docs(pipeline): add author-side gates from contribution retrospective - #2672
Conversation
Derived from a week-long retrospective of the AI-contribution pipeline (Gemini, /review-impl, /pr-contribution-handler) over ~2,000 PR comments. Stops the most-flagged author-side defects at the source. - engine-implementation-executor: blocking discriminating-test gate (the #1 finding across every reviewer was AST-shape-only tests that pass when the fix is reverted) and CR-annotation diff gate (catches hallucinated subparts like CR 702.808); extend the parser-diff-gate regex to rfind/split/split_once (check-parser-combinators.sh blind spots). - engine-implementer: Step 4 confirms both new gates before commit. - pr-contribution-handler: duplicate-PR + scope-contamination intake gate and enqueue item; post comments via --body-file (shell backtick mangling) and batch comment-fetch (rate-limit). - review-impl: codify CR +// compound-annotation convention so Gemini's format pedantry is refuted, not echoed. - AI-CONTRIBUTOR: contributor-side duplicate-PR check (3.1), branch off current upstream/main (4), and test-discrimination + rfind/split in the 5 cross-check.
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
mike-theDude
left a comment
There was a problem hiding this comment.
Architecture Review
[LOW] Prose over-claims the parser-script blind spot by including split_once/rsplit. Evidence: .claude/agents/engine-implementation-executor.md:133 ("The rfind/split/split_once/rsplit arms are deliberate: scripts/check-parser-combinators.sh does not catch them"). Why it matters: scripts/check-parser-combinators.sh:56 already flags \.split_once\( and \.rsplit_once\(, so naming them as uncaught blind spots is inaccurate and could erode trust in the gate or imply a hole that isn't there — the genuinely-uncaught arms are bare rfind/split(/splitn/rsplit(. Suggested fix: narrow the prose to "the rfind/split/splitn/rsplit arms are deliberate (split_once/rsplit_once are already covered by the script)". The AI-CONTRIBUTOR.md (c) edit already scopes this correctly (".rfind/.split are not caught"); mirror that precision here.
Derived from a week-long retrospective of the AI-contribution pipeline
(Gemini, /review-impl, /pr-contribution-handler) over ~2,000 PR comments.
Stops the most-flagged author-side defects at the source.
finding across every reviewer was AST-shape-only tests that pass when the
fix is reverted) and CR-annotation diff gate (catches hallucinated subparts
like CR 702.808); extend the parser-diff-gate regex to rfind/split/split_once
(check-parser-combinators.sh blind spots).
enqueue item; post comments via --body-file (shell backtick mangling) and
batch comment-fetch (rate-limit).
pedantry is refuted, not echoed.
upstream/main (4), and test-discrimination + rfind/split in the 5 cross-check.