delegate ยท diff

git:20260507.7f08bbe to git:20260507.35dad1e

1 added, 4 removed. Audit A to A.

---
name: delegate
description: Delegate a sub-task to OpenAI Codex CLI. Use when a task is better handled by Codex (e.g. complex file editing, multi-step coding tasks, automated code generation).
---
Use the `codex` MCP tools to delegate tasks to OpenAI Codex CLI.
## Prerequisites
Codex must be installed and signed in:
```bash
npm i -g @openai/codex
# sign in via the Codex app โ€” no API key needed
```
## Synchronous delegation (need result now)
Use `codex_run` when you need the output before continuing:
```
codex_run(
prompt="<full task description with all context Codex needs>",
cwd="/path/to/project" # optional, defaults to current directory
)
```
Returns the final Codex message as a string.
## Background delegation (parallel work)
Use `codex_task_submit` to fire off tasks and continue working:
1. Submit one or more tasks:
```
codex_task_submit(name="Refactor auth module", prompt="...", cwd="...")
codex_task_submit(name="Write tests for API", prompt="...", cwd="...")
```
2. Check status:
```
codex_task_status(task_id="<id>")
```
3. Fetch result when completed:
```
codex_task_result(task_id="<id>")
```
4. See all tasks:
```
codex_tasks_list()
```
## Environment variables
| Variable | Purpose |
|----------|---------|
- | `OPENAI_API_KEY` | API key for Codex (required unless using browser auth) |
| `CODEX_PATH` | Override path to the `codex` binary |
- | `CODEX_YOLO=1` | Bypass all approvals and sandbox (fastest, least safe) |
- | `CODEX_APPROVAL` | `never` (default) or `untrusted` |
- | `CODEX_SANDBOX` | `workspace-write` (default), `read-only`, or `danger-full-access` |
+ | `CODEX_SANDBOX` | `read-only`, `workspace-write`, or `danger-full-access` โ€” if unset, uses `--dangerously-bypass-approvals-and-sandbox` for headless operation |
## Writing good prompts for Codex
- Include full context โ€” Codex starts fresh with no knowledge of prior work
- Specify the working directory if the task involves files
- Be explicit about what "done" looks like (files to edit, tests to pass, output expected)
- If the task spans multiple files, list the key ones