CI: non-hermetic pi-fork build breaks `check` (models.generated.ts drifts against pinned OpenAIResponsesCompat)

[pdettori]

Summary

The check CI job fails at the Build pi-fork (type declarations) step on main (and consequently on every PR branch) because pi-fork/packages/ai regenerates src/models.generated.ts from the live models.dev API at build time, and the freshly generated file no longer typechecks against pi-fork's pinned types.

src/models.generated.ts(10662,13): error TS2353: Object literal may only specify
known properties, and '"supportsReasoningEffort"' does not exist in type 'OpenAIResponsesCompat'.

Root cause

The build script runs:

npm run generate-models && npm run generate-image-models && tsgo -p tsconfig.build.json

scripts/generate-models.ts fetches from models.dev on every build (Loaded 633 tool-capable models from models.dev) and overwrites src/models.generated.ts. models.dev has since added a supportsReasoningEffort field that pi-fork's pinned OpenAIResponsesCompat type does not declare — so tsgo rejects the generated literal. Because the file is regenerated from a live external API, this broke retroactively with no code change: the build is non-hermetic.

Evidence / timeline

main CI (workflow CI):

Date	Result
2026-08-04	✅ success
2026-08-20 (run 32408167817)	❌ failure — same supportsReasoningEffort error

The models.dev drift landed in that window. The job dies at this step, so the downstream Typecheck and Test steps never run — meaning harness unit tests are currently not executed in CI at all.

Impact

• Red CI on main and every PR (e.g. feat: multi-protocol model provider (Anthropic + OpenAI-compat) — support LiteLLM and RITS from main #159), independent of PR content.
• Loss of test/typecheck signal in CI (the check job aborts before those steps).

Proposed fixes (pick one)

1. Make the build hermetic — commit models.generated.ts (and image-models.generated.ts) and drop the live generate-models call from the build script; regenerate deliberately via a separate, reviewed step. (preferred — removes the whole class of failure)
2. Widen the pinned type — add supportsReasoningEffort?: boolean to OpenAIResponsesCompat so generated data validates.
3. Pin the models.dev snapshot — vendor a fixed models.dev dataset instead of fetching live at build time.

References

• Failing job: https://github.com/rossoctl/serverless-harness/actions/runs/32485738448/job/96781706009
• Surfaced during review of feat: multi-protocol model provider (Anthropic + OpenAI-compat) — support LiteLLM and RITS from main #159 (the PR is unaffected by / not the cause of this failure).