pr-comment-fact-check

name	pr-comment-fact-check
description	STOP and fact-check PR review comments before applying or dismissing reviewer or bot feedback. Use on CodeRabbit, Copilot, Cursor bot, dependabot, or human reviewer comments.

PR comment fact-check

STOP before applying or dismissing any PR comment. Each comment is a claim — verify against the actual code and this repo's authoritative sources (.agents/, docs/architecture.md, JSDoc, tests) before acting. Workflow: WORKFLOW.md.

Verdict	Default action
✅ Correct	Apply + resolve
⚠️ Partial	Apply salvageable part + explain nuance
❌ Hallucinated	Push back with evidence; resolve-on-merge-gate exception in WORKFLOW
🕒 Outdated	Point at fix commit + resolve
💭 Style	Apply if cheap, else defer

Hallucination catalog (scrutinize harder)

Common LLM-reviewer patterns on this repo:

1. "This isn't tested" without checking siblings — core contracts are pinned across packages/core/src/layerStack.test.ts (open/dismiss/scope/gcTime/loadFn cancel, transitions enteringDelay/exitingDelay/settle, blockers addBlocker/dismissing/async veto, dismissAll modes); each framework adapter has co-located tests under packages/<fw>/src/; real-renderer paths span packages/*/tests-dom/** (vitest, rerender/detach — e.g. React confirm.test.tsx, blocker.test.tsx, transition.test.tsx). Import isolation is enforced by package manifests, sherif, and workspace-aware knip, not an itImportsOnlyFromCore test. Verify coverage before accepting.
2. Type-safety alarms — if bun run typecheck passes, the claim is almost always wrong, or about runtime behavior the type system can't see (then the reviewer must justify with the runtime case).
3. Generic "best practice" claims unsupported by our rules — "always destructure", "prefer interfaces over types", "add useMemo/useCallback" — stylistic; we either have a rule or we don't. Grep .agents/ for the convention.
4. Convention citations that don't exist — "this breaks the library's API conventions" — grep .agents/ + docs/architecture.md. If not codified, it's preference, not rule.
5. Memory-leak / race-condition claims with no concrete trigger — "this could leak" without a scenario is speculation; ask for the path. Real candidates here: notifyManager.batch flushing, AbortController cancel on dismiss, gcTime cache teardown, useSyncExternalStore unmount-detach, the scope queue draining on dismissAll — demand the specific path.
6. Wrong API for our seam — bot suggests reaching into LayerStack internals from an adapter when the seam is subscribe/getSnapshot; suggests a per-adapter core fork when the core/adapter seam is meant to keep one engine across all frameworks.
7. Public-API suggestions that leak internals — createCallContext is public and intentionally generic; Layer's #state and LayerStack's #layers / #snapshot / #queuedSnapshot / #scopeQueue / #gcCache / #blockers are deliberately private. Push back on "export this helper" / "tighten this to the internal type".

Anti-patterns

• ❌ Applying every suggestion to clear the queue.
• ❌ Replying "fixed!" without verifying.
• ❌ Dismissing without evidence.
• ❌ Resolving a thread you rejected as hallucinated (the reviewer needs to see the receipts).

Reference

• verify-after-each-step — run after applying a fix.
• harden-pr — optional full pass on the branch once comments are triaged.
• docs/architecture.md — seam model, public-surface policy, test matrix (authoritative for "is this the right seam / is this tested" claims).