morluto/rea

Fork 0

mirror of https://github.com/morluto/rea.git synced 2026-11-05 07:47:29 +00:00

[GH-ISSUE #3510] fix(ci): include authored GitHub modules in static checker cache inputs #1307

New issue

Closed

opened 2026-10-16 21:04:02 +00:00 by bsakaguchi · 3 comments

bsakaguchi commented

2026-10-16 21:04:02 +00:00

Owner

Originally created by @N0zoM1z0 on GitHub (Oct 14, 2026).
Original GitHub issue: https://github.com/morluto/rea/issues/3510

Originally assigned to: @N0zoM1z0 on GitHub.

Problem

Static checker cache inputs omit authored modules under .github/. Main now contains .github/actions/publish-mcp/registry-state.mjs and its .d.mts declaration, introduced by release-recovery commit 790657dc0. Oxlint scans both files, but the explicit lint and lint:built input inventories in turbo.json do not include them.

This is a real invalidation boundary omission. The existing repository inventory guard correctly blocks CI; removing that assertion would hide the problem. A passing cached lint result can otherwise be reused after one of these checker inputs changes.

Executed baseline and reproduction

Repository: morluto/rea; main eab789aed74acd6ef7c15d509e0f3de350be2733.
Linux, Node 24.18.0, npm 11.16.0, Turbo 2.11.7, Vitest 5.0.3.
Hosted CI 38570471404, failed shard: tests/boundary/filesystem/staticCheckReports.test.ts fails covers the lint inventory and root tool configs in resolved cache inputs.
Local reproduction: npm run test:focused -- tests/boundary/filesystem/staticCheckReports.test.ts — 1 failed, 7 passed, 2.34 s, exit 1.

Observed assertion:

lint omits actual checker inputs:
.github/actions/publish-mcp/registry-state.d.mts
.github/actions/publish-mcp/registry-state.mjs

The guard runs the actual oxlint --debug=files inventory and compares it with actual Turbo dry-run resolved inputs; this is not an assertion against an invented file list. Expected: every file consumed by the selected checker participates in its cache key.

Scope and attribution

The triggering files were added by subsequent release-recovery work. This is separate from #3281's completed mutation-feedback/report changes. Latest-main CI is not being described as passing. No product-runtime defect or published-package failure is claimed from this cache finding.

Imported .github declarations may also affect TypeScript; imported module contents may affect Knip. Those are hypotheses pending actual checker/cache controls, not established findings yet. The investigation will follow these consumers before deciding the complete fix scope.

Proposed correction and verification

Extend each affected task's own input ownership to authored GitHub modules and relevant package/type boundaries. Preserve existing generated-file exclusions and cache reuse for unrelated Markdown/YAML.
Preserve the existing whole-inventory guard.
Use isolated local cache directories: establish a healthy cold/warm pass, inject an actual checker error only under .github, demonstrate baseline stale reuse versus a forced checker failure, then prove the corrected key misses and reports the error. Restore every injected change.
Extend existing regression coverage where it exercises distinct missing invalidation; avoid blanket cache disabling, global invalidation, weaker assertions, or duplicate implementation-shape tests.
Retest checker/report/cache families and current-head CI. Report elapsed/CPU/memory for measured experiments, with cache state and limits explicit.

No corrective source/config changes have been committed at issue creation.

Originally created by @N0zoM1z0 on GitHub (Oct 14, 2026). Original GitHub issue: https://github.com/morluto/rea/issues/3510 Originally assigned to: @N0zoM1z0 on GitHub. ## Problem Static checker cache inputs omit authored modules under `.github/`. Main now contains `.github/actions/publish-mcp/registry-state.mjs` and its `.d.mts` declaration, introduced by release-recovery commit `790657dc0`. Oxlint scans both files, but the explicit `lint` and `lint:built` input inventories in `turbo.json` do not include them. This is a real invalidation boundary omission. The existing repository inventory guard correctly blocks CI; removing that assertion would hide the problem. A passing cached lint result can otherwise be reused after one of these checker inputs changes. ## Executed baseline and reproduction - Repository: `morluto/rea`; main `eab789aed74acd6ef7c15d509e0f3de350be2733`. - Linux, Node 24.18.0, npm 11.16.0, Turbo 2.11.7, Vitest 5.0.3. - Hosted CI [38570471404, failed shard](https://github.com/morluto/rea/actions/runs/38570471404): `tests/boundary/filesystem/staticCheckReports.test.ts` fails `covers the lint inventory and root tool configs in resolved cache inputs`. - Local reproduction: `npm run test:focused -- tests/boundary/filesystem/staticCheckReports.test.ts` — **1 failed, 7 passed**, 2.34 s, exit 1. Observed assertion: ```text lint omits actual checker inputs: .github/actions/publish-mcp/registry-state.d.mts .github/actions/publish-mcp/registry-state.mjs ``` The guard runs the actual `oxlint --debug=files` inventory and compares it with actual Turbo dry-run resolved inputs; this is not an assertion against an invented file list. Expected: every file consumed by the selected checker participates in its cache key. ## Scope and attribution The triggering files were added by subsequent release-recovery work. This is separate from #3281's completed mutation-feedback/report changes. Latest-main CI is **not** being described as passing. No product-runtime defect or published-package failure is claimed from this cache finding. Imported `.github` declarations may also affect TypeScript; imported module contents may affect Knip. Those are **hypotheses pending actual checker/cache controls**, not established findings yet. The investigation will follow these consumers before deciding the complete fix scope. ## Proposed correction and verification - Extend each affected task's own input ownership to authored GitHub modules and relevant package/type boundaries. Preserve existing generated-file exclusions and cache reuse for unrelated Markdown/YAML. - Preserve the existing whole-inventory guard. - Use isolated local cache directories: establish a healthy cold/warm pass, inject an actual checker error only under `.github`, demonstrate baseline stale reuse versus a forced checker failure, then prove the corrected key misses and reports the error. Restore every injected change. - Extend existing regression coverage where it exercises distinct missing invalidation; avoid blanket cache disabling, global invalidation, weaker assertions, or duplicate implementation-shape tests. - Retest checker/report/cache families and current-head CI. Report elapsed/CPU/memory for measured experiments, with cache state and limits explicit. No corrective source/config changes have been committed at issue creation.

bsakaguchi closed this issue

2026-10-16 21:04:02 +00:00

bsakaguchi commented

2026-10-16 21:04:03 +00:00

Author

Owner

@N0zoM1z0 commented on GitHub (Oct 14, 2026):

Confirmed consumers and current follow-up

Upstream c0aa2e32b fixed the original .github/actions code omission in all four checker tasks. This follow-up retains that fix and addresses remaining producer inputs on main 6060de329de9f3636f63e24a2419847fd9d3f889:

TypeScript follows imported GitHub declarations through package scopes. An added/changed .github/**/package.json can change module interpretation and diagnostics. Its actual --listFilesOnly inventory consumes registry-state.d.mts, not the JS implementation; the follow-up therefore keeps typed sources and package scopes without invalidating typechecking for JS-only body edits.
Knip's installed GitHub Actions plugin reads .github/workflows/*.{yml,yaml} and .github/**/action.{yml,yaml} to discover reachable modules. Changing either entrypoint manifest can change unlisted dependencies without changing any previously hashed code. Package scopes also participate in resolution.

Actual failures and controls

On the original eab789aed baseline, isolated real repository caches for lint, TypeScript and Knip each reused success after a relevant module/declaration fault. A write-only local-cache execution (no cache reads or remote activity) rejected each fault with its expected parser / TS2345 / unlisted-dependency diagnostic.
On current parent 6060de329, the five added compiler/Knip regression cases yield 3 failed / 2 passed without the follow-up: package scope plus both entrypoint manifests fail; source/declaration cases already covered by upstream pass.
Independently omitting Knip module ownership makes both manifest scenarios fail after a previously reachable healthy module changes dependencies. Omitting manifest ownership also fails both scenarios. These are actual checker/cache flows, not glob-shape assertions.
Final affected families: 2 files / 23 tests pass (64.04 s on the shared host). A broader integration also passed CI selection, registry/release recovery and the real installed-tarball MCP canary. check:ci, 1,398-source architecture guards and normal commit/pre-push checks pass.
Existing whole lint inventory and cross-checkout report-restoration coverage are retained unchanged. The extra lint scenario initially added during investigation was removed as redundant. Unrelated notes/YAML still reuse caches; runtime compilation stays cached for imported GitHub test-only changes.

Current real repository three-checker probe under the corrected input configuration:

Task	Healthy warm cache	Warm wall	Relevant key changes	Fault run exit	Intended forced diagnostic
lint	Cached: 1 cached, 1 total	0.416 s	True	1	True
typecheck	Cached: 1 cached, 1 total	0.418 s	True	1	True
knip	Cached: 3 cached, 3 total	0.417 s	True	1	True

The complete experiment includes cold runs, warm runs, deliberately invalid states and uncached controls: 75.87 s wall / 276.20 s user / 33.84 s system CPU / 4,194.60 MiB sampled summed process-family RSS. Sampling is 200 ms; shared pages can be counted twice and short-lived/detached peaks can be missed. This is not normal healthy CI latency, an equivalent-workload speedup comparison, or an agent-token measurement. Every injected file was restored. CLI-usage-error and disabled-lint-rule probes were explicitly discarded, not counted as checker detections.

Implementation head: c892fb223028628cde4b302a30c013936b757e32, two files only. Anonymous author/committer and requested Kaoru co-author trailer retained. Hosted current-head CI/review is being started; no passing result is claimed yet.

@N0zoM1z0 commented on GitHub (Oct 14, 2026): ## Confirmed consumers and current follow-up Upstream `c0aa2e32b` fixed the original `.github/actions` code omission in all four checker tasks. This follow-up retains that fix and addresses remaining producer inputs on main `6060de329de9f3636f63e24a2419847fd9d3f889`: 1. TypeScript follows imported GitHub declarations through package scopes. An added/changed `.github/**/package.json` can change module interpretation and diagnostics. Its actual `--listFilesOnly` inventory consumes `registry-state.d.mts`, not the JS implementation; the follow-up therefore keeps typed sources and package scopes without invalidating typechecking for JS-only body edits. 2. Knip's installed GitHub Actions plugin reads `.github/workflows/*.{yml,yaml}` and `.github/**/action.{yml,yaml}` to discover reachable modules. Changing either entrypoint manifest can change unlisted dependencies without changing any previously hashed code. Package scopes also participate in resolution. ## Actual failures and controls - On the original `eab789aed` baseline, isolated real repository caches for **lint, TypeScript and Knip** each reused success after a relevant module/declaration fault. A write-only local-cache execution (no cache reads or remote activity) rejected each fault with its expected parser / TS2345 / unlisted-dependency diagnostic. - On current parent `6060de329`, the five added compiler/Knip regression cases yield **3 failed / 2 passed** without the follow-up: package scope plus both entrypoint manifests fail; source/declaration cases already covered by upstream pass. - Independently omitting Knip module ownership makes both manifest scenarios fail after a previously reachable healthy module changes dependencies. Omitting manifest ownership also fails both scenarios. These are actual checker/cache flows, not glob-shape assertions. - Final affected families: **2 files / 23 tests pass** (64.04 s on the shared host). A broader integration also passed CI selection, registry/release recovery and the real installed-tarball MCP canary. `check:ci`, 1,398-source architecture guards and normal commit/pre-push checks pass. - Existing whole lint inventory and cross-checkout report-restoration coverage are retained unchanged. The extra lint scenario initially added during investigation was removed as redundant. Unrelated notes/YAML still reuse caches; runtime compilation stays cached for imported GitHub test-only changes. Current real repository three-checker probe under the corrected input configuration: | Task | Healthy warm cache | Warm wall | Relevant key changes | Fault run exit | Intended forced diagnostic | | --- | --- | ---: | --- | ---: | --- | | lint | Cached: 1 cached, 1 total | 0.416 s | True | 1 | True | | typecheck | Cached: 1 cached, 1 total | 0.418 s | True | 1 | True | | knip | Cached: 3 cached, 3 total | 0.417 s | True | 1 | True | The complete experiment includes cold runs, warm runs, deliberately invalid states and uncached controls: **75.87 s wall / 276.20 s user / 33.84 s system CPU / 4,194.60 MiB sampled summed process-family RSS**. Sampling is 200 ms; shared pages can be counted twice and short-lived/detached peaks can be missed. This is not normal healthy CI latency, an equivalent-workload speedup comparison, or an agent-token measurement. Every injected file was restored. CLI-usage-error and disabled-lint-rule probes were explicitly discarded, not counted as checker detections. Implementation head: `c892fb223028628cde4b302a30c013936b757e32`, two files only. Anonymous author/committer and requested Kaoru co-author trailer retained. Hosted current-head CI/review is being started; no passing result is claimed yet.

bsakaguchi commented

2026-10-16 21:04:04 +00:00

Author

Owner

@N0zoM1z0 commented on GitHub (Oct 14, 2026):

Current exact head c892fb223028628cde4b302a30c013936b757e32 has completed hosted code validation: CI run 38578694538, 35 jobs successful. Downloaded merged artifact 11903757172 parses to 1,281 unique files / 14,599 passed / 125 skipped / zero failed / no duplicates. Exact-head Greptile source review is 5/5 with no inline findings; it did not run tests itself.

The candidate merge tree currently matches the qualified PR tree. The PR is ready (not draft) as requested. Initial human ownership/diff/evidence/readiness declarations for this separate change are still pending; they have not been inferred from the earlier #3281 confirmation or fabricated. This issue remains open until an actual merge.

@N0zoM1z0 commented on GitHub (Oct 14, 2026): Current exact head `c892fb223028628cde4b302a30c013936b757e32` has completed hosted code validation: [CI run 38578694538](https://github.com/morluto/rea/actions/runs/38578694538), **35 jobs successful**. Downloaded merged artifact `11903757172` parses to **1,281 unique files / 14,599 passed / 125 skipped / zero failed / no duplicates**. Exact-head Greptile source review is 5/5 with no inline findings; it did not run tests itself. The candidate merge tree currently matches the qualified PR tree. The PR is ready (not draft) as requested. Initial human ownership/diff/evidence/readiness declarations for this separate change are still pending; they have not been inferred from the earlier #3281 confirmation or fabricated. This issue remains open until an actual merge.

bsakaguchi commented

2026-10-16 21:04:04 +00:00

Author

Owner

@N0zoM1z0 commented on GitHub (Oct 14, 2026):

Merged as eb44921c4c3682e33c78508285629926e8ef241e in #3522 after all 35 current-head CI jobs passed, exact-head Greptile review scored 5/5 with no inline findings, and the human owner's actual initial declarations passed readiness validation.

Downloaded and parsed current-head test artifact 11905109208: 1,284 unique test files, 14,650 passed, 125 explicitly skipped, zero failures and no duplicate files. Current-head CI: https://github.com/morluto/rea/actions/runs/38585311071 . Latest main integration also passed the 35-test cache/report/evaluation check and all eight static tasks plus architecture guards; the preceding Ghidra integration passed 165 tests with two explicit skips.

Post-merge verification: actual parent ca1b2f8608c52ed721d3576ba5eeb1d66717b08c; actual tree 94968171e2ad7fc22b129c5245c1daac90202dac exactly equals the qualified PR tree. Verified the anonymous author email and requested Kaoru co-author trailer in the actual squash commit. The change is limited to turbo.json and the real cache regression fixture: stale success is rejected while healthy warm-cache reuse is retained. No runtime changes, live-test caching, global invalidation or weakened test oracles. No CI-wide speedup is claimed from the cold/warm/error experiment.

@N0zoM1z0 commented on GitHub (Oct 14, 2026): Merged as `eb44921c4c3682e33c78508285629926e8ef241e` in #3522 after all 35 current-head CI jobs passed, exact-head Greptile review scored 5/5 with no inline findings, and the human owner's actual initial declarations passed readiness validation. Downloaded and parsed current-head test artifact `11905109208`: 1,284 unique test files, 14,650 passed, 125 explicitly skipped, zero failures and no duplicate files. Current-head CI: https://github.com/morluto/rea/actions/runs/38585311071 . Latest main integration also passed the 35-test cache/report/evaluation check and all eight static tasks plus architecture guards; the preceding Ghidra integration passed 165 tests with two explicit skips. Post-merge verification: actual parent `ca1b2f8608c52ed721d3576ba5eeb1d66717b08c`; actual tree `94968171e2ad7fc22b129c5245c1daac90202dac` exactly equals the qualified PR tree. Verified the anonymous author email and requested Kaoru co-author trailer in the actual squash commit. The change is limited to `turbo.json` and the real cache regression fixture: stale success is rejected while healthy warm-cache reuse is retained. No runtime changes, live-test caching, global invalidation or weakened test oracles. No CI-wide speedup is claimed from the cold/warm/error experiment.

bsakaguchi referenced this issue

2026-10-17 21:10:14 +00:00

[PR #1307] [CLOSED] fix(browser): compare complete frame projections deterministically #2787

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

morluto/rea#1307

No description provided.

Rows
Columns