review model: opencode-go/claude-haiku-5-5#medium instead of
deepseek-v4.1-flash#max; flash has no medium variant
merge lint and typecheck into one job with a single pnpm install
cache the pnpm store and opencode binary via actions/cache@v4.1.0,
keyed on a sha256 of the lockfile (hashFiles is empty on this runner)
update the decision records and log
- review model: opencode-go/claude-haiku-5-5#medium instead of
deepseek-v4.1-flash#max; flash has no medium variant
- merge lint and typecheck into one job with a single pnpm install
- cache the pnpm store and opencode binary via actions/cache@v4.1.0,
keyed on a sha256 of the lockfile (hashFiles is empty on this runner)
- update the decision records and log
- review model: opencode-go/claude-haiku-5-5#medium instead of
deepseek-v4.1-flash#max; flash has no medium variant
- merge lint and typecheck into one job with a single pnpm install
- cache the pnpm store and opencode binary via actions/cache@v4.1.0,
keyed on a sha256 of the lockfile (hashFiles is empty on this runner)
- update the decision records and log
build · claude-haiku-5-5
[0m→ [0mRead .pr-review-diff.patch
build · claude-haiku-5-5
[0m✱ [0mGrep "name: (Lint|Typecheck|Restore|Install|Add opencode|Get PR|Resolve)|continue-on-error|pnpm lint|pnpm typecheck|cache-hit|actions/cache" [90min .gitea/workflows/pr-checker.yml · 22 matches[0m
build · claude-haiku-5-5
Reviewed the diff. Changes are CI workflow YAML plus docs; the workflow is the only executable logic, so findings are limited to it.
[WARN] .gitea/workflows/pr-checker.yml:73 - Typecheck is now a later step in the same job as lint. Without if: always(), a failing pnpm lint skips pnpm typecheck entirely. The comment at lines 13-16 says the split "still report[s] as separate steps, so a failure says which check failed," but the old independent signal is lost. Fix: add if: always() to the Typecheck step, or restore the separate job.
[WARN] .gitea/workflows/pr-checker.yml:52-58 - The pnpm store cache is saved and restored across runs with a restore-keys prefix fallback. Its contents come from pnpm install on PR-controlled code, including lifecycle scripts. A PR run that writes the store can poison the entry restored by later runs, including main. Fix: only save the cache from trusted refs (e.g., if: github.ref == 'refs/heads/main' on a save step), or include a PR-specific component in the key and drop the broad restore-keys fallback.
[WARN] .gitea/workflows/pr-checker.yml:248-259 - The cached ~/.opencode/bin is restored under opencode-v2.0.32-${{ runner.os }} with no integrity check. The review job has OPENCODE_API_KEY in its environment, so a tampered binary in that cache would receive the key. Restricting cache writes to trusted refs (as above) and verifying a checksum after restore before use would close this.
[WARN] OKF_Nuxt/decisions/ai-review-setup.md:22 - The text says the model is "covered by the Go plan subscription, no per-token cost," but the workflow comment (pr-checker.yml, the review step) and the log describe claude-haiku-5-5 as priced per token ($0.10/$0.50 per M). Reconcile the decision record with the actual billing model.
[NIT] OKF_Nuxt/decisions/pr-pipeline-layering.md (Caching section) - This PR adds continue-on-error: true to two cache steps, but the same document's table says "continue-on-error removed everywhere; a failed check is red." Either document the exception (cache outages are not check failures) or note it explicitly in that table.
VERDICT: APPROVE
[0m
> build · claude-haiku-5-5
[0m→ [0mRead .pr-review-diff.patch
> build · claude-haiku-5-5
[0m✱ [0mGrep "name: (Lint|Typecheck|Restore|Install|Add opencode|Get PR|Resolve)|continue-on-error|pnpm lint|pnpm typecheck|cache-hit|actions/cache" [90min .gitea/workflows/pr-checker.yml · 22 matches[0m
> build · claude-haiku-5-5
Reviewed the diff. Changes are CI workflow YAML plus docs; the workflow is the only executable logic, so findings are limited to it.
[WARN] .gitea/workflows/pr-checker.yml:73 - Typecheck is now a later step in the same job as lint. Without `if: always()`, a failing `pnpm lint` skips `pnpm typecheck` entirely. The comment at lines 13-16 says the split "still report[s] as separate steps, so a failure says which check failed," but the old independent signal is lost. Fix: add `if: always()` to the Typecheck step, or restore the separate job.
[WARN] .gitea/workflows/pr-checker.yml:52-58 - The pnpm store cache is saved and restored across runs with a `restore-keys` prefix fallback. Its contents come from `pnpm install` on PR-controlled code, including lifecycle scripts. A PR run that writes the store can poison the entry restored by later runs, including `main`. Fix: only save the cache from trusted refs (e.g., `if: github.ref == 'refs/heads/main'` on a save step), or include a PR-specific component in the key and drop the broad `restore-keys` fallback.
[WARN] .gitea/workflows/pr-checker.yml:248-259 - The cached `~/.opencode/bin` is restored under `opencode-v2.0.32-${{ runner.os }}` with no integrity check. The review job has `OPENCODE_API_KEY` in its environment, so a tampered binary in that cache would receive the key. Restricting cache writes to trusted refs (as above) and verifying a checksum after restore before use would close this.
[WARN] OKF_Nuxt/decisions/ai-review-setup.md:22 - The text says the model is "covered by the Go plan subscription, no per-token cost," but the workflow comment (pr-checker.yml, the review step) and the log describe `claude-haiku-5-5` as priced per token ($0.10/$0.50 per M). Reconcile the decision record with the actual billing model.
[NIT] OKF_Nuxt/decisions/pr-pipeline-layering.md (Caching section) - This PR adds `continue-on-error: true` to two cache steps, but the same document's table says "`continue-on-error` removed everywhere; a failed check is red." Either document the exception (cache outages are not check failures) or note it explicitly in that table.
VERDICT: APPROVE
Noel
merged commit db04d441c9 into main2026-10-20 19:05:35 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
deepseek-v4.1-flash#max; flash has no medium variant
keyed on a sha256 of the lockfile (hashFiles is empty on this runner)
[0m
[WARN] .gitea/workflows/pr-checker.yml:73 - Typecheck is now a later step in the same job as lint. Without
if: always(), a failingpnpm lintskipspnpm typecheckentirely. The comment at lines 13-16 says the split "still report[s] as separate steps, so a failure says which check failed," but the old independent signal is lost. Fix: addif: always()to the Typecheck step, or restore the separate job.[WARN] .gitea/workflows/pr-checker.yml:52-58 - The pnpm store cache is saved and restored across runs with a
restore-keysprefix fallback. Its contents come frompnpm installon PR-controlled code, including lifecycle scripts. A PR run that writes the store can poison the entry restored by later runs, includingmain. Fix: only save the cache from trusted refs (e.g.,if: github.ref == 'refs/heads/main'on a save step), or include a PR-specific component in the key and drop the broadrestore-keysfallback.[WARN] .gitea/workflows/pr-checker.yml:248-259 - The cached
~/.opencode/binis restored underopencode-v2.0.32-${{ runner.os }}with no integrity check. The review job hasOPENCODE_API_KEYin its environment, so a tampered binary in that cache would receive the key. Restricting cache writes to trusted refs (as above) and verifying a checksum after restore before use would close this.[WARN] OKF_Nuxt/decisions/ai-review-setup.md:22 - The text says the model is "covered by the Go plan subscription, no per-token cost," but the workflow comment (pr-checker.yml, the review step) and the log describe
claude-haiku-5-5as priced per token ($0.10/$0.50 per M). Reconcile the decision record with the actual billing model.[NIT] OKF_Nuxt/decisions/pr-pipeline-layering.md (Caching section) - This PR adds
continue-on-error: trueto two cache steps, but the same document's table says "continue-on-errorremoved everywhere; a failed check is red." Either document the exception (cache outages are not check failures) or note it explicitly in that table.VERDICT: APPROVE