claude-review.yml requires the first-attempt and retry steps to carry byte-identical prompt: blocks, and says so in a comment ("the two must not diverge"), but nothing enforces it. The sibling lane already has exactly this test.
Current state
claude-review.yml — invariant stated in a comment only. .github/scripts/claude-review-superseded-guard.test.cjs has four tests, none covering parity.claude-security-review.yml — enforced. .github/scripts/claude-security-review-fail-closed.test.cjs:417 asserts with-block equality across the two attempt steps, plus action-pin parity, continue-on-error, and the retry if:.
Why it matters
The two prompt blocks are maintained by hand and are long. A change to the first attempt that misses the retry produces a lane whose retry reviews under different instructions than the first attempt — and the divergence is invisible until someone diffs the blocks manually. Every PR touching this lane's prompt currently relies on the author remembering, and on a reviewer checking by eye.
This was checked by hand during #338 (both blocks identical, 810 bytes) precisely because no test does it.
Suggested fix
Mirror the security lane's assertion into claude-review's test file. The security lane's version compares the whole with: map rather than just prompt, which also catches pin drift and a diverging timeout-minutes — worth keeping that shape rather than narrowing it to the prompt string.
Not folded into #338: that PR is a behavior fix to the lane's arg composition, and this is test coverage for a pre-existing invariant it happened to surface.
Context: #338.
claude-review.ymlrequires the first-attempt and retry steps to carry byte-identicalprompt:blocks, and says so in a comment ("the two must not diverge"), but nothing enforces it. The sibling lane already has exactly this test.Current state
claude-review.yml— invariant stated in a comment only..github/scripts/claude-review-superseded-guard.test.cjshas four tests, none covering parity.claude-security-review.yml— enforced..github/scripts/claude-security-review-fail-closed.test.cjs:417assertswith-block equality across the two attempt steps, plus action-pin parity,continue-on-error, and the retryif:.Why it matters
The two prompt blocks are maintained by hand and are long. A change to the first attempt that misses the retry produces a lane whose retry reviews under different instructions than the first attempt — and the divergence is invisible until someone diffs the blocks manually. Every PR touching this lane's prompt currently relies on the author remembering, and on a reviewer checking by eye.
This was checked by hand during #338 (both blocks identical, 810 bytes) precisely because no test does it.
Suggested fix
Mirror the security lane's assertion into
claude-review's test file. The security lane's version compares the wholewith:map rather than justprompt, which also catches pin drift and a divergingtimeout-minutes— worth keeping that shape rather than narrowing it to the prompt string.Not folded into #338: that PR is a behavior fix to the lane's arg composition, and this is test coverage for a pre-existing invariant it happened to surface.
Context: #338.