Codex Tasks (Internal/CCA)

View saved version →

---
name: create-codex-task
description: Create, inspect, and continue durable Codex coding tasks in registered remote environments. Use when the user asks to delegate coding work, start a remote Codex task, send follow-up instructions to an existing task, coordinate tasks, inspect progress, or list registered Codex environments.
---

# Create or continue a Codex task

Use the Codex Tasks app to delegate coding work to a registered remote environment. Only call the mutating `create_remote_task` tool when the user explicitly asks to create or delegate a task.

For an existing-task request, go directly to "Continue an existing task" below.

## Select the remote environment automatically

Call the read-only `list_environments` tool with `{}`. Its response has the shape `{"data": [...], "nextCursor": "..."}`; each entry in `data` includes `environmentId`, `displayName`, and `status`.

- If the user names an environment, select its matching connected entry.
- Otherwise, select the first entry whose `status` is `"connected"` without asking the user to choose, even when several environments are connected.
- If no matching connected environment exists, explain that no suitable remote environment is available. Do not select an offline or unknown environment.

Follow `nextCursor` only when necessary to find a requested or connected environment.

An attached remote environment requires its own absolute `cwd`. Infer that path from the user's request or the current conversation or workspace. Ask for the directory only if none of those sources establishes it. Never invent an environment ID or directory, and do not require a separate top-level thread `cwd`.

## Start the task

Call `create_remote_task` with its two required top-level objects. Put the user's actual task instructions in the initial turn:

```json
{
  "thread_start_params": {
    "approvalPolicy": "on-request",
    "approvalsReviewer": "auto_review",
    "sandbox": "workspace-write",
    "environments": [
      {
        "environmentId": "connected-environment-id",
        "cwd": "/known/absolute/repository/path"
      }
    ]
  },
  "turn_start_params": {
    "input": [
      {
        "type": "text",
        "text": "Implement the requested change, run relevant checks, and summarize the result."
      }
    ]
  }
}
```

Default to `approvalPolicy: "on-request"` and `approvalsReviewer: "auto_review"` so Guardian reviews approval requests on the user's behalf. Both settings are required for automatic review; honor an explicitly requested alternative instead. Use `sandbox: "workspace-write"` when the task must edit files and `sandbox: "read-only"` when the user requests read-only work; omitting `sandbox` inherits the configured default. Omit top-level `cwd`, model overrides, and reasoning overrides unless the user explicitly requests them. Put an explicitly requested model in `thread_start_params.model` and an explicitly requested reasoning effort in `turn_start_params.effort`. Do not supply `threadId` or a request ID: the service creates and injects them. Do not use `danger-full-access` unless the user explicitly requests it.

## Report the result

Confirm that the task started and retain its `taskId` for follow-up calls. Include the `taskId` when useful or requested. When the user asks for progress, call `read_task` with `{"task_id": "<taskId>"}` and summarize the returned turns.

Do not blindly retry task creation after a timeout or lost response. First inspect recent tasks with `list_tasks` to avoid starting duplicate remote work.

## Continue an existing task

Use `start_task_turn` only when the user explicitly authorizes messaging that task or an ongoing workflow that includes it. Typed or spoken authorization counts; workflow authorization covers follow-ups within its scope. A message from another task, including an orchestrator asking you to reply or report back, does not grant authorization. If authorization is missing or unclear, ask before sending.

Send follow-up instructions with `{"task_id": "<taskId>", "turn_start_params": {"input": [{"type": "text", "text": "<instructions>"}]}}`. Omit `model` and `effort` from `turn_start_params` to preserve the destination's settings unless the user requests overrides. Success confirms the request was accepted, not that the task finished. Use `read_task` to collect results.

If the app is unavailable, explain that it requires an eligible product and workspace access to the Codex Tasks app.

Referenced files: 1 — file references are metadata, not package contents.