code-review is clobbered by Claude built-in skill #1234

<html>
<body>
<!--StartFragment--><h3 dir="ltr">Comparison</h3><div><div>
  | Built-in /code-review | Your on-disk code-review
-- | -- | --
Question asked | “Is this diff buggy or messy?” | “Does it follow our standards, and does it match the spec?”
Scope | @{upstream}...HEAD plus uncommitted changes, or a PR, branch or path | <fixed-point>...HEAD, which you must supply. It asks if missing and checks the ref resolves and the diff isn’t empty.
Structure | Many angles in one finder pool: line-by-line scan, removed-behavior auditor, cross-file tracer, language pitfalls, wrapper correctness, plus reuse, simplification, efficiency, altitude and CLAUDE.md conventions | Exactly two sub-agents, Standards and Spec, run in parallel
Verification | Each candidate gets one verifier vote (CONFIRMED / PLAUSIBLE / REFUTED), then a sweep for gaps | None. Sub-agent output goes into the report mostly as-is, capped at 400 words each.
Output | Ranked findings (file, line, summary, failure_scenario), capped at 4 (low), 8 (medium) or 10 (high) | Two sections, ## Standards and ## Spec, deliberately never merged or reranked
Spec awareness | None. It never asks what the change was meant to do. | Central. It finds the spec via commit refs, an argument, or docs/, specs/ and .scratch/.
Standards | Only CLAUDE.md rules, and only when it can quote the rule and the violating line | Repo standards docs plus a fixed 12-smell Fowler baseline, always treated as judgement calls
Bug hunting | Its main job, with explicit angles for removed guards and call-site breakage | Not covered
Effort levels / flags | low to max, plus --fix, --comment, --post, ultra | None, and it pins model: sonnet
Dependencies | Self-contained | Needs docs/agents/issue-tracker.md from /setup-agentic-project

</div></div><!--EndFragment-->
</body>
</html>