Review

Unified Review

Run review in two phases: review phase then optional fix phase. $ARGUMENTS

Parameters

Required

• scope determines review target:
  • scope=branch → review branch diff against detected base
  • scope=changes → review uncommitted changes (git diff + git diff --cached)
  • scope=pr or scope=pr:<id|#id|url> → review PR diff

If scope missing or invalid, ask the user.

Optional

• Coverage parameter decides reviewer set:
  • full → 7 reviewers
  • partial or half → 4 reviewers

Default coverage: full

• autofix decides whether fix phase runs:
  • explicit autofix or autofix=true → force ON
  • explicit autofix=false → force OFF
  • omitted → auto mode (ON when total findings < 10, otherwise OFF)

Reviewer Sets

Full (7)

• mp-reviewer-code-quality
• mp-reviewer-best-practices
• mp-reviewer-spec-alignment
• mp-reviewer-test-quality
• mp-reviewer-security
• mp-reviewer-performance
• mp-reviewer-error-handling

Partial/Half (4)

• mp-reviewer-code-quality
• mp-reviewer-best-practices
• mp-reviewer-spec-alignment
• mp-reviewer-test-quality

Workflow

Step 1: Resolve scope

Branch scope

1. Determine current branch (git branch --show-current)
2. Run node ${CLAUDE_PLUGIN_ROOT}/scripts/detect-base-branch.js [explicit-branch] (pass user hint as arg if given). It checks origin/<branch> existence and commits-ahead for dev > develop > main > master and prints one branch name to stdout (falls back to main).
3. Use the printed branch as <base>
4. Build diff scope: <base>...HEAD using git diff <base>...HEAD

Changes scope

1. Collect git diff and git diff --cached
2. If both empty, report "no changes" and stop

PR scope

1. Resolve PR from scope=pr:<...>; otherwise use current branch PR
2. Fetch metadata + diff (gh pr view, gh pr diff, draft allowed)
3. If PR cannot be resolved, report blocker and stop

Step 2: Review phase

1. Select reviewer set from coverage param
2. Spawn reviewers in parallel with resolved scope context:
   • diff body/scope
   • changed files
   • original task/spec text when available
   • stack/conventions context
3. Accept high-confidence findings only
4. Merge findings and classify exactly in this order:
   1. Critical
   2. Important
   3. Minor (optional)
5. Write report only if findings exist into REVIEW.md

Report shape:

• Actionable Checklist
  • Critical items with location, why, concrete fix
  • Important items with location, why, concrete fix
• Nice-to-Have
  • Minor observations

If no findings, skip the report file and return a clean summary in conversation.

Step 3: Autofix decision

Compute total issues from merged findings (Critical + Important + Minor).

• If user explicitly set autofix param, honor it always.
• If omitted:
  • total findings < 10 → autofix ON
  • total findings >= 10 → autofix OFF

Step 4: Fix phase (conditional)

Run only when autofix ON and at least one finding exists.

Analyze each finding and determine the concrete fix — identify exact file paths, lines, and the specific code change needed. Then spawn mp-executor sub-agent with pre-analyzed fix instructions:

• concrete fix per finding: file path, current code, and exact change to apply (the executor applies — it does not diagnose)
• resolved scope summary (branch/changes/pr)
• report path (REVIEW.md)
• requirement: fix only in-scope findings

After executor completes, run the same selected reviewer set again on updated scope.

• If issues remain, append ## Post-Fix Review section to the same report and repeat Fix Phase up to 3 iterations or until clean
• If no issues remain, append ## Post-Fix Review with clean status

Behavior Contract

• Read-only when fix phase does not run
• Source edits only in fix phase via mp-executor
• No commits
• No GitHub comments/reviews

Output

Mode: review
Scope: [branch|changes|pr]
Coverage: [full|partial]
Autofix: [true|false|auto-resolved]

Findings:

- total: [N]
- critical: [N]
- important: [N]
- minor: [N]

Report:

- [path or "none"]

Fix Step:

- executed: [yes/no]
- result: [clean/remaining issues/not requested]