unified-code-review

Agent skill for PR / branch review that forces a fixed order: risk → agent-authored checks → call-graph pincer → structure. It is a sensor (PASS / ADVISORY / BLOCKERS), not a merge button — humans still own the merge. In agent pipelines, treat BLOCKERS as actionable; leave ADVISORY / nits for human triage.

The problem

LLM reviews often fail in two predictable ways:

1. Green CI = ship — the agent rubber-stamps after local tests pass, including tests it weakened or gamed.
2. Single-file blindness — it audits one function and never opens the callee. Classic miss: callers assume getUser throws on missing users; the callee returns null; wrong branch / silent failure downstream.

Structure-only review (code golf / “make it cleaner”) without risk triage makes that worse.

What it does

Pass	Focus
1. Risk	Blast radius first (auth, payments, migrations ≠ copy tweaks). What could go wrong, what to read line-by-line, empirical checks.
1b. Ops laws	Only if the repo defines them (REVIEWS.md / task / deploy gates). Else skip.
2. Agent-authored	Intent evidence; tests first; treat agent output as unreviewed external code.
2b / 2c. Pincer	Trace one level deeper before BLOCKERS. Bidirectional wiring check: what callers assume vs what callees do. Default Lite, not Full.
3. Structure	Code judo — delete branches/layers that can disappear; presumptive blockers (no tests, assertion gaming, spaghetti, …).

The full rubric, output template, and Full-tier pincer appendix live in SKILL.md. That file is what agents load.

Install

npx skills add dancingteeth/unified-code-review

Or copy SKILL.md into your agent skills directory (e.g. ~/.cursor/skills/unified-code-review/, .agents/skills/unified-code-review/).

Use it for PR/branch audits and agent-authored diffs — not as an always-on rule (too large for every chat).

Tested with (author dogfood)

Host	Models
Cursor	Composer 2.5, Grok 4.5, GPT 5.6 Terra
Kilo	Tencent HY3, Ling 3.0 Flash (inclusionAI)
Cline	Qwen 3.5, DeepSeek Flash, DeepSeek Pro

Observed use, not a certification matrix. Weaker models should stay on the TL;DR path (Lite pincer by default).

If it helped

Star the repo — cheap signal that this is worth keeping public.

Better: ask your agent to open a short GitHub issue with what it caught — risk tier, one failure mode, verdict (PASS / ADVISORY / BLOCKERS).

Optional tip: Ko-fi. More agent tooling notes: Vibing Agents.