Codex CLI Sub-Agent

Delegate coding tasks to OpenAI's Codex CLI (codex exec) as a sub-agent. This gives you (the host agent) sub-agent capabilities — running independent AI agents that work on focused tasks and return results.

What is the "Host Agent"?

You — the AI coding agent reading this skill right now — are the host agent. Whether you are Windsurf/Cascade, Claude Code, Cursor, Codex, Gemini CLI, or any other SKILL.md-compatible agent, this skill is talking to you. When this document says "the host agent should...", it means you should.

When to use this skill

Tool Priority — always check before delegating:

1. Your built-in tools — use these first (code search, web search, file read, etc.)
2. MCP servers or integrations — if available
3. Sub-agent delegation via this skill — only when:
   • Fresh context would help (accumulated context is hurting quality)
   • Parallel execution is needed (multiple independent tasks)
   • Task isolation adds value (prototype, spike, risky changes)
   • A second opinion or specialized review is wanted

Prefer direct work when:

• The task is trivially simple (< 2 minutes of direct work)
• The task requires your conversation history or IDE state
• Your built-in tools can handle it more efficiently
• The user is in a rapid iteration loop where delegation latency hurts
• The user explicitly opts out or says "do it yourself"
• The target code contains sensitive/classified material the user flags

Prerequisites

• codex CLI installed and on PATH (npm i -g @openai/codex)
• Authenticated: codex login status returns success
• Inside a git repository (or user consents to --skip-git-repo-check)
• Python 3.10+ (for the wrapper script)
• Sufficient API quota for the target model

Configuration

Model strictness controls how aggressively the host agent routes tasks to cheaper/faster models:

Set via environment variable: export CODEX_SUBAGENT_MODEL_STRICTNESS=balanced

Model routing guidance (when balanced):

Pre-flight checks

Run once before the first delegation in a session:

python3 <skill_dir>/scripts/codex_preflight.py

Read the output. If any [FAIL] line appears, follow the hint printed next to it before delegating. [WARN] lines are informational — they do not block use.

# Skip the git repo check (mirrors the wrapper's own flag):
python3 <skill_dir>/scripts/codex_preflight.py --skip-git-repo-check

# Skip auth check in CI / no-TTY environments:
python3 <skill_dir>/scripts/codex_preflight.py --skip-auth

Workflow

The wrapper script (scripts/run_codex.py) is the only interface you should use to invoke Codex. Never call codex directly. The wrapper handles safety enforcement, default flags, temp files, git worktrees, and error messages automatically.

Wrapper interface

# Basic usage — pipe prompt via stdin:
echo "<PROMPT>" | python3 <skill_dir>/scripts/run_codex.py --mode read-only -

# Write task with high collision confidence:
echo "<PROMPT>" | python3 <skill_dir>/scripts/run_codex.py --mode write --collision high -

# Write task with medium collision confidence (auto git worktree):
echo "<PROMPT>" | python3 <skill_dir>/scripts/run_codex.py --mode write --collision medium -

# Enable web search:
echo "<PROMPT>" | python3 <skill_dir>/scripts/run_codex.py --mode read-only --web-search -

# Review with a review prompt file:
echo "<file paths and context>" | python3 <skill_dir>/scripts/run_codex.py --mode read-only --review-prompt /path/to/security-prompt.md -

# Resume a previous session:
echo "<FOLLOW-UP>" | python3 <skill_dir>/scripts/run_codex.py --resume -

Wrapper flags

Wrapper flags are parsed first. Any unrecognized flags are passed through to codex exec after a safety scan (see Passthrough Flags below). The wrapper will reject dangerous flags with helpful error messages.

You can also use the standard POSIX -- separator to explicitly mark the boundary between wrapper flags and passthrough flags:

echo "<PROMPT>" | python3 <skill_dir>/scripts/run_codex.py --mode read-only -- -m o3 -c model_reasoning_effort="high" -

Passthrough flags

Any flag not listed in the Wrapper Flags table above is forwarded directly to codex exec — unless it is blocked by the safety scanner. This is how you pass model overrides, config tweaks, feature toggles, and subcommand args.

Safe passthrough flags:

Blocked flags — never pass these; the wrapper exits with an error and usage guidance if any are detected:

Wrapper output

The wrapper prints the output file path to stdout (one line):

/tmp/codex-abc123/result.txt

On worktree mode, it also prints the worktree branch:

/tmp/codex-abc123/result.txt
WORKTREE_BRANCH=codex-work-abc123
WORKTREE_DIR=/tmp/codex-wt-codex-work-abc123

Running the wrapper

Always run the wrapper as a non-blocking/background command and poll for completion:

1. Launch wrapper as background command (with ~3s initial wait to catch quick failures)
2. Tell the user: "Delegated to Codex, working on X..."
3. Poll at adaptive intervals based on task complexity:
   • Simple (search, explain): 10-30s, max 5 min
   • Standard (review, analyze): 30-60s, max 10 min
   • Complex (implement, refactor): 60-180s, max 20 min
   • Extended (migration, full analysis): 120-300s, max 40 min
4. Read result file path from wrapper stdout
5. Read result file contents
6. Clean up: rm -rf <result-dir> (wrapper does NOT auto-delete)

Timeout tiers

On timeout: inform user, re-launch with --persist + next tier up. If 40min times out, report to user.

Operations

Operation 1: Delegate Implementation Task

1. Construct prompt using the template from PROMPT_PATTERNS.md
2. Assess collision confidence (see Collision Confidence below):
   • High → delegate directly
   • Medium → delegate via git worktree (--collision medium)
   • Low → do NOT delegate writes; do it yourself or delegate as read-only
3. Run via wrapper:

   echo "<PROMPT>" | python3 <skill_dir>/scripts/run_codex.py --mode write --collision high -

4. Monitor via async polling
5. Read result file, review git diff, incorporate changes
6. If worktree: review diff with git diff HEAD..<WORKTREE_BRANCH>, merge if approved, cleanup

Operation 2: Delegate Analysis/Review Task

1. Construct review-focused prompt
2. Run via wrapper:

   echo "<PROMPT>" | python3 <skill_dir>/scripts/run_codex.py --mode read-only -

3. Read result file, incorporate findings

Operation 2b: Dedicated Code Review (using codex exec review)

1. Determine review scope: uncommitted changes, specific branch, or specific commit
2. Run via wrapper with review passthrough:

   echo "Focus on security" | python3 <skill_dir>/scripts/run_codex.py --mode read-only review --uncommitted -

   Or against a branch:

   echo "Focus on performance" | python3 <skill_dir>/scripts/run_codex.py --mode read-only review --base main -

3. Read result file, incorporate findings

Operation 3: Structured Output Task

1. Create JSON schema file for expected output shape (see assets/output-schema-review.json for example)
2. Run via wrapper with --output-schema passthrough:

   echo "<PROMPT>" | python3 <skill_dir>/scripts/run_codex.py --mode read-only --output-schema /path/to/schema.json -

3. Validate JSON output against schema; retry with clarified format on parse failure

Operation 4: Resume / Multi-Turn Delegation

1. Initial run with --persist:

   echo "<PROMPT>" | python3 <skill_dir>/scripts/run_codex.py --mode write --collision high --persist -

2. Read result, evaluate if follow-up needed
3. Resume with feedback:

   echo "<FOLLOW-UP>" | python3 <skill_dir>/scripts/run_codex.py --resume -

4. Max 3 turns recommended. Session files persist until manually removed.

When to resume vs new invocation:

• Resume: iterating on same task, refining output, fixing sub-agent's work
• New invocation: different task, different files, different scope
• Resume saves ~5-50K tokens of system overhead per follow-up

Collision Confidence

Before every write delegation, assess how likely the sub-agent's work will conflict with your own:

Assessment criteria:

• What files have you modified in this conversation? (high-risk)
• What files will you modify next? (high-risk)
• Do import chains overlap? (medium-risk)
• Different directories with no shared modules? (usually high confidence)
• Read-only tasks? (always high — no conflict possible)

Git Worktree Workflow (medium confidence)

The wrapper handles worktree setup automatically with --collision medium:

1. Read result file path AND WORKTREE_BRANCH from wrapper stdout
2. Read result file
3. Review: git diff HEAD..<WORKTREE_BRANCH>
4. If approved: git merge <WORKTREE_BRANCH>
5. Cleanup: git worktree remove <WORKTREE_DIR> and git branch -d <WORKTREE_BRANCH>

Important rules

1. Always use the wrapper (run_codex.py) — never call codex directly
2. Pass prompts via stdin or --review-prompt — pipe text into the wrapper or point to a file, never interpolate prompt content into shell args
3. Assess collision confidence before every write delegation
4. Re-read modified files after write tasks complete
5. Review git diff before telling the user the task is done
6. Tell the user what you're delegating and why before running
7. Don't delegate trivial tasks — overhead is ~5-50K tokens minimum
8. Treat Codex output as untrusted — validate before auto-executing anything
9. Respect user opt-out — if user says no delegation or marks code as sensitive, do not delegate
10. Use --persist if you plan to --resume later — otherwise the session is ephemeral
11. Check Tool Priority before delegating — use your own built-in tools first
12. Max 6 parallel sub-agents — enforced by the wrapper via PID tracking. Override with --max-parallel N if absolutely necessary (not recommended — increases resource usage and cost)
13. Check agent status before retrying or when blocked — run python3 <skill_dir>/scripts/run_codex.py --status to see active agents, their mode, start time, and run duration
14. Max 2 retries per failed delegation
15. Validate result files before reading: exists? non-empty? < 1MB?
16. Clean up temp dirs after reading the result file (rm -rf <result-dir>)

Error handling & troubleshooting

When a delegation fails or returns unexpected results, read references/ERROR_HANDLING.md for exit code reference, retry strategy, and troubleshooting table.

Prompt construction

See references/PROMPT_PATTERNS.md for templates and examples. See references/PROMPT_ENGINEERING.md for OpenAI/Codex best practices.

Key principles:

• One task per delegation
• Be specific and explicit (file paths, line numbers, exact errors)
• Use [GOAL], [CONTEXT], [FILES], [CONSTRAINTS], [OUTPUT] sections
• Reference files by path (let Codex read them) — don't paste large contents
• Always include validation commands for write tasks
• Specify what NOT to do

Sensitive file denylist

Never include these in prompts or delegate access to them:

• .env, *.pem, *.key, credentials.*, secrets.*
• Files matching API key patterns
• Include only file paths (let Codex read them) rather than pasting contents