AgentSkillsCN

auto-drive

依托协调器决策、可选子代理、验证关卡与历史压缩,实现自主任务编排。当用户希望端到端运行一项长期任务,或在目标达成前不断迭代、反复验证,又或需要协调多个代理的尝试,并在最终完成时进行一次全面检查时,此技能将助您事半功倍。

SKILL.md
--- frontmatter
name: auto-drive
description: Autonomous task orchestration with coordinator decisions, optional sub-agents, verification pass, and history compaction. Use when a user asks to run a long task end-to-end, keep iterating until a goal is verified complete, or coordinate multiple agent attempts with a final completion check.

Auto Drive

What This Skill Does

Drive a long-running task by looping through: plan -> execute -> summarize -> coordinator decision -> verify. You are the executor: you run the actual commands, edits, and research. The script only tells you what to do next.

How To Use

  1. Create a session with a goal.
bash
node {{SKILL_DIR}}/scripts/auto_drive.js --goal "Your goal here"

Expected output:

code
Auto Drive Session Created
Call `node skills/auto-drive/scripts/auto_drive.js --id <id>` repeatedly until it says you are complete
  1. Ask for the next instruction.
bash
node {{SKILL_DIR}}/scripts/auto_drive.js --id <id>

If it asks for a plan, create a plan and send it back:

bash
node {{SKILL_DIR}}/scripts/auto_drive.js --id <id> --plan "Your plan here"
  1. Execute the task you are given. When the script returns a task prompt, do the work (run commands, edit files, launch agents). Then summarize what you actually did:
bash
node {{SKILL_DIR}}/scripts/auto_drive.js --id <id> --summary "Work completed since last call"
  1. Repeat. Keep calling --id or sending --summary until the script says the session is complete.

Notes

  • If the task prompt includes an <agents> block, use native agents if available, then merge their results into your work.
  • Provide concrete summaries: commands run, files changed, tests run, and results.
  • The script runs codex exec under the hood. You do not need to call codex directly.
  • Default model timeout is 3 minutes. Set AUTO_DRIVE_CLI_TIMEOUT_MS (milliseconds) to override (use 0 to disable).