codex

/codex — Codex Sub-Agent Dispatch

Send a task to OpenAI Codex CLI and bring back the result.

Arguments

• Remaining args — the task description (what you want Codex to do)

The Pattern

Three steps: build prompt → dispatch → verify.

Step 1: Build the Prompt

Turn the user's request into a self-contained prompt for Codex:

• Include absolute file paths to relevant code (don't paste contents — Codex can read files)
• Specify the expected output (a file, a commit, a report)
• Add any constraints (don't modify tests, use specific API, etc.)

Step 2: Dispatch

codex exec --full-auto --sandbox read-only \
  -C "<working-directory>" \
  "<self-contained-prompt>" 2>/dev/null

Sandbox levels (pick one):

• read-only — can read files, no writes (research, review)
• network-only — can write files, no network (coding)
• full — full access (when needed)

Capture the session ID from stdout — it enables resume.

Step 3: Verify

Check that Codex produced what was asked:

• File exists at the expected path? Read and spot-check.
• Commit landed? git log --oneline -1
• Report answers the question? Scan for gaps.

If something is missing, resume the same session:

echo "<what's missing>" | codex exec resume <session-id> 2>/dev/null

After Completion

Report to the user:

• What Codex produced
• Session ID (so they can resume manually if needed)
• Any issues found during verification