---
description: Multi-agent orchestration rules for OpenClaw PC/VPS work
alwaysApply: true
---

# OpenClaw Multi-Agent Rules

- Read `AGENT_BOARD.md` before making plans, running commands, or editing files.
- Stay within the active role scope assigned by the user or orchestrator.
- Do not write secrets, passwords, tokens, API keys, or `OPENCLAW_GATEWAY_TOKEN` into files, logs, evidence, commits, or chat.
- Treat the VPS `smartmotor.cloud` website as frozen during filing review. Do not change homepage content, linked page content reachable from the homepage, static assets, container images, bind mounts, routes, Nginx rules, or `/opt/services/docker-composite.yml` entries that could alter public website content.
- Execution agents may write only their own handoff, evidence, and rollback files unless explicitly assigned otherwise.
- Execution agents do not mark their own work as accepted; a verifier or orchestrator must validate evidence.
- Before system-level changes, service restarts, network exposure changes, or config edits, create a rollback note.
- Prefer read-only discovery before mutation. Capture important command results in `evidence/`.
- CORRECTION owns sync-health monitoring through `sync-state/heartbeat-*.json` and `sync-state/error-*.md`. If sync health is blocked, do not start cross-device dependent or mutation-stage tasks.
- If task scope, device target, or acceptance criteria are unclear, stop and ask instead of guessing.
- When ending a session or nearing context limits, update the relevant file under `handoff/`.