Coordinate Codex Tasks

A Codex skill for delegating work to separate Codex Tasks while the original Task stays responsible for coordination and verification. It creates one dedicated, state-aware heartbeat per delegated Task, keeps each monitor scoped to its own Task, and pauses it when the work is verified complete or needs a user decision.

繁體中文 · Prompt examples · Skill instructions · Verification record

This is an independent community skill. It is not an official OpenAI product.

Why use it?

A separate Codex Task can keep working after its coordinator yields. The coordinator needs a durable way to check progress without repeatedly interrupting the worker, losing the assignment boundary, or treating idle as proof of completion. This skill describes that workflow using Codex Task and automation tools.

Behavior

• Creates a separate Task only when the user explicitly requests one.
• Distinguishes an independent, user-owned Task from a Task delegated to complete part of the coordinator's assignment.
• For delegated work, maintains a one-to-one mapping between worker Task and heartbeat. Multiple workers can have distinct heartbeats attached to the same coordinating Task.
• Names the coordinating and executing Tasks in cross-Task messages, so a worker cannot mistake a forwarded "this Task" for the coordinator.
• Lets the coordinator choose and adjust each monitor's interval based on the worker's stage and next useful checkpoint, with a 30-minute maximum for the user's cache-related constraint. Unchanged checks stay quiet; active coordination uses wait_threads instead of faster schedules.
• Reads the worker's current state and relevant artifacts before updating the monitor prompt, giving a scoped correction, or pausing the monitor. A finished turn is not necessarily a finished assignment.
• Leaves the worker focused on its work. The coordinator uses this skill; workers do not need to load it to manage their monitors.

The skill is guidance for the model, not an event hook or a guaranteed background daemon. Scheduled runs require the relevant Codex app tools and an available host. See verification limits.

Requirements

• Codex with Skills enabled.
• Access to the Codex app Task tools (create_thread, wait_threads, read_thread, send_message_to_thread) and automation_update for scheduled monitoring.
• The user's authorization to create a separate Task. Ordinary internal subtasks should use the available subagent mechanism.

Install

git clone https://github.com/easyvibecoding/coordinate-codex-tasks.git
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills/coordinate-codex-tasks"
cp coordinate-codex-tasks/SKILL.md "${CODEX_HOME:-$HOME/.codex}/skills/coordinate-codex-tasks/SKILL.md"

Then start a new Codex Task so it can discover the skill. A plain-language request to create another Task can let the model select it from its description; you do not need to name the skill. For example:

Create a separate Codex Task to implement the parser change. Keep this Task responsible for integration and review, and follow the new Task's progress.

The skill is model guidance, not a guaranteed hook on every Task creation. For an independent Task that you will follow directly, say so; that case does not call for a coordinator monitor.

For ready-to-copy prompts covering one delegated Task, multiple Tasks, an existing Task, and independent handoff, see Prompt examples.

State policy

Worker evidence	Coordinator action
Active, no verified change	Keep its heartbeat active; do not rewrite the prompt or send a status message.
New stage or recoverable blocker	Update only its heartbeat checkpoint; send one specific correction if needed.
Idle or turn completed, assignment unfinished	Inspect the result and artifacts; continue the same Task when authorized.
Outcome verified, terminal blocker, or user decision required	Pause only its heartbeat, read back the saved status, and report once.

The actual SKILL.md is the operational source. The table summarizes its decisions.

Development

Run the repository's dependency-free package check:

python3 scripts/check.py

The check validates packaging and the social preview asset; it does not prove model behavior. See CONTRIBUTING.md for change guidance.

License

MIT. The social preview artwork was generated with Codex image generation and is included in this repository.