Walkie Openclaw Adapter

Run Walkie + OpenClaw in 托管模式（v2）: stable, observable, and low-noise.

Core rules (must follow)

1. Single reader: only one process may run walkie read --wait.
2. Single sender path: all outbound messages must go through adapter path (no manual direct send in production mode).
3. Audit is source of truth: group sync reads audit stream (recv/send/error) with cursor dedupe.
4. Structured protocol first: prefer JSON task/result envelopes over free text.

Setup

npm install -g walkie-sh
walkie --version
cp skills/walkie-openclaw-adapter/references/config.example.json skills/walkie-openclaw-adapter/references/config.json

Init confirmation (required)

Before first run, explicitly confirm with user:

1. Which Walkie channel to use (channel)
2. Which chat/group should receive callback sync (platform + chat id)

Do not assume callback destination. Bind it only after user confirms.

Edit config.json:

• channel
• autoReply / autoExecute
• autoReplyMode (ack-task if needed)
• policy.levels (L1/L2/L3)
• policy.stopConditions (maxRounds / maxTaskMinutes / maxConsecutiveErrors)
• syncHookCmd + syncGroupId

Self-check (required after install)

After editing config.json, run the self-check script to verify all requirements:

node skills/walkie-openclaw-adapter/scripts/self-check.js

Acceptance criteria:

• Exit 0: All checks passed ✅ → Ready for production
• Exit 1: Warnings only ⚠️ → OK to proceed, but review warnings
• Exit 2: Critical failures ❌ → Must fix before production

Critical checks:

1. Config valid + channel set
2. Audit log writable
3. Single reader enforcement (no duplicate processes)
4. System service supervision (systemd/launchd) configured
5. Walkie CLI available

If any critical check fails, do not run adapter in production mode. Fix issues first or inform user of risks.

Run (dev)

node skills/walkie-openclaw-adapter/scripts/adapter.js skills/walkie-openclaw-adapter/references/config.json

Run (recommended prod)

• Linux: use systemd to supervise adapter
• macOS: use launchd to supervise adapter
• Keep cron/timer for audit-forwarder.py only (sync/announce), not as primary process supervision

Install systemd service (Linux)

# Copy and edit service file
cp skills/walkie-openclaw-adapter/references/walkie-adapter.service.example ~/.config/systemd/user/walkie-adapter.service

# Edit and adjust paths:
# - WorkingDirectory: your workspace path
# - ExecStart: path to node + adapter.js path
# - Environment: PATH to your node (e.g., ~/.nvm/versions/node/v22.22.0/bin)

# Enable and start
systemctl --user daemon-reload
systemctl --user enable walkie-adapter
systemctl --user start walkie-adapter

# Check status
systemctl --user status walkie-adapter

Why systemd matters:

• Auto-restart on crash (your previous issue)
• Survives reboot
• Proper logging via journalctl

Group sync format (fixed)

Use concise 3-line updates:

• [walkie][task_id] 收到: ...
• [walkie][task_id] 动作: ...
• [walkie][task_id] 结果: OK/ERR/WAIT ...

Protocol (minimum fields)

{"type":"task","task_id":"t-001","intent":"summarize_links","args":{},"trace_id":"tr-001","ts":"..."}

{"type":"result","task_id":"t-001","ok":true,"output":{},"trace_id":"tr-001","ts":"..."}

Decision policy

• L1/L2 auto-execute: routine chat, retrieval, summarize, translation, progress updates.
• L3 human approval: irreversible/delete/publish/payment/sensitive data externalization.

Anti-loop safety

• event dedupe key
• task TTL/max rounds
• ignore self-origin loops
• legal state transitions only
• cooldown/rate limit
• circuit breaker after repeated failures

Privacy and repo hygiene

Never commit runtime files:

• references/config.json
• references/state.json
• references/audit.log
• references/audit.cursor

Troubleshooting

• Missing messages -> check for extra readers first.
• Sync delay -> check forwarder schedule and cursor state.
• Adapter exits -> check supervisor logs (systemd/launchd) + audit error lines.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kaibuildsai
• Source: kaibuildsai/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.