mirror of
https://github.com/hanzoai/ui.git
synced 2026-11-21 00:18:41 +00:00
The value is a pk- prefixed PUBLISHABLE key (the Stripe vocabulary, and what the build-time gate already tests for). EVENT_INGEST_KEY hid that. One substitution covers all three spellings, since the framework prefixes wrap the same token: EVENT_INGEST_KEY -> PUBLISHABLE_KEY, NEXT_PUBLIC_* and VITE_* follow. KMS already carries deploy/PUBLISHABLE_KEY (env prod) with the same value, written and read back first, so no build can reach a name that does not exist yet. Co-authored-by: Hanzo Dev <[email protected]>
123 lines
6.3 KiB
YAML
123 lines
6.3 KiB
YAML
|
|
# This repo is BOTH a library and a container. `pkg/ui` publishes to npm from
|
|
# .hanzo/workflows/publish.yml on a version change — still the only thing allowed
|
|
# to reach npmjs. `app/` is ui.hanzo.ai: a Next static export served by
|
|
# hanzoai/static as ghcr.io/hanzoai/ui, pinned live in hanzoai/universe.
|
|
#
|
|
# The `images:` block below is what makes the second half true here. This file
|
|
# carried a `test:` block and nothing else, and said "a library, not a container"
|
|
# — accurate about pkg/ui and wrong about the site, which has been live on
|
|
# ghcr.io/hanzoai/ui since before any workflow could build it.
|
|
#
|
|
# The gate is pkg/ui — the package this repo actually publishes — and it is
|
|
# the same three commands publish.yml already proves on this runner, moved
|
|
# EARLIER. That ordering is the whole point: publish.yml is the LAST place a
|
|
# defect costs nothing, because the version it ships is immutable. Here it
|
|
# costs a re-push.
|
|
#
|
|
# The trailing `...` on the build filter is load-bearing: it builds @hanzo/ui's
|
|
# WORKSPACE DEPENDENCIES first, in topological order. Without it the siblings
|
|
# have no dist/, tsc reads them through their own `"types": "dist/index.d.ts"`,
|
|
# and they come back as missing modules — `Cannot find module '@hanzo/tokens'`,
|
|
# `'@hanzo/products'`, `'@hanzo/cd'` — plus a wave of implicit-any in every file
|
|
# that imported them. Those errors name @hanzo/ui and are not its fault.
|
|
#
|
|
# `pnpm install` without --frozen-lockfile is deliberate and is the literal
|
|
# command measured green here: pnpm already defaults frozen-lockfile to true
|
|
# under CI, so the flag adds nothing but a second way to spell the same thing.
|
|
#
|
|
# WHAT THIS DOES NOT GATE, and why the ci.yml it replaces was red on every run
|
|
# for months. That workflow had four jobs and three of them could not pass:
|
|
#
|
|
# Build `cd pkgs/ui && pnpm run build:full`
|
|
# Test `cd pkgs/ui && pnpm test:coverage`
|
|
# → both: `cd: pkgs/ui: No such file or directory`. The package moved
|
|
# to pkg/ (singular) — the same split publish.yml documents at
|
|
# length — and these two steps were never followed across.
|
|
# Type `cd app && pnpm run typecheck`
|
|
# → ~10x `error TS7016: Could not find a declaration file for module
|
|
# '@hanzo/ui/animation/*'`. app/ is the shadcn registry this repo
|
|
# grew from, and its package.json aliases
|
|
# `"@hanzo/ui": "npm:@hanzo/ui-shadcn@^5"`, so those imports resolve
|
|
# to a PUBLISHED package that ships no declarations. Nothing in this
|
|
# tree can fix it, and the app is stale anyway.
|
|
#
|
|
# A gate that cannot go green teaches nobody anything — it is read once, filed
|
|
# as "CI is just red here", and from then on it hides real breakage. Gating
|
|
# pkg/ui is not a smaller gate than that one. It is the first real one.
|
|
|
|
# ui.hanzo.ai. The recipe is the root Dockerfile and lives nowhere else; this
|
|
# block only declares WHAT is built and what it needs to be valid.
|
|
#
|
|
# THE BUILD SECRET IS WHY THIS BLOCK EXISTS. hanzoai/ci (which
|
|
# .hanzo/workflows/cicd.yml already imports) is the ONLY lane that implements
|
|
# `build_secrets`: it reads each name from KMS `deploy/<NAME>?env=prod` and passes
|
|
# it as `--build-arg NAME=value`, refusing the run when the value is empty. The
|
|
# hand-rolled image job in .hanzo/workflows/deploy.yml ran a bare `docker build`
|
|
# with no build-arg and no KMS hop, so no lane in this repo could put a key in the
|
|
# bundle. The one key that ever reached production got there on somebody's laptop,
|
|
# and it has since gone dead — api.hanzo.ai 401s it — which is what a credential
|
|
# with no automated source always does eventually. That job is now
|
|
# workflow_dispatch-only; this is the lane.
|
|
#
|
|
# The KMS name IS the build-arg name IS the Dockerfile ARG. One name, one place.
|
|
# Supply it or the Dockerfile's gate refuses the build — the two halves either
|
|
# both hold or both fail loudly.
|
|
#
|
|
# No `version:` — imgver derives it, and deriving is correct here. The root
|
|
# package.json is a 0.0.1 workspace stub and pkg/ui's 8.0.58 is the NPM package's
|
|
# number, not the site's; neither describes the image. imgver takes
|
|
# max(declared, published)+1 patch, so the series continues from what is already
|
|
# in the registry (5.7.6 live) instead of restarting or colliding with npm.
|
|
images:
|
|
- name: ui
|
|
context: .
|
|
dockerfile: Dockerfile
|
|
repo: ghcr.io/hanzoai/ui
|
|
platforms: [linux/amd64]
|
|
build_secrets: [PUBLISHABLE_KEY]
|
|
|
|
# No `deploy:` — ON PURPOSE. hanzoai/universe names what is live, in
|
|
# charts/app/values/hanzo/ui.yaml, and the in-cluster reconcile restores it — so a
|
|
# CI-side patch does not stick. It would also LOOK like it worked: the patch
|
|
# applies, the pod rolls, and the declared tag comes back a minute later. A build
|
|
# emits an image; git declares desired state; CD applies it.
|
|
#
|
|
# That file pins `tag:` AND `digest:` together, and the digest is what is actually
|
|
# pulled: moving the tag alone changes nothing. Read the tag off the build you
|
|
# fired — this lane publishes `sha-<sha7>-amd64`, `latest` and the imgver semver,
|
|
# none of which is the `v5.7.6` shape pinned there today.
|
|
|
|
# Two jobs, not one, and the split is not cosmetic. `pnpm run test` is now
|
|
# vitest + the Vite consumer + the WEBPACK consumer, and the two consumer
|
|
# suites need a browser and a real npm install that unit tests do not. Running
|
|
# them as one step means a playwright download sits in front of every unit
|
|
# failure.
|
|
#
|
|
# The `...` on the build filter stays and is load-bearing: it builds @hanzo/ui's
|
|
# WORKSPACE DEPENDENCIES first, in topological order. @hanzo/cd is still one of
|
|
# them. Without it the sibling has no dist/, tsc reads it through its own
|
|
# `"types": "dist/index.d.ts"`, and it comes back as a missing module — an error
|
|
# that names @hanzo/ui and is not its fault.
|
|
test:
|
|
- name: ui-unit
|
|
run: |
|
|
set -e
|
|
corepack enable
|
|
pnpm install
|
|
pnpm --filter @hanzo/ui... build
|
|
pnpm --filter @hanzo/ui run test:unit
|
|
- name: ui-consumer
|
|
run: |
|
|
set -e
|
|
corepack enable
|
|
pnpm install
|
|
pnpm --filter @hanzo/ui... build
|
|
cd pkg/ui
|
|
pnpm exec playwright install --with-deps chromium
|
|
# Both bundlers. They disagree about css url() asset resolution, and a
|
|
# Vite-only gate is how 8.0.46 shipped a stylesheet pointing at a font
|
|
# file the package does not pack.
|
|
pnpm run test:consumer
|
|
pnpm run test:consumer:webpack
|