sees nothing — the same id is unreachable.
+	other, err := TargetsForOrg(t.Context(), "other")
+	if err != nil {
+		t.Fatalf("TargetsForOrg(other): %v", err)
+	}
+	if len(other) != 0 {
+		t.Fatalf("CROSS-TENANT LEAK: org 'other' enumerated %+v", other)
+	}
+
+	// A blank/oversized org fails closed on both.
+	for _, bad := range []string{"", "   "} {
+		if _, err := TargetsForOrg(t.Context(), bad); err == nil {
+			t.Fatalf("TargetsForOrg(%q) must fail closed", bad)
+		}
+		if _, err := LoadOn(t.Context(), bad, created.ID, ""); err == nil {
+			t.Fatalf("LoadOn(%q) must fail closed", bad)
+		}
+	}
+
+	// LoadOn is org-keyed too: the foreign tenant resolves no load for the id.
+	load, err := LoadOn(t.Context(), "other", created.ID, "secret.local")
+	if err != nil {
+		t.Fatalf("LoadOn(other): %v", err)
+	}
+	if load.Sessions != 0 || load.Running != 0 {
+		t.Fatalf("CROSS-TENANT LEAK: foreign load %+v", load)
+	}
+}

clients/aihttp.go

@@ -11,7 +11,7 @@ import (
 	"strings"
 	"time"
 
-	openai "github.com/sashabaranov/go-openai"
+	openai "github.com/hanzoai/go-openai"
 	"go.opentelemetry.io/otel"
 	"go.opentelemetry.io/otel/attribute"
 	"go.opentelemetry.io/otel/codes"

clients/analytics/analytics.go

@@ -100,15 +100,21 @@ func routes(app *zip.App, s *cloud.Service[state]) {
 	app.Get("/v1/analytics/timeseries", cloud.Handle(s, timeseries))
 	app.Get("/v1/analytics/top", cloud.Handle(s, top))
 
-	// Capture (WRITE) side — the ingest that fills hanzo.events (capture.go). All
-	// POST, all tenant-gated in-handler. /v1/tracker is the page-unload beacon
-	// alias (bare route; never collides with the /v1/tracker/projects* issue tracker).
+	// Capture (WRITE) side — the ingest that fills hanzo.events. POST /v1/event
+	// (event.go) is the ONE canonical front door: body Event | [Event], org
+	// resolved IAM-only and fail-closed, into the ONE write core (ingestEvents).
+	app.Post("/v1/event", cloud.Handle(s, eventIngest))
+
+	// DEPRECATED ingest aliases — thin wire adapters that normalize onto the SAME
+	// write core (log a one-shot deprecation, keep working). /v1/analytics{,/batch}
+	// and /v1/tracker speak the Segment/beacon CaptureBatch wire; /v1/tracker is a
+	// bare route (never collides with the /v1/tracker/projects* issue tracker).
 	app.Post("/v1/analytics", cloud.Handle(s, capture))
 	app.Post("/v1/analytics/batch", cloud.Handle(s, capture))
 	app.Post("/v1/tracker", cloud.Handle(s, capture))
 
-	// /v1/insights — the unified native surface (insights.go): PostHog-wire
-	// ingest + console reads over the SAME engine. Flags live at /v1/flags.
+	// /v1/insights — console reads over the SAME engine + the DEPRECATED PostHog-
+	// wire ingest adapter (/v1/insights/e → the ONE write core). Flags live at /v1/flags.
 	app.Get("/v1/insights/health", cloud.Handle(s, insightsHealth))
 	app.Post("/v1/insights/e", cloud.Handle(s, insightsIngest))
 	app.Get("/v1/insights/events", cloud.Handle(s, insightsEvents))

clients/analytics/capture.go

@@ -52,6 +52,7 @@ import (
 	"os"
 	"regexp"
 	"strings"
+	"sync"
 	"sync/atomic"
 	"time"
 
@@ -498,19 +499,72 @@ func publicCaptureEnabled() bool {
 	}
 }
 
-// captureTenant resolves the tenant a batch is attributed to. A VALIDATED
-// principal always wins (authenticated product traffic → its own org). Otherwise,
-// for anonymous marketing traffic, the tenant is the PUBLIC brand org derived
-// SERVER-SIDE from the request Host via the white-label registry — never a
-// client-claimed org, so the isolation invariant holds: a caller with no bearer
-// can only ever write into the brand-public partition of the Host it actually
-// reached, and a forged X-Org-Id is ignored exactly as on the read path. An
-// unrecognized Host is refused (we never dump anonymous events into a default
-// org). Returns ("", false) when the caller must be answered 403.
+// resolveKeyOrg maps a presented project/API key to its org through the ONE IAM
+// key seam (cloud.OrgForKey). It is a package var ONLY so a test can substitute a
+// resolver without standing up IAM; production is always cloud.OrgForKey.
+var resolveKeyOrg = cloud.OrgForKey
+
+// projectKey returns the project/API key a keyed SDK presents OUT-OF-BAND of the
+// Authorization header — the transports SanitizeIdentity does NOT mint identity
+// from, so they never reach tenant() as a principal. In priority order: the
+// ?api_key= query, the x-api-key / api-key headers, and the PostHog-wire body
+// field `api_key` (posthog-js and the insights-go batch envelope put it there).
+// "" when none is present.
+//
+// The body is PEEKED via c.Body() — fasthttp buffers the full body, so the later
+// c.Bind in the handler re-reads the same bytes; peeking does not consume it. Only
+// the api_key field is decoded (a bad/unrelated JSON body simply yields "").
+func projectKey(c *zip.Ctx) string {
+	if k := trim(c.Query("api_key")); k != "" {
+		return k
+	}
+	if k := trim(c.Header("x-api-key")); k != "" {
+		return k
+	}
+	if k := trim(c.Header("api-key")); k != "" {
+		return k
+	}
+	if body := c.Body(); len(body) > 0 {
+		var probe struct {
+			APIKey string `json:"api_key"`
+		}
+		if json.Unmarshal(body, &probe) == nil {
+			if k := trim(probe.APIKey); k != "" {
+				return k
+			}
+		}
+	}
+	return ""
+}
+
+// captureTenant resolves the tenant a batch is attributed to, in strict trust
+// order:
+//
+//  1. A VALIDATED principal always wins (authenticated product traffic → its own
+//     org; this also covers a Hanzo key sent as a bearer, which SanitizeIdentity
+//     has already resolved to a principal upstream).
+//  2. Otherwise, if the caller PRESENTS a project key out-of-band (posthog-js /
+//     insights-go: api_key in the body/query/x-api-key), resolve it to its org
+//     through the ONE IAM key seam. This FAILS CLOSED: a presented-but-unresolvable
+//     key is refused (→ 403), NEVER falling through to the brand-host fallback —
+//     attributing a keyed request to the wrong (brand-public) partition would be a
+//     cross-tenant write.
+//  3. Only for TRULY anonymous traffic (no principal, no key) is the tenant the
+//     PUBLIC brand org derived SERVER-SIDE from the request Host via the
+//     white-label registry — never a client-claimed org. An unrecognized Host is
+//     refused (we never dump anonymous events into a default org).
+//
+// Returns ("", false) when the caller must be answered 403.
 func captureTenant(c *zip.Ctx) (string, bool) {
 	if org, ok := tenant(c); ok {
 		return org, true
 	}
+	if key := projectKey(c); key != "" {
+		if org, ok := resolveKeyOrg(c.Context(), key); ok {
+			return org, true
+		}
+		return "", false // presented key that does not resolve → fail CLOSED
+	}
 	if !publicCaptureEnabled() {
 		return "", false
 	}
@@ -520,36 +574,75 @@ func captureTenant(c *zip.Ctx) (string, bool) {
 	return "", false
 }
 
-// capture ingests one batch into hanzo.events, tenant-scoped. Shared by
-// /v1/analytics, /v1/analytics/batch, and /v1/tracker.
-func capture(s *cloud.Service[state], c *zip.Ctx) error {
-	org, ok := captureTenant(c)
-	if !ok {
-		return zip.ErrForbidden("valid bearer or a recognized brand host required")
+// ── ONE write core ───────────────────────────────────────────────────────────
+
+// event source tags — the ingest adapter each row arrived through. Stamped into
+// properties.$source by ingestEvents so the ONE hanzo.events table stays honest
+// about origin (canonical vs. deprecated wire) WITHOUT a second table or a schema
+// migration: the read lenses are unchanged and $source is queryable in the
+// properties JSON, which is exactly the migration signal for the alias sunset.
+const (
+	sourceEvent   = "event"   // canonical POST /v1/event (native Event wire)
+	sourcePostHog = "posthog" // POST /v1/insights/e (PostHog wire adapter, deprecated)
+	sourceCapture = "capture" // POST /v1/analytics{,/batch}, /v1/tracker (Segment/beacon, deprecated)
+)
+
+// withSource returns a copy of p carrying $source=source (the ingest adapter), so
+// normalizeEvent's scrub+store path records origin as a property. nil-safe; never
+// mutates the caller's map (the adapters share their event structs).
+func withSource(p map[string]any, source string) map[string]any {
+	if source == "" {
+		return p
 	}
-	var batch CaptureBatch
-	if err := c.Bind(&batch); err != nil {
-		return zip.ErrBadRequest("malformed capture batch")
+	out := make(map[string]any, len(p)+1)
+	for k, v := range p {
+		out[k] = v
 	}
-	evs := batch.events()
+	out["$source"] = source
+	return out
+}
+
+// deprecatedOnce records one deprecation log per alias path per process, so a
+// high-volume ingest alias signals its sunset exactly once instead of flooding.
+var deprecatedOnce sync.Map
+
+// deprecated logs (once per path) that a superseded ingest alias was hit, pointing
+// callers at the canonical front door. It NEVER changes behavior — the alias keeps
+// working — it only records the migration signal (also visible as $source in the
+// warehouse).
+func deprecated(s *cloud.Service[state], c *zip.Ctx, canonical string) {
+	p := c.Path()
+	if _, seen := deprecatedOnce.LoadOrStore(p, struct{}{}); seen {
+		return
+	}
+	s.Log.Warn("deprecated analytics ingest endpoint; migrate to the canonical event front door",
+		"path", p, "canonical", canonical)
+}
+
+// ingestEvents is the ONE write core: normalize → scrub → batch INSERT into the
+// ONE hanzo.events table. org is the SERVER-resolved tenant (never client input);
+// source tags the ingest adapter. Every front door — the canonical /v1/event and
+// the deprecated PostHog / Segment / beacon adapters — funnels here, so there is
+// exactly one write path. Returns the honest accepted/dropped receipt; the errors
+// it returns are already HTTP-shaped (zip) for the handler to pass straight up.
+func ingestEvents(ctx context.Context, org, source string, evs []CaptureEvent) (CaptureResult, error) {
 	if len(evs) == 0 {
-		return c.JSON(http.StatusOK, CaptureResult{})
+		return CaptureResult{}, nil
 	}
 	if len(evs) > maxBatch {
-		return zip.ErrBadRequest("batch too large")
+		return CaptureResult{}, zip.ErrBadRequest("batch too large")
 	}
 	if err := requireDatastore(); err != nil {
-		return err
+		return CaptureResult{}, err
 	}
-	ctx := c.Context()
 	if err := EnsureEventsTable(ctx); err != nil {
-		return zip.Errorf(http.StatusServiceUnavailable, "analytics warehouse unavailable: %v", err)
+		return CaptureResult{}, zip.Errorf(http.StatusServiceUnavailable, "analytics warehouse unavailable: %v", err)
 	}
-
 	now := time.Now().UTC()
 	rows := make([]eventRow, 0, len(evs))
 	dropped := 0
 	for _, e := range evs {
+		e.Properties = withSource(e.Properties, source)
 		row, ok := normalizeEvent(org, now, e)
 		if !ok {
 			dropped++
@@ -558,12 +651,33 @@ func capture(s *cloud.Service[state], c *zip.Ctx) error {
 		rows = append(rows, row)
 	}
 	if len(rows) == 0 {
-		return c.JSON(http.StatusOK, CaptureResult{Dropped: dropped})
+		return CaptureResult{Dropped: dropped}, nil
 	}
-
 	stmt, args := buildEventsInsert(rows)
 	if err := aiobject.DatastoreExec(ctx, stmt, args...); err != nil {
-		return warehouseErr("capture", err)
+		return CaptureResult{}, warehouseErr("capture", err)
 	}
-	return c.JSON(http.StatusOK, CaptureResult{Accepted: len(rows), Dropped: dropped})
+	return CaptureResult{Accepted: len(rows), Dropped: dropped}, nil
+}
+
+// capture ingests a Segment/beacon batch into hanzo.events, tenant-scoped. It is
+// the DEPRECATED wire adapter behind /v1/analytics, /v1/analytics/batch, and
+// /v1/tracker: a thin CaptureBatch decoder over the ONE write core (ingestEvents).
+// New callers post the canonical Event to /v1/event; this alias keeps working and
+// keeps captureTenant's brand-host path for anonymous marketing traffic.
+func capture(s *cloud.Service[state], c *zip.Ctx) error {
+	deprecated(s, c, "/v1/event")
+	org, ok := captureTenant(c)
+	if !ok {
+		return zip.ErrForbidden("valid bearer or a recognized brand host required")
+	}
+	var batch CaptureBatch
+	if err := c.Bind(&batch); err != nil {
+		return zip.ErrBadRequest("malformed capture batch")
+	}
+	res, err := ingestEvents(c.Context(), org, sourceCapture, batch.events())
+	if err != nil {
+		return err
+	}
+	return c.JSON(http.StatusOK, res)
 }

clients/analytics/capture_keyorg_test.go

@@ -0,0 +1,120 @@
+// Copyright 2023-2026 Hanzo AI Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package analytics
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/zap-proto/zip"
+)
+
+// These tests cover the project-API-key → org resolution added to captureTenant so
+// keyed, bearer-less SDK traffic (posthog-js / insights-go batch) maps to a tenant.
+// They drive the REAL /v1/insights/e handler through the injectable resolveKeyOrg
+// seam, so no IAM is needed. The observable proxy for "resolved to a tenant" is
+// "passed the tenant gate" — i.e. NOT 403; without a datastore the handler then
+// returns 503, so any non-403 status means captureTenant admitted the request.
+
+// stubResolver swaps resolveKeyOrg for the test and records the key it was handed,
+// so a test asserts BOTH that projectKey extracted the right key AND that
+// captureTenant honored the resolution. Restored via t.Cleanup.
+func stubResolver(t *testing.T, fn func(key string) (string, bool)) *string {
+	t.Helper()
+	var got string
+	orig := resolveKeyOrg
+	resolveKeyOrg = func(_ context.Context, key string) (string, bool) {
+		got = key
+		return fn(key)
+	}
+	t.Cleanup(func() { resolveKeyOrg = orig })
+	return &got
+}
+
+// postKeyed issues POST path (with optional ?query) to the mounted app, setting an
+// optional Host and headers and a raw JSON body — no middleware, mirroring the
+// SanitizeIdentity-minted-header harness the other analytics tests use.
+func postKeyed(t *testing.T, app *zip.App, path, host, body string, hdr map[string]string) int {
+	t.Helper()
+	req := httptest.NewRequest(http.MethodPost, path, strings.NewReader(body))
+	req.Header.Set("Content-Type", "application/json")
+	if host != "" {
+		req.Host = host
+	}
+	for k, v := range hdr {
+		req.Header.Set(k, v)
+	}
+	resp, err := app.Fiber().Test(req)
+	if err != nil {
+		t.Fatalf("Test POST %s: %v", path, err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+	_, _ = io.Copy(io.Discard, resp.Body)
+	return resp.StatusCode
+}
+
+// TestCaptureTenant_KeyExtractionReachesResolver: a project key presented in the
+// body, the ?api_key= query, or the x-api-key header is extracted and handed to the
+// resolver, and a resolved key passes the tenant gate (never 403).
+func TestCaptureTenant_KeyExtractionReachesResolver(t *testing.T) {
+	app := mountApp(t)
+	cases := []struct {
+		name, path, body string
+		hdr              map[string]string
+		wantKey          string
+	}{
+		{"body", "/v1/insights/e", `{"api_key":"hk-body","event":"e","distinct_id":"d"}`, nil, "hk-body"},
+		{"query", "/v1/insights/e?api_key=hk-query", `{"event":"e","distinct_id":"d"}`, nil, "hk-query"},
+		{"x-api-key", "/v1/insights/e", `{"event":"e","distinct_id":"d"}`, map[string]string{"x-api-key": "hk-hdr"}, "hk-hdr"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := stubResolver(t, func(string) (string, bool) { return "acme", true })
+			code := postKeyed(t, app, tc.path, "", tc.body, tc.hdr)
+			if *got != tc.wantKey {
+				t.Fatalf("resolver handed key %q, want %q", *got, tc.wantKey)
+			}
+			if code == http.StatusForbidden {
+				t.Fatalf("a resolved key must pass the tenant gate, got 403")
+			}
+		})
+	}
+}
+
+// TestCaptureTenant_UnresolvableKeyFailsClosed: a PRESENTED key that does not
+// resolve is refused (403) EVEN on a recognized brand host — it must never fall
+// through to the brand-public partition (that would be a cross-tenant write).
+func TestCaptureTenant_UnresolvableKeyFailsClosed(t *testing.T) {
+	app := mountApp(t)
+	stubResolver(t, func(string) (string, bool) { return "", false }) // nothing resolves
+	code := postKeyed(t, app, "/v1/insights/e", "hanzo.ai",
+		`{"api_key":"hk-bad","event":"e","distinct_id":"d"}`, nil)
+	if code != http.StatusForbidden {
+		t.Fatalf("presented-but-unresolvable key on a brand host must 403 (fail closed), got %d", code)
+	}
+}
+
+// TestCaptureTenant_AnonBrandHostFallsBack: with NO key presented, anonymous
+// traffic on a recognized brand host still resolves to the brand org (not 403), and
+// the key resolver is never consulted — the key path only triggers on a real key.
+func TestCaptureTenant_AnonBrandHostFallsBack(t *testing.T) {
+	app := mountApp(t)
+	stubResolver(t, func(key string) (string, bool) {
+		t.Fatalf("resolver consulted for a keyless request (key=%q)", key)
+		return "", false
+	})
+	code := postKeyed(t, app, "/v1/insights/e", "hanzo.ai",
+		`{"event":"e","distinct_id":"d"}`, nil)
+	if code == http.StatusForbidden {
+		t.Fatalf("anonymous capture on a recognized brand host must pass the tenant gate, got 403")
+	}
+}

clients/analytics/event.go

@@ -0,0 +1,139 @@
+// Copyright 2023-2026 Hanzo AI Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// event.go — the ONE canonical event-ingestion front door.
+//
+//	POST /v1/event   body: Event | [Event]   ->  {accepted, dropped}
+//
+// A JSON object is one event; a JSON array IS the batch (there is deliberately no
+// /v1/event/batch). Every other ingest surface (the PostHog wire at
+// /v1/insights/e, the Segment/beacon wire at /v1/analytics{,/batch} and
+// /v1/tracker) is a thin DEPRECATED adapter that normalizes its own wire shape
+// onto CaptureEvent and funnels through the SAME write core (ingestEvents) into
+// the SAME hanzo.events table. One write path, many adapters.
+//
+// AUTH — IAM ONLY, FAIL-CLOSED: the tenant is resolved SERVER-SIDE from a
+// validated bearer principal (its owner org) or, for a keyed bearer-less SDK, an
+// access key resolved through the ONE IAM key seam (cloud.OrgForKey). There is NO
+// brand-host fallback on this endpoint: an unauthenticated or unresolvable caller
+// is refused (403), so the canonical door never writes an event into a tenant IAM
+// did not vouch for. The org is NEVER read from the body.
+package analytics
+
+import (
+	"encoding/json"
+	"net/http"
+
+	"github.com/hanzoai/cloud"
+	"github.com/zap-proto/zip"
+)
+
+// Event is the canonical analytics event — the entire ingest contract in four
+// fields. Only these are first-class; everything else a caller wants to record
+// travels in Properties (the scrubber runs over it downstream, same as every
+// event). The tenant is NOT a field: it is resolved server-side from IAM, so a
+// caller can only ever write into its OWN org's partition.
+type Event struct {
+	Event      string         `json:"event"`      // event name (required; empty ⇒ dropped as unroutable)
+	DistinctID string         `json:"distinctId"` // the person/visitor id the caller owns
+	Time       string         `json:"time"`       // optional RFC3339; clamped to server-now on skew/absent
+	Properties map[string]any `json:"properties"` // everything non-core
+}
+
+// toCapture adapts the canonical Event onto the internal CaptureEvent the write
+// core consumes. Type is left empty (canonicalType ⇒ "event"); no $-property is
+// promoted to a column here — /v1/event stays a strict four-field contract, and
+// every non-core field the caller sent stays in Properties.
+func (e Event) toCapture() CaptureEvent {
+	return CaptureEvent{
+		Event:      e.Event,
+		DistinctID: e.DistinctID,
+		Timestamp:  e.Time,
+		Properties: e.Properties,
+	}
+}
+
+// eventTenant resolves the tenant for POST /v1/event — IAM ONLY, FAIL-CLOSED. A
+// validated bearer principal wins (its owner org); otherwise a presented access
+// key is resolved to its org through the ONE IAM key seam (resolveKeyOrg →
+// cloud.OrgForKey). There is NO brand-host fallback: an unauthenticated or
+// unresolvable caller returns ("", false) → 403. (The deprecated aliases keep
+// captureTenant's brand-host path for anonymous marketing traffic; the canonical
+// endpoint is deliberately stricter — IAM is the only tenant authority here.)
+func eventTenant(c *zip.Ctx) (string, bool) {
+	if org, ok := tenant(c); ok {
+		return org, true
+	}
+	if key := projectKey(c); key != "" {
+		if org, ok := resolveKeyOrg(c.Context(), key); ok {
+			return org, true
+		}
+	}
+	return "", false
+}
+
+// decodeEvents decodes a request body as Event | []Event. The first non-space
+// byte decides: '[' ⇒ the array batch, anything else ⇒ a single Event. An empty
+// body yields no events (an honest empty receipt, not an error). Pure over the
+// raw bytes (the handler passes c.Body() — fasthttp-buffered, the same bytes
+// projectKey peeked) so the decode is driven directly by tests.
+func decodeEvents(body []byte) ([]Event, error) {
+	i := 0
+	for i < len(body) {
+		if b := body[i]; b == ' ' || b == '\t' || b == '\r' || b == '\n' {
+			i++
+			continue
+		}
+		break
+	}
+	if i >= len(body) {
+		return nil, nil
+	}
+	if body[i] == '[' {
+		var evs []Event
+		if err := json.Unmarshal(body, &evs); err != nil {
+			return nil, err
+		}
+		return evs, nil
+	}
+	var e Event
+	if err := json.Unmarshal(body, &e); err != nil {
+		return nil, err
+	}
+	return []Event{e}, nil
+}
+
+// eventIngest answers POST /v1/event — the ONE canonical ingestion front door.
+// Org is IAM-derived and fail-closed (eventTenant); the body is Event | [Event];
+// every event flows through the ONE write core (ingestEvents) into the ONE
+// hanzo.events table, tagged source=event.
+func eventIngest(s *cloud.Service[state], c *zip.Ctx) error {
+	org, ok := eventTenant(c)
+	if !ok {
+		return zip.ErrForbidden("valid bearer or a resolvable access key required")
+	}
+	evs, err := decodeEvents(c.Body())
+	if err != nil {
+		return zip.ErrBadRequest("malformed event payload")
+	}
+	caps := make([]CaptureEvent, len(evs))
+	for i, e := range evs {
+		caps[i] = e.toCapture()
+	}
+	res, err := ingestEvents(c.Context(), org, sourceEvent, caps)
+	if err != nil {
+		return err
+	}
+	return c.JSON(http.StatusOK, res)
+}

clients/analytics/event_test.go

@@ -0,0 +1,240 @@
+// Copyright 2023-2026 Hanzo AI Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package analytics
+
+import (
+	"net/http"
+	"testing"
+	"time"
+)
+
+// ── decodeEvents: Event | []Event (single + batch) ──────────────────────────
+
+func TestDecodeEvents_Single(t *testing.T) {
+	evs, err := decodeEvents([]byte(`{"event":"signup","distinctId":"u1","properties":{"plan":"pro"}}`))
+	if err != nil {
+		t.Fatalf("decode single: %v", err)
+	}
+	if len(evs) != 1 {
+		t.Fatalf("single want 1 event, got %d", len(evs))
+	}
+	if evs[0].Event != "signup" || evs[0].DistinctID != "u1" {
+		t.Fatalf("single decoded = %+v", evs[0])
+	}
+	if evs[0].Properties["plan"] != "pro" {
+		t.Fatalf("single properties = %v", evs[0].Properties)
+	}
+}
+
+func TestDecodeEvents_Batch(t *testing.T) {
+	evs, err := decodeEvents([]byte(`[{"event":"a","distinctId":"d"},{"event":"b","distinctId":"d"}]`))
+	if err != nil {
+		t.Fatalf("decode batch: %v", err)
+	}
+	if len(evs) != 2 || evs[0].Event != "a" || evs[1].Event != "b" {
+		t.Fatalf("batch decoded = %+v", evs)
+	}
+}
+
+func TestDecodeEvents_BatchLeadingWhitespace(t *testing.T) {
+	// The array is detected past leading whitespace, not only at byte 0.
+	evs, err := decodeEvents([]byte("  \n\t [{\"event\":\"a\"}]"))
+	if err != nil {
+		t.Fatalf("decode ws-batch: %v", err)
+	}
+	if len(evs) != 1 || evs[0].Event != "a" {
+		t.Fatalf("ws-batch decoded = %+v", evs)
+	}
+}
+
+func TestDecodeEvents_EmptyIsNoEvents(t *testing.T) {
+	// Empty / whitespace-only body ⇒ zero events, NOT an error (honest empty receipt).
+	for _, b := range []string{"", "   ", "\n\t"} {
+		evs, err := decodeEvents([]byte(b))
+		if err != nil || len(evs) != 0 {
+			t.Fatalf("empty %q ⇒ evs=%v err=%v", b, evs, err)
+		}
+	}
+}
+
+func TestDecodeEvents_Malformed(t *testing.T) {
+	for _, b := range []string{`{"event":`, `[{"event":"a"},`, `not json`} {
+		if _, err := decodeEvents([]byte(b)); err == nil {
+			t.Fatalf("malformed %q want error, got nil", b)
+		}
+	}
+}
+
+// ── adapter → Event/CaptureEvent normalization ──────────────────────────────
+
+// TestEventToCapture: the canonical Event maps onto CaptureEvent with ONLY the
+// four core fields promoted; Type is left empty (⇒ "event") and everything else
+// stays in Properties (nothing is lifted to a column).
+func TestEventToCapture(t *testing.T) {
+	e := Event{
+		Event:      "purchase",
+		DistinctID: "u9",
+		Time:       "2026-07-18T00:00:00Z",
+		Properties: map[string]any{"amount": 42, "$current_url": "https://x/y"},
+	}
+	ce := e.toCapture()
+	if ce.Event != "purchase" || ce.DistinctID != "u9" || ce.Timestamp != "2026-07-18T00:00:00Z" {
+		t.Fatalf("core fields = %+v", ce)
+	}
+	if ce.Type != "" {
+		t.Fatalf("Type must be empty (⇒ canonicalType event), got %q", ce.Type)
+	}
+	if ce.URL != "" {
+		t.Fatalf("no $-property is promoted to a column on the canonical wire; URL=%q", ce.URL)
+	}
+	// non-core stays in properties
+	if ce.Properties["amount"] != 42 || ce.Properties["$current_url"] != "https://x/y" {
+		t.Fatalf("properties passthrough = %v", ce.Properties)
+	}
+}
+
+// TestEventNormalizeThroughCore: a canonical Event, adapted and normalized, yields
+// a row stamped with the SERVER org and the resolved event name.
+func TestEventNormalizeThroughCore(t *testing.T) {
+	row, ok := normalizeEvent("acme", time.Now(), Event{Event: "signup", DistinctID: "u1"}.toCapture())
+	if !ok {
+		t.Fatal("want routable")
+	}
+	if row.tenant != "acme" || row.event != "signup" || row.eventType != "event" {
+		t.Fatalf("row = tenant %q event %q type %q", row.tenant, row.event, row.eventType)
+	}
+}
+
+// TestInsightsAdapterNormalization: the PostHog wire adapter lifts well-known
+// $-properties to columns (that is its job) — the OTHER adapter feeding the ONE
+// core, distinct from the canonical Event wire.
+func TestInsightsAdapterNormalization(t *testing.T) {
+	ce := insightsEvent{
+		Event:      "$pageview",
+		DistinctID: "d",
+		Properties: map[string]any{"$current_url": "https://x/y", "$session_id": "s1"},
+	}.toCapture()
+	if ce.Type != "pageview" {
+		t.Fatalf("posthog $pageview ⇒ type pageview, got %q", ce.Type)
+	}
+	if ce.URL != "https://x/y" || ce.SessionID != "s1" {
+		t.Fatalf("posthog $-props ⇒ columns: url=%q session=%q", ce.URL, ce.SessionID)
+	}
+}
+
+// TestCaptureBatchAdapter: the Segment/beacon adapter prefers `batch`, falling
+// back to `events`.
+func TestCaptureBatchAdapter(t *testing.T) {
+	if got := (CaptureBatch{Batch: []CaptureEvent{{Event: "a"}}, Events: []CaptureEvent{{Event: "b"}}}).events(); len(got) != 1 || got[0].Event != "a" {
+		t.Fatalf("batch preferred over events, got %+v", got)
+	}
+	if got := (CaptureBatch{Events: []CaptureEvent{{Event: "b"}}}).events(); len(got) != 1 || got[0].Event != "b" {
+		t.Fatalf("events fallback, got %+v", got)
+	}
+}
+
+// ── source tagging (the $source property, one-table origin discriminator) ────
+
+func TestWithSource(t *testing.T) {
+	// stamps $source
+	got := withSource(nil, sourceEvent)
+	if got["$source"] != "event" {
+		t.Fatalf("withSource(nil,event) = %v", got)
+	}
+	// does not mutate the caller's map, and preserves existing keys
+	orig := map[string]any{"a": 1}
+	out := withSource(orig, sourcePostHog)
+	if out["a"] != 1 || out["$source"] != "posthog" {
+		t.Fatalf("withSource copy = %v", out)
+	}
+	if _, leaked := orig["$source"]; leaked {
+		t.Fatalf("withSource mutated the caller's map: %v", orig)
+	}
+	// empty source is a no-op passthrough (same map)
+	if got := withSource(orig, ""); got["$source"] != nil {
+		t.Fatalf("empty source must not stamp, got %v", got)
+	}
+}
+
+// TestSourceStampedIntoProperties: source flows through withSource → normalizeEvent
+// → the stored properties JSON, so the ONE hanzo.events table carries origin
+// WITHOUT a schema column.
+func TestSourceStampedIntoProperties(t *testing.T) {
+	e := CaptureEvent{Event: "x", Properties: withSource(nil, sourceEvent)}
+	row, ok := normalizeEvent("acme", time.Now(), e)
+	if !ok {
+		t.Fatal("want routable")
+	}
+	props := decodeProps(t, row.properties)
+	if props["$source"] != "event" {
+		t.Fatalf("row.properties $source = %v (props=%v)", props["$source"], props)
+	}
+}
+
+// ── POST /v1/event: IAM-only, fail-closed auth ──────────────────────────────
+//
+// Observable proxy (mirrors capture_keyorg_test): a REFUSED request is 403; an
+// ADMITTED one reaches requireDatastore and returns 503 (no datastore in tests).
+// So "not 403" ⇒ the tenant gate admitted the request.
+
+func TestEvent_NoPrincipalNoKeyForbidden(t *testing.T) {
+	app := mountApp(t)
+	if code, _ := doBody(t, app, http.MethodPost, "/v1/event", "", "", `{"event":"e","distinctId":"d"}`); code != http.StatusForbidden {
+		t.Fatalf("no-principal no-key /v1/event want 403, got %d", code)
+	}
+}
+
+func TestEvent_BearerPrincipalAdmitted(t *testing.T) {
+	app := mountApp(t)
+	// A validated principal (X-User/X-Org) is admitted → 503 (datastore down), not 403.
+	if code, _ := doBody(t, app, http.MethodPost, "/v1/event", "user-dave", "acme", `{"event":"signup","distinctId":"u1"}`); code != http.StatusServiceUnavailable {
+		t.Fatalf("bearer /v1/event want 503 (admitted, datastore down), got %d", code)
+	}
+	// A batch body is admitted the same way.
+	if code, _ := doBody(t, app, http.MethodPost, "/v1/event", "user-dave", "acme", `[{"event":"a","distinctId":"d"},{"event":"b","distinctId":"d"}]`); code != http.StatusServiceUnavailable {
+		t.Fatalf("bearer /v1/event batch want 503, got %d", code)
+	}
+}
+
+func TestEvent_ResolvedKeyAdmitted(t *testing.T) {
+	app := mountApp(t)
+	got := stubResolver(t, func(string) (string, bool) { return "acme", true })
+	code := postKeyed(t, app, "/v1/event", "", `{"api_key":"hk-k","event":"e","distinctId":"d"}`, nil)
+	if *got != "hk-k" {
+		t.Fatalf("resolver handed key %q, want hk-k", *got)
+	}
+	if code == http.StatusForbidden {
+		t.Fatalf("a resolved access key must pass the /v1/event gate, got 403")
+	}
+}
+
+func TestEvent_UnresolvableKeyFailsClosedEvenOnBrandHost(t *testing.T) {
+	app := mountApp(t)
+	stubResolver(t, func(string) (string, bool) { return "", false })
+	code := postKeyed(t, app, "/v1/event", "hanzo.ai", `{"api_key":"hk-bad","event":"e","distinctId":"d"}`, nil)
+	if code != http.StatusForbidden {
+		t.Fatalf("presented-but-unresolvable key on /v1/event must 403 (fail closed), got %d", code)
+	}
+}
+
+// TestEvent_NoBrandHostFallback is THE distinguishing invariant: anonymous traffic
+// on a recognized brand host is ADMITTED by the deprecated /v1/analytics alias
+// (brand-public partition) but REFUSED by the canonical /v1/event — IAM is the
+// only tenant authority on the canonical door.
+func TestEvent_NoBrandHostFallback(t *testing.T) {
+	app := mountApp(t)
+	body := `{"event":"e","distinctId":"d"}`
+	if code, _ := doHost(t, app, "/v1/event", "", "", "hanzo.ai", body); code != http.StatusForbidden {
+		t.Fatalf("anonymous brand-host /v1/event must 403 (no brand fallback), got %d", code)
+	}
+	// Contrast: the deprecated alias still admits the same anonymous brand-host
+	// traffic (503 = admitted, datastore down), proving the difference is by design.
+	if code, _ := doHost(t, app, "/v1/analytics", "", "", "hanzo.ai", `{"batch":[{"type":"pageview"}]}`); code != http.StatusServiceUnavailable {
+		t.Fatalf("deprecated alias still brand-admits (want 503), got %d", code)
+	}
+}

clients/analytics/insights.go

@@ -80,9 +80,13 @@ func (e insightsEvent) toCapture() CaptureEvent {
 	}
 }
 
-// insightsIngest answers POST /v1/insights/e — the PostHog-compatible front
-// door. Same tenant gate, same normalize/scrub, same warehouse as /v1/analytics.
+// insightsIngest answers POST /v1/insights/e — the DEPRECATED PostHog-wire
+// adapter. It normalizes the PostHog single/batch shape onto CaptureEvent and
+// funnels through the ONE write core (ingestEvents, source=posthog); it keeps
+// captureTenant's brand-host path so anonymous PostHog-wire traffic is unbroken.
+// New callers post the canonical Event to /v1/event.
 func insightsIngest(s *cloud.Service[state], c *zip.Ctx) error {
+	deprecated(s, c, "/v1/event")
 	org, ok := captureTenant(c)
 	if !ok {
 		return zip.ErrForbidden("valid bearer or a recognized brand host required")
@@ -95,39 +99,15 @@ func insightsIngest(s *cloud.Service[state], c *zip.Ctx) error {
 	if len(events) == 0 && body.Event != "" {
 		events = []insightsEvent{body.insightsEvent}
 	}
-	if len(events) == 0 {
-		return c.JSON(http.StatusOK, CaptureResult{})
+	caps := make([]CaptureEvent, len(events))
+	for i, e := range events {
+		caps[i] = e.toCapture()
 	}
-	if len(events) > maxBatch {
-		return zip.ErrBadRequest("batch too large")
-	}
-	if err := requireDatastore(); err != nil {
+	res, err := ingestEvents(c.Context(), org, sourcePostHog, caps)
+	if err != nil {
 		return err
 	}
-	ctx := c.Context()
-	if err := EnsureEventsTable(ctx); err != nil {
-		return zip.Errorf(http.StatusServiceUnavailable, "analytics warehouse unavailable: %v", err)
-	}
-
-	now := time.Now().UTC()
-	rows := make([]eventRow, 0, len(events))
-	dropped := 0
-	for _, e := range events {
-		row, ok := normalizeEvent(org, now, e.toCapture())
-		if !ok {
-			dropped++
-			continue
-		}
-		rows = append(rows, row)
-	}
-	if len(rows) == 0 {
-		return c.JSON(http.StatusOK, CaptureResult{Dropped: dropped})
-	}
-	stmt, args := buildEventsInsert(rows)
-	if err := aiobject.DatastoreExec(ctx, stmt, args...); err != nil {
-		return zip.Errorf(http.StatusServiceUnavailable, "analytics warehouse write failed: %v", err)
-	}
-	return c.JSON(http.StatusOK, CaptureResult{Accepted: len(rows), Dropped: dropped})
+	return c.JSON(http.StatusOK, res)
 }
 
 // insightsEvents answers GET /v1/insights/events — the console's recent-events

clients/authors/commerce.go

@@ -1,152 +1,44 @@
 package authors
 
 import (
-	"bytes"
 	"context"
-	"encoding/json"
-	"errors"
-	"fmt"
-	"io"
-	"net/http"
-	"net/url"
-	"strings"
-	"time"
 
-	"github.com/hanzoai/cloud/clients/commerceinproc"
+	"github.com/hanzoai/cloud/clients/payout"
 )
 
 // commerce is the narrow money seam the author royalty loop needs: read a deploying
-// org's metered spend (the accrual base) and grant a promo credit to a wallet (a
-// payout made in credits). It is an INTERFACE so the store/handler logic is testable
-// with a fake ledger — the HTTP impl below is the ONE production binding.
+// org's metered spend (the royalty accrual base) and grant a promo credit to a wallet
+// (a payout made in credits, ledger tag grant:author). It is an INTERFACE so the
+// store/handler logic is testable with a fake ledger; the production binding is
+// clients/payout, reached through the thin adapter below.
 //
-// This mirrors clients/affiliates/commerce.go EXACTLY: the same
-// COMMERCE_SERVICE_TOKEN S2S path, the same X-Org-Id=<org> namespace + bare org
-// `user` subject — so an author payout-in-credits lands in precisely the wallet the
-// balance panel reads, indistinguishable from an admin/affiliate grant except by its
-// ledger tag (grant:author, → the commerce Credit/trial bucket per DepositKind's
-// grant:* rule).
+// The S2S impl (COMMERCE_SERVICE_TOKEN path, X-Org-Id=<org> namespace, bare-org
+// `user` subject) was three byte-identical commerce.go copies; it now lives ONCE in
+// clients/payout. An author payout-in-credits still lands in precisely the wallet the
+// balance panel reads, indistinguishable from an admin grant except by its
+// grant:author tag.
 type commerce interface {
 	configured() bool
-	// deposit grants amountCents to org's wallet (Credit/trial bucket via the
-	// grant:author tag) and returns the ledger transaction id.
 	deposit(ctx context.Context, org, user string, amountCents int64, currency, notes, tags string) (txnID string, err error)
-	// spendCents is a deploying org's month-to-date metered consumption — the royalty
-	// accrual base (spend × the author's share).
 	spendCents(ctx context.Context, org, user string) (int64, error)
 }
 
-// errUnconfigured is returned by a deposit against an unwired commerce so the caller
-// records an honest failure rather than reporting a phantom payout.
-var errUnconfigured = errors.New("authors: commerce endpoint not configured")
+// errUnconfigured is the shared sentinel a deposit against an unwired commerce
+// returns, so the caller records an honest failure rather than a phantom payout.
+var errUnconfigured = payout.ErrUnconfigured
 
-// httpCommerce is the production commerce binding (COMMERCE_SERVICE_TOKEN S2S).
-type httpCommerce struct {
-	base  string
-	token string
-	http  *http.Client
+// commerceSeam adapts the shared payout.Client onto this program's lowercase seam
+// (Go package-scoped interface methods cannot cross packages). Zero logic — pure
+// delegation; the money path lives in clients/payout.
+type commerceSeam struct{ c *payout.Client }
+
+func (s commerceSeam) configured() bool { return s.c.Configured() }
+func (s commerceSeam) deposit(ctx context.Context, org, user string, amountCents int64, currency, notes, tags string) (string, error) {
+	return s.c.Deposit(ctx, org, user, amountCents, currency, notes, tags)
+}
+func (s commerceSeam) spendCents(ctx context.Context, org, user string) (int64, error) {
+	return s.c.SpendCents(ctx, org, user)
 }
 
-func newCommerceClient(base, token string) *httpCommerce {
-	return &httpCommerce{
-		base:  strings.TrimRight(strings.TrimSpace(base), "/"),
-		token: strings.TrimSpace(token),
-		http:  commerceinproc.Client(15 * time.Second),
-	}
-}
-
-func (c *httpCommerce) configured() bool { return c != nil && c.base != "" && c.token != "" }
-
-// deposit posts POST /v1/billing/deposit — the ONE money-in primitive (identical to
-// affiliates.httpCommerce.deposit). Commerce's EdgeAuth pins the body `user` to the
-// X-Org-Id subject, so a payout can never be mis-targeted to another wallet.
-func (c *httpCommerce) deposit(ctx context.Context, org, user string, amountCents int64, currency, notes, tags string) (string, error) {
-	if !c.configured() {
-		return "", errUnconfigured
-	}
-	if currency == "" {
-		currency = "usd"
-	}
-	body, err := json.Marshal(map[string]any{
-		"user":     user,
-		"currency": currency,
-		"amount":   amountCents,
-		"notes":    notes,
-		"tags":     tags,
-	})
-	if err != nil {
-		return "", err
-	}
-	raw, err := c.do(ctx, http.MethodPost, "/v1/billing/deposit", nil, org, body)
-	if err != nil {
-		return "", err
-	}
-	var out struct {
-		TransactionID string `json:"transactionId"`
-	}
-	if err := json.Unmarshal(raw, &out); err != nil {
-		return "", fmt.Errorf("commerce deposit decode: %w", err)
-	}
-	return out.TransactionID, nil
-}
-
-// spendCents reads GET /v1/billing/usage-rollup and returns consumedCents. Zero (not
-// an error) when commerce is unconfigured so a partial deploy degrades to "no spend
-// to accrue yet" rather than a 5xx.
-func (c *httpCommerce) spendCents(ctx context.Context, org, user string) (int64, error) {
-	if !c.configured() {
-		return 0, nil
-	}
-	q := url.Values{"user": {user}}
-	raw, err := c.do(ctx, http.MethodGet, "/v1/billing/usage-rollup", q, org, nil)
-	if err != nil {
-		return 0, err
-	}
-	var out struct {
-		ConsumedCents int64 `json:"consumedCents"`
-	}
-	if err := json.Unmarshal(raw, &out); err != nil {
-		return 0, fmt.Errorf("commerce rollup decode: %w", err)
-	}
-	return out.ConsumedCents, nil
-}
-
-// do performs one admin-S2S commerce request. X-Org-Id=<org> is the per-org
-// namespace selector commerce's EdgeAuth trusts only behind the service token.
-func (c *httpCommerce) do(ctx context.Context, method, path string, q url.Values, org string, body []byte) ([]byte, error) {
-	u := c.base + path
-	if enc := q.Encode(); enc != "" {
-		u += "?" + enc
-	}
-	var r io.Reader
-	if body != nil {
-		r = bytes.NewReader(body)
-	}
-	req, err := http.NewRequestWithContext(ctx, method, u, r)
-	if err != nil {
-		return nil, err
-	}
-	req.Header.Set("Accept", "application/json")
-	if body != nil {
-		req.Header.Set("Content-Type", "application/json")
-	}
-	if c.token != "" {
-		req.Header.Set("Authorization", "Bearer "+c.token)
-	}
-	if org != "" {
-		req.Header.Set("X-Org-Id", org)
-	}
-	resp, err := c.http.Do(req)
-	if err != nil {
-		return nil, fmt.Errorf("commerce unreachable: %w", err)
-	}
-	defer func() { _ = resp.Body.Close() }()
-	out, err := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
-	if err != nil {
-		return nil, err
-	}
-	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		return nil, fmt.Errorf("commerce status %d", resp.StatusCode)
-	}
-	return out, nil
-}
+// newCommerceClient builds the production binding, delegating to clients/payout.
+func newCommerceClient(base, token string) commerce { return commerceSeam{payout.NewClient(base, token)} }

clients/automations/automations.go

@@ -43,6 +43,7 @@ import (
 
 	"github.com/hanzoai/cloud"
 	"github.com/hanzoai/cloud/audit"
+	"github.com/hanzoai/cloud/clients/connectorruntime"
 	"github.com/hanzoai/cloud/clients/principal"
 	"github.com/hanzoai/cloud/clients/tools"
 	"github.com/zap-proto/zip"
@@ -138,6 +139,15 @@ func Mount(app *zip.App, deps cloud.Deps) error {
 	tools.Register(connectorToolProvider{})
 
 	b.Log.Info("automations mounted", "connectors", catalog.ConnectorCount, "runtime", len(registry), "brand", deps.Brand)
+
+	// Native single-connector execution (HIP-0126): POST /v1/automations/connectors/:id/run,
+	// the in-process goja runner paired with the connector catalogue above. It mounts one
+	// route DISTINCT from every automations route (no /v1/automations/* wildcard here, so no
+	// shadow), and was a separate Wire entry purely for that one route — fold it in as a
+	// terminal sub-mount so connector catalogue + execution are ONE automations subsystem.
+	if err := connectorruntime.Mount(app, deps); err != nil {
+		return err
+	}
 	return nil
 }
 
@@ -676,7 +686,7 @@ func recordRunEnd(s *cloud.Service[state], ctx context.Context, in RunEndInput)
 
 // meterUnit records one metered unit for an HTTP caller's org. Nil/disabled meter → no-op.
 func meterUnit(s *cloud.Service[state], org string, c *zip.Ctx) {
-	s.Bill.Meter(principal.Payer(c), principal.Project(c), meterKind, cloud.ResourceFeeCents(feeEnvPrefix, meterKind), c.RequestID(), cloud.ClientIP(c))
+	s.Bill.Meter(principal.HomeOrg(c), principal.Project(c), meterKind, cloud.ResourceFeeCents(feeEnvPrefix, meterKind), c.RequestID(), cloud.ClientIP(c))
 }
 
 // meterRun records one metered unit for a flow run from the durable path (no HTTP

clients/base/base.go

@@ -15,7 +15,7 @@
 //	LANE 2 — managed Base hosting (what superbase/PocketHost provided). ONE Base
 //	app PER ORG, opened lazily and pooled, each on its OWN SQLite under
 //	{DataDir}/base/{orgSegment}/ — the same "prod = SQLite per tenant" model
-//	(HIP-0302) the gojabase leaves (captable/sign/dataroom) use, so an org's
+//	(HIP-0302) the NewBase leaves (captable/sign/dataroom) use, so an org's
 //	collections/records are PHYSICALLY isolated. Served AUTHENTICATED under
 //	/v1/base/*, the org resolved from the VALIDATED cloud principal (never a
 //	client header). This is the console Bases manager's backend.

clients/base/base_test.go

@@ -18,7 +18,7 @@ import (
 	baseapp "github.com/hanzoai/base"
 	"github.com/hanzoai/base/core"
 	"github.com/hanzoai/cloud"
-	"github.com/hanzoai/cloud/clients/gojabase"
+	"github.com/hanzoai/cloud/clients/goja"
 	luxlog "github.com/luxfi/log"
 	fiber "github.com/zap-proto/fiber/v3"
 	"github.com/zap-proto/zip"
@@ -170,8 +170,8 @@ func TestPerOrgIsolatedCRUD(t *testing.T) {
 	}
 
 	// The isolation is physical: distinct on-disk data dirs per org segment.
-	acmeDir := filepath.Join(dataDir, "base", gojabase.TenantSegment("acme"))
-	globexDir := filepath.Join(dataDir, "base", gojabase.TenantSegment("globex"))
+	acmeDir := filepath.Join(dataDir, "base", goja.TenantSegment("acme"))
+	globexDir := filepath.Join(dataDir, "base", goja.TenantSegment("globex"))
 	if acmeDir == globexDir {
 		t.Fatalf("orgs share a data dir: %s", acmeDir)
 	}

clients/base/pool.go

@@ -17,7 +17,7 @@ import (
 	baseapp "github.com/hanzoai/base"
 	"github.com/hanzoai/base/apis"
 	"github.com/hanzoai/cloud"
-	"github.com/hanzoai/cloud/clients/gojabase"
+	"github.com/hanzoai/cloud/clients/goja"
 )
 
 // Pool sizing (env-overridable). A full Base app is heavier than a bare *sql.DB
@@ -35,7 +35,7 @@ const (
 // ({DataDir}/base/{TenantSegment}/), the "prod = SQLite per tenant" rule
 // (HIP-0302). Apps open lazily on first request, migrate once, and are pooled
 // (LRU-capped, idle-evicted). Concurrent opens of the same org are
-// single-flighted under mu. The org→segment encoding is gojabase.TenantSegment —
+// single-flighted under mu. The org→segment encoding is goja.TenantSegment —
 // the ONE injective, traversal-safe tenant→path encoder, shared so an org maps
 // to exactly one physical identity everywhere in the binary.
 type pool struct {
@@ -85,7 +85,7 @@ func (p *pool) acquire(org string) (http.Handler, func(), error) {
 	if strings.TrimSpace(org) == "" {
 		return nil, nil, fmt.Errorf("base: empty org")
 	}
-	seg := gojabase.TenantSegment(org)
+	seg := goja.TenantSegment(org)
 
 	p.mu.Lock()
 	defer p.mu.Unlock()
@@ -111,7 +111,7 @@ func (p *pool) acquire(org string) (http.Handler, func(), error) {
 // for in-process callers that drive the engine's Go API directly (collection
 // provisioning, seeding) rather than the HTTP path.
 func (p *pool) appFor(org string) (*baseapp.Base, error) {
-	seg := gojabase.TenantSegment(org)
+	seg := goja.TenantSegment(org)
 	p.mu.Lock()
 	defer p.mu.Unlock()
 	if e, ok := p.m[seg]; ok {

clients/billing/balance.go

@@ -4,8 +4,9 @@ import (
 	"context"
 	"strings"
 
-	aiobject "github.com/hanzoai/ai/object"
+	"github.com/hanzoai/account"
 	"github.com/hanzoai/cloud/clients/finance"
+	"github.com/hanzoai/cloud/clients/principal"
 	"github.com/zap-proto/zip"
 )
 
@@ -30,12 +31,14 @@ import (
 // self-dispatch, and the number shown is the number that admits or refuses a request.
 // The commerce S2S read stays as the split-deploy fallback, unchanged.
 
-// balanceSubject resolves the wallet key for the caller by CALLING ai/object.BillingSubject
-// — the same function the ai prepaid gate (routers/filter_balance.go resolveBillingKey) and
-// the usage debit resolve. It is deliberately not re-implemented here: cloud and ai each
-// keeping their own copy of this rule is what let them drift apart (cloud's console view
-// scoping to the org while the gate scoped to "org/user"), so the view showed a funded org
-// while the gate refused the member. One function, one rule, one wallet.
+// subjectFor resolves the billing subject for the caller by CALLING the ONE rule,
+// ai/object.Payer — the same function the ai prepaid gate (routers/filter_balance.go
+// resolveBillingKey) and the usage debit resolve. It is the SHARED resolver for every
+// commerce-projected read in this package (the balance read AND the finance reads), so
+// there is exactly one copy of the rule. cloud and ai each keeping their own copy is
+// what let them drift apart (cloud's console view scoping to the org while the gate
+// scoped to "org/user"), so the view showed a funded org while the gate refused the
+// member. One function, one rule, one wallet.
 //
 // org is the VALIDATED principal org. The name half uses the SAME precedence as
 // clients/account.resolveCaller: X-User-Name (the IAM username the identity boundary mints
@@ -43,22 +46,28 @@ import (
 // X-User-Id. Both are authorityHeaders — stripped on ingress and re-injected only from
 // verified claims — so neither is a client value.
 //
-// The X-User-Id fallback goes through ai's own BillingSubjectFromUserKey because that
-// header's shape is path-dependent: the gateway historically minted X-User-Id == the
-// username, while the in-binary direct-Bearer path mints the UUID subject, and callers
-// hold it as an "<owner>/<name>" key. FromUserKey is the function ai already uses to fold
-// that key form back to a subject, so the two agree by construction instead of by a
-// re-implemented split here.
+// The X-User-Id fallback goes through the shared account.PayerOf because that header's shape is
+// path-dependent: the gateway historically minted X-User-Id == the username, while the
+// in-binary direct-Bearer path mints the UUID subject, and callers hold it as an
+// "<owner>/<name>" key. PayerOf is the parse ai already uses to fold that key form back
+// to the payer, so the two agree by construction instead of by a re-implemented split.
 //
 // KNOWN RESIDUAL: a validated principal carrying NEITHER X-User-Name NOR an "<owner>/<name>"
 // X-User-Id (i.e. a bare username id) folds to the org pool. That requires a JWT with no
 // `name` and no `preferred_username`, since the boundary mints X-User-Name from either;
 // production tokens carry one. Called out for review rather than papered over.
-func balanceSubject(c *zip.Ctx, org string) string {
+func subjectFor(c *zip.Ctx, org string) string {
 	if name := strings.TrimSpace(c.Header("X-User-Name")); name != "" {
-		return aiobject.BillingSubject(org, name)
+		// Hand Payer the account the credential NAMES (the validated `billing_account`
+		// claim, minted into X-Billing-Account-Id). The ai gate reads the same claim,
+		// so this view and that gate resolve one wallet. Reading it here is what keeps
+		// them from drifting the way the org-vs-"org/user" split once did — except
+		// that split was two rules, and this would be one rule fed two different
+		// credentials, which reads the same to a user: a funded balance the gate
+		// refuses. Absent ⟹ Payer's legacy rule, exactly today's answer.
+		return account.Payer(account.Credential{Owner: org, Name: name, Account: principal.BillingAccount(c)}).Subject()
 	}
-	return aiobject.BillingSubjectFromUserKey(org, strings.TrimSpace(c.User()))
+	return account.PayerOf(org, strings.TrimSpace(c.User())).Subject()
 }
 
 // availableCents returns the caller's spendable prepaid balance from the co-resident

clients/billing/balance_test.go

@@ -4,11 +4,11 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"github.com/hanzoai/account"
 	"net/http"
 	"net/http/httptest"
 	"testing"
 
-	aiobject "github.com/hanzoai/ai/object"
 	"github.com/hanzoai/cloud/clients/finance"
 	"github.com/hanzoai/cloud/clients/money"
 	"github.com/hanzoai/cloud/types"
@@ -19,11 +19,19 @@ import (
 // shows, and can be made to fail so a test can prove an unreadable balance is never
 // rendered as zero.
 type fakeFinance struct {
-	wallets map[string]int64 // "org|subject" -> cents
-	err     error
-	gotOrg  string
-	gotSubj string
-	calls   int
+	wallets   map[string]int64 // "org|subject" -> cents
+	usageRows []finance.UsageRow
+	err       error
+	gotOrg    string
+	gotSubj   string
+	calls     int
+}
+
+// ListUsage satisfies the optional co-resident usage-read capability coResidentUsage
+// resolves; returns the seeded rows so a test can prove the usage view answers from the
+// ledger instead of the self-dispatching commerce hop.
+func (f *fakeFinance) ListUsage(context.Context, string, int) ([]finance.UsageRow, error) {
+	return f.usageRows, f.err
 }
 
 func (f *fakeFinance) Balance(_ context.Context, org, subject, _ string, _ bool) (money.Amount, error) {
@@ -81,21 +89,21 @@ func TestBalance_ReadsFinanceLedgerNotCommerce(t *testing.T) {
 
 // TestBalance_SubjectIsTheGateSubject pins the invariant this incident broke: the wallet
 // the console SHOWS must be the wallet the ai prepaid gate READS. Both derive it from the
-// one function, ai/object.BillingSubject, so they cannot drift apart again — cloud keeping
-// its own copy of this rule is exactly what let the console show a funded org while the
-// gate refused the member.
+// one function, ai/object.Payer, so they cannot drift apart again — cloud keeping its own
+// copy of this rule is exactly what let the console show a funded org while the gate
+// refused the member.
 func TestBalance_SubjectIsTheGateSubject(t *testing.T) {
-	// Reproduce the live pod env: both allowlists present-but-empty. Under ai v1.809.2
-	// PERSONAL_BILLING_ORGS defaults to "hanzo", so "hanzo" is a PERSONAL-billing org and
-	// the gate reads the per-user wallet.
+	// The allowlists are gone; set them to values that WOULD have flipped the resolution
+	// to prove they are inert — nothing reads them, the signup org still bills per-person.
 	t.Setenv("PERSONAL_BILLING_ORGS", "")
-	t.Setenv("ORG_BILLING_ORGS", "")
+	t.Setenv("ORG_BILLING_ORGS", "hanzo")
 
 	// The gate's subject for this principal, from ai itself — not a value this test invents.
-	// This is what routers/filter_balance.go resolveBillingKey computes from the JWT claims.
-	want := aiobject.BillingSubject("hanzo", "z")
+	// This is what routers/filter_balance.go resolveBillingKey computes from the JWT claims:
+	// a person in the signup org bills their OWN account, hanzo/z.
+	want := account.Payer(account.Credential{Owner: "hanzo", Name: "z"}).Subject()
 	if want != "hanzo/z" {
-		t.Fatalf("precondition: ai v1.809.2 resolves the live env to %q, want hanzo/z", want)
+		t.Fatalf("precondition: ai resolves a signup-org person to %q, want hanzo/z", want)
 	}
 
 	// Every identity shape a validated principal can arrive in must land on that ONE wallet.
@@ -104,7 +112,7 @@ func TestBalance_SubjectIsTheGateSubject(t *testing.T) {
 		{"gateway mints X-User-Name", "z", "8f14e45f-ea1b-4c2a-9f3d-000000000001"},
 		// In-binary direct-Bearer: X-User-Id is the UUID subject, X-User-Name carries the name.
 		{"in-binary direct bearer", "z", "hanzo/z"},
-		// No X-User-Name: the "<owner>/<name>" key form folds back via FromUserKey.
+		// No X-User-Name: the "<owner>/<name>" key form folds back via PayerOf.
 		{"owner/name id, no X-User-Name", "", "hanzo/z"},
 	} {
 		t.Run(tc.name, func(t *testing.T) {

clients/billing/billing.go

@@ -246,6 +246,24 @@ func usage(s *cloud.Service[state], c *zip.Ctx) error {
 	if !ok {
 		return zip.ErrUnauthorized("sign in to view billing")
 	}
+
+	// Co-resident, read the usage ledger DIRECTLY from cloud's own finance ledger
+	// (usage_coresident.go explains why this is NOT a commerce proxy: proxying
+	// "/v1/billing/usage" re-enters THIS handler — commerce's own /v1/billing/usage
+	// route is behind //go:build cloud and never compiled here, so the only
+	// registration of that path is this handler — and the in-proc S2S hop carries no
+	// validated principal, so usage() self-answered "sign in to view billing"; that
+	// self-dispatch is the 500 a valid caller saw). This is the exact move balance()
+	// already makes. Off the co-resident path the commerce S2S proxy is unchanged.
+	if body, coResident, err := coResidentUsage(c.Context(), org, strings.TrimSpace(c.Query("product")), strings.TrimSpace(c.Query("groupBy"))); err != nil {
+		s.Log.Warn("finance usage read failed", "org", org, "err", err)
+		return zip.Errorf(http.StatusBadGateway, "billing upstream unreachable")
+	} else if coResident {
+		c.SetHeader("Content-Type", "application/json")
+		c.SetHeader("Cache-Control", "no-store")
+		return c.Bytes(http.StatusOK, body)
+	}
+
 	if !s.State.commerce.configured() {
 		return zip.Errorf(http.StatusNotImplemented, "billing is not configured")
 	}
@@ -286,7 +304,7 @@ func balance(s *cloud.Service[state], c *zip.Ctx) error {
 		// true "not signed in" (401), matching usage/gpuCharge.
 		return zip.ErrUnauthorized("sign in to view billing")
 	}
-	cents, coResident, err := availableCents(c.Context(), org, balanceSubject(c, org))
+	cents, coResident, err := availableCents(c.Context(), org, subjectFor(c, org))
 	if err != nil {
 		// A balance that cannot be READ is unknown — surface it as an upstream failure.
 		// It must never render as a zero balance: unknown is not "broke".

clients/billing/finance.go

@@ -202,7 +202,7 @@ func financeBalance(s *cloud.Service[state], c *zip.Ctx) error {
 	// The ONE balance read (balance.go) — the same wallet /v1/billing/balance answers, so
 	// the two surfaces can never disagree. Co-resident this is the finance ledger; only a
 	// split deploy falls through to the commerce S2S read below.
-	if cents, coResident, err := availableCents(c.Context(), org, balanceSubject(c, org)); err != nil {
+	if cents, coResident, err := availableCents(c.Context(), org, subjectFor(c, org)); err != nil {
 		s.Log.Warn("finance balance read failed", "org", org, "err", err)
 		return zip.Errorf(http.StatusBadGateway, "billing upstream unreachable")
 	} else if coResident {
@@ -354,7 +354,7 @@ func financePaymentMethods(s *cloud.Service[state], c *zip.Ctx) error {
 		return zip.Errorf(http.StatusNotImplemented, "billing is not configured")
 	}
 	// Portal read filters on customerId; the subject is pinned to the caller's own org.
-	body, status, err := s.State.commerce.get(c.Context(), "/v1/billing/portal/payment-methods", org, financeSubject(org, nil))
+	body, status, err := s.State.commerce.get(c.Context(), "/v1/billing/portal/payment-methods", org, financeSubject(subjectFor(c, org), nil))
 	if err != nil {
 		s.Log.Warn("commerce payment-methods read failed", "org", org, "err", err)
 		return zip.Errorf(http.StatusBadGateway, "billing upstream unreachable")
@@ -436,12 +436,12 @@ func financeCaller(s *cloud.Service[state], c *zip.Ctx) (string, bool) {
 	return principal.Org(c)
 }
 
-// financeSubject builds the commerce query with every billing-subject key PINNED to org
-// (the client can never widen scope), plus any extra passthrough params.
-func financeSubject(org string, extra url.Values) url.Values {
+// financeSubject builds the commerce query with every billing-subject key PINNED to
+// subject (the client can never widen scope), plus any extra passthrough params.
+func financeSubject(subject string, extra url.Values) url.Values {
 	q := url.Values{}
 	for _, k := range billingSubjectKeys {
-		q.Set(k, org)
+		q.Set(k, subject)
 	}
 	for k, vs := range extra {
 		for _, v := range vs {
@@ -454,7 +454,7 @@ func financeSubject(org string, extra url.Values) url.Values {
 // financeGet does one org-scoped commerce GET and decodes the 2xx body into out. A
 // non-2xx or unreachable upstream is surfaced honestly (never masked as empty data).
 func financeGet(s *cloud.Service[state], c *zip.Ctx, path, org string, extra url.Values, out any) error {
-	body, status, err := s.State.commerce.get(c.Context(), path, org, financeSubject(org, extra))
+	body, status, err := s.State.commerce.get(c.Context(), path, org, financeSubject(subjectFor(c, org), extra))
 	if err != nil {
 		s.Log.Warn("commerce finance read failed", "org", org, "path", path, "err", err)
 		return zip.Errorf(http.StatusBadGateway, "billing upstream unreachable")
@@ -472,7 +472,7 @@ func financeGet(s *cloud.Service[state], c *zip.Ctx, path, org string, extra url
 // credits/usage/ledger projections share). Tolerates the wrapped {transactions:[…]}
 // shape and a bare array.
 func financeTxns(s *cloud.Service[state], c *zip.Ctx, org string) ([]commerceTxn, error) {
-	body, status, err := s.State.commerce.get(c.Context(), "/v1/billing/transactions", org, financeSubject(org, url.Values{"limit": {"2000"}}))
+	body, status, err := s.State.commerce.get(c.Context(), "/v1/billing/transactions", org, financeSubject(subjectFor(c, org), url.Values{"limit": {"2000"}}))
 	if err != nil {
 		s.Log.Warn("commerce transactions read failed", "org", org, "err", err)
 		return nil, zip.Errorf(http.StatusBadGateway, "billing upstream unreachable")

clients/billing/usage_coresident.go

@@ -0,0 +1,84 @@
+package billing
+
+// usage_coresident.go — the co-resident source for GET /v1/billing/usage.
+//
+// WHY IT EXISTS. balance() already reads cloud's OWN finance ledger directly rather
+// than proxying "/v1/billing/balance" through commerceinproc, because co-resident the
+// ONLY registration of that path is balance() itself (commerce's api.Route() is behind
+// //go:build cloud and never compiled here), so the S2S proxy re-dispatches BY PATH
+// straight back into the same handler, which self-answers "sign in to view billing"
+// (the in-proc hop carries no validated principal). usage() had the SAME defect on
+// "/v1/billing/usage" — a valid caller's usage read re-entered usage() and failed. This
+// file gives usage() the co-resident answer balance() already has: the usage ledger read
+// straight from finance (the wallet→revenue debits RecordUsage wrote), off the
+// self-dispatching hop. Split deploy (no co-resident finance) falls back to the commerce
+// S2S read, unchanged.
+
+import (
+	"context"
+	"encoding/json"
+	"time"
+
+	"github.com/hanzoai/cloud/clients/finance"
+)
+
+// coResidentUsage builds the customer usage envelope from cloud's OWN finance ledger
+// when the money plane is co-resident (finance.Current() published AND exposes the
+// usage read). It returns (body, true, nil) with the commerce-shaped
+// {user,count,usage:[...]} envelope — enriched + optionally ?product=filtered /
+// ?groupBy=product-reduced exactly like the proxied path — or (nil, false, nil) when
+// finance is not co-resident (split deploy), so the caller falls back to the commerce
+// S2S read. This is what keeps /v1/billing/usage off the self-dispatching commerceinproc
+// hop; a real read failure surfaces as a non-nil error (never a masked-empty ledger).
+func coResidentUsage(ctx context.Context, org, product, groupBy string) ([]byte, bool, error) {
+	fin := finance.Current()
+	if fin == nil {
+		return nil, false, nil // split deploy → commerce S2S read
+	}
+	// The usage read is an OPTIONAL capability (the base FinanceClient is
+	// Balance+Deposit+RecordUsage); a finance impl without it falls back to the proxy.
+	lister, ok := fin.(interface {
+		ListUsage(context.Context, string, int) ([]finance.UsageRow, error)
+	})
+	if !ok {
+		return nil, false, nil
+	}
+	rows, err := lister.ListUsage(ctx, org, 2000)
+	if err != nil {
+		return nil, false, err
+	}
+	env := usageEnvelope(org, rows)
+	if out, ok := enrichUsageLedger(env, product, groupBy); ok {
+		return out, true, nil
+	}
+	return env, true, nil
+}
+
+// usageEnvelope renders finance usage rows as commerce's GetUsage envelope
+// ({user,count,usage:[{transactionId,amount,metadata,createdAt}]}) — the exact shape the
+// console's normalizeUsageRecords + this package's enrichUsageLedger already parse.
+// amount is USD cents; metadata carries the metered unit (model) the debit recorded, so
+// enrichUsageLedger can still attribute a product where the unit implies one.
+func usageEnvelope(org string, rows []finance.UsageRow) []byte {
+	type usageRow struct {
+		TransactionID string         `json:"transactionId"`
+		Amount        int64          `json:"amount"`
+		Metadata      map[string]any `json:"metadata"`
+		CreatedAt     string         `json:"createdAt"`
+	}
+	out := make([]usageRow, 0, len(rows))
+	for _, r := range rows {
+		md := map[string]any{}
+		if r.Model != "" {
+			md["model"] = r.Model
+		}
+		out = append(out, usageRow{
+			TransactionID: r.ID,
+			Amount:        r.Cents,
+			Metadata:      md,
+			CreatedAt:     time.Unix(r.CreatedAt, 0).UTC().Format(time.RFC3339),
+		})
+	}
+	body, _ := json.Marshal(map[string]any{"user": org, "count": len(out), "usage": out})
+	return body
+}

clients/billing/usage_coresident_test.go

@@ -0,0 +1,89 @@
+package billing
+
+import (
+	"context"
+	"encoding/json"
+	"testing"
+
+	"github.com/hanzoai/cloud/clients/finance"
+)
+
+// TestCoResidentUsage proves usage() answers from the finance ledger (never the
+// self-dispatching commerce hop) and shapes the commerce GetUsage envelope the
+// console parses. fakeFinance + publishFinance live in balance_test.go.
+func TestCoResidentUsage(t *testing.T) {
+	publishFinance(t, &fakeFinance{usageRows: []finance.UsageRow{
+		{ID: "u1", Cents: 150, Model: "gpt-x", CreatedAt: 1_700_000_000},
+		{ID: "u2", Cents: 75, Model: "embed-y", CreatedAt: 1_700_000_100},
+	}})
+
+	body, coResident, err := coResidentUsage(context.Background(), "acme", "", "")
+	if err != nil {
+		t.Fatalf("coResidentUsage: %v", err)
+	}
+	if !coResident {
+		t.Fatal("want coResident=true when finance is published")
+	}
+	var env struct {
+		User  string `json:"user"`
+		Count int    `json:"count"`
+		Usage []struct {
+			TransactionID string         `json:"transactionId"`
+			Amount        int64          `json:"amount"`
+			Metadata      map[string]any `json:"metadata"`
+			CreatedAt     string         `json:"createdAt"`
+		} `json:"usage"`
+	}
+	if err := json.Unmarshal(body, &env); err != nil {
+		t.Fatalf("envelope not valid JSON: %v\n%s", err, body)
+	}
+	if env.User != "acme" || env.Count != 2 || len(env.Usage) != 2 {
+		t.Fatalf("bad envelope: user=%q count=%d rows=%d", env.User, env.Count, len(env.Usage))
+	}
+	if env.Usage[0].TransactionID != "u1" || env.Usage[0].Amount != 150 {
+		t.Errorf("row0 = %+v", env.Usage[0])
+	}
+	if env.Usage[0].Metadata["model"] != "gpt-x" {
+		t.Errorf("row0 metadata missing model: %+v", env.Usage[0].Metadata)
+	}
+	if env.Usage[0].CreatedAt == "" {
+		t.Error("row0 createdAt should be RFC3339, got empty")
+	}
+}
+
+// TestCoResidentUsageSplitDeploy proves that with no co-resident finance, usage()
+// falls through to the commerce S2S proxy (coResident=false), unchanged.
+func TestCoResidentUsageSplitDeploy(t *testing.T) {
+	finance.Publish(nil)
+	body, coResident, err := coResidentUsage(context.Background(), "acme", "", "")
+	if err != nil {
+		t.Fatalf("coResidentUsage: %v", err)
+	}
+	if coResident || body != nil {
+		t.Fatalf("want fall-through (coResident=false, nil body); got coResident=%v body=%s", coResident, body)
+	}
+}
+
+// TestCoResidentUsageGroupBy proves the ?groupBy=product reduction still runs on the
+// co-resident path (a token-metered row attributes to product "inference").
+func TestCoResidentUsageGroupBy(t *testing.T) {
+	publishFinance(t, &fakeFinance{usageRows: []finance.UsageRow{{ID: "u1", Cents: 150, Model: "gpt-x", CreatedAt: 1_700_000_000}}})
+	body, ok, err := coResidentUsage(context.Background(), "acme", "", "product")
+	if err != nil || !ok {
+		t.Fatalf("coResidentUsage groupBy: ok=%v err=%v", ok, err)
+	}
+	var grouped struct {
+		GroupBy string `json:"groupBy"`
+		Groups  []struct {
+			Product     string `json:"product"`
+			Requests    int    `json:"requests"`
+			AmountCents int64  `json:"amountCents"`
+		} `json:"groups"`
+	}
+	if err := json.Unmarshal(body, &grouped); err != nil {
+		t.Fatalf("grouped not valid JSON: %v\n%s", err, body)
+	}
+	if grouped.GroupBy != "product" || len(grouped.Groups) != 1 {
+		t.Fatalf("bad grouped envelope: %s", body)
+	}
+}

clients/captable/captable.go

@@ -8,12 +8,12 @@
 // business LOGIC (ported to a self-contained goja bundle in github.com/hanzoai/
 // captable) and gives it PERSISTENCE over per-tenant Base/SQLite. The bundle
 // carries logic; the Go host carries storage. The seam between them is the
-// REUSABLE clients/gojabase binding (the RW-Base goja host), which esign (#100)
+// REUSABLE clients/goja binding (the RW-Base goja host), which esign (#100)
 // and dataroom (#101) reuse unchanged — this leaf is just:
 //
 //	captable bundle (github.com/hanzoai/captable.Bundle)  +  the per-tenant Schema
 //	                         │
-//	                  clients/gojabase.New(...)   ← injects __db/__newId/__now,
+//	                  clients/goja.NewBase(...)   ← injects __db/__newId/__now,
 //	                         │                       one SQLite file per tenant,
 //	                  /v1/captable/* zip routes     one transaction per request
 //
@@ -37,7 +37,7 @@ import (
 
 	hcaptable "github.com/hanzoai/captable"
 	"github.com/hanzoai/cloud"
-	"github.com/hanzoai/cloud/clients/gojabase"
+	"github.com/hanzoai/cloud/clients/goja"
 	"github.com/hanzoai/cloud/clients/principal"
 	"github.com/zap-proto/zip"
 )
@@ -48,7 +48,7 @@ const maxBody = 1 << 20 // 1 MiB
 
 // state is captable's own data; shared deps live in the embedded cloud.Base.
 type state struct {
-	host *gojabase.Host
+	host *goja.BaseHost
 }
 
 // mounted is the active service so Shutdown can release the per-tenant stores.
@@ -72,7 +72,7 @@ func Mount(app *zip.App, deps cloud.Deps) error {
 	if err != nil {
 		return fmt.Errorf("captable.Mount: load bundle: %w", err)
 	}
-	host, err := gojabase.New(gojabase.Config{
+	host, err := goja.NewBase(goja.BaseConfig{
 		Name:    "captable",
 		Bundle:  bundle,
 		Schema:  schema,
@@ -80,7 +80,7 @@ func Mount(app *zip.App, deps cloud.Deps) error {
 		OnOpen:  seedCompany,
 	})
 	if err != nil {
-		return fmt.Errorf("captable.Mount: gojabase host: %w", err)
+		return fmt.Errorf("captable.Mount: goja NewBase host: %w", err)
 	}
 	s := &cloud.Service[state]{Base: cloud.NewBase(deps, "captable"), State: state{host: host}}
 	mounted = s
@@ -176,7 +176,7 @@ func dispatch(s *cloud.Service[state], c *zip.Ctx, route string, params map[stri
 			}
 		}
 	}
-	resp, err := s.State.host.Dispatch(c.Context(), org, gojabase.Request{
+	resp, err := s.State.host.Dispatch(c.Context(), org, goja.BaseRequest{
 		Route:  route,
 		Params: params,
 		Body:   body,

clients/captable/captable_test.go

@@ -6,22 +6,22 @@ import (
 	"testing"
 
 	hcaptable "github.com/hanzoai/captable"
-	"github.com/hanzoai/cloud/clients/gojabase"
+	"github.com/hanzoai/cloud/clients/goja"
 )
 
 // TestFullLifecycle drives the REAL embedded captable bundle against a REAL
-// per-tenant SQLite through gojabase — the end-to-end proof that the bundle's SQL
+// per-tenant SQLite through NewBase — the end-to-end proof that the bundle's SQL
 // matches the Go host schema and that the whole cap-table fold round-trips
 // through Base: company → stakeholders → share class → equity plan → share
 // issuance → options → SAFE → priced round + investment (dilution) → transfer →
 // computed cap table. Any column/route drift fails here, not in production.
-func newHost(t *testing.T) *gojabase.Host {
+func newHost(t *testing.T) *goja.BaseHost {
 	t.Helper()
 	bundle, err := hcaptable.Bundle()
 	if err != nil {
 		t.Fatal(err)
 	}
-	h, err := gojabase.New(gojabase.Config{
+	h, err := goja.NewBase(goja.BaseConfig{
 		Name:    "captable",
 		Bundle:  bundle,
 		Schema:  schema,
@@ -36,9 +36,9 @@ func newHost(t *testing.T) *gojabase.Host {
 }
 
 // do dispatches a route and returns (status, decoded body).
-func do(t *testing.T, h *gojabase.Host, org, route string, params map[string]string, body any) (int, any) {
+func do(t *testing.T, h *goja.BaseHost, org, route string, params map[string]string, body any) (int, any) {
 	t.Helper()
-	resp, err := h.Dispatch(context.Background(), org, gojabase.Request{Route: route, Params: params, Body: body})
+	resp, err := h.Dispatch(context.Background(), org, goja.BaseRequest{Route: route, Params: params, Body: body})
 	if err != nil {
 		t.Fatalf("dispatch %s: %v", route, err)
 	}
@@ -300,7 +300,7 @@ func TestDilutiveOptionsExcludeTerminal(t *testing.T) {
 }
 
 // firstShareID returns the founder's original certificate share id.
-func firstShareID(t *testing.T, h *gojabase.Host, org string) string {
+func firstShareID(t *testing.T, h *goja.BaseHost, org string) string {
 	t.Helper()
 	_, body := do(t, h, org, "shares.list", nil, nil)
 	data := body.(map[string]any)["data"].([]any)

clients/captable/facade.go

@@ -7,7 +7,7 @@ import (
 	"fmt"
 	"time"
 
-	"github.com/hanzoai/cloud/clients/gojabase"
+	"github.com/hanzoai/cloud/clients/goja"
 )
 
 // facade.go is the in-process cap-table seam: it lets a sibling subsystem (Hanzo
@@ -74,7 +74,7 @@ type RoundInput struct {
 // response. The body is round-tripped through JSON to a generic value so the goja
 // bundle sees the SAME wire shape (lower-case json keys) the HTTP path produces —
 // passing a typed Go struct straight to goja would expose Go field names instead.
-func facadeDispatch(ctx context.Context, org, route string, params map[string]string, body any) (*gojabase.Response, error) {
+func facadeDispatch(ctx context.Context, org, route string, params map[string]string, body any) (*goja.Response, error) {
 	if mounted == nil || mounted.State.host == nil {
 		return nil, ErrNotMounted
 	}
@@ -85,7 +85,7 @@ func facadeDispatch(ctx context.Context, org, route string, params map[string]st
 	if err != nil {
 		return nil, err
 	}
-	return mounted.State.host.Dispatch(ctx, org, gojabase.Request{Route: route, Params: params, Body: wire})
+	return mounted.State.host.Dispatch(ctx, org, goja.BaseRequest{Route: route, Params: params, Body: wire})
 }
 
 // toWire normalizes a typed value to a generic JSON value (map[string]any /
@@ -107,7 +107,7 @@ func toWire(body any) (any, error) {
 
 // okBody checks the response is 2xx and returns the body bytes, else a descriptive
 // error carrying the bundle's own message.
-func okBody(resp *gojabase.Response, route string) ([]byte, error) {
+func okBody(resp *goja.Response, route string) ([]byte, error) {
 	if resp.Status/100 != 2 {
 		return nil, fmt.Errorf("captable %s: status %d: %s", route, resp.Status, string(resp.Body))
 	}

clients/captable/http_test.go

@@ -19,7 +19,7 @@ import (
 // mountApp builds a bare zip.App (no SanitizeIdentity middleware, so X-Org-Id +
 // X-User-Id are trusted verbatim — the standard cloud leaf test harness) and
 // mounts the captable leaf on it. This exercises the REAL HTTP path: routing →
-// body decode → principal gate → gojabase dispatch → per-tenant Base → response,
+// body decode → principal gate → NewBase dispatch → per-tenant Base → response,
 // the same path the live binary serves under CLOUD_ENABLE=captable.
 func mountApp(t *testing.T) *zip.App {
 	t.Helper()

clients/captable/schema.go

@@ -9,7 +9,7 @@ import (
 // schema is the per-tenant SQLite DDL — the Go host owns migrations; the goja
 // bundle only issues SQL against these tables. Column names MUST match the SQL in
 // the captable bundle (github.com/hanzoai/captable goja/src/routes/*). Idempotent
-// (IF NOT EXISTS), so it runs on every tenant DB open via gojabase.
+// (IF NOT EXISTS), so it runs on every tenant DB open via NewBase.
 //
 // This is the Prisma data model (prisma/schema.prisma) translated to SQLite:
 // DateTime → TEXT (ISO strings stored verbatim; the bundle never parses them),
@@ -217,7 +217,7 @@ CREATE INDEX IF NOT EXISTS ix_investment_company ON investment(company_id);
 CREATE INDEX IF NOT EXISTS ix_investment_round   ON investment(round_id);
 `
 
-// seedCompany is the gojabase OnOpen hook: it ensures the tenant's cap-table
+// seedCompany is the NewBase OnOpen hook: it ensures the tenant's cap-table
 // company row exists (id == the validated tenant), so the bundle's companyId
 // always resolves. The name defaults to the tenant and is renamed via
 // PUT /v1/captable/company. INSERT OR IGNORE makes it idempotent across reopens.

clients/chat/capability.go

@@ -1,125 +0,0 @@
-package chat
-
-import (
-	"encoding/json"
-	"strings"
-
-	openai "github.com/sashabaranov/go-openai"
-)
-
-// capability is a named chat mode. It frames one tool-calling round: the system
-// prompt that instructs the model, the builtin tool defs offered alongside the
-// caller's and the org's registered tools, and whether the model's tool calls are
-// EXECUTED server-side (registry Dispatch → actions) or handed back to the client
-// as ops (a graph/UI mutation the server cannot perform). Adding a capability is
-// one entry in the map below — no other file changes.
-type capability struct {
-	Name         string
-	SystemPrompt string
-	BuiltinTools []openai.Tool
-	// ServerExecuted gates the tool-call split: when true, a call the tool registry
-	// knows is dispatched server-side and reported in actions; when false the round
-	// is advisory — every call is returned as an op for the client to apply.
-	ServerExecuted bool
-}
-
-// defaultCapability is used when a request omits `capability`.
-const defaultCapability = "graph"
-
-// capabilities is the registry. Small and closed: more modes add cleanly here.
-var capabilities = map[string]capability{
-	"graph": {
-		Name:           "graph",
-		ServerExecuted: false,
-		SystemPrompt: strings.TrimSpace(`
-You are Hanzo Studio Copilot. You read and mutate a node workflow graph
-(LiteGraph / ComfyUI-compatible). The user describes an outcome; you respond with
-tool calls that describe the graph operations to apply. You never execute the
-graph yourself — the Studio client applies each operation and re-queues the graph.
-Reference nodes you create by the id you give them in add_node. Prefer the fewest
-operations that achieve the request; explain briefly in text when helpful.`),
-		BuiltinTools: []openai.Tool{
-			toolDef("add_node", "Add a node to the graph.", `{
-				"type":"object",
-				"properties":{
-					"id":{"type":"string","description":"caller label to reference this node in later ops"},
-					"type":{"type":"string","description":"node class, e.g. KSampler"},
-					"pos":{"type":"array","items":{"type":"number"},"description":"[x,y] canvas position"}
-				},
-				"required":["type"]
-			}`),
-			toolDef("set_widget", "Set a widget value on a node.", `{
-				"type":"object",
-				"properties":{
-					"node_id":{"type":"string"},
-					"name":{"type":"string","description":"widget name"},
-					"value":{}
-				},
-				"required":["node_id","name","value"]
-			}`),
-			toolDef("set_prompt", "Set the prompt text on a node.", `{
-				"type":"object",
-				"properties":{"node_id":{"type":"string"},"text":{"type":"string"}},
-				"required":["node_id","text"]
-			}`),
-			toolDef("connect", "Connect an output slot to an input slot.", `{
-				"type":"object",
-				"properties":{
-					"from":{"type":"string","description":"source node id"},
-					"from_slot":{"description":"output slot index or name"},
-					"to":{"type":"string","description":"target node id"},
-					"to_slot":{"description":"input slot index or name"}
-				},
-				"required":["from","to"]
-			}`),
-			toolDef("move_node", "Move a node to a new position.", `{
-				"type":"object",
-				"properties":{"node_id":{"type":"string"},"pos":{"type":"array","items":{"type":"number"}}},
-				"required":["node_id","pos"]
-			}`),
-			toolDef("delete_node", "Remove a node from the graph.", `{
-				"type":"object",
-				"properties":{"node_id":{"type":"string"}},
-				"required":["node_id"]
-			}`),
-			toolDef("layout", "Auto-arrange the graph.", `{"type":"object","properties":{}}`),
-			toolDef("queue", "Queue the graph for execution.", `{"type":"object","properties":{}}`),
-		},
-	},
-	"create": {
-		Name:           "create",
-		ServerExecuted: true,
-		SystemPrompt: strings.TrimSpace(`
-You are Hanzo Create, a fashion and product render assistant. You help the user
-generate, edit, and refine product and fashion imagery. Use the tools available to
-this workspace — the org's connected render and MCP services — to run and iterate
-on renders. Call a tool when it advances the request; describe the result briefly.`),
-		// No builtin tools: the callable set is the org's registered MCP/registry
-		// tools, resolved per request and dispatched server-side.
-		BuiltinTools: nil,
-	},
-}
-
-// capabilityFor resolves a capability by name, defaulting an empty name to graph.
-// ok is false for an unknown name so the handler answers 400 rather than guessing.
-func capabilityFor(name string) (capability, bool) {
-	name = strings.TrimSpace(name)
-	if name == "" {
-		name = defaultCapability
-	}
-	c, ok := capabilities[name]
-	return c, ok
-}
-
-// toolDef builds an OpenAI function tool from a JSON-Schema string. The schema is
-// held as json.RawMessage so it reaches the model verbatim.
-func toolDef(name, desc, schema string) openai.Tool {
-	return openai.Tool{
-		Type: openai.ToolTypeFunction,
-		Function: &openai.FunctionDefinition{
-			Name:        name,
-			Description: desc,
-			Parameters:  json.RawMessage(schema),
-		},
-	}
-}

clients/chat/chat.go

@@ -1,334 +0,0 @@
-// Package chat is the ONE canonical chat orchestrator for Hanzo Cloud: POST
-// /v1/chat runs a single LLM tool-calling round that lets a model manage a system
-// through tools. It COMPOSES existing cloud pieces rather than reinventing them —
-//
-//   - LLM routing + per-org reserve/settle billing: the ai subsystem's
-//     /v1/chat/completions, invoked in-process (the ONLY path that both returns
-//     tool_calls and carries the billing gate);
-//   - the unified tool plane (clients/tools): the org's registered MCP/registry
-//     tools are offered to the model and dispatched server-side through the ONE
-//     per-principal plane (activation + price gates included);
-//   - capabilities (capability.go): the system prompt + builtin tool defs that
-//     frame a round for a use case (graph editing, product/fashion render).
-//
-// A round yields three things: the assistant's text (reply), the tool calls the
-// server executed against the registry (actions), and the tool calls the client
-// must apply itself — a graph/UI mutation the server cannot run (ops).
-//
-// Route ownership: chat OWNS /v1/chat. The ai module also registers a beego
-// /v1/chat alias behind its /v1/* catch-all, but chat mounts BEFORE ai in Wire, so
-// Fiber's first-match resolves /v1/chat here and the alias is shadowed (never
-// reached). ai keeps /v1/chat/completions and /v1/completions.
-package chat
-
-import (
-	"bytes"
-	"context"
-	"encoding/json"
-	"fmt"
-	"io"
-	"net/http"
-	"net/http/httptest"
-	"strings"
-
-	"github.com/hanzoai/cloud"
-	"github.com/hanzoai/cloud/clients/tools"
-	openai "github.com/sashabaranov/go-openai"
-	fiber "github.com/zap-proto/fiber/v3"
-	"github.com/zap-proto/zip"
-)
-
-// maxCompletionResponse bounds the in-process completion body read so a hostile or
-// broken upstream cannot balloon memory.
-const maxCompletionResponse = 8 << 20
-
-// completeFn runs one chat completion (with tools) and returns the parsed response.
-// It is the seam onto the ai subsystem's /v1/chat/completions: the real one replays
-// the request in-process carrying the caller's credential (so ai's per-org billing
-// runs); tests inject a fake so the round is exercised without a live model.
-type completeFn func(ctx context.Context, cred map[string]string, req openai.ChatCompletionRequest) (openai.ChatCompletionResponse, error)
-
-// state is the chat subsystem's own data. The three registry seams default to the
-// process-wide tool plane (tools.Default) and are overridable in tests, mirroring
-// the guide subsystem's invoke/toolOK seams.
-type state struct {
-	model    string
-	complete completeFn
-	known    func(ctx context.Context, scope tools.Scope, name string) bool
-	list     func(ctx context.Context, scope tools.Scope) []tools.Tool
-	dispatch func(ctx context.Context, p tools.Principal, name string, args map[string]any) (any, error)
-}
-
-// mounted is the active service, retained so tests can inject the seams above.
-var mounted *cloud.Service[state]
-
-// Mount wires POST /v1/chat onto app. It composes the ai completion path and the
-// tool plane; it owns no store, so there is no Shutdown.
-func Mount(app *zip.App, deps cloud.Deps) error {
-	if app == nil {
-		return fmt.Errorf("chat.Mount: nil zip.App")
-	}
-	if deps.Logger == nil {
-		return fmt.Errorf("chat.Mount: nil deps.Logger")
-	}
-	s := &cloud.Service[state]{Base: cloud.NewBase(deps, "chat"), State: state{
-		model:    strings.TrimSpace(deps.AIDefaultModel),
-		complete: appCompleter(app),
-		known: func(ctx context.Context, scope tools.Scope, name string) bool {
-			return tools.Default().Exists(ctx, scope, name)
-		},
-		list: func(ctx context.Context, scope tools.Scope) []tools.Tool {
-			return tools.Default().List(ctx, scope)
-		},
-		dispatch: func(ctx context.Context, p tools.Principal, name string, args map[string]any) (any, error) {
-			return tools.Default().Dispatch(ctx, p, name, args)
-		},
-	}}
-	mounted = s
-	app.Post("/v1/chat", cloud.Handle(s, handleChat))
-	s.Log.Info("chat mounted", "route", "/v1/chat", "capabilities", len(capabilities), "brand", deps.Brand)
-	return nil
-}
-
-// ── request / response ──────────────────────────────────────────────────────────
-
-type message struct {
-	Role    string `json:"role"`
-	Content string `json:"content"`
-}
-
-type chatRequest struct {
-	Messages   []message       `json:"messages"`
-	Capability string          `json:"capability"`
-	Tools      []openai.Tool   `json:"tools"`
-	System     string          `json:"system"`
-	State      json.RawMessage `json:"state"`
-}
-
-// action is one tool call the server executed against the registry.
-type action struct {
-	Name   string         `json:"name"`
-	Args   map[string]any `json:"args,omitempty"`
-	Result any            `json:"result,omitempty"`
-	Error  string         `json:"error,omitempty"`
-}
-
-// op is one tool call the client must apply itself (a graph/UI mutation).
-type op struct {
-	Name string         `json:"name"`
-	Args map[string]any `json:"args,omitempty"`
-}
-
-type chatResponse struct {
-	Reply   string   `json:"reply"`
-	Actions []action `json:"actions"`
-	Ops     []op     `json:"ops"`
-}
-
-// ── handler ─────────────────────────────────────────────────────────────────────
-
-// handleChat runs one tool-calling round. It resolves the caller, assembles the
-// tool set (capability builtins + caller tools + the org's registry tools), calls
-// the ai completion in-process, then splits the model's tool calls: a call the
-// registry knows (under a server-executed capability) is dispatched server-side and
-// reported as an action; every other call is returned as an op for the client.
-func handleChat(s *cloud.Service[state], c *zip.Ctx) error {
-	p, ok := tools.PrincipalFrom(c)
-	if !ok {
-		return zip.ErrForbidden("a validated principal is required")
-	}
-	var body chatRequest
-	if err := c.Bind(&body); err != nil {
-		return err
-	}
-	if len(body.Messages) == 0 {
-		return zip.ErrBadRequest("messages required")
-	}
-	mode, ok := capabilityFor(body.Capability)
-	if !ok {
-		return zip.ErrBadRequest("unknown capability: " + strings.TrimSpace(body.Capability))
-	}
-
-	scope := tools.Scope{Org: p.Org, Project: p.Project}
-	req := openai.ChatCompletionRequest{
-		Model:    s.State.model,
-		Messages: assembleMessages(mode, body),
-		Tools:    assembleTools(c.Context(), s, mode, body.Tools, scope),
-	}
-
-	resp, err := s.State.complete(c.Context(), credential(c), req)
-	if err != nil {
-		return zip.Errorf(http.StatusBadGateway, "chat: completion: %v", err)
-	}
-	if len(resp.Choices) == 0 {
-		return zip.Errorf(http.StatusBadGateway, "chat: completion returned no choices")
-	}
-	msg := resp.Choices[0].Message
-
-	out := chatResponse{Reply: msg.Content, Actions: []action{}, Ops: []op{}}
-	for _, tc := range msg.ToolCalls {
-		name := tc.Function.Name
-		args := decodeArgs(tc.Function.Arguments)
-		if mode.ServerExecuted && s.State.known(c.Context(), scope, name) {
-			a := action{Name: name, Args: args}
-			if result, derr := s.State.dispatch(c.Context(), p, name, args); derr != nil {
-				a.Error = derr.Error()
-			} else {
-				a.Result = result
-			}
-			out.Actions = append(out.Actions, a)
-			continue
-		}
-		out.Ops = append(out.Ops, op{Name: name, Args: args})
-	}
-	return c.JSON(http.StatusOK, out)
-}
-
-// ── assembly ────────────────────────────────────────────────────────────────────
-
-// assembleMessages prepends the round's system prompt (the capability's, plus any
-// caller-supplied system text) to the caller's messages.
-func assembleMessages(mode capability, body chatRequest) []openai.ChatCompletionMessage {
-	system := mode.SystemPrompt
-	if extra := strings.TrimSpace(body.System); extra != "" {
-		system = system + "\n\n" + extra
-	}
-	msgs := make([]openai.ChatCompletionMessage, 0, len(body.Messages)+1)
-	if system != "" {
-		msgs = append(msgs, openai.ChatCompletionMessage{Role: openai.ChatMessageRoleSystem, Content: system})
-	}
-	for _, m := range body.Messages {
-		role := strings.TrimSpace(m.Role)
-		if role == "" {
-			role = openai.ChatMessageRoleUser
-		}
-		msgs = append(msgs, openai.ChatCompletionMessage{Role: role, Content: m.Content})
-	}
-	return msgs
-}
-
-// assembleTools builds the tool array offered to the model: the capability's
-// builtin tools, then the caller-passed tool defs, then the org's ACTIVATED
-// registry tools — deduped by function name (an earlier, more-specific def wins a
-// name collision). Only activated + dispatchable registry tools are offered, the
-// same agent-facing projection the unified MCP endpoint uses (tools/http.go), so a
-// chat sees exactly the tools the org turned on (its registered MCP servers for the
-// create capability) rather than the whole route table.
-func assembleTools(ctx context.Context, s *cloud.Service[state], mode capability, caller []openai.Tool, scope tools.Scope) []openai.Tool {
-	out := make([]openai.Tool, 0, len(mode.BuiltinTools)+len(caller))
-	seen := map[string]bool{}
-	add := func(t openai.Tool) {
-		if t.Function == nil || t.Function.Name == "" || seen[t.Function.Name] {
-			return
-		}
-		seen[t.Function.Name] = true
-		out = append(out, t)
-	}
-	for _, t := range mode.BuiltinTools {
-		add(t)
-	}
-	for _, t := range caller {
-		add(t)
-	}
-	for _, t := range s.State.list(ctx, scope) {
-		if !t.Activated || !t.Dispatchable || seen[t.Name] {
-			continue
-		}
-		seen[t.Name] = true
-		out = append(out, openai.Tool{
-			Type: openai.ToolTypeFunction,
-			Function: &openai.FunctionDefinition{
-				Name:        t.Name,
-				Description: t.Description,
-				Parameters:  schemaOrEmpty(t.Schema),
-			},
-		})
-	}
-	return out
-}
-
-// schemaOrEmpty returns the tool's JSON-Schema, or the empty-object schema when a
-// source published none (the model requires a parameters object).
-func schemaOrEmpty(schema json.RawMessage) json.RawMessage {
-	if len(schema) == 0 {
-		return json.RawMessage(`{"type":"object","properties":{}}`)
-	}
-	return schema
-}
-
-// decodeArgs parses a tool call's JSON argument string into a map. A model may emit
-// empty or malformed arguments; that yields an empty map rather than an error, so a
-// no-arg tool call still dispatches / returns as an op.
-func decodeArgs(raw string) map[string]any {
-	raw = strings.TrimSpace(raw)
-	if raw == "" {
-		return map[string]any{}
-	}
-	var args map[string]any
-	if err := json.Unmarshal([]byte(raw), &args); err != nil || args == nil {
-		return map[string]any{}
-	}
-	return args
-}
-
-// ── in-process completion ───────────────────────────────────────────────────────
-
-// credentialHeaders are the caller's own credential headers replayed on the
-// in-process completion, so cloud's SanitizeIdentity re-derives the SAME validated
-// identity for the ai handler. No minted authority header (X-Org-Id/…) is replayed —
-// those are stripped on ingress and re-minted from the credential. This mirrors the
-// tool plane's builtin-replay contract (clients/tools).
-var credentialHeaders = []string{"Authorization", "X-Authorization", "Cookie", "Accept-Language", "X-Forwarded-For"}
-
-func credential(c *zip.Ctx) map[string]string {
-	cred := map[string]string{}
-	for _, h := range credentialHeaders {
-		if v := c.Header(h); v != "" {
-			cred[h] = v
-		}
-	}
-	return cred
-}
-
-// appCompleter is the real completeFn: it replays the request against the SAME
-// Fiber app in-process, so it flows through the whole middleware chain and lands on
-// the ai subsystem's /v1/chat/completions — the one path that returns tool_calls
-// AND carries per-org reserve/settle billing. Non-streaming, so the upstream
-// forwards a single JSON ChatCompletionResponse. This is the tool plane's in-process
-// dispatch pattern (clients/tools/builtin.go) applied to the LLM round.
-func appCompleter(app *zip.App) completeFn {
-	return func(ctx context.Context, cred map[string]string, req openai.ChatCompletionRequest) (openai.ChatCompletionResponse, error) {
-		req.Stream = false
-		b, err := json.Marshal(req)
-		if err != nil {
-			return openai.ChatCompletionResponse{}, err
-		}
-		hreq := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", bytes.NewReader(b)).WithContext(ctx)
-		hreq.Header.Set("Content-Type", "application/json")
-		for k, v := range cred {
-			hreq.Header.Set(k, v)
-		}
-		resp, err := app.Fiber().Test(hreq, fiber.TestConfig{Timeout: 0})
-		if err != nil {
-			return openai.ChatCompletionResponse{}, err
-		}
-		defer func() { _ = resp.Body.Close() }()
-		raw, _ := io.ReadAll(io.LimitReader(resp.Body, maxCompletionResponse))
-		if resp.StatusCode/100 != 2 {
-			return openai.ChatCompletionResponse{}, fmt.Errorf("upstream status %d: %s", resp.StatusCode, clip(raw))
-		}
-		var out openai.ChatCompletionResponse
-		if err := json.Unmarshal(raw, &out); err != nil {
-			return openai.ChatCompletionResponse{}, fmt.Errorf("decode completion: %w", err)
-		}
-		return out, nil
-	}
-}
-
-func clip(b []byte) string {
-	const max = 256
-	if len(b) > max {
-		return string(b[:max])
-	}
-	return string(b)
-}

clients/chat/chat_test.go

@@ -1,243 +0,0 @@
-package chat
-
-import (
-	"bytes"
-	"context"
-	"encoding/json"
-	"fmt"
-	"io"
-	"net/http"
-	"net/http/httptest"
-	"sort"
-	"testing"
-
-	"github.com/hanzoai/cloud"
-	"github.com/hanzoai/cloud/clients/tools"
-	luxlog "github.com/luxfi/log"
-	openai "github.com/sashabaranov/go-openai"
-	fiber "github.com/zap-proto/fiber/v3"
-	"github.com/zap-proto/zip"
-)
-
-// newApp mounts chat on a fresh zip.App. The ai completion path and the tool
-// registry are stubbed on the mounted service per test — no live model or plane is
-// wired, so the tests exercise the round assembly + the action/ops split in
-// isolation. list is stubbed to empty so the round never depends on the process-wide
-// registry.
-func newApp(t *testing.T) *zip.App {
-	t.Helper()
-	app := zip.New(zip.Config{Logger: luxlog.New("test")})
-	deps := cloud.Deps{Logger: luxlog.New("test"), AIDefaultModel: "zen"}
-	if err := Mount(app, deps); err != nil {
-		t.Fatalf("Mount: %v", err)
-	}
-	mounted.State.list = func(context.Context, tools.Scope) []tools.Tool { return nil }
-	return app
-}
-
-// req issues one POST /v1/chat. org != "" sets the validated identity headers
-// (X-Org-Id + X-User-Id) exactly as SanitizeIdentity would; org == "" sends none —
-// the anonymous path the 403 test needs.
-func req(t *testing.T, app *zip.App, org string, body any) (int, chatResponse) {
-	t.Helper()
-	var r io.Reader
-	if body != nil {
-		b, _ := json.Marshal(body)
-		r = bytes.NewReader(b)
-	}
-	rq := httptest.NewRequest(http.MethodPost, "/v1/chat", r)
-	rq.Header.Set("Content-Type", "application/json")
-	if org != "" {
-		rq.Header.Set("X-Org-Id", org)
-		rq.Header.Set("X-User-Id", "u-"+org)
-	}
-	resp, err := app.Fiber().Test(rq, fiber.TestConfig{Timeout: 0})
-	if err != nil {
-		t.Fatalf("Test POST /v1/chat: %v", err)
-	}
-	defer func() { _ = resp.Body.Close() }()
-	raw, _ := io.ReadAll(resp.Body)
-	var out chatResponse
-	_ = json.Unmarshal(raw, &out)
-	return resp.StatusCode, out
-}
-
-// completion builds a fake assistant completion: optional text plus zero or more
-// tool calls (each with a trivial argument object).
-func completion(content string, toolCalls ...string) openai.ChatCompletionResponse {
-	msg := openai.ChatCompletionMessage{Role: openai.ChatMessageRoleAssistant, Content: content}
-	for i, name := range toolCalls {
-		msg.ToolCalls = append(msg.ToolCalls, openai.ToolCall{
-			ID:       fmt.Sprintf("call_%d", i),
-			Type:     openai.ToolTypeFunction,
-			Function: openai.FunctionCall{Name: name, Arguments: `{"x":1}`},
-		})
-	}
-	return openai.ChatCompletionResponse{Choices: []openai.ChatCompletionChoice{{Message: msg}}}
-}
-
-func names[T any](xs []T, name func(T) string) []string {
-	out := make([]string, 0, len(xs))
-	for _, x := range xs {
-		out = append(out, name(x))
-	}
-	sort.Strings(out)
-	return out
-}
-
-func eq(a, b []string) bool {
-	if len(a) != len(b) {
-		return false
-	}
-	for i := range a {
-		if a[i] != b[i] {
-			return false
-		}
-	}
-	return true
-}
-
-// TestChatRound is the table test for the tool-calling round: a registry tool call
-// dispatches to actions; an unknown call falls to ops; plain content is the reply; a
-// graph-capability call is always an op (advisory); an unknown capability is 400.
-func TestChatRound(t *testing.T) {
-	cases := []struct {
-		name           string
-		capability     string
-		completion     openai.ChatCompletionResponse
-		known          map[string]bool
-		wantStatus     int
-		wantReply      string
-		wantActions    []string
-		wantOps        []string
-		wantDispatched []string
-	}{
-		{
-			name:           "known registry tool dispatched to actions",
-			capability:     "create",
-			completion:     completion("", "render_image"),
-			known:          map[string]bool{"render_image": true},
-			wantStatus:     http.StatusOK,
-			wantActions:    []string{"render_image"},
-			wantOps:        []string{},
-			wantDispatched: []string{"render_image"},
-		},
-		{
-			name:           "unknown tool falls to ops",
-			capability:     "create",
-			completion:     completion("", "frobnicate"),
-			known:          map[string]bool{},
-			wantStatus:     http.StatusOK,
-			wantActions:    []string{},
-			wantOps:        []string{"frobnicate"},
-			wantDispatched: []string{},
-		},
-		{
-			name:           "plain content is the reply",
-			capability:     "graph",
-			completion:     completion("here is your graph plan"),
-			known:          map[string]bool{},
-			wantStatus:     http.StatusOK,
-			wantReply:      "here is your graph plan",
-			wantActions:    []string{},
-			wantOps:        []string{},
-			wantDispatched: []string{},
-		},
-		{
-			name:           "graph capability calls are advisory ops, never dispatched",
-			capability:     "graph",
-			completion:     completion("", "add_node"),
-			known:          map[string]bool{"add_node": true}, // known, but graph never dispatches
-			wantStatus:     http.StatusOK,
-			wantActions:    []string{},
-			wantOps:        []string{"add_node"},
-			wantDispatched: []string{},
-		},
-		{
-			name:       "unknown capability is 400",
-			capability: "bogus",
-			completion: completion(""),
-			wantStatus: http.StatusBadRequest,
-		},
-	}
-
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			app := newApp(t)
-			var dispatched []string
-			mounted.State.complete = func(context.Context, map[string]string, openai.ChatCompletionRequest) (openai.ChatCompletionResponse, error) {
-				return tc.completion, nil
-			}
-			mounted.State.known = func(_ context.Context, _ tools.Scope, name string) bool { return tc.known[name] }
-			mounted.State.dispatch = func(_ context.Context, _ tools.Principal, name string, _ map[string]any) (any, error) {
-				dispatched = append(dispatched, name)
-				return map[string]any{"ok": true, "tool": name}, nil
-			}
-
-			status, out := req(t, app, "acme", map[string]any{
-				"capability": tc.capability,
-				"messages":   []message{{Role: "user", Content: "do the thing"}},
-			})
-			if status != tc.wantStatus {
-				t.Fatalf("status: want %d, got %d", tc.wantStatus, status)
-			}
-			if tc.wantStatus != http.StatusOK {
-				return
-			}
-			if out.Reply != tc.wantReply {
-				t.Fatalf("reply: want %q, got %q", tc.wantReply, out.Reply)
-			}
-			gotActions := names(out.Actions, func(a action) string { return a.Name })
-			if !eq(gotActions, tc.wantActions) {
-				t.Fatalf("actions: want %v, got %v", tc.wantActions, gotActions)
-			}
-			gotOps := names(out.Ops, func(o op) string { return o.Name })
-			if !eq(gotOps, tc.wantOps) {
-				t.Fatalf("ops: want %v, got %v", tc.wantOps, gotOps)
-			}
-			sort.Strings(dispatched)
-			if dispatched == nil {
-				dispatched = []string{}
-			}
-			if !eq(dispatched, tc.wantDispatched) {
-				t.Fatalf("dispatched: want %v, got %v", tc.wantDispatched, dispatched)
-			}
-			// A dispatched action carries the registry result.
-			for _, a := range out.Actions {
-				if a.Error != "" || a.Result == nil {
-					t.Fatalf("action %s should carry a result, got %+v", a.Name, a)
-				}
-			}
-		})
-	}
-}
-
-// TestChatRequiresPrincipal: a request with no validated identity is refused 403 —
-// a forged X-Org-Id with no credential never runs a round.
-func TestChatRequiresPrincipal(t *testing.T) {
-	app := newApp(t)
-	mounted.State.complete = func(context.Context, map[string]string, openai.ChatCompletionRequest) (openai.ChatCompletionResponse, error) {
-		t.Fatal("completion must not run without a principal")
-		return openai.ChatCompletionResponse{}, nil
-	}
-	status, _ := req(t, app, "", map[string]any{
-		"capability": "graph",
-		"messages":   []message{{Role: "user", Content: "hi"}},
-	})
-	if status != http.StatusForbidden {
-		t.Fatalf("no principal: want 403, got %d", status)
-	}
-}
-
-// TestChatRequiresMessages: an empty messages array is 400 before any completion.
-func TestChatRequiresMessages(t *testing.T) {
-	app := newApp(t)
-	mounted.State.complete = func(context.Context, map[string]string, openai.ChatCompletionRequest) (openai.ChatCompletionResponse, error) {
-		t.Fatal("completion must not run with no messages")
-		return openai.ChatCompletionResponse{}, nil
-	}
-	status, _ := req(t, app, "acme", map[string]any{"capability": "graph", "messages": []message{}})
-	if status != http.StatusBadRequest {
-		t.Fatalf("empty messages: want 400, got %d", status)
-	}
-}

clients/cloudflare/cloudflare.go

@@ -0,0 +1,400 @@
+// Package cloudflare is the per-org Cloudflare asset plane for the unified Hanzo
+// Cloud binary — the /v1/integrations/cloudflare/* surface that manages an org's Cloudflare
+// Pages, Workers, and (Phase 2) R2/KV/D1 through the SAME per-org, KMS-sealed API
+// token the org connected via clients/integrations. It is a sibling of hanzodns
+// (which owns /v1/dns as a separate CoreDNS process): both drive Cloudflare with an
+// org's own scoped token, so the platform never reaches Cloudflare with a global
+// env token again — one token, one custody boundary, one org.
+//
+// TENANT ISOLATION (the crown jewel). Every handler resolves the caller's org from
+// the VALIDATED principal (principal.Org → the X-Org-Id the identity boundary minted
+// from a verified credential, HIP-0026 / SanitizeIdentity), NEVER from a body or
+// query field. The org is then the ONLY input to token custody: the per-org token is
+// read in-process through the ONE seam integrations.TokenFor, which keys KMS on that
+// org (/orgs/{org}/integrations/cloudflare/api_token). So a request can ONLY ever
+// address its own org's Cloudflare account:
+//   - no validated principal ⟹ principal.Org fails ⟹ 403 (a forged X-Org-Id with no
+//     bearer is refused by the identity boundary, then again here);
+//   - a non-SuperAdmin bearer has X-Org-Id pinned to its own owner (SanitizeIdentity),
+//     so it cannot name another org;
+//   - cross-org token reach is structurally impossible — the token path is derived
+//     from the validated org, not from any caller-controlled field.
+//
+// The token rides ONLY the Authorization header on the outbound Cloudflare request;
+// it is never logged, echoed in an error, or stored by this subsystem.
+//
+// FAIL-CLOSED. An org that has not connected Cloudflare, an unmounted integrations
+// plane, or a KMS that is not Ready each yield an error and a 503 — never another
+// org's data and never a silent success.
+package cloudflare
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"os"
+	"regexp"
+	"strings"
+	"time"
+
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/integrations"
+	"github.com/hanzoai/cloud/clients/principal"
+	"github.com/zap-proto/zip"
+)
+
+const (
+	// providerCloudflare is the integrations provider slug the token is custodied
+	// under, and secretAPIToken the secret name — the SAME coordinate the connector
+	// (clients/integrations/cloudflare.go) seals BOTH the apikey and OAuth paths to,
+	// and hanzodns reads for DNS. One coordinate, auth-method-agnostic.
+	providerCloudflare = "cloudflare"
+	secretAPIToken     = "api_token"
+)
+
+// tokenFor is the ONE door to per-org Cloudflare token custody. It defaults to
+// integrations.TokenFor (KMS-sealed, fail-closed, org-validated). It is a package
+// var ONLY so a test can inject per-org tokens and prove every fetch scopes to the
+// caller's org; production never reassigns it.
+var tokenFor = integrations.TokenFor
+
+// connectionFor reads an org's NON-secret connection metadata (the account id captured
+// at connect time, ExternalID) from the integrations plane. Also a package var ONLY
+// for test injection; production never reassigns it.
+var connectionFor = integrations.ConnectionFor
+
+// cfAPIBase is Cloudflare's API v4 origin. Overridable via CLOUDFLARE_API_BASE for
+// tests (an httptest server) and CF-compatible endpoints; read at call time. The
+// default is the real Cloudflare API. (Same knob hanzodns uses, so a test harness
+// points both planes at one stub.)
+func cfAPIBase() string {
+	if v := strings.TrimSpace(os.Getenv("CLOUDFLARE_API_BASE")); v != "" {
+		return strings.TrimRight(v, "/")
+	}
+	return "https://api.cloudflare.com/client/v4"
+}
+
+// cfHTTPClient is the shared client for every Cloudflare call. A bounded timeout so a
+// slow/hung Cloudflare never wedges a request goroutine.
+var cfHTTPClient = &http.Client{Timeout: 30 * time.Second}
+
+var (
+	// nameRE bounds a Cloudflare NAME path segment (Pages project, Worker script,
+	// custom-domain name/id, bucket, namespace, database). It is validated before it
+	// is folded into an upstream URL so a hostile value can never smuggle path
+	// structure (a `/` or `..`) into the Cloudflare request.
+	nameRE = regexp.MustCompile(`^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$`)
+	// idRE bounds a Cloudflare 32-hex ID path segment (account id, zone id, route id).
+	idRE = regexp.MustCompile(`^[0-9a-fA-F]{32}$`)
+)
+
+// state is this subsystem's own data — none. Token custody lives in integrations;
+// the account id is resolved per request. So the value is the shared Base only.
+type state struct{}
+
+// Mount wires /v1/integrations/cloudflare/* onto app. The subsystem is stateless (no store, no
+// goroutine): it reads the per-org token in-process per request and proxies to
+// Cloudflare, so there is nothing to build or tear down.
+func Mount(app *zip.App, deps cloud.Deps) error {
+	return cloud.Mount(app, deps, "cloudflare",
+		func(cloud.Base) (state, error) { return state{}, nil },
+		routes)
+}
+
+// routes registers the /v1/integrations/cloudflare surface. Every route runs through authClient
+// (validated-org gate + fail-closed per-org token) FIRST, so no route is a softer
+// target than another. Pages + Workers are WIRED; R2/KV/D1 are typed Phase-2 stubs
+// that answer an honest 501 (never a fake success).
+func routes(app *zip.App, s *cloud.Service[state]) {
+	// Pages (wired) — account-scoped.
+	app.Get("/v1/integrations/cloudflare/pages/projects", cloud.Handle(s, pagesList))
+	app.Post("/v1/integrations/cloudflare/pages/projects", cloud.Handle(s, pagesCreate))
+	app.Get("/v1/integrations/cloudflare/pages/projects/:project", cloud.Handle(s, pagesGet))
+	app.Delete("/v1/integrations/cloudflare/pages/projects/:project", cloud.Handle(s, pagesDelete))
+	app.Post("/v1/integrations/cloudflare/pages/projects/:project/deployments", cloud.Handle(s, pagesDeploy))
+	app.Post("/v1/integrations/cloudflare/pages/projects/:project/domains", cloud.Handle(s, pagesDomainAdd))
+	app.Delete("/v1/integrations/cloudflare/pages/projects/:project/domains/:domain", cloud.Handle(s, pagesDomainDelete))
+
+	// Workers (wired) — scripts + workers.dev subdomain are account-scoped; routes
+	// are zone-scoped.
+	app.Get("/v1/integrations/cloudflare/workers/scripts", cloud.Handle(s, workersScriptList))
+	app.Put("/v1/integrations/cloudflare/workers/scripts/:script", cloud.Handle(s, workersScriptPut))
+	app.Delete("/v1/integrations/cloudflare/workers/scripts/:script", cloud.Handle(s, workersScriptDelete))
+	app.Post("/v1/integrations/cloudflare/workers/scripts/:script/subdomain", cloud.Handle(s, workersScriptSubdomainSet))
+	app.Get("/v1/integrations/cloudflare/workers/subdomain", cloud.Handle(s, workersSubdomainGet))
+	app.Get("/v1/integrations/cloudflare/workers/zones/:zone/routes", cloud.Handle(s, workersRouteList))
+	app.Post("/v1/integrations/cloudflare/workers/zones/:zone/routes", cloud.Handle(s, workersRouteCreate))
+	app.Delete("/v1/integrations/cloudflare/workers/zones/:zone/routes/:route", cloud.Handle(s, workersRouteDelete))
+
+	// R2 / KV / D1 (Phase-2 stubs) — routes + typed provider methods exist; bodies
+	// ship in Phase 2. Each answers an honest 501, never a misleading 200.
+	app.Get("/v1/integrations/cloudflare/r2/buckets", cloud.Handle(s, r2BucketList))
+	app.Post("/v1/integrations/cloudflare/r2/buckets", cloud.Handle(s, r2BucketCreate))
+	app.Delete("/v1/integrations/cloudflare/r2/buckets/:bucket", cloud.Handle(s, r2BucketDelete))
+	app.Get("/v1/integrations/cloudflare/kv/namespaces", cloud.Handle(s, kvNamespaceList))
+	app.Post("/v1/integrations/cloudflare/kv/namespaces", cloud.Handle(s, kvNamespaceCreate))
+	app.Delete("/v1/integrations/cloudflare/kv/namespaces/:namespace", cloud.Handle(s, kvNamespaceDelete))
+	app.Get("/v1/integrations/cloudflare/d1/databases", cloud.Handle(s, d1DatabaseList))
+	app.Post("/v1/integrations/cloudflare/d1/databases", cloud.Handle(s, d1DatabaseCreate))
+	app.Delete("/v1/integrations/cloudflare/d1/databases/:database", cloud.Handle(s, d1DatabaseDelete))
+}
+
+// ── client (the cfDo shape, reused verbatim from hanzodns) ──────────────────────
+
+// client drives the Cloudflare API v4 with a per-org scoped token. The token rides
+// only the Authorization header — never a query parameter, error, or log line.
+type client struct {
+	token string
+	base  string
+}
+
+// cfEnvelope is the shared Cloudflare API v4 response envelope.
+type cfEnvelope struct {
+	Success bool `json:"success"`
+	Errors  []struct {
+		Code    int    `json:"code"`
+		Message string `json:"message"`
+	} `json:"errors"`
+}
+
+// cfError carries the upstream Cloudflare HTTP status so a proxied not-found/bad
+// request is reported with a recognizable code rather than a blanket 502. Its
+// message is Cloudflare's own — token-free by construction.
+type cfError struct {
+	upstream int
+	code     int
+	msg      string
+}
+
+func (e *cfError) Error() string { return e.msg }
+
+func (e cfEnvelope) err(status int) error {
+	if e.Success {
+		return nil
+	}
+	if len(e.Errors) > 0 {
+		return &cfError{upstream: status, code: e.Errors[0].Code,
+			msg: fmt.Sprintf("cloudflare API error (%d): [%d] %s", status, e.Errors[0].Code, e.Errors[0].Message)}
+	}
+	return &cfError{upstream: status, msg: fmt.Sprintf("cloudflare API error (status %d)", status)}
+}
+
+// cfDo performs a Cloudflare API v4 call, JSON-encoding body, unwrapping the
+// {success, errors, result} envelope, and decoding result into out. It fails closed:
+// a transport error, an unsuccessful envelope, or a non-2xx status yields an error,
+// and the error NEVER contains the token. (The hanzodns cfDo shape, verbatim.)
+func (cl *client) cfDo(ctx context.Context, method, path string, body, out any) error {
+	var reader io.Reader
+	var contentType string
+	if body != nil {
+		b, err := json.Marshal(body)
+		if err != nil {
+			return err
+		}
+		reader = bytes.NewReader(b)
+		contentType = "application/json"
+	}
+	return cl.do(ctx, method, path, contentType, reader, out)
+}
+
+// cfUpload performs a Cloudflare call with a pre-built body + content type (the
+// multipart Worker-script upload), sharing cfDo's fail-closed envelope handling.
+func (cl *client) cfUpload(ctx context.Context, method, path, contentType string, body []byte, out any) error {
+	return cl.do(ctx, method, path, contentType, bytes.NewReader(body), out)
+}
+
+// do is the shared request core for cfDo and cfUpload: Bearer-only auth, bounded
+// response read, envelope unwrap, fail-closed. Token appears ONLY in the header.
+func (cl *client) do(ctx context.Context, method, path, contentType string, body io.Reader, out any) error {
+	req, err := http.NewRequestWithContext(ctx, method, cl.base+path, body)
+	if err != nil {
+		return err
+	}
+	req.Header.Set("Authorization", "Bearer "+cl.token)
+	if contentType != "" {
+		req.Header.Set("Content-Type", contentType)
+	}
+
+	resp, err := cfHTTPClient.Do(req)
+	if err != nil {
+		// Never wrap err: a transport error can echo the request URL but never the
+		// header. Keep the message token-free regardless.
+		return fmt.Errorf("cloudflare request failed")
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	if resp.StatusCode == http.StatusNoContent {
+		return nil
+	}
+
+	data, _ := io.ReadAll(io.LimitReader(resp.Body, 1<<20))
+
+	var env cfEnvelope
+	if len(data) > 0 {
+		_ = json.Unmarshal(data, &env)
+	}
+	if err := env.err(resp.StatusCode); err != nil {
+		return err
+	}
+	if out != nil {
+		var wrap struct {
+			Result json.RawMessage `json:"result"`
+		}
+		if err := json.Unmarshal(data, &wrap); err != nil {
+			return fmt.Errorf("cloudflare: malformed response")
+		}
+		if len(wrap.Result) > 0 {
+			if err := json.Unmarshal(wrap.Result, out); err != nil {
+				return fmt.Errorf("cloudflare: malformed result")
+			}
+		}
+	}
+	return nil
+}
+
+// pass runs a Cloudflare call and relays its result to the caller VERBATIM (as raw
+// JSON), so the upstream shape reaches the platform without field loss — the ONE
+// response path for every wired handler. An empty result (e.g. a 204 delete) becomes
+// {"success":true}.
+func (cl *client) pass(c *zip.Ctx, method, path string, body any) error {
+	var out json.RawMessage
+	if err := cl.cfDo(c.Context(), method, path, body, &out); err != nil {
+		return cfErr(err)
+	}
+	return writeResult(c, out)
+}
+
+func writeResult(c *zip.Ctx, out json.RawMessage) error {
+	if len(out) == 0 {
+		return c.JSON(http.StatusOK, map[string]any{"success": true})
+	}
+	c.SetHeader("Content-Type", "application/json; charset=utf-8")
+	return c.Bytes(http.StatusOK, out)
+}
+
+// ── request gate + token custody ────────────────────────────────────────────────
+
+// actingOrgHeader is stamped on every SERVED /v1/integrations/cloudflare response with the org
+// whose Cloudflare token was actually used. A per-org caller (the platform BFF) MUST
+// assert it equals the org it requested: if a misdeployed, non-org-switch-capable
+// service credential made the identity boundary PIN X-Org-Id to the token's OWN
+// owner, this header exposes the mismatch so the caller fails LOUD instead of
+// silently reading/writing another tenant's Cloudflare account.
+const actingOrgHeader = "X-Hanzo-Org"
+
+// authClient is the READ front door: it resolves the caller's validated org (403 if
+// unvalidated — a forged X-Org-Id with no bearer never gets past this) and builds a
+// Cloudflare client bound to THAT org's KMS-sealed token (503 if the org has not
+// connected Cloudflare or KMS is down). On success it stamps actingOrgHeader with the
+// served org. The token detail is logged token-free and never surfaced to the client.
+func authClient(s *cloud.Service[state], c *zip.Ctx) (*client, string, error) {
+	org, ok := principal.Org(c)
+	if !ok {
+		return nil, "", zip.ErrForbidden("a validated principal is required")
+	}
+	tok, err := tokenFor(c.Context(), org, providerCloudflare, secretAPIToken)
+	if err != nil || len(bytes.TrimSpace(tok)) == 0 {
+		// err is custody-authored and token-free (not-connected / invalid-org /
+		// KMS-down). Log the reason, tell the client only that CF is unavailable.
+		s.Log.Warn("cloudflare token unavailable", "org", org, "err", err)
+		return nil, org, zip.Errorf(http.StatusServiceUnavailable, "cloudflare is not connected for this org")
+	}
+	// Stamp the org actually served so a per-org caller can prove no tenant comingling.
+	c.SetHeader(actingOrgHeader, org)
+	return &client{token: string(bytes.TrimSpace(tok)), base: cfAPIBase()}, org, nil
+}
+
+// authWrite is the MUTATION front door (POST/PUT/DELETE): it additionally requires the
+// caller be an admin of its OWN org (principal.IsOrgAdmin — NOT SuperAdmin), parity
+// with the AdminOnly connector that seals the token. Wielding the token's dangerous
+// verbs (a Worker script PUT is arbitrary code on the org's Cloudflare account/domains;
+// a Pages project DELETE is production destruction) must match connecting it. The admin
+// check is FIRST, so a non-admin is refused before any KMS token read. Reads stay
+// validated-org-only via authClient — org members may look, only org admins may change.
+func authWrite(s *cloud.Service[state], c *zip.Ctx) (*client, string, error) {
+	if !principal.IsOrgAdmin(c) {
+		return nil, "", zip.ErrForbidden("this action requires org admin")
+	}
+	return authClient(s, c)
+}
+
+// resolveAccount resolves the Cloudflare account id for account-scoped endpoints
+// (Pages / Workers). Order: (1) an explicit, validated ?account= override wins (for an
+// org whose token spans multiple accounts); (2) the account captured at connect time
+// (the connection's ExternalID) — no per-call round-trip and deterministic for a
+// multi-account token; (3) only if none is stored, discover it live from the token's
+// own /accounts. Every candidate is validated 32-hex so it can never inject path
+// structure. Fails closed (400) when nothing yields a usable account.
+func (cl *client) resolveAccount(ctx context.Context, org string, c *zip.Ctx) (string, error) {
+	if a := strings.TrimSpace(c.Query("account")); a != "" {
+		if !idRE.MatchString(a) {
+			return "", zip.ErrBadRequest("account must be a 32-character hex id")
+		}
+		return url.PathEscape(a), nil
+	}
+	if conn, ok := connectionFor(org, providerCloudflare); ok {
+		if id := strings.TrimSpace(conn.ExternalID); idRE.MatchString(id) {
+			return url.PathEscape(id), nil
+		}
+	}
+	var accts []struct {
+		ID string `json:"id"`
+	}
+	if err := cl.cfDo(ctx, http.MethodGet, "/accounts?per_page=1", nil, &accts); err != nil {
+		return "", cfErr(err)
+	}
+	for _, a := range accts {
+		if idRE.MatchString(a.ID) {
+			return url.PathEscape(a.ID), nil
+		}
+	}
+	return "", zip.ErrBadRequest("no cloudflare account is resolvable for this token; pass ?account=<id>")
+}
+
+// pathSeg reads a route param, rejects anything not matching re (so it can never
+// smuggle path structure into the upstream Cloudflare URL), and returns the
+// url.PathEscape'd value ready to concatenate into a CF path.
+func pathSeg(c *zip.Ctx, name string, re *regexp.Regexp) (string, error) {
+	v := strings.TrimSpace(c.Param(name))
+	if !re.MatchString(v) {
+		return "", zip.ErrBadRequest(name + " is invalid")
+	}
+	return url.PathEscape(v), nil
+}
+
+// cfErr maps a Cloudflare call failure to a client-facing HTTP error, propagating a
+// recognizable upstream status (404/400/403/409) so a proxied not-found is not
+// mis-reported as a 502, and defaulting to 502 Bad Gateway otherwise. The message is
+// Cloudflare's own (token-free by construction), never this process's token.
+func cfErr(err error) error {
+	var ce *cfError
+	if errors.As(err, &ce) {
+		switch ce.upstream {
+		case http.StatusNotFound, http.StatusBadRequest, http.StatusForbidden, http.StatusConflict:
+			return zip.Errorf(ce.upstream, "%s", ce.msg)
+		}
+	}
+	return zip.Errorf(http.StatusBadGateway, "%s", err.Error())
+}
+
+// ── Phase-2 stub plumbing ───────────────────────────────────────────────────────
+
+// errPhase2 marks a provider capability whose route + typed method exist but whose
+// body ships in Phase 2. stubResult maps it to an honest 501 — never a fake 200, so
+// a caller is never misled into thinking a no-op succeeded.
+var errPhase2 = errors.New("cloudflare: not yet implemented (phase 2)")
+
+// stubResult maps a Phase-2 provider seam's result to the wire: errPhase2 → 501, any
+// real error surfaces as-is, and even a nil still yields 501 (a stub can never report
+// success). This guarantees a stub route NEVER returns a misleading 200.
+func stubResult(c *zip.Ctx, capability string, err error) error {
+	if err != nil && !errors.Is(err, errPhase2) {
+		return cfErr(err)
+	}
+	return zip.Errorf(http.StatusNotImplemented, "cloudflare %s is not yet wired (phase 2)", capability)
+}

clients/cloudflare/cloudflare_test.go

@@ -0,0 +1,446 @@
+package cloudflare
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"testing"
+
+	luxlog "github.com/luxfi/log"
+	"github.com/zap-proto/zip"
+
+	"github.com/hanzoai/cloud"
+	"github.com/hanzoai/cloud/clients/integrations"
+)
+
+// testAccountID is a valid 32-hex Cloudflare account id the stub discovers.
+const testAccountID = "0123456789abcdef0123456789abcdef"
+
+// capture records every request the fake Cloudflare API received, so a test can
+// assert WHICH token (Authorization) and WHICH path a handler used.
+type capture struct {
+	mu   sync.Mutex
+	reqs []capturedReq
+}
+
+type capturedReq struct {
+	method, path, auth, ctype string
+	body                      []byte
+}
+
+func (c *capture) add(r *http.Request) {
+	body, _ := io.ReadAll(r.Body)
+	c.mu.Lock()
+	c.reqs = append(c.reqs, capturedReq{r.Method, r.URL.Path, r.Header.Get("Authorization"), r.Header.Get("Content-Type"), body})
+	c.mu.Unlock()
+}
+
+// find returns the first captured request whose path CONTAINS sub.
+func (c *capture) find(sub string) (capturedReq, bool) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	for _, r := range c.reqs {
+		if strings.Contains(r.path, sub) {
+			return r, true
+		}
+	}
+	return capturedReq{}, false
+}
+
+// hasExact reports whether any captured request hit EXACTLY path p — used to detect
+// the account-discovery call (GET /accounts), which "/accounts/{id}/..." contains as
+// a substring but is not.
+func (c *capture) hasExact(p string) bool {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	for _, r := range c.reqs {
+		if r.path == p {
+			return true
+		}
+	}
+	return false
+}
+
+// fakeCF is a minimal Cloudflare API v4 stub: it answers account discovery and
+// echoes a success envelope for any account/zone-scoped call, recording every
+// request. resultFor lets a test control the result body per path substring.
+func fakeCF(rec *capture, resultFor func(path string) (int, string)) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		rec.add(r)
+		w.Header().Set("Content-Type", "application/json")
+		if r.URL.Path == "/accounts" { // discovery
+			io.WriteString(w, `{"success":true,"errors":[],"result":[{"id":"`+testAccountID+`","name":"acme"}]}`)
+			return
+		}
+		if resultFor != nil {
+			if status, body := resultFor(r.URL.Path); body != "" {
+				w.WriteHeader(status)
+				io.WriteString(w, body)
+				return
+			}
+		}
+		io.WriteString(w, `{"success":true,"errors":[],"result":{"ok":true}}`)
+	}
+}
+
+// harness mounts the subsystem against a fake CF and per-org token seam.
+func harness(t *testing.T, tokens map[string]string, rec *capture, resultFor func(string) (int, string)) *zip.App {
+	t.Helper()
+	srv := httptest.NewServer(fakeCF(rec, resultFor))
+	t.Cleanup(srv.Close)
+	t.Setenv("CLOUDFLARE_API_BASE", srv.URL)
+
+	prev := tokenFor
+	tokenFor = func(_ context.Context, org, provider, name string) ([]byte, error) {
+		if provider != providerCloudflare || name != secretAPIToken {
+			return nil, fmt.Errorf("seam called with unexpected coordinate %s/%s", provider, name)
+		}
+		tok, ok := tokens[org]
+		if !ok {
+			return nil, fmt.Errorf("cloudflare not connected for org %q", org)
+		}
+		return []byte(tok), nil
+	}
+	t.Cleanup(func() { tokenFor = prev })
+
+	app := zip.New(zip.Config{Logger: luxlog.New("test")})
+	if err := Mount(app, cloud.Deps{Logger: luxlog.New("test"), DataDir: t.TempDir()}); err != nil {
+		t.Fatalf("Mount: %v", err)
+	}
+	return app
+}
+
+// doReq drives one request with the given minted identity headers (as
+// SanitizeIdentity would set them): user!="" makes a validated principal; admin sets
+// X-User-IsOrgAdmin. Returns status + body + response headers.
+func doReq(t *testing.T, app *zip.App, method, path, user, org string, admin bool, body string) (int, string, http.Header) {
+	t.Helper()
+	var rdr io.Reader
+	if body != "" {
+		rdr = strings.NewReader(body)
+	}
+	req := httptest.NewRequest(method, path, rdr)
+	if user != "" {
+		req.Header.Set("X-User-Id", user)
+	}
+	if org != "" {
+		req.Header.Set("X-Org-Id", org)
+	}
+	if admin {
+		req.Header.Set("X-User-IsOrgAdmin", "true")
+	}
+	if body != "" {
+		req.Header.Set("Content-Type", "application/json")
+	}
+	resp, err := app.Fiber().Test(req)
+	if err != nil {
+		t.Fatalf("Test(%s %s): %v", method, path, err)
+	}
+	defer func() { _ = resp.Body.Close() }()
+	b, _ := io.ReadAll(resp.Body)
+	return resp.StatusCode, string(b), resp.Header
+}
+
+// do is the common non-admin read/driver: validated principal, no admin bit.
+func do(t *testing.T, app *zip.App, method, path, user, org, body string) (int, string) {
+	status, b, _ := doReq(t, app, method, path, user, org, false, body)
+	return status, b
+}
+
+// ── tenant isolation (the crown jewel) ──────────────────────────────────────────
+
+// A request with no validated principal (no X-User-Id) is refused 403 and NEVER
+// reaches token custody or Cloudflare — the forged-X-Org-Id path is dead.
+func TestForgedOrgWithoutPrincipalIs403(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"victim": "tok-victim"}, rec, nil)
+
+	status, body := do(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects", "", "victim", "")
+	if status != http.StatusForbidden {
+		t.Fatalf("status = %d, want 403; body=%s", status, body)
+	}
+	if len(rec.reqs) != 0 {
+		t.Fatalf("Cloudflare was contacted %d time(s) for an unvalidated request; must be 0", len(rec.reqs))
+	}
+	if strings.Contains(body, "tok-victim") {
+		t.Fatalf("victim token leaked into response: %s", body)
+	}
+}
+
+// Each org's request uses ONLY its own org's token; there is no input by which one
+// org's request can carry another org's token (org is derived solely from the
+// validated principal, never from a body/query field).
+func TestTokenScopedToCallerOrg(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"orga": "tok-A", "orgb": "tok-B"}, rec, nil)
+
+	if status, body := do(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects", "ua", "orga", ""); status != 200 {
+		t.Fatalf("orgA status=%d body=%s", status, body)
+	}
+	rA, ok := rec.find("/pages/projects")
+	if !ok {
+		t.Fatal("orgA: no /pages/projects request reached Cloudflare")
+	}
+	if rA.auth != "Bearer tok-A" {
+		t.Fatalf("orgA used %q, want Bearer tok-A (cross-org token reach!)", rA.auth)
+	}
+
+	rec.reqs = nil
+	if status, _ := do(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects", "ub", "orgb", ""); status != 200 {
+		t.Fatalf("orgB status=%d", status)
+	}
+	rB, ok := rec.find("/pages/projects")
+	if !ok {
+		t.Fatal("orgB: no /pages/projects request reached Cloudflare")
+	}
+	if rB.auth != "Bearer tok-B" {
+		t.Fatalf("orgB used %q, want Bearer tok-B", rB.auth)
+	}
+	if rB.auth == "Bearer tok-A" {
+		t.Fatal("orgB reached orgA's token — cross-tenant break")
+	}
+}
+
+// A body/query field naming another org is IGNORED: the token is the caller-org's,
+// proving the handler derives the org only from the validated principal. (Mutation →
+// org-admin driver.)
+func TestBodyOrgFieldCannotRedirectToken(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"orga": "tok-A", "orgb": "tok-B"}, rec, nil)
+
+	body := `{"name":"site","org":"orgb","organizationId":"orgb","account":"ffffffffffffffffffffffffffffffff"}`
+	if status, resp, _ := doReq(t, app, http.MethodPost, "/v1/integrations/cloudflare/pages/projects", "ua", "orga", true, body); status != 200 {
+		t.Fatalf("status=%d resp=%s", status, resp)
+	}
+	r, ok := rec.find("/pages/projects")
+	if !ok {
+		t.Fatal("no create request reached Cloudflare")
+	}
+	if r.auth != "Bearer tok-A" {
+		t.Fatalf("hostile body redirected the token to %q; must stay Bearer tok-A", r.auth)
+	}
+}
+
+// An org that has not connected Cloudflare fails closed with 503, never another
+// org's data and never a fake success.
+func TestNotConnectedIs503(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"orga": "tok-A"}, rec, nil)
+	status, body := do(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects", "ux", "stranger", "")
+	if status != http.StatusServiceUnavailable {
+		t.Fatalf("status=%d, want 503; body=%s", status, body)
+	}
+	if len(rec.reqs) != 0 {
+		t.Fatalf("Cloudflare contacted for an unconnected org; must be 0 (got %d)", len(rec.reqs))
+	}
+}
+
+// Every served response stamps X-Hanzo-Org with the org whose token was used, so a
+// per-org caller can detect a pinned/comingled org (fix for the non-SuperAdmin
+// platform-token misdeploy).
+func TestResponseStampsActingOrg(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"orga": "tok-A"}, rec, nil)
+	status, _, hdr := doReq(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects", "ua", "orga", false, "")
+	if status != 200 {
+		t.Fatalf("status=%d", status)
+	}
+	if got := hdr.Get(actingOrgHeader); got != "orga" {
+		t.Fatalf("%s = %q, want orga", actingOrgHeader, got)
+	}
+}
+
+// ── mutation authorization (least privilege) ────────────────────────────────────
+
+// Mutations (POST/PUT/DELETE) require org admin; reads do not. A refused non-admin
+// mutation never reaches Cloudflare (rejected before the token read).
+func TestMutationRequiresOrgAdmin(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"orga": "tok-A"}, rec, nil)
+
+	mutations := []struct{ method, path, body string }{
+		{http.MethodPost, "/v1/integrations/cloudflare/pages/projects", `{"name":"site"}`},
+		{http.MethodDelete, "/v1/integrations/cloudflare/pages/projects/site", ""},
+		{http.MethodPost, "/v1/integrations/cloudflare/pages/projects/site/deployments", ""},
+		{http.MethodPut, "/v1/integrations/cloudflare/workers/scripts/hello", `{"script":"export default {}"}`},
+		{http.MethodDelete, "/v1/integrations/cloudflare/workers/scripts/hello", ""},
+		{http.MethodPost, "/v1/integrations/cloudflare/r2/buckets", `{"name":"b"}`},
+	}
+	for _, m := range mutations {
+		if s, _, _ := doReq(t, app, m.method, m.path, "member", "orga", false, m.body); s != http.StatusForbidden {
+			t.Fatalf("non-admin %s %s: status=%d, want 403", m.method, m.path, s)
+		}
+	}
+	// A refused non-admin mutation must not have reached Cloudflare at all.
+	if len(rec.reqs) != 0 {
+		t.Fatalf("refused non-admin mutations reached Cloudflare %d time(s); must be 0", len(rec.reqs))
+	}
+
+	// An org ADMIN is allowed through to Cloudflare.
+	if s, b, _ := doReq(t, app, http.MethodPost, "/v1/integrations/cloudflare/pages/projects", "admin", "orga", true, `{"name":"site"}`); s != 200 {
+		t.Fatalf("org-admin create: status=%d body=%s, want 200", s, b)
+	}
+	// A read stays open to a non-admin member.
+	if s, _ := do(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects", "member", "orga", ""); s != 200 {
+		t.Fatalf("non-admin read: status=%d, want 200", s)
+	}
+}
+
+// ── wired behavior ──────────────────────────────────────────────────────────────
+
+func TestPagesListHappyPath(t *testing.T) {
+	rec := &capture{}
+	resultFor := func(path string) (int, string) {
+		if strings.HasSuffix(path, "/pages/projects") {
+			return 200, `{"success":true,"errors":[],"result":[{"id":"p1","name":"marketing"}]}`
+		}
+		return 0, ""
+	}
+	app := harness(t, map[string]string{"orga": "tok-A"}, rec, resultFor)
+	status, body := do(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects", "ua", "orga", "")
+	if status != 200 {
+		t.Fatalf("status=%d body=%s", status, body)
+	}
+	if !strings.Contains(body, `"name":"marketing"`) {
+		t.Fatalf("result not relayed verbatim: %s", body)
+	}
+	r, _ := rec.find("/pages/projects")
+	if r.path != "/accounts/"+testAccountID+"/pages/projects" {
+		t.Fatalf("addressed %q, want the resolved-account path", r.path)
+	}
+	if strings.Contains(body, "tok-A") {
+		t.Fatalf("token leaked into response body: %s", body)
+	}
+}
+
+// Worker script PUT sends the modern multipart module upload with the caller-org
+// token (org-admin driver).
+func TestWorkersScriptPutMultipart(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"orga": "tok-A"}, rec, nil)
+	body := `{"script":"export default { fetch(){ return new Response('hi') } }","mainModule":"worker.js"}`
+	status, resp, _ := doReq(t, app, http.MethodPut, "/v1/integrations/cloudflare/workers/scripts/hello", "ua", "orga", true, body)
+	if status != 200 {
+		t.Fatalf("status=%d resp=%s", status, resp)
+	}
+	r, ok := rec.find("/workers/scripts/hello")
+	if !ok {
+		t.Fatal("no script PUT reached Cloudflare")
+	}
+	if !strings.HasPrefix(r.ctype, "multipart/form-data") {
+		t.Fatalf("content-type = %q, want multipart/form-data", r.ctype)
+	}
+	if !strings.Contains(string(r.body), `"main_module":"worker.js"`) {
+		t.Fatalf("multipart metadata missing main_module: %s", r.body)
+	}
+	if !strings.Contains(string(r.body), "export default") {
+		t.Fatal("multipart body missing the module source")
+	}
+	if r.auth != "Bearer tok-A" {
+		t.Fatalf("script PUT used %q, want Bearer tok-A", r.auth)
+	}
+}
+
+// Zone routes are zone-scoped and do NOT resolve an account (org-admin driver).
+func TestWorkersRouteBindZoneScoped(t *testing.T) {
+	rec := &capture{}
+	zone := "abcdef0123456789abcdef0123456789"
+	app := harness(t, map[string]string{"orga": "tok-A"}, rec, nil)
+	body := `{"pattern":"example.com/*","script":"hello"}`
+	status, resp, _ := doReq(t, app, http.MethodPost, "/v1/integrations/cloudflare/workers/zones/"+zone+"/routes", "ua", "orga", true, body)
+	if status != 200 {
+		t.Fatalf("status=%d resp=%s", status, resp)
+	}
+	if r, ok := rec.find("/zones/" + zone + "/workers/routes"); !ok {
+		t.Fatalf("route bind did not address the zone path; got %+v", rec.reqs)
+	} else if r.auth != "Bearer tok-A" {
+		t.Fatalf("route bind used %q, want Bearer tok-A", r.auth)
+	}
+	if rec.hasExact("/accounts") {
+		t.Fatal("route bind resolved an account; zone routes must not")
+	}
+}
+
+// ── stubs never lie ─────────────────────────────────────────────────────────────
+
+// R2/KV/D1 stub READ routes answer an honest 501 for a CONNECTED org — never a fake
+// 200 — and still enforce the validated-principal gate.
+func TestStubRoutesReturn501NeverSuccess(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"orga": "tok-A"}, rec, nil)
+	for _, path := range []string{
+		"/v1/integrations/cloudflare/r2/buckets",
+		"/v1/integrations/cloudflare/kv/namespaces",
+		"/v1/integrations/cloudflare/d1/databases",
+	} {
+		status, body := do(t, app, http.MethodGet, path, "ua", "orga", "")
+		if status != http.StatusNotImplemented {
+			t.Fatalf("%s: status=%d, want 501; body=%s", path, status, body)
+		}
+		if strings.Contains(strings.ToLower(body), `"success":true`) || strings.Contains(strings.ToLower(body), `"ok":true`) {
+			t.Fatalf("%s: stub returned a misleading success: %s", path, body)
+		}
+		if s, _ := do(t, app, http.MethodGet, path, "", "orga", ""); s != http.StatusForbidden {
+			t.Fatalf("%s: unvalidated status=%d, want 403", path, s)
+		}
+	}
+}
+
+// ── input hardening + account resolution ────────────────────────────────────────
+
+// An explicit ?account= override must be a 32-hex id (defense against path
+// injection), and it skips discovery.
+func TestAccountOverrideValidated(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"orga": "tok-A"}, rec, nil)
+
+	if status, _ := do(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects?account=../../evil", "ua", "orga", ""); status != http.StatusBadRequest {
+		t.Fatalf("hostile account override status=%d, want 400", status)
+	}
+	rec.reqs = nil
+	override := "ffffffffffffffffffffffffffffffff"
+	if status, _ := do(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects?account="+override, "ua", "orga", ""); status != 200 {
+		t.Fatalf("valid account override status=%d", status)
+	}
+	r, _ := rec.find("/pages/projects")
+	if r.path != "/accounts/"+override+"/pages/projects" {
+		t.Fatalf("override not honored: addressed %q", r.path)
+	}
+	if rec.hasExact("/accounts") {
+		t.Fatal("discovery call made despite an explicit account override")
+	}
+}
+
+// The account captured at connect time (ConnectionFor.ExternalID) is used without a
+// live discovery round-trip.
+func TestStoredAccountSkipsDiscovery(t *testing.T) {
+	rec := &capture{}
+	app := harness(t, map[string]string{"orga": "tok-A"}, rec, nil)
+	stored := "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
+	prev := connectionFor
+	connectionFor = func(org, provider string) (integrations.Connection, bool) {
+		if org == "orga" && provider == providerCloudflare {
+			return integrations.Connection{ExternalID: stored}, true
+		}
+		return integrations.Connection{}, false
+	}
+	t.Cleanup(func() { connectionFor = prev })
+
+	if s, _ := do(t, app, http.MethodGet, "/v1/integrations/cloudflare/pages/projects", "ua", "orga", ""); s != 200 {
+		t.Fatalf("status=%d", s)
+	}
+	r, ok := rec.find("/pages/projects")
+	if !ok {
+		t.Fatal("no pages request reached Cloudflare")
+	}
+	if r.path != "/accounts/"+stored+"/pages/projects" {
+		t.Fatalf("addressed %q, want the stored-account path", r.path)
+	}
+	if rec.hasExact("/accounts") {
+		t.Fatal("live discovery happened despite a stored account id")
+	}
+}

clients/cloudflare/d1.go

@@ -0,0 +1,64 @@
+package cloudflare
+
+// d1.go — Cloudflare D1 (account-scoped: /accounts/{id}/d1/database). PHASE-2 STUB:
+// routes + typed provider methods exist so the surface is complete and typed; the
+// bodies ship in Phase 2, and each handler answers an honest 501 (via stubResult)
+// after the authClient gate — never a fake success.
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+
+	"github.com/hanzoai/cloud"
+	"github.com/zap-proto/zip"
+)
+
+// D1Database is the Phase-2 D1 database shape.
+type D1Database struct {
+	UUID string `json:"uuid"`
+	Name string `json:"name"`
+}
+
+// listD1Databases / createD1Database / deleteD1Database are the typed Phase-2
+// provider seams; the wired bodies land in Phase 2.
+func (cl *client) listD1Databases(context.Context, string) ([]D1Database, error) {
+	return nil, errPhase2
+}
+func (cl *client) createD1Database(context.Context, string, string) (*D1Database, error) {
+	return nil, errPhase2
+}
+func (cl *client) deleteD1Database(context.Context, string, string) error {
+	return errPhase2
+}
+
+func d1DatabaseList(s *cloud.Service[state], c *zip.Ctx) error {
+	cl, _, err := authClient(s, c)
+	if err != nil {
+		return err
+	}
+	_, err = cl.listD1Databases(c.Context(), strings.TrimSpace(c.Query("account")))
+	return stubResult(c, "d1", err)
+}
+
+func d1DatabaseCreate(s *cloud.Service[state], c *zip.Ctx) error {
+	cl, _, err := authWrite(s, c)
+	if err != nil {
+		return err
+	}
+	var in struct {
+		Name string `json:"name"`
+	}
+	_ = json.Unmarshal(c.Body(), &in)
+	_, err = cl.createD1Database(c.Context(), strings.TrimSpace(c.Query("account")), strings.TrimSpace(in.Name))
+	return stubResult(c, "d1", err)
+}
+
+func d1DatabaseDelete(s *cloud.Service[state], c *zip.Ctx) error {
+	cl, _, err := authWrite(s, c)
+	if err != nil {
+		return err
+	}
+	err = cl.deleteD1Database(c.Context(), strings.TrimSpace(c.Query("account")), strings.TrimSpace(c.Param("database")))
+	return stubResult(c, "d1", err)
+}

clients/cloudflare/kv.go

@@ -0,0 +1,64 @@
+package cloudflare
+
+// kv.go — Cloudflare Workers KV (account-scoped: /accounts/{id}/storage/kv/
+// namespaces). PHASE-2 STUB: routes + typed provider methods exist so the surface is
+// complete and typed; the bodies ship in Phase 2, and each handler answers an honest
+// 501 (via stubResult) after the authClient gate — never a fake success.
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+
+	"github.com/hanzoai/cloud"
+	"github.com/zap-proto/zip"
+)
+
+// KVNamespace is the Phase-2 KV namespace shape.
+type KVNamespace struct {
+	ID    string `json:"id"`
+	Title string `json:"title"`
+}
+
+// listKVNamespaces / createKVNamespace / deleteKVNamespace are the typed Phase-2
+// provider seams; the wired bodies land in Phase 2.
+func (cl *client) listKVNamespaces(context.Context, string) ([]KVNamespace, error) {
+	return nil, errPhase2
+}
+func (cl *client) createKVNamespace(context.Context, string, string) (*KVNamespace, error) {
+	return nil, errPhase2
+}
+func (cl *client) deleteKVNamespace(context.Context, string, string) error {
+	return errPhase2
+}
+
+func kvNamespaceList(s *cloud.Service[state], c *zip.Ctx) error {
+	cl, _, err := authClient(s, c)
+	if err != nil {
+		return err
+	}
+	_, err = cl.listKVNamespaces(c.Context(), strings.TrimSpace(c.Query("account")))
+	return stubResult(c, "kv", err)
+}
+
+func kvNamespaceCreate(s *cloud.Service[state], c *zip.Ctx) error {
+	cl, _, err := authWrite(s, c)
+	if err != nil {
+		return err
+	}
+	var in struct {
+		Title string `json:"title"`
+	}
+	_ = json.Unmarshal(c.Body(), &in)
+	_, err = cl.createKVNamespace(c.Context(), strings.TrimSpace(c.Query("account")), strings.TrimSpace(in.Title))
+	return stubResult(c, "kv", err)
+}
+
+func kvNamespaceDelete(s *cloud.Service[state], c *zip.Ctx) error {
+	cl, _, err := authWrite(s, c)
+	if err != nil {
+		return err
+	}
+	err = cl.deleteKVNamespace(c.Context(), strings.TrimSpace(c.Query("account")), strings.TrimSpace(c.Param("namespace")))
+	return stubResult(c, "kv", err)
+}