Codex

Codex Delegation Skill

Delegate to Codex CLI for autonomous repo-scale execution tasks.

When to delegate to Codex

Delegate only when ALL of these hold:

1. Execution task — requires writing/modifying files, not just reasoning
2. Clear done condition — can be verified (tests pass, lint clean, etc.)
3. Repo-scale or repetitive — spans multiple files or involves bulk changes

Examples that qualify:

• Writing test files for multiple tools at once
• File-level refactoring with explicit before/after spec
• Code review (codex review)
• Bulk edits: adding type hints, docstrings, or formatting across a directory

When NOT to delegate

• Design or architecture decisions (Claude reasons better here)
• Small edits under ~20 lines (delegation overhead exceeds benefit)
• Explanation-only or analysis requests (no file writes needed)
• Tasks requiring contextual judgment before execution

Pre-flight check

codex login status   # expect "Logged in"

If it fails, instruct the user to run codex login.

Default execution pattern

codex exec \
  --cd /path/to/repo \
  --sandbox workspace-write \
  --full-auto \
  --ephemeral \
  --json \
  -o /tmp/codex-result.txt \
  "task instruction"

Fixed option rules:

• --cd — always set to project root
• --sandbox workspace-write — default; allows writes while minimising system impact
• --ephemeral — prevents session state from leaking between runs
• -o /tmp/codex-result.txt — saves output for Claude to review

Commands by task type

Code tasks (tests, refactoring, bulk edits)

codex exec \
  --cd "$(pwd)" \
  --sandbox workspace-write \
  --full-auto \
  --ephemeral \
  -o /tmp/codex-result.txt \
  "task instruction here"

Code review

codex review --base main "Review for security, performance, and regression risk. Cite file path and line number."

Or non-interactive:

codex exec review \
  --base main \
  --ephemeral \
  -o /tmp/codex-review.txt \
  "Review for security, performance, and regression risk."

Long prompts (stdin)

cat prompt.txt | codex exec - \
  --cd "$(pwd)" \
  --sandbox workspace-write \
  --full-auto \
  --ephemeral \
  -o /tmp/codex-result.txt

Workflow

1. Identify the task scope and target files from the user request.
2. Write a concrete prompt (more specific = better results).
3. Pick the matching pattern above and run it via Bash.
4. After completion, Read the -o file and summarise results for the user.
5. For any follow-up fixes, either handle them directly or re-run Codex.

Prompt writing tips

A good Codex prompt:

• Names the target files/directories: inside tests/mcp_server/
• References existing patterns: follow the style of test_apartment_trades.py
• States the done condition: all tests must pass
• States constraints: do not modify any other files

Troubleshooting