codex-review
Send code to Codex for deep review and bug finding. Use this PROACTIVELY after writing significant code, completing implementations, or when complex code needs analysis. Also use when user explicitly says codex review, delegate review, or wants thorough code analysis.
When & Why to Use This Skill
The Codex Review skill facilitates deep, automated code analysis by delegating review tasks to a specialized Codex agent via tmux-based IPC. It is designed to proactively identify complex bugs, security vulnerabilities, and performance issues, providing a comprehensive second opinion on significant code changes or implementations to ensure high-quality software delivery.
Use Cases
- Deep Logic Review: Analyzing complex algorithms or intricate business logic to identify subtle edge cases and potential logic errors.
- Security Auditing: Proactively scanning new code implementations for security vulnerabilities and unsafe coding patterns before deployment.
- Post-Implementation Quality Check: Performing a thorough, automated review after completing a major feature or significant refactoring to maintain code health.
- Multi-Agent Collaboration: Leveraging a secondary specialized agent (Codex) in a tmux environment to provide high-fidelity feedback without interrupting the primary development workflow.
| name | codex-review |
|---|---|
| description | "[TMUX MODE] Send code to Codex via tmux file-based IPC. Only use when user explicitly runs /codex-review command. For natural language requests like 'review with codex', use the delegate_codex_review MCP tool instead." |
Codex Review Skill
Send code to the Codex agent (running in tmux pane 1) for deep code review.
When to Use
- User wants thorough code review
- User says "codex review" or "delegate review"
- Complex code needs security/bug analysis
Steps
0. Check for Tmux
First, verify we're running in tmux. Run this check:
[ -n "$TMUX" ] && echo "TMUX_OK" || echo "NOT_IN_TMUX"
If NOT_IN_TMUX: Stop immediately and tell the user:
"This skill requires tmux dual-pane mode. You're not in tmux.
Instead, just ask me naturally: 'review this code with codex' and I'll use the MCP tool which works without tmux."
Do not proceed with the remaining steps if not in tmux.
1. Resolve .agent-collab Directory
Before any file operations, resolve the .agent-collab directory so commands work outside the project root:
AGENT_COLLAB_DIR="${AGENT_COLLAB_DIR:-}"
if [ -n "$AGENT_COLLAB_DIR" ]; then
if [ -d "$AGENT_COLLAB_DIR/.agent-collab" ]; then
AGENT_COLLAB_DIR="$AGENT_COLLAB_DIR/.agent-collab"
elif [ ! -d "$AGENT_COLLAB_DIR" ]; then
AGENT_COLLAB_DIR=""
fi
fi
if [ -z "$AGENT_COLLAB_DIR" ]; then
AGENT_COLLAB_DIR="$(pwd)"
while [ "$AGENT_COLLAB_DIR" != "/" ] && [ ! -d "$AGENT_COLLAB_DIR/.agent-collab" ]; do
AGENT_COLLAB_DIR="$(dirname "$AGENT_COLLAB_DIR")"
done
AGENT_COLLAB_DIR="$AGENT_COLLAB_DIR/.agent-collab"
fi
If $AGENT_COLLAB_DIR does not exist, stop and ask for the project root.
2. Gather Code to Review
Ask user what to review if not specified:
- Specific file(s)
- Recent changes (git diff)
- A code block they provide
3. Write Task Request
Write to $AGENT_COLLAB_DIR/requests/task.md:
# Task Request for Codex
## Task Type: CODE_REVIEW
## Timestamp
[Current timestamp]
## Files to Review
[List files with FULL ABSOLUTE paths, e.g. /Users/antor/project/src/main.cpp]
**NOTE: Codex runs in the same working directory and CAN read these files directly.
Do NOT copy file contents here. Just list the paths and Codex will read them.**
## Review Focus
- Look for bugs, edge cases, logic errors
- Check for security vulnerabilities
- Identify performance issues
- Suggest improvements
## Specific Concerns
[Any areas user wants examined]
4. Update Status
Write pending to $AGENT_COLLAB_DIR/status
5. Trigger Codex
Run this bash command to trigger Codex in the other pane:
tmux send-keys -t 1 '$read-task' && sleep 0.5 && tmux send-keys -t 1 Enter Enter
6. Notify User
Tell user briefly that the review was delegated to Codex.
7. Wait for Codex (Background Polling)
Start a background polling loop to wait for Codex to complete. Run this EXACT bash command (with $AGENT_COLLAB_DIR/status) using the Bash tool with run_in_background: true:
while [ "$(cat "$AGENT_COLLAB_DIR/status")" != "done" ]; do sleep 3; done; echo "CODEX_COMPLETE"
CRITICAL: Use the resolved $AGENT_COLLAB_DIR/status path so polling works outside the project root. Use background execution so you can continue helping the user while waiting.
8. Auto-Read Response
When the background poll completes (returns "CODEX_COMPLETE"), automatically:
- Read
$AGENT_COLLAB_DIR/responses/response.md - Present findings to user with clear formatting
- Reset
$AGENT_COLLAB_DIR/statustoidle
This should happen seamlessly - user sees the delegation message, then later sees the results appear automatically.