subagent-orchestrator

# subagent-orchestrator

Run continuous task lines through the main controller, not by accident.

## Core rules

1. **Treat `autopilot` as a continuous line**
   - Do not stop at “one run finished”.
   - A finished run means **handoff**, not “whole line done”.
   - Continue until one of: complete, blocked, paused by user.

2. **The main controller must directly re-dispatch**
   - Do not wait for the user to remind you to continue.
   - For every autopilot line, the main controller decides the next slice and spawns it.

3. **Keep the task board current**
   - Update `AGENT_TASK_BOARD.md` before or when dispatching the next round.
   - Record: goal, current round, completed items, remaining items, next slice, follow-up roles.

4. **Keep subagent replies short**
   - Prefer: goal / completed / files / status / next / risk.
   - Avoid long background restatements.

5. **Clean finished sessions, keep memory**
   - Finished subagent sessions do not need to be kept.
   - Before forgetting them, write the durable record into `memory/YYYY-MM-DD.md`.
   - Include at least: task goal, changed files, result status, risk, next step.
   - Promote stable collaboration rules into `MEMORY.md` when needed.

6. **Do not keep unnecessary artifacts**
   - Temporary screenshots, temp files, and intermediate outputs can be deleted when no longer needed.
   - If the user wants screenshot proof, prefer sending it directly instead of long-term local retention.

## Role routing

Do not hard-code project-specific model rules into this skill.

Use the current session defaults unless the user or project rules override them. If a project has role-specific model preferences, keep them in project memory, task board rules, or project docs — not as universal orchestrator behavior.

## Dispatch pattern

When an autopilot line is active:

1. Read the current line from `AGENT_TASK_BOARD.md`
2. Pick the next **small closed slice**
3. Spawn the correct role
4. Require short structured return
5. On complet...

# subagent-orchestrator

Version: 1.0.0

A lightweight orchestration skill for OpenClaw subagent workflows.

## What it does
- Treats `autopilot` as a continuous task line instead of a one-shot run
- Requires structured subagent return headers: `tag / line / goal_status / next_role`
- Helps the main controller decide whether to continue, hand off, wait, block, or finish
- Encourages cleaning finished sessions while preserving durable memory records

## Return protocol
Subagents should return these headers before the short report:
- `tag:` `autopilot | done | idle | blocked | handoff | need_user`
- `line:` task-line name
- `goal_status:` `partial | complete | waiting | blocked`
- `next_role:` free-form next owner/role label

## Notes
- This skill is intentionally generic.
- Project-specific role defaults and collaboration rules should live in project memory or task boards, not in the skill itself.