
Review
- 393 installs
- 1 repo stars
- Updated May 26, 2026
- camacho/ai-skills
review is an agent skill that performs structured code and plan reviews on branches, PRs, or local diffs so developers can catch defects, security issues, and design problems before merge.
About
review is a unified camacho/ai-skills agent skill that auto-detects whether to review a plan or code, assembles the right reviewer panel, and runs a bounded review-fix loop with severity gating. Plan mode triggers when a non-done plan exists in ai-workspace/plans/ without commits beyond main; code mode triggers on git diff main...HEAD output. Default round cap is 3 with thorough at 5 and quick at 1, hard-capped at 5 rounds. Conditional reviewers add architect-reviewer, security-auditor, codex-specialist, or design-reviewer based on signals in the diff. Code mode fixes P0–P3 findings with a terminal P3 sweep that skips re-dispatch when only nits remain, recording deferred items via gh issue create or .branch-context.md. Developers reach for review when a plan or implementation needs structured feedback—not for greenfield coding without reviewable artifacts.
- Diff and branch review workflows
- Severity-ranked finding reports
- Security and correctness checks
- Actionable fix recommendations
- Standards-aligned quality rubrics
Review by the numbers
- 393 all-time installs (skills.sh)
- Ranked #262 of 1,352 Code Review & Quality skills by installs in the Skillselion catalog
- Data as of Jul 28, 2026 (Skillselion catalog sync)
npx skills add https://github.com/camacho/ai-skills --skill reviewAdd your badge
Show developers this skill is listed on Skillselion. Paste this into your README.
| Installs | 393 |
|---|---|
| repo stars | ★ 1 |
| Last updated | May 26, 2026 |
| Repository | camacho/ai-skills ↗ |
How do you run structured pre-merge code review?
Perform structured code reviews on branches, PRs, or local diffs to catch defects, security issues, and design problems before merge.
Who is it for?
Developers with a plan file or branch diff who want an agent-orchestrated review-fix loop with P0–P3 triage before merge.
Skip if: Greenfield implementation with no plan or diff yet, or teams wanting a one-shot lint pass without iterative fix-and-re-review governance.
When should I use this skill?
User requests /review, asks to review a plan or branch diff, or needs pre-merge defect and security review with severity gating.
What you get
Severity-tagged review findings, fix commits, deferred issue records, and an APPROVE or ESCALATE verdict summary.
- Review verdict summary
- Fix commits
- Deferred finding records
By the numbers
- Default review round cap of 3 with hard maximum of 5 rounds
- Severity scale P0 through P3 with mode-specific default gates
- Conditional panel adds up to 4 specialist reviewers based on diff signals
Files
/review
Detects whether to review a plan or code, assembles the right panel, runs the review-fix loop.
1. Detect Mode
From args or context:
1. Explicit: /review plan or /review code — use that. 2. Plan file exists in ai-workspace/plans/ (non-.done.md) and no commits beyond main — plan mode. 3. git diff main...HEAD produces output — code mode. 4. Ambiguous — ask: "Reviewing the plan or the code?"
Optional overrides after mode:
- Round cap: default 3. "thorough" (5) or "quick" (1).
- Severity gate: defaults differ by mode — plan mode = P2, code mode = P3 (local-merge flow assumes nits are cheaper to fix now than to file as issues). "strict" forces P1, "p2" forces P2, "lenient" forces P3.
2. Assemble Panel
Plan mode (Step 4)
Lead: technical-editor (always). Target: the plan file.
Code mode (Step 6)
Lead: code-reviewer (always). Target: changed files (git diff main...HEAD).
Conditional reviewers (both modes)
Scan the target for signals. Add matching reviewers:
| Signal | Plan mode | Code mode |
|---|---|---|
| Architecture, new modules, DI patterns | architect-reviewer | architect-reviewer |
| AGENTS.md, config.toml, sync.sh, skills | codex-specialist | codex-specialist |
| Auth, credentials, tokens, secrets | security-auditor | security-auditor |
| UI, CSS, components, accessibility | ui-designer | design-reviewer |
No signal match — run with lead only.
Invoke /assemble-panel if available for RETAIN, EXPAND, CONVERGE, and ESCALATE_RECURRING governance. Do not duplicate its policy algebra.
Codex/Cursor: read .agents/skills/assemble-panel/SKILL.md and apply inline.
3. Dispatch
- Plan mode: invoke
/plan-reviewvia Skill tool. Pass panel and round cap. Its internal P0 loop counts as R1. - Code mode: dispatch panel via Agent tool. Request severity-tagged findings (P0-P3).
Untagged findings: treat as P1.
4. Loop: Fix, Record, Re-review
For each round (R1..R{cap}):
a. Triage by severity gate:
- Fix (P0 through P{gate}): fix immediately. Commit:
fix: address R{N} review findingswithCo-Authored-Bytrailer. - Record (P{gate+1}+): record, do not fix.
b. Record below-gate findings:
- Plans: append
### Deferred Findings (R{N})to the plan file. - Code:
gh issue create --label review-finding --title "R{N}: <summary>". Cap 10/round, 20 total. Ifghunavailable, append to.branch-context.md.
c–f. Panel governance — invoke /assemble-panel for RETAIN, EXPAND, CONVERGE, ESCALATE_RECURRING.
Terminal P3 Sweep (code mode, gate = P3)
P3 nits multiply: fixing 15 P3s produces a diff that re-enters the panel and surfaces 15 new P3s. To bound this, the last fix round in code mode is a terminal sweep:
1. Apply P3 fixes from the round. 2. Commit: fix: address R{N} review findings (P3 sweep). 3. Run validation (typecheck + tests). If any fail, treat the failure as a new round at gate=P2 — the sweep is voided and the loop continues normally. 4. Do NOT re-dispatch the panel for another round. 5. Print the summary and exit with verdict APPROVE.
A round is the terminal sweep when all of: gate is P3, no P0/P1/P2 findings remain unfixed, and the round surfaced only P3 findings. P0–P2 findings always re-enter the loop normally (terminal sweep applies only to nit-only rounds).
5. Summary
Review Summary
───────────────────────────────────────
Mode: plan | code
Target: <file or diff summary>
Rounds: <N of cap>
Gate: P<N>
│ Round │ P0 │ P1 │ P2 │ P3+ │ Fixed │ Recorded │
│───────│────│────│────│─────│───────│──────────│
Verdict: APPROVE | ESCALATE- APPROVE — all above-gate findings resolved. Next step:
- Plan mode → Build (Step 5)
- Code mode → Archive (Step 7)
- ESCALATE — unresolved above-gate findings after cap. Present to human.
Guardrails
- Never modify protected files (see AGENTS.md).
- Hard max: 5 rounds even if requested higher.
- Never auto-merge or auto-approve. Produces a verdict, not an action.
- Plain text output only — cross-surface safe.
Fallbacks
/plan-reviewunavailable → dispatchtechnical-editor+architect-reviewerdirectly./assemble-panelunavailable → fall back to[technical-editor, code-reviewer], gate=P2, cap=3.ghnot authenticated → degrade to.branch-context.md.- Reviewer returns DROP → exit immediately, print summary.
Related skills
How it compares
Use review for agent-orchestrated plan and code review-fix loops; use single-purpose security or bugbot skills when you only need one focused audit pass.
FAQ
How does review choose plan mode versus code mode?
The review skill uses explicit /review plan or /review code arguments first. Otherwise plan mode applies when a non-done file exists in ai-workspace/plans/ without commits beyond main, and code mode applies when git diff main...HEAD produces output.
What is the default review round cap?
The review skill defaults to 3 rounds with optional thorough mode at 5 and quick mode at 1. A hard maximum of 5 rounds applies even if a higher cap is requested, and the skill never auto-merges or auto-approves.
How does review handle low-severity P3 nits?
In code mode with a P3 gate, the review skill runs a terminal P3 sweep on the last fix round: apply nits, commit, run validation, and skip re-dispatch when only P3 findings remain. Failures void the sweep and continue at gate P2.