name: codex description: Delegate coding tasks to OpenAI Codex via codex-listener allowed-tools: Bash(python3 *)

Codex Skill

Delegate coding tasks to OpenAI Codex CLI through the codex-listener daemon.

IMPORTANT RULES

1. Submit only. After submitting a task, immediately move on to other work. Do NOT poll, wait, or check the task status.
2. No polling. Do NOT call status.py or list_tasks.py after submitting unless the user explicitly asks you to check a task's status.
3. Notification is automatic. The daemon will notify the user through the configured messaging channel(s) (Feishu/Telegram) when the task finishes. You will NOT receive the result — just move on.
4. Use official submit path only. Always use scripts/submit.py (or POST /tasks for manual HTTP). Do NOT use /submit or any other unofficial endpoint.
5. Status source of truth. Task status must come only from /tasks APIs via scripts/status.py or scripts/list_tasks.py.
6. No inferred excuses. If status query fails, report the raw JSON error and stop guessing. Do NOT claim "权限受限/系统拦截" unless the tool output explicitly says so.
7. No shell fallbacks for status. Do NOT append 2>/dev/null, || echo, or pipes that change script output.
8. Do not use artifacts as status proxy. Do NOT inspect .codex/sessions or output files to infer task state unless the user explicitly asks to verify deliverables.
9. Use explicit sandbox when needed. Default sandbox follows server-side model defaults (currently workspace-write). For system-level tasks, explicitly pass --sandbox danger-full-access.
10. System tasks must include acceptance checks in the same prompt. For installs/services/users/permissions, require "execute + verify + report verification output".
11. Complex tasks must enter PlanMode first. Trigger PlanMode if any of: write/delete files >=2, any delete/overwrite/batch replace, estimated steps >=5, or system-level changes (packages/services/permissions/env).
12. Use Plan Bridge for multi-turn planning. Submit stage-A with --workflow-mode plan_bridge; if result is bridge_stage=needs_input, collect user answers and continue with --resume-session (stage-B). Do not execute implementation until plan is ready.

Prerequisites

The daemon must be running:

codex-listener start

All scripts are in the scripts/ directory relative to this skill.

Workflow

# 1. Submit a task (canonical path: POST /tasks via submit.py)
python3 scripts/submit.py --prompt "fix the type error in auth.py" --cwd /path/to/project
# Returns: {"task_id": "a1b2c3d4", "status": "pending", ...}

# 2. Done. Move on to other work. The user will be notified through their configured channels when codex finishes.

Plan Bridge (two-stage):

# Stage A: ask planning questions only
python3 scripts/submit.py \
  --workflow-mode plan_bridge \
  --prompt "Complex task: ask only clarifying questions first, then emit planmode.v1 JSON with stage=needs_input." \
  --cwd /home/Hera/.nanobot/workspace

# Stage B: continue same session after user answers
python3 scripts/submit.py \
  --workflow-mode plan_bridge \
  --resume-session <session_id> \
  --parent-task-id <task_id> \
  --prompt "User answers: ... Continue planning and emit planmode.v1 JSON." \
  --cwd /home/Hera/.nanobot/workspace

System-task prompt template (required):

Install tmux on Debian/Ubuntu, then verify with:
1) tmux -V
2) dpkg -l tmux
Return the exact verification output in your final response.

Scripts

Submit a task

python3 scripts/submit.py --prompt "fix the bug in auth.py" --cwd /path/to/project
python3 scripts/submit.py --prompt "refactor this module" --model o3-mini --cwd .
python3 scripts/submit.py --prompt "quick fix" --reasoning-effort low --cwd .
python3 scripts/submit.py --prompt "install tmux and verify with tmux -V + dpkg -l tmux" --cwd /home/Hera/.nanobot/workspace
python3 scripts/submit.py --prompt "code-only task" --sandbox workspace-write --cwd /home/Hera/.nanobot/workspace
python3 scripts/submit.py --workflow-mode plan_bridge --prompt "ask questions first" --cwd /home/Hera/.nanobot/workspace
python3 scripts/submit.py --workflow-mode plan_bridge --resume-session <session_id> --parent-task-id <task_id> --prompt "answers: ..." --cwd /home/Hera/.nanobot/workspace

The script above sends requests to POST /tasks. Do not hand-write /submit.

Options: --prompt (required), --model, --cwd, --sandbox, --reasoning-effort (high/medium/low, default: high), --workflow-mode, --resume-session, --parent-task-id

Cancel a task

python3 scripts/cancel.py --task-id <id>

Health check

python3 scripts/health.py

Check task status (only when user asks)

python3 scripts/status.py --task-id <id>
python3 scripts/list_tasks.py

Status handling:

• If success: report status directly from JSON.
• If error: copy the error field verbatim, then run python3 scripts/health.py once and report result.
• Do not switch to guessed narratives.

Plan Bridge handling:

• If bridge_stage=needs_input: ask user for answers. Preferred reply format is /plan-reply <task_id> <answer>.
• Natural-language reply is allowed only when there is exactly one pending needs_input task; otherwise require explicit /plan-reply.
• Continue by resubmitting with --resume-session <session_id> and --parent-task-id <task_id>.
• Telegram button semantics:
  • ✍️ 回复问题: only pre-fills /plan-reply <task_id> , user must still send the final answer text.
  • ✅ 执行计划: requires second confirmation before creating execution task.
  • 📝 继续修改: route back to /plan-reply <task_id> ....
  • ❌ 取消: cancel current execution intent and do not auto-submit implementation.

Output Format

All scripts output a single JSON object to stdout. Exit code 0 = success, 1 = error.

Submitted task:

{"task_id": "a1b2c3d4", "status": "pending", ...}

Daemon not running:

{"error": "codex-listener is not running. Start it with: codex-listener start"}