review

Quick code review for uncommitted changes. Fast, focused on critical issues.

Step 1: Gather Context (Haiku Agent)

Detect what needs to be reviewed:

1. Run `git status` and `git diff` / `git diff --staged`
2. If no changes → inform user and stop
3. Extract:
   - Files changed (list)
   - Languages detected (from extensions)
   - Line counts (additions/deletions)
   - Has CLAUDE.md? (root or in changed directories)

Step 2: Load Agents (Progressive)

Only load agents relevant to detected context:

Condition	Load Agent
Always	@agents/bugs.md
Always	@agents/security.md
CLAUDE.md exists	@agents/compliance.md
.ts/.tsx/.js/.jsx files	@agents/language-typescript.md
.py files	@agents/language-python.md

See @agents/index.md for full routing logic.

Step 3: Run Review (Parallel Sonnet Agents)

Launch loaded agents in parallel. Each agent:

1. Reads full file context for changed files
2. Analyzes only the diff (not pre-existing code)
3. Returns structured issues with diff-style fixes

Output format per agent (see @agents/bugs.md for example):

### 🐛 {{issue_title}}
**Location:** `{{file}}:{{line}}`
**Confidence:** {{score}}/100

**Problem:** {{reason}}

**Suggested Fix:**
```diff
- {{old_code}}
+ {{new_code}}

## Step 4: Score & Filter (Haiku Agents)

For each issue, score confidence 0-100:

| Factor | Points |
|--------|--------|
| In the diff (new code) | +20 |
| Would cause failure | +30 |
| In CLAUDE.md rules | +20 |
| Senior engineer would flag | +20 |
| Has ignore comment | -50 |

Apply filters from `@templates/false-positive-rules.md`:
- Filter issues with score < 80
- Track filtered count by reason

## Step 5: Present Results

Format output using `@templates/output-format.md`:

Code Review

Summary: Reviewed X files, Y lines changed

Found	Reported	Filtered
total	≥80 score	<80 score

Critical (95-100) 🔴

[Issues with diff-style fixes]

Warning (80-94) 🟠

[Issues with diff-style fixes]

Filtered Issues 🔇

[Count by reason, expandable details]