code-review

name	code-review
description	Review local code changes: uncommitted work (staged/unstaged) or recent unpushed commits. Use when code needs quality review during local development. Do NOT use for PR review or branch comparison.

Review Flow

Create a task for each step and complete them in order:

1. Get diff — determine review target and collect changes
2. Assess diff size — count lines to select checklist depth
3. Gather context — collect module descriptions, original requirements
4. Load checklist — read references based on diff size (see below)
5. Load output format — read references/output-format.md
6. Assemble review prompt — fill assets/review-prompt-template.md with {context}, {diff}, {checklist}, {output_format}
7. Select executor — choose execution method (see Executor Selection)
8. Dispatch review — send to executor
9. Collect result — parse structured output and Verdict
10. Present to user — show review result; user decides next action
11. Fix (if requested) — initiator executes fixes per user's decision
12. Re-review (if requested) — user may request another review cycle → go to step 1

Review Target

Scenario	Command
Uncommitted changes	git diff + git diff --cached
Recent N commits	git diff HEAD~N..HEAD (default N=1)

Diff Size & Checklist Loading

Count diff output total lines (including context lines) to determine size:

Size	Lines	Action
Small	<50	Read references/review-checklist-core.md
Medium	50-300	Read references/review-checklist-core.md + references/review-checklist-extended.md
Large	>300	Same as Medium; review file-by-file in batches

Executor Selection

Default: Sub agent. Switch to tmux only when user explicitly requests.

Sub agent (default)

Use Claude Code Agent tool to dispatch review in an isolated sub-context. Pass the assembled prompt as the agent's task.

omp dispatch — custom prompt (codex / claude)

Write the assembled review prompt to a temp file, then dispatch:

omp dispatch run codex --prompt-file /tmp/code-review-prompt.md --timeout 600
# or
omp dispatch run claude --prompt-file /tmp/code-review-prompt.md --timeout 600

ANSI-clean output goes to stdout. Exit codes: 0 success, 124 timeout, 1 worker error.

Live observation (optional)

Spawn first, then tail/wait separately if you want to watch progress:

SID=$(omp dispatch spawn codex --prompt-file /tmp/code-review-prompt.md --session-name code-review)
omp dispatch tail "$SID" --follow &
omp dispatch wait "$SID" --timeout 600

codex built-in review (no custom prompt)

codex exec review is a Codex-native subcommand and not driven by a prompt file, so dispatch does not wrap it. Run it directly via tmux when needed:

# Uncommitted changes:
tmux new-session -d -s code-review-builtin \
  'codex exec review --uncommitted --ephemeral --dangerously-bypass-approvals-and-sandbox; exit'

# Specific commit:
tmux new-session -d -s code-review-builtin \
  'codex exec review --commit <SHA> --ephemeral --dangerously-bypass-approvals-and-sandbox; exit'

Cleanup: tmux kill-session -t code-review-builtin 2>/dev/null; rm -f /tmp/code-review-prompt.md

Degradation Strategy

When an executor fails, degrade and inform the user:

omp dispatch fails (spawn error / timeout / worker error)
  → Inform user, degrade to sub agent
    → Sub agent fails
      → Inform user, initiator performs review directly (load checklist, review inline)

Key Constraints

• Initiator prepares context — executor receives a self-contained prompt with all necessary information
• Executor works independently — no access to initiator's conversation history
• User decides fixes — never auto-fix; present results and let user choose
• Every degradation notifies user — state the reason and current execution method