fix: stop committed generated artifacts drifting from their generators

[josecarneiro]

Replaces #466 — identical tree, recommitted under the correct author. #466 can be closed.

Summary

pnpm install && pnpm build on a clean main left 68 tracked files dirty. CI was building fine — it just never committed or checked most of what it regenerated.

Root cause, in the auto-release job: the SDK branch stages packages/epilot-sdk-v2/ wholesale, but the CLI branch stages only

FILES_TO_ADD="$FILES_TO_ADD packages/cli/package.json packages/cli-wrapper/package.json"

packages/cli commits ~160 generated artifacts. CI regenerated them on every push, tested them, and threw them away. They had not been updated since 2026-08-26, and no git diff --exit-code existed anywhere in .github/workflows/ to notice.

Two concrete symptoms:

• packages/cli/src/generated/api-list.ts advertised 12 operations for ai-agents against the spec's 21 — executeAgentStream, getExecutionFeedback and seven others were missing. The CLI resolves operations at runtime from definitions/*.json, so running it from source exposed the stale set.
• packages/cli/src/index.ts reported version 0.1.51 while package.json was at 0.1.140.

Published tarballs were never affected — CI rebuilds before publish.

This is still happening

While #466 was open, @epilot/cli@0.1.141 (5e39d9ce) released off the back of the entity-mapping spec update in #467. It again staged only packages/cli/package.json, so main immediately had:

• packages/cli/definitions/entity-mapping.json stale against the spec it is copied from
• packages/cli/src/index.ts reporting 0.1.51 against a package.json at 0.1.141

The drift recurred within the hour, unprompted. That is the failure mode reproducing on its own, not a reconstruction.

Changes

Commit	What
79fe8c7	Align 15 committed client openapi.d.ts with generator output — all cosmetic (stale /* eslint-disable */ header from the pre-biome era, blank lines, trailing whitespace). No type changed.
372e632	journey-client and configuration-hub-client were the only 2 of 52 clients running codegen from build. journey's npm run openapi fetched the live spec from docs.api.epilot.io and overwrote committed source, so a build had a network dependency. Both now match the other 50: build only compiles, typegen moved to prepublishOnly.
df5b78e	Refresh journey-config spec 1.4.2 → 1.4.3 (JourneyActivationGuarantee). Stale precisely because the only thing refreshing it was the build removed above.
acc7a57	Regenerate the stale CLI and SDK artifacts. 30 of 51 definitions/*.json → 0 stale.
37dba9e	The CI fixes (below).
4a57254	Release the journey refresh via changeset, matching every other client update.
18fafb4	Move the build-cleanliness gate to sit directly after pnpm build rather than after pnpm test, so a dirty tree can only be the build's doing.
7725c2e + d393608	Merge main and regenerate, absorbing the 0.1.141 drift described above.

CI

Two gates in the test job:

1. After running both generators, fail if the tree is dirty — a PR that changes a spec without regenerating is rejected with the command to run.
2. Directly after pnpm build, fail if the tree is dirty — a build can never again quietly regenerate or re-fetch committed source.

And auto-release now regenerates the CLI after the version bump and stages all of packages/cli/. Ordering was the reason src/index.ts kept the wrong version: Build and test CLI runs before the bump, and Rebuild CLI with bumped version runs after the push, so neither produced output the release commit could carry.

Test plan

• pnpm lint — clean (524 files)
• pnpm generate-sdk && pnpm --filter @epilot/cli generate → no diff (gate 1 passes)
• pnpm build → exit 0, tree clean (gate 2 passes; was 68 dirty files)
• All 51 packages/cli/definitions/*.json byte-identical to the clients/*/src/openapi.json they are copied from
• CI green on chore: bring committed generated artifacts back in sync with their generators (1/2) #466's head — Test, Lint, CodeQL and Analyze all pass, with both new gates passing in CI
• Gate 1 demonstrated working: after merging main, it flagged the four artifacts the 0.1.141 release had left stale
• Locally pnpm test fails identically to unmodified main (cli 19/100, sdk-v2 25/334) — both failing suites spin up a local HTTP server and hit ERR_INVALID_URL from proxy-prefixed URLs in the dev sandbox. Zero new failures, and green in CI.
• git diff between this branch and chore: bring committed generated artifacts back in sync with their generators (1/2) #466's head is empty — the tree is byte-identical, only author/committer metadata changed

Review notes

• The regeneration commits are large but entirely generator output — reviewing 37dba9e, 372e632 and 18fafb4 is what matters.
• The journey spec change is type-level only: no new paths or operations, so src/openapi-runtime.json is unchanged.
• Merging this triggers an SDK auto-release, since clients/journey-client/src/openapi.json changed.
• Time-sensitive: every auto-release re-introduces this drift, so the branch needs main merged and artifacts regenerated each time one lands. The longer it stays open, the more of those merges it collects.

---

Generated by Claude Code