Bloodbank Sdk Generation

Bloodbank SDK Generation

Deterministic recipe for generating typed bindings from bloodbank/schemas/bloodbank/v1/**. Outputs are plain artifacts the consumer commits or vendors — this skill produces code, not an installable package.

Prereqs

• Run from the Bloodbank repo root (the directory containing schemas/ and docs/event-naming.md), or pass the schema path explicitly.
• uvx (from uv) for Python tooling, npx for TypeScript tooling — both are project-less, no install required.
• Validate the tree is well-formed first: mise run validate:schemas (or bash scripts/validate_schemas.sh). Skip generation if it fails — fix schemas before generating.

Python (Pydantic v2 models)

uvx --from 'datamodel-code-generator>=0.25' datamodel-codegen \
  --input schemas/bloodbank/v1 \
  --input-file-type jsonschema \
  --output-model-type pydantic_v2.BaseModel \
  --target-python-version 3.12 \
  --use-schema-description \
  --use-field-description \
  --use-double-quotes \
  --enum-field-as-literal one \
  --output bloodbank_contracts/models.py

Flag rationale:

• --input-file-type jsonschema — required; the default auto mis-detects when the tree has multiple files.
• --enum-field-as-literal one — properties.type.const becomes Literal["bloodbank.v1.<...>"], matching the 5-token contract.
• --use-schema-description / --use-field-description — preserves the human-readable descriptions from the JSON Schemas as Pydantic field docstrings.
• --target-python-version 3.12 — matches the project floor; bump if the project floor moves.

The generator resolves $ref: "../../../_common/cloudevent_base.v1.json" correctly because all schemas live under one input root.

TypeScript (interfaces + type unions)

npx json-schema-to-typescript \
  'schemas/bloodbank/v1/**/*.v1.json' \
  --cwd schemas \
  --bannerComment '' \
  --unreachableDefinitions \
  --additionalProperties false \
  > bloodbank_contracts/types.ts

Flag rationale:

• --cwd schemas — sets the resolver root so $ref: "../../../_common/..." resolves from each file's directory through the same root the schemas were authored against.
• --bannerComment '' — drops the autogenerated header that fights with linters.
• --unreachableDefinitions — emits the _common types even if no top-level schema references them by name, so consumers can import CloudEventBase directly.
• --additionalProperties false — matches the strict-shape default of the runtime validator.

Verify the output

After generation, run a one-shot import smoke:

# Python
python3 -c "from bloodbank_contracts.models import *; import bloodbank_contracts.models as m; print(f'{sum(1 for n in dir(m) if not n.startswith(\"_\")):>3} symbols')"

# TypeScript (requires tsc)
npx tsc --noEmit --strict bloodbank_contracts/types.ts

If either fails, regenerate from a clean validate:schemas pass before debugging the generator output.

Common combinations

• Refresh + commit on schema change: re-run both generators whenever you edit anything under bloodbank/schemas/bloodbank/v1/. Treat the generated files as committed artifacts the consumer holds — git diff is the drift signal.
• One-off consumer scaffold: a downstream project that needs the types runs both commands once, vendors the outputs, and pins via the schema-tree commit hash recorded in their changelog.
• Future packaging: when a second external consumer materializes, promote bloodbank_contracts/ into a real package (uv build for Python, tsc --declaration + package.json for TS). See RECOMMENDATION.md step 5 in the Bloodbank repo.

Gotchas

Read references/gotchas.md only when generation fails with a non-obvious error, output looks wrong (missing fields, broken refs, unexpected types), or you're integrating the generated code into an existing project with its own Pydantic/Zod setup.

Out of scope

This skill produces SDK artifacts from the canonical schema tree. It does NOT cover:

• Authoring or modifying schemas. Edit JSON files under bloodbank/schemas/ directly per bloodbank/docs/event-naming.md §12.
• Validating envelopes at runtime. Set BLOODBANK_HOOK_VALIDATE=1 and rely on services/agent-hooks/core/validate.py, which uses a pre-loaded referencing.Registry.
• Checking schema-tree consistency. Run mise run validate:schemas (well-formedness + ref resolution) and mise run smoketest:schema-contract-consistency (validator allowlist drift).
• The deprecated Holyfields generators (tools/generators/generate_pydantic.py, tools/generators/generate_typescript.sh). Those are hand-rolled, locked to Holyfields' packages/python workspace, and superseded by the third-party tools above. Do not port them.
• Apicurio registry uploads. The apicurio://holyfields/<type>/versions/1 dataschema URI is a registry artifact id, separate from this skill's filesystem generation path.
• Publishing the SDK (PyPI, npm, GitHub Packages). This skill writes files; distribution is a separate decision.
• Zod schemas. json-schema-to-typescript emits TypeScript types + interfaces, not Zod validators. If runtime TS validation is needed, layer a tool like json-schema-to-zod on top — but evaluate whether the type-only output is sufficient first.