run-codex-subagents（RUN Codex subagents 搜索）

Orchestrate Codex Work with codex-worker

Use this skill when delegating coding, research, verification, or follow-up work through the released codex-worker CLI.

This skill intentionally tracks the public installed surface, not unreleased repo-local commands. Re-check codex-worker --help before relying on anything outside the references below.

Trigger Boundaries

Use this skill for:

• launching file-backed Codex work through codex-worker run
• continuing an existing thread through send, thread resume, or turn steer
• monitoring progress through read, logs, and wait
• answering blocked runtime requests through request list, request read, and request respond
• recovering failed or interrupted work from local artifacts and daemon state

Do not use for:

• MCP server setup or mcpc testing
• changing codex-worker source code unless the user explicitly asks
• general Codex/OpenAI documentation questions

Prerequisites

1. Verify the CLI: codex-worker --help
2. If it is missing, install it first: sudo -v; curl -fsSL https://github.com/yigitkonur/codex-worker/releases/latest/download/install.sh | sudo bash Or user-local: curl -fsSL https://github.com/yigitkonur/codex-worker/releases/latest/download/install.sh | bash -s -- --install-dir "$HOME/.local/bin" Fallback: npm install -g codex-worker
3. Confirm runtime health: codex-worker --output json doctor
4. Prompts are file-backed. Write a .md file before dispatching.

Current Command Families

There is no task layer, no session layer, no monitor, no review, no exec, and no request answer helper in the released CLI this skill targets.

Canonical Workflow

1. Write the prompt file

Each prompt file should define:

• what the worker should do
• which files matter
• what must not be touched
• what success looks like
• which commands prove success

2. Start the work

Blocking one-shot:

codex-worker run task.md

Async launch:

codex-worker --output json run task.md --async

Supported run flags in the released CLI:

• --cwd <dir>
• --model <id>
• --async
• --timeout <ms>

3. Monitor and inspect

Use the returned threadId:

codex-worker wait --thread-id <thread-id> --timeout 300000
codex-worker read <thread-id> --tail 50
codex-worker logs <thread-id> --tail 100

Use:

• wait to block on terminal state
• read for thread state plus local artifact paths
• logs for readable execution output

4. Answer blocked requests

When the thread status becomes waiting_request:

codex-worker request list --status pending
codex-worker request read <request-id>
codex-worker request respond <request-id> --decision accept

For free-text answers:

codex-worker request respond <request-id> --answer "Yes" --question-id <id>

For unusual payloads, inspect first and then use raw JSON:

codex-worker request respond <request-id> --json '{"result":{"approved":true}}'

5. Continue or recover

Friendly continuation:

codex-worker send <thread-id> followup.md

Protocol-first continuation:

codex-worker turn steer <thread-id> <turn-id> followup.md

Interrupt an active turn:

codex-worker turn interrupt <thread-id> <turn-id>

Inspect before retrying:

codex-worker read <thread-id> --tail 50
codex-worker logs <thread-id> --tail 100
codex-worker doctor

Quick Reference

Status Meanings

Use read or thread list output to reason about state:

Machine-Readable Output

Use --output json for shell composition:

codex-worker --output json run task.md --async
codex-worker --output json read <thread-id>
codex-worker --output json thread list --cwd /abs/project
codex-worker --output json request list --status pending

Prefer JSON output over parsing text.

State Layout

Default state root: ~/.codex-worker/ (override: CODEX_WORKER_STATE_DIR).

Key paths:

• registry.json — local thread/turn/request state
• daemon.json / daemon.sock — daemon connection info
• workspaces/<hash>/threads/<id>.jsonl — transcript events
• workspaces/<hash>/logs/<id>.output — readable log
• workspaces/<hash>/logs/<id>.raw.ndjson — raw protocol log when available

Reference Files

Guardrails

• Do not invent a task or session subcommand layer.
• Do not use request answer; the released CLI only exposes request respond.
• Do not assume unreleased flags such as --follow, --plan, --skip-plan, --label, or --compact unless --help shows them.
• Do not rely on monitor, review, exec, or thread rollback unless you have verified you are on a newer unreleased build; they are not part of the released surface this skill targets.
• Do not reuse stale turn ids without re-reading current thread state.
• Do not treat waiting_request as failure.
• Do not claim a run is irrecoverable until you have checked read, logs, and doctor.