review

name	review
description	Use when completing tasks, implementing major features, or before merging. Spawns parallel review agents for architecture, security, C quality, embedded, and test coverage. Invoke when the user says "review code", "code review", "check my work", or before creating a PR.
argument-hint	[commit-range or file...]
allowed-tools	Read, Bash, Task, Grep

Code Review with Specialized Agents

Argument received (optional): $0 — a commit range (e.g. HEAD~3..HEAD, master..HEAD), specific files, or empty (defaults to uncommitted changes).

This skill collects the patch, then spawns parallel sub-agents that each review from a different angle. Results are synthesised into a single report.

Step 1 — Collect the patch context

Use the review context collection script instead of manual git commands:

scripts/collect-review-context.sh [git-ref] [file...]

Examples:

scripts/collect-review-context.sh                  # uncommitted changes
scripts/collect-review-context.sh HEAD~1..HEAD     # last commit
scripts/collect-review-context.sh master..HEAD     # commits since master

Output: JSON to stdout with patch content, changed files list, commit messages, pre-commit results, and file headers (first 30 lines of each changed file).

JSON output schema:

{
  "script": "collect-review-context",
  "timestamp": "2026-05-25T12:00:00Z",
  "verdict": "READY|EMPTY",
  "files_changed": ["lib/foo/src/foo.c", "lib/foo/include/foo/foo.h"],
  "file_count": 2,
  "patch_lines": 150,
  "diff_stat": "...",
  "patch_content": "...",
  "commit_messages": "...",
  "pre_commit_output": "...",
  "file_headers": [
    {"path": "lib/foo/src/foo.c", "header": "...first 30 lines..."},
    {"path": "lib/foo/include/foo/foo.h", "header": "...first 30 lines..."}
  ],
  "summary": "Collected context for 2 files, 150 lines of diff."
}

Exit codes: 0 = context collected, 1 = no changes to review

If the patch is empty (verdict: "EMPTY"), tell the user there is nothing to review and stop.

Step 2 — Spawn review agents in parallel

Launch all agents in a single message so they run in parallel. Use the fastest available model for each agent to keep cost low and speed high.

Agent prompt templates are in references/. Load the appropriate template and insert the patch, file context, and architecture rules as placeholders.

Agent 1: Architecture Review

See references/arch-review-prompt for the full prompt template. Insert {architecture_rules}, {patch_content}, and {file_headers}.

Agent 2: Security Review

See references/security-review-prompt for the full prompt template. Insert {patch_content}.

Agent 3: Coding Standards & C Quality

See references/c-quality-review-prompt for the full prompt template. Insert {patch_content} and {pre_commit_output}.

Agent 4: Embedded & Resource Review

See references/embedded-review-prompt for the full prompt template. Insert {patch_content}.

Agent 5: Test Coverage Review

See references/test-coverage-review-prompt for the full prompt template. Insert {patch_content}.

Step 3 — Synthesise the report

After all agents complete, combine their findings into a single structured report. Group by severity:

## Code Review Summary

### CRITICAL (must fix before merge)
- <issue from any agent>

### WARNING (should fix)
- <issue from any agent>

### NOTE (consider)
- <issue from any agent>

### Test gaps
- <from test coverage agent>

### Verdict
<PASS / PASS WITH WARNINGS / NEEDS CHANGES>

Deduplicate: if two agents flag the same line, merge into one finding. Attribute each finding to its source angle (arch / security / C quality / embedded / tests).

Red Flags — STOP and re-check

Feeling	Reality
"I can review in my head"	Reviewing your own code misses blind spots. Use agents.
"The patch is small, skip some agents"	All 5 agents run. Small patches are fast.
"Pre-commit already caught everything"	Pre-commit catches style, not arch/security/embedded issues.
"I'll skip the test coverage agent"	Missing test gaps is how regressions ship.

Rules

• Always spawn all 5 agents in a single message (parallel execution)
• Use the fastest available model for each review agent
• Never modify code during review — this skill is read-only
• If the patch is large (>500 lines), mention that the review may miss issues and suggest splitting the change
• Keep individual agent responses under 300 words to avoid noise
• The prompt templates are in references/. Load them, insert context, and dispatch.
• Do not hardcode model names in the skill — use the fastest available model for your platform.

Terminal state: Synthesised review report presented to the user with a clear verdict (PASS / PASS WITH WARNINGS / NEEDS CHANGES).