Bring the AI-review callers up to main, and stop any label triggering a review

[kriszyp]

This updates the Claude and Gemini review callers from the July ai-review-prompts revision to the September revision, restricts label events to each provider's own opt-in label, and replays a labeled draft when it becomes ready. Review feedback removed every pnpm change from this PR, leaving package-manager policy to #844 and the final diff limited to the two review callers plus their validator.

For the human reviewer

1. Cancellation ownership: cancellation moves from the caller workflow to the reusable's provider-and-PR-scoped review job after authorization. This prevents an unrelated or unauthorized event from cancelling an admitted review. The pinned Claude reusable and Gemini reusable implement the claimed job-level groups. The tradeoff is deliberate: in opt-in mode, a later push does not invoke the reusable and therefore does not cancel an already-admitted review of the prior SHA. Restoring caller concurrency is easy, but would restore the unrelated-event cancellation bug.
2. Opt-in label lifetime: claude-review and gemini-review remain one-shot gestures, with ready_for_review as the draft replay point, rather than sticky subscriptions to every later push or reopen. Making them sticky is a one-line expression change per provider, but increases paid review frequency and should remain symmetric.
3. Atomic reusable upgrade: one SHA upgrades workflow logic, authorization, scripts, and review layers together instead of selectively adopting pieces. The 19-commit comparison is the review surface, and af81872e was verified as identical to ai-review-prompts main. Rollback is a paired one-SHA change in each caller.
4. Verification strategy: this PR relies on the reusable repository's workflow-contract tests and this repository's live caller validator rather than adding a second truth-table harness. The pinned validator covers both claude-* and gemini-* callers; adding consumer-side tests later is straightforward if that upstream contract proves insufficient.

Changes

• The Claude trigger adds ready_for_review, the caller gate rejects unrelated labels, and removal of caller-level concurrency delegates cancellation to the reusable after authorization. The reusable ref and its paired checkout ref advance together.
• The Gemini caller mirrors the same trigger, provider gate, caller-concurrency removal, reusable ref, and paired checkout ref, keeping the two providers behaviorally aligned.
• The caller validator advances its reusable ref and paired script ref to the same revision, so validation runs against the caller contract shipped by that revision.

Verification

• pnpm build — production bundle and native binding built successfully.
• pnpm check — type-check, oxlint, and oxfmt checks passed.
• pnpm test — 65 files passed, 1 skipped; 906 tests passed, 9 skipped.
• pnpm test:native — 194 tests passed across 30 suites.
• Static workflow checks — all three changed workflows parse as YAML; six reusable/input refs match; both providers contain the provider-specific label and ready_for_review gates; no caller-level concurrency block remains.
• Scope check — all six pnpm workflow files are byte-for-byte identical to the merge base, and the final PR diff contains only the three AI-review workflow files described above.
• Upstream contract check — af81872e is identical to ai-review-prompts main; both reusable review jobs use distinct provider-and-PR concurrency groups after authorization.
• Final independent pre-push review (Claude + Gemini + Harper-domain) covered 6aad6c36 and converged with no unaddressed defect. Its one in-scope tradeoff and one pre-existing policy item are captured above.

The end-to-end route is the live Validate caller workflows check on this pushed head; the runtime/library tests are corroboration because this PR does not change shipped code.

Complexity: medium

🤖 Generated with Claude Code; updated with Codex

_{Review-Coverage: authored=codex; ran=gemini,claude; adjudicated=domain; declined=cursor-grok,cursor-composer; rounds=6 @ 6aad6c3}

_{Human-Review-Need: 3 (decisions: cancellation-ownership, opt-in-label-lifetime, atomic-reusable-upgrade, verification-strategy) @ 6aad6c3}

[gemini-code-assist]

Note

Gemini is unable to generate a review for this pull request due to the file types involved not being currently supported.

[github-actions]

📊 Benchmark Results

get-sync.bench.ts

getSync() > random keys - small key size (100 records)

Implementation	Rank	Operations/sec	Mean (ms)	Min (ms)	Max (ms)	RME (%)	Samples
🥇 lmdb	1	24.51K ops/sec	40.81	39.41	545.023	0.112	122,529
🥈 rocksdb	2	11.13K ops/sec	89.81	86.11	24,207.151	0.979	55,674

getSync() > sequential keys - small key size (100 records)

Implementation	Rank	Operations/sec	Mean (ms)	Min (ms)	Max (ms)	RME (%)	Samples
🥇 lmdb	1	28.66K ops/sec	34.89	33.70	536.907	0.104	143,302
🥈 rocksdb	2	11.87K ops/sec	84.25	81.79	3,710.212	0.149	59,349

ranges.bench.ts

getRange() > small range (100 records, 50 range)

Implementation	Rank	Operations/sec	Mean (ms)	Min (ms)	Max (ms)	RME (%)	Samples
🥇 lmdb	1	24.22K ops/sec	41.30	37.87	1,731.846	0.285	121,076
🥈 rocksdb	2	15.25K ops/sec	65.57	57.69	1,128.905	0.123	76,260

realistic-load.bench.ts

Realistic write load with workers > write variable records with transaction log

Implementation	Rank	Operations/sec	Mean (ms)	Min (ms)	Max (ms)	RME (%)	Samples
🥇 rocksdb	1	345.78 ops/sec	2,892.009	54.86	65,559.929	18.85	692
🥈 lmdb	2	26.64 ops/sec	37,537.981	52.72	1,179,294.532	136.72	64.00

transaction-log.bench.ts

Transaction log > read 100 iterators while write log with 100 byte records

Implementation	Rank	Operations/sec	Mean (ms)	Min (ms)	Max (ms)	RME (%)	Samples
🥇 rocksdb	1	37.48K ops/sec	26.68	12.28	20,899.456	0.851	187,406
🥈 lmdb	2	443.97 ops/sec	2,252.406	172.395	14,656.027	1.40	2,220

Transaction log > read one entry from random position from log with 1000 100 byte records

Implementation	Rank	Operations/sec	Mean (ms)	Min (ms)	Max (ms)	RME (%)	Samples
🥇 rocksdb	1	704.48K ops/sec	1.42	1.22	4,116.191	0.229	3,522,413
🥈 lmdb	2	475.76K ops/sec	2.10	1.07	946.604	0.284	2,378,808

worker-put-sync.bench.ts

putSync() > random keys - small key size (100 records, 10 workers)

Implementation	Rank	Operations/sec	Mean (ms)	Min (ms)	Max (ms)	RME (%)	Samples
🥇 rocksdb	1	836.90 ops/sec	1,194.881	1,049.6	2,055.418	0.279	1,674
🥈 lmdb	2	1.16 ops/sec	864,078.691	816,431.742	927,677.786	3.04	10.00

worker-transaction-log.bench.ts

Transaction log with workers > write log with 100 byte records

Implementation	Rank	Operations/sec	Mean (ms)	Min (ms)	Max (ms)	RME (%)	Samples
🥇 rocksdb	1	21.45K ops/sec	46.62	29.42	20,372.085	2.07	42,899
🥈 lmdb	2	801.87 ops/sec	1,247.079	142.573	13,923.424	5.40	1,605

---

Results from commit ef6fd3d

[cb1kenobi]

I'm fine with the ai-review-prompts updates, but as evident by #844, we don't need to touch the pnpm stuff.