pkg/hanzoai/cloud goes +2186 -2126 ~175: 1700 paths / 2354 operations / 2186 schemas, 182 api modules and 2172 model modules, against the 239/1116 it carried from hanzo.yaml. The document is hanzoai/cloud's own emission now, pinned in a new .spec-lock by commit and sha256, and generate.py reaches it by value — --skip-validate-spec means the 1012 missing-`responses` errors no longer write zero files, so a client can be cut from the authority instead of a projection of it. Two renamings come with that, and both are the fix rather than the damage. IAM's types are namespace-qualified (iam.Role, iam.Application, 95 of them) because a bare `Role` was two unrelated shapes wearing one name. And the <svc>_ prefix is gone from every operationId, so every generated method lost it. The prefix is what broke the examples, and the gate saw only two thirds of it. Four flows failed on their imports; `money` and `tools` PASSED while every call in them named a method that no longer existed — an import resolves the names in the `from … import` line and a method is looked up at call time. All five flows now name operations that exist, checked by resolving each one as an attribute, and hanzo.yml records that ceiling so the next reader does not trust the gate for more than it says. `chat` is removed, which is a measurement and not a preference: cloud declares POST /v1/chat/completions with no requestBody and no responses at all, so the generated method takes no body and returns None — the one call a chat example exists to make cannot be expressed. Inventing the type, or hand-rolling the HTTP inside a generated client, is the drift these SDKs exist to prevent; js-sdk dropped its own chat flow at 2.0.7 for exactly this. It returns the release cloud gives that route a body. Two smoke assertions were already red before this regeneration and are now true again: they pinned AIApi/APIKeysApi/MCPApi and AdminApi.plugin_admin_*, names from the retired lineage. The surface is all still there under AiApi/KeysApi/ McpApi and AdminApi.admin_plugins. tests/test_zap_transport.py still errors, exactly as it did before: it imports `Hanzo` from hanzoai, which the hand-written package does not define. That is not this document's business and is untouched. Minor rather than patch: every generated method changed name. Co-authored-by: Hanzo Dev <dev@hanzo.ai>
8.9 KiB
LLM.md — hanzoai/python-sdk
What this is: the flagship, most-complete Hanzo SDK — a uv workspace of 60+
packages: the typed cloud client (hanzoai), agents, MCP server + tools, memory,
distributed compute, and the hanzo CLI. pip install hanzo.
Canonical role (one-way SDK model): Hanzo ships two SDK lines per language —
(1) the full cloud SDK generated from OpenAPI, (2) the AI/agents library. This repo
is the Python flagship of line 2, the reference for every other language.
Completeness: Python → Rust → C++ → Go. One impl, one place; discovery repos link
OUT, never duplicate. Full spec: ~/work/hanzo/SDK-ARCHITECTURE.md.
Install / run
pip install hanzoai # typed cloud client
pip install hanzo # agents + MCP + orchestration helpers
uv sync --all-packages # dev: whole workspace
uv run pytest tests/ -v # tests
Console-script law — only the native binary is called hanzo
The hanzo command is the Rust CLI (curl -fsSL https://hanzo.sh | sh). No
package in this workspace may declare a hanzo console script.
Both hanzo and hanzo-cli used to, and hanzo depends on hanzo-cli, so
pip install hanzo installed two distributions fighting over one name — whichever
landed last won. Measured on a clean venv at 0.4.3: hanzo --help printed
hanzo-cli's program (bot/deploy/iam/k8s/kms/login/logout/paas/s3/whoami), not
the one pkg/hanzo/README.md documented. And if the native CLI was already on
PATH, hanzo login meant one of three different things depending on install
order — the real CLI reads a bare hanzo login as an AI task, since the verb
there is hanzo auth login.
Now: hanzo ships hanzo-py, hanzo-cli ships hanzo-cli. Script named after
its distribution, one canonical hanzo.
hanzo-node on PyPI is also not the hanzo-node command. That command is a
symlink to the Hanzo CLI, installed by hanzo.sh. The PyPI package fetches a
different Rust binary from hanzoai/node — a private repo, so its release assets
404 for anyone outside the org. Both READMEs say so rather than implying one
product.
Brand rules (hard — enforce in all docs)
- Never "LLM gateway"; never position against LiteLLM. Hanzo is a full AI SDK / AI cloud, not a proxy.
- Zen models are our own family — never name upstream models.
- Paths are
/v1/…, never/api/…. Base host:https://api.hanzo.ai. - Voice: "Hanzo — the Open AI Cloud." Crisp, developer-first, no emoji-spam.
Codegen — this repo PULLS, it never pushes
hanzoai/cloud emits its own router spec -> cloud/openapi.yaml [the ONE SDK input]
hanzoai/openapi generate.py + sdks.yaml -> the invocation, as data
this repo owns its test + bump + release, and pins what it projected
The client is a projection of cloud's document directly, and .spec-lock
names the commit and sha256 it was cut from. generate.py passes
--skip-validate-spec, so the 1012 missing-responses errors that once made
cloud's emission write zero files no longer stop it; hanzo.yaml is out of this
SDK's path (it still feeds the doc site and the skills plane). Regenerate with
the document by value:
cd ~/work/hanzo/openapi && uv run --with pyyaml python3 generate.py python \
--repo ~/work/hanzo/python-sdk --spec ~/work/hanzo/cloud/openapi.yaml
Current pkg/hanzoai/cloud/ is 1700 paths / 2354 operations / 2186 schemas →
182 api modules + 2172 model modules.
Two renamings arrived with the lineage, and neither is a defect to undo.
IAM's types are namespace-qualified — iam.Role, iam.Application, 95 of them
— because a bare Role had been two unrelated shapes under one name (IAM's
14-property role, and a 2-property {role, user} row from another service).
Both exist now and each says which it is. And the <svc>_ operationId prefix is
gone, so every method lost it: cloud_get_v1_tools → get_v1_tools,
AIApi/APIKeysApi/MCPApi → AiApi/KeysApi/McpApi,
AdminApi.plugin_admin_plugins → admin_plugins.
pkg/hanzoai/cloud/ is generated — never hand-edit it. generate.py does
rmtree(dst) + copytree(src), so anything written there dies on the next run. Regenerate
only from hanzoai/openapi (python3 generate.py python) — never from here. The old
scripts/generate.sh was a second, destructive driver (rm -rf pkg/hanzoai, which would
have eaten the hand-written config/mcp/protocols/session/zap modules); it is deleted.
Consumed at 3.1.3: cloud 8143fc0e, openapi 2861089. Regenerate whenever either moves —
openapi 3300cda dropped {org} from the KMS secrets routes (/v1/kms/orgs/{org}/secrets
-> /v1/kms/secrets; the org is read from the token), which silently stranded 3.1.2 on a
path the server no longer serves.
The case-variant tag defect is CLOSED. hanzo.yaml used to carry 23 tag groups
differing only by case (AI/ai, Users/users, …); openapi-generator mapped both
spellings onto one module and 127 of the 411 operations in those groups never reached the
client. Fixed upstream in the per-service specs, as that note predicted. Verified on the
current spec: 239 distinct tags → 239 api modules, 1:1, so nothing collapses, and
generate.py python --check reports [python] clean with no local strip of any kind.
Two spec defects were found and fixed upstream while regenerating at 3.1.5. Neither was
patched here; both are in hanzoai/openapi main:
fc0c17a— 35/v1/platformoperations carried noresponses. OAS 3.x requires it and openapi-generator aborts the entire document, sohanzo.yamlwas producing no client in any language, not just Python.07783f5—ChatCompletionResponse.choiceswasitems: {type: object}, sochoices[0].message.contentwasList[object]and unusable without a cast. Now anai_ChatChoiceschema.
The rule holds in both directions: a generated tree is never hand-repaired, and a defect found by generating is fixed in the spec, where every other language gets the fix too.
Examples — the six canonical flows
examples/{hello,chat,money,store,agent,tools}, one directory each, plus
examples/client.py as the single place a base URL or an env var is resolved. The same six
exist in every Hanzo SDK. Run one with python -m examples.hello from the repo root.
Each flow's call sits behind if __name__ == "__main__": on purpose. That is what lets CI
import all six to prove every from hanzoai.cloud import X still resolves, without an API
key and without opening a socket — so a spec change that renames or drops an operation goes
red in the gate instead of in a user's app.
They are a gate, not decoration. The TypeScript twin of the chat flow is what surfaced the
choices defect above: the generated tree imported and built perfectly, because building
generated code only proves it is internally consistent. Only calling it proves the surface
is usable.
CI
Fleet convention, added at 3.1.5: root hanzo.yml (the test: gate) plus a 7-line
.github/workflows/cicd.yml importing hanzoai/ci. The gate is two blocks — import every
generated module (for generated code that IS the build step; there is no compiler to catch a
bad $ref), then import the six flows. Both provision an interpreter with uv when the arc
runner lacks one.
Scope is deliberate: the cloud client and its flows, not all 65 packages. A red gate should mean "the client the spec just produced is broken", not "something, somewhere".
Publishing is not here. .hanzo/workflows/publish-pypi.yml on our own runners stays the
canonical path because it reads the PyPI token from KMS like every other publish credential
in the fleet. hanzo.yml gates only; a second publish path would be one too many.
Key entry points
pkg/hanzoai/— typed OpenAPI client (ApiClient,Configuration,Ai*Api). Two surfaces live here:pkg/hanzoai/{api,models}(older, frozen — nothing regenerates it now) andpkg/hanzoai/cloud/(current, spec-driven). New work targetscloud/.pkg/hanzo/src/hanzo/cli.py— thehanzoCLI command tree.pkg/hanzo-mcp/— MCP server; tools via[project.entry-points."hanzo.tools"].pkg/hanzo-tools-*/— one concern each, exports aTOOLSlist.pkg/hanzo-{agents,agent,network,memory}/— agent/compute/memory libraries.pkg/hanzo-kms/— KMS client. The server is luxfi/kms (kms.hanzo.ai,kms.lux.cloud) and its whole surface is/v1/kms/auth/loginplus/v1/kms/orgs/{org}/secrets[/{path}/{name}]./api/*is Infisical's and was never served — it looked like a decode error rather than a 404 only because old builds answered every unmatched path with the console SPA (200 text/html). A secret is (org, path, name, env), one value each — no versions. The server splits the trailing URL at its LAST slash into (path, name), so escape each segment individually.pkg/hanzo-kms/tests/pins all of it.
Rules for agents: update THIS file with significant discoveries; never write random summary files; keep the README cross-link block intact.