openclaw-config

OpenClaw 运维手册

诊断并解决实际问题。这里的每条命令都经过测试，可以正常运行。

---

快速健康检查

出现任何异常时，先运行此检查。复制并粘贴整个代码块：

echo "=== GATEWAY ===" && \
ps aux | grep -c "[o]penclaw" && \
echo "=== CONFIG JSON ===" && \
python3 -m json.tool ~/.openclaw/openclaw.json > /dev/null 2>&1 && echo "JSON: OK" || echo "JSON: BROKEN" && \
echo "=== CHANNELS ===" && \
cat ~/.openclaw/openclaw.json | jq -r '.channels | to_entries[] | "\(.key): policy=\(.value.dmPolicy // "n/a") enabled=\(.value.enabled // "implicit")"' && \
echo "=== PLUGINS ===" && \
cat ~/.openclaw/openclaw.json | jq -r '.plugins.entries | to_entries[] | "\(.key): \(.value.enabled)"' && \
echo "=== CREDS ===" && \
ls ~/.openclaw/credentials/whatsapp/default/ 2>/dev/null | wc -l | xargs -I{} echo "WhatsApp keys: {} files" && \
for d in ~/.openclaw/credentials/telegram/*/; do bot=$(basename "$d"); [ -f "$d/token.txt" ] && echo "Telegram $bot: OK" || echo "Telegram $bot: MISSING"; done && \
[ -f ~/.openclaw/credentials/bird/cookies.json ] && echo "Bird cookies: OK" || echo "Bird cookies: MISSING" && \
echo "=== CRON ===" && \
cat ~/.openclaw/cron/jobs.json | jq -r '.jobs[] | "\(.name): enabled=\(.enabled) status=\(.state.lastStatus // "never") \(.state.lastError // "")"' && \
echo "=== RECENT ERRORS ===" && \
tail -10 ~/.openclaw/logs/gateway.err.log 2>/dev/null && \
echo "=== MEMORY DB ===" && \
sqlite3 ~/.openclaw/memory/main.sqlite "SELECT COUNT(*) || ' chunks, ' || (SELECT COUNT(*) FROM files) || ' files indexed' FROM chunks;" 2>/dev/null

---

文件一览

~/.openclaw/
├── openclaw.json                    # MAIN CONFIG — channels, auth, gateway, plugins, skills
├── openclaw.json.bak*               # Auto-backups (.bak, .bak.1, .bak.2 ...)
├── exec-approvals.json              # Exec approval socket config
│
├── agents/main/
│   ├── agent/auth-profiles.json     # Anthropic auth tokens
│   └── sessions/
│       ├── sessions.json            # SESSION INDEX — keys are like agent:main:whatsapp:+1234
│       └── *.jsonl                  # Session transcripts (one JSON per line)
│
├── workspace/                       # Agent workspace (git-tracked)
│   ├── SOUL.md                      # Personality, writing style, tone rules
│   ├── IDENTITY.md                  # Name, creature type, vibe
│   ├── USER.md                      # Owner context and preferences
│   ├── AGENTS.md                    # Session behavior, memory rules, safety
│   ├── BOOT.md                      # Boot instructions (autopilot notification protocol)
│   ├── HEARTBEAT.md                 # Periodic task checklist (empty = skip heartbeat)
│   ├── MEMORY.md                    # Curated long-term memory (main session only)
│   ├── TOOLS.md                     # Contacts, SSH hosts, device nicknames
│   ├── memory/                      # Daily logs: YYYY-MM-DD.md, topic-chat.md
│   └── skills/                      # Workspace-level skills
│
├── memory/main.sqlite               # Vector memory DB (Gemini embeddings, FTS5 search)
│
├── logs/
│   ├── gateway.log                  # Runtime: startup, channel init, config reload, shutdown
│   ├── gateway.err.log              # Errors: connection drops, API failures, timeouts
│   └── commands.log                 # Command execution log
│
├── cron/
│   ├── jobs.json                    # Job definitions (schedule, payload, delivery target)
│   └── runs/                        # Per-job run logs: {job-uuid}.jsonl
│
├── credentials/
│   ├── whatsapp/default/            # Baileys session: ~1400 app-state-sync-key-*.json files
│   ├── telegram/{botname}/token.txt # Bot tokens (one per bot account)
│   └── bird/cookies.json            # X/Twitter auth cookies
│
├── extensions/{name}/               # Custom plugins (TypeScript)
│   ├── openclaw.plugin.json         # {"id", "channels", "configSchema"}
│   ├── index.ts                     # Entry point
│   └── src/                         # channel.ts, actions.ts, runtime.ts, types.ts
│
├── identity/                        # device.json, device-auth.json
├── devices/                         # paired.json, pending.json
├── media/inbound/                   # Received images, audio files
├── media/browser/                   # Browser screenshots
├── browser/openclaw/user-data/      # Chromium profile (~180MB)
├── tools/signal-cli/                # Signal CLI binary
├── subagents/runs.json              # Sub-agent execution log
├── canvas/index.html                # Web canvas UI
└── telegram/
    ├── update-offset-coder.json     # {"lastUpdateId": N} — Telegram polling cursor
    └── update-offset-sales.json     # Reset these to 0 to replay missed messages

---

故障排查：WhatsApp

“我发了消息，但没有收到回复”

这是最常见的问题。消息已送达，但机器人没有回复。按以下顺序检查：

# 1. Is the bot actually running?
grep -i "whatsapp.*starting\|whatsapp.*listening" ~/.openclaw/logs/gateway.log | tail -5

# 2. Check for 408 timeout drops (WhatsApp web disconnects frequently)
grep -i "408\|499\|retry" ~/.openclaw/logs/gateway.err.log | tail -10
# If you see "Web connection closed (status 408). Retry 1/12" — this is normal,
# it auto-recovers. But if retries reach 12/12, the session dropped completely.

# 3. Check for cross-context messaging blocks
grep -i "cross-context.*denied" ~/.openclaw/logs/gateway.err.log | tail -10
# Common: "Cross-context messaging denied: action=send target provider "whatsapp" while bound to "signal""
# This means the agent was in a Signal session and tried to reply on WhatsApp.
# FIX: The message needs to come through in the WhatsApp session context, not Signal.

# 4. Check the session exists for that contact
cat ~/.openclaw/agents/main/sessions/sessions.json | jq -r 'to_entries[] | select(.key | test("whatsapp")) | "\(.key) | \(.value.origin.label // "?")"'

# 5. Check if the sender is allowed
cat ~/.openclaw/openclaw.json | jq '.channels.whatsapp | {dmPolicy, allowFrom, selfChatMode, groupPolicy}'
# If dmPolicy is "allowlist" and the sender isn't in allowFrom, message is silently dropped.

# 6. Check if it's a group message (groups are disabled by default)
cat ~/.openclaw/openclaw.json | jq '.channels.whatsapp.groupPolicy'
# "disabled" means ALL group messages are ignored.

# 7. Check for lane congestion (agent busy with another task)
grep "lane wait exceeded" ~/.openclaw/logs/gateway.err.log | tail -5
# If the agent is stuck on a long LLM call, new messages queue up.

# 8. Check for agent run timeout
grep "embedded run timeout" ~/.openclaw/logs/gateway.err.log | tail -5
# Hard limit is 600s (10 min). If the agent's response takes longer, it's killed.

“WhatsApp 完全断开连接”

# Check credential files exist (should be ~1400 files)
ls ~/.openclaw/credentials/whatsapp/default/ | wc -l

# If 0 files: session was never created or got wiped
# Fix: re-pair with `openclaw configure`

# Check for QR/pairing events
grep -i "pair\|link\|qr\|scan\|logged out" ~/.openclaw/logs/gateway.log | tail -10

# Check for Baileys errors
grep -i "baileys\|DisconnectReason\|logout\|stream:error" ~/.openclaw/logs/gateway.err.log | tail -20

# Nuclear fix: delete credentials and re-pair
# rm -rf ~/.openclaw/credentials/whatsapp/default/
# openclaw configure

---

故障排查：Telegram

“机器人出了问题／会忘事”

有两个看起来相同的独立问题：

# 1. Check for config validation errors (THE COMMON ONE)
grep -i "telegram.*unrecognized\|telegram.*invalid\|telegram.*policy" ~/.openclaw/logs/gateway.err.log | tail -10
# Known issue: the keys "token" and "username" under accounts are not recognized.
# The correct field is "botToken", not "token".

# 2. Check the actual config
cat ~/.openclaw/openclaw.json | jq '.channels.telegram'
# Verify each bot has "botToken" (not "token") and "name" fields.

# 3. Check polling status — bots die after getUpdates timeout
grep -i "telegram.*exit\|telegram.*timeout\|getUpdates" ~/.openclaw/logs/gateway.err.log | tail -10
# "[telegram] [sales] channel exited: Request to 'getUpdates' timed out after 500 seconds"
# This means the bot lost connection to Telegram's API and stopped listening.
# Fix: restart gateway — `openclaw gateway restart`

# 4. Check the polling offset (if bot "forgets" or replays old messages)
cat ~/.openclaw/telegram/update-offset-coder.json
cat ~/.openclaw/telegram/update-offset-sales.json
# If lastUpdateId is stuck or 0, the bot will reprocess old messages.
# To skip to latest: the gateway sets this automatically on restart.

# 5. Check if both bots are starting
grep -i "telegram.*starting\|telegram.*coder\|telegram.*sales" ~/.openclaw/logs/gateway.log | tail -10

# 6. "Bot forgets" — this is usually a session issue, not Telegram
# Each Telegram user gets their own session in sessions.json.
# Check if the session exists:
cat ~/.openclaw/agents/main/sessions/sessions.json | jq -r 'to_entries[] | select(.key | test("telegram")) | "\(.key) | \(.value.origin.label // "?")"'

# 7. Check if compaction happened (context window pruned = "forgot")
SESS_ID="paste-session-id"
grep '"compaction"' ~/.openclaw/agents/main/sessions/$SESS_ID.jsonl | wc -l
# If compaction count > 0, old messages were pruned from context.
# The agent's compaction mode is:
cat ~/.openclaw/openclaw.json | jq '.agents.defaults.compaction'

Telegram 配置修复模板

# Correct Telegram config structure:
cat ~/.openclaw/openclaw.json | jq '.channels.telegram = {
  "enabled": true,
  "accounts": {
    "coder": {
      "name": "Bot Display Name",
      "enabled": true,
      "botToken": "your-bot-token-here"
    },
    "sales": {
      "name": "Sales Bot Name",
      "enabled": true,
      "botToken": "your-bot-token-here"
    }
  },
  "dmPolicy": "pairing",
  "groupPolicy": "disabled"
}' > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

---

故障排查：Signal

“Signal RPC 发送消息失败”

这会导致 cron 任务和跨渠道通知无法运行。

# 1. Check if signal-cli process is alive
ps aux | grep "[s]ignal-cli"

# 2. Check the RPC endpoint
grep -i "signal.*starting\|signal.*8080\|signal.*rpc" ~/.openclaw/logs/gateway.log | tail -10
# Should see: "[signal] [default] starting provider (http://127.0.0.1:8080)"

# 3. Check for connection instability
grep -i "HikariPool\|reconnecting\|SSE stream error\|terminated" ~/.openclaw/logs/gateway.err.log | tail -10
# "HikariPool-1 - Thread starvation or clock leap detected" = signal-cli internal DB issue
# "SSE stream error: TypeError: terminated" = lost connection to signal-cli daemon

# 4. Check for rate limiting
grep -i "signal.*rate" ~/.openclaw/logs/gateway.err.log | tail -5
# "Signal RPC -5: Failed to send message due to rate limiting"

# 5. Check for wrong target format
grep -i "unknown target" ~/.openclaw/logs/gateway.err.log | tail -5
# "Unknown target "adi" for Signal. Hint: <E.164|uuid:ID|...>"
# The agent must use phone numbers (+1...) or uuid: format, not names.

# 6. Fix profile name warning spam
grep -c "No profile name set" ~/.openclaw/logs/gateway.err.log
# If high count: run signal-cli updateProfile to set a name

# 7. Test signal-cli directly
ACCT=$(cat ~/.openclaw/openclaw.json | jq -r '.channels.signal.account')
echo "Account: $ACCT"
# signal-cli -a $ACCT send -m "test" +TARGET_NUMBER

# 8. Check if the signal-cli daemon needs restart
# The gateway manages signal-cli as a subprocess.
# Restart the whole gateway: openclaw gateway restart

---

故障排查：Cron 任务

# 1. Overview of all jobs
cat ~/.openclaw/cron/jobs.json | jq -r '.jobs[] | "\(.enabled | if . then "ON " else "OFF" end) \(.state.lastStatus // "never" | if . == "error" then "FAIL" elif . == "ok" then "OK  " else .  end) \(.name)"'

# 2. Failing jobs with error details
cat ~/.openclaw/cron/jobs.json | jq '.jobs[] | select(.state.lastStatus == "error") | {name, error: .state.lastError, lastRun: (.state.lastRunAtMs | . / 1000 | todate), id}'

# 3. Read the actual run log for a failing job
JOB_ID="paste-job-uuid-here"
tail -20 ~/.openclaw/cron/runs/$JOB_ID.jsonl | python3 -c "
import sys, json
for line in sys.stdin:
    try:
        obj = json.loads(line)
        if obj.get('type') == 'message':
            role = obj['message']['role']
            text = ''.join(c.get('text','') for c in obj['message'].get('content',[]) if isinstance(c,dict))
            if text.strip():
                print(f'[{role}] {text[:300]}')
    except: pass
"

# 4. Common cron failure causes:
#    - "Signal RPC -1" → Signal daemon down, see Signal section above
#    - "gateway timeout after 10000ms" → gateway was restarting when cron fired
#    - "Brave Search 429" → free tier rate limit hit (2000 req/month)
#    - "embedded run timeout" → job took longer than 600s

# 5. Next scheduled run times
cat ~/.openclaw/cron/jobs.json | jq -r '.jobs[] | select(.enabled) | "\(.name): \((.state.nextRunAtMs // 0) | . / 1000 | todate)"'

# 6. Disable a broken job temporarily
cat ~/.openclaw/cron/jobs.json | jq '(.jobs[] | select(.name == "JOB_NAME")).enabled = false' > /tmp/cron.json && mv /tmp/cron.json ~/.openclaw/cron/jobs.json

---

故障排查：记忆／“它忘了”

记忆系统分为 3 层。当 Agent “忘了”时，其中一层出现了故障：

第 1 层：上下文窗口（会话内）

# Check compaction count for a session (compaction = old messages pruned)
grep -c '"compaction"' ~/.openclaw/agents/main/sessions/SESSION_ID.jsonl
# 7 compactions = the agent has "forgotten" its earliest messages 7 times.

# Check compaction mode
cat ~/.openclaw/openclaw.json | jq '.agents.defaults.compaction'
# "safeguard" = only compacts when hitting context limit

第 2 层：工作区记忆文件

# What daily memory files exist
ls -la ~/.openclaw/workspace/memory/

# What's in MEMORY.md (long-term curated)
cat ~/.openclaw/workspace/MEMORY.md

# Search memory files for something specific
grep -ri "KEYWORD" ~/.openclaw/workspace/memory/

第 3 层：向量记忆数据库（SQLite + Gemini 嵌入）

# What files are indexed
sqlite3 ~/.openclaw/memory/main.sqlite "SELECT path, size, datetime(mtime/1000, 'unixepoch') as modified FROM files;"

# How many chunks (text fragments) exist
sqlite3 ~/.openclaw/memory/main.sqlite "SELECT COUNT(*) FROM chunks;"

# Search chunks by text (FTS5 full-text search)
sqlite3 ~/.openclaw/memory/main.sqlite "SELECT substr(text, 1, 200) FROM chunks_fts WHERE chunks_fts MATCH 'KEYWORD' LIMIT 5;"

# Check embedding config
sqlite3 ~/.openclaw/memory/main.sqlite "SELECT value FROM meta WHERE key='memory_index_meta_v1';" | python3 -m json.tool

# Check for Gemini embedding rate limits (breaks indexing)
grep -i "gemini.*batch.*failed\|RESOURCE_EXHAUSTED\|429" ~/.openclaw/logs/gateway.err.log | tail -10
# "embeddings: gemini batch failed (2/2); disabling batch" = indexing degraded

# Rebuild memory index (re-index all workspace files)
# Delete the DB and restart gateway — it will rebuild:
# rm ~/.openclaw/memory/main.sqlite && openclaw gateway restart

---

搜索会话

查找某人的对话

# Search session index by name (case-insensitive)
cat ~/.openclaw/agents/main/sessions/sessions.json | jq -r 'to_entries[] | select(.value.origin.label // "" | test("NAME"; "i")) | "\(.value.sessionId) | \(.value.lastChannel) | \(.value.origin.label)"'

按渠道查找会话

cat ~/.openclaw/agents/main/sessions/sessions.json | jq -r 'to_entries[] | select(.value.lastChannel == "whatsapp") | "\(.value.sessionId) | \(.value.origin.label // .key)"'
# Replace "whatsapp" with: signal, telegram, or check .key for cron sessions

最近的会话

cat ~/.openclaw/agents/main/sessions/sessions.json | jq -r '[to_entries[] | {id: .value.sessionId, updated: .value.updatedAt, label: (.value.origin.label // .key), ch: (.value.lastChannel // "cron")}] | sort_by(.updated) | reverse | .[:10][] | "\(.updated | . / 1000 | todate) | \(.ch) | \(.label)"'

搜索所有会话中的消息内容

# Quick: find which session files contain a keyword
grep -l "KEYWORD" ~/.openclaw/agents/main/sessions/*.jsonl

# Detailed: show matching messages with timestamps
grep "KEYWORD" ~/.openclaw/agents/main/sessions/*.jsonl | python3 -c "
import sys, json
for line in sys.stdin:
    path, data = line.split(':', 1)
    try:
        obj = json.loads(data)
        if obj.get('type') == 'message':
            role = obj['message']['role']
            text = ''.join(c.get('text','') for c in obj['message'].get('content',[]) if isinstance(c,dict))
            if text.strip():
                sid = path.split('/')[-1].replace('.jsonl','')[:8]
                ts = obj.get('timestamp','')[:19]
                print(f'{ts} [{sid}] [{role}] {text[:200]}')
    except: pass
" | head -30

阅读特定会话的记录

# Last 30 messages from a session
tail -50 ~/.openclaw/agents/main/sessions/SESSION_ID.jsonl | python3 -c "
import sys, json
for line in sys.stdin:
    try:
        obj = json.loads(line)
        if obj.get('type') == 'message':
            role = obj['message']['role']
            text = ''.join(c.get('text','') for c in obj['message'].get('content',[]) if isinstance(c,dict))
            if text.strip() and role != 'toolResult':
                print(f'[{role}] {text[:300]}')
                print()
    except: pass
"

---

编辑配置

安全编辑流程

始终先备份，再用 jq 编辑，然后重启。

cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.bak.manual
jq 'YOUR_EDIT_HERE' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json
openclaw gateway restart

常见编辑

# Switch WhatsApp to allowlist
jq '.channels.whatsapp.dmPolicy = "allowlist" | .channels.whatsapp.allowFrom = ["+1XXXXXXXXXX"]' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

# Enable WhatsApp autopilot (bot responds as you to everyone)
jq '.channels.whatsapp += {dmPolicy: "open", selfChatMode: false, allowFrom: ["*"]}' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

# Add number to Signal allowlist
jq '.channels.signal.allowFrom += ["+1XXXXXXXXXX"]' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

# Change model
jq '.agents.defaults.models = {"anthropic/claude-sonnet-4": {"alias": "sonnet"}}' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

# Set concurrency
jq '.agents.defaults.maxConcurrent = 10 | .agents.defaults.subagents.maxConcurrent = 10' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

# Disable a plugin
jq '.plugins.entries.imessage.enabled = false' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

从备份恢复

# Latest backup
cp ~/.openclaw/openclaw.json.bak ~/.openclaw/openclaw.json

# List all backups by date
ls -lt ~/.openclaw/openclaw.json.bak*

# Validate JSON before restart
python3 -m json.tool ~/.openclaw/openclaw.json > /dev/null && echo "OK" || echo "BROKEN"

# Nuclear reset
openclaw configure

---

渠道安全模式

模式	行为	风险
open + allowFrom: ["*"]	任何人都能发消息，机器人会回复所有人	高 — 消耗 API 额度，机器人会以你的身份发言
allowlist + allowFrom: ["+1..."]	只有列出的号码可以通过	低 — 控制明确
pairing	未知发件人会收到一个代码，由你批准	低 — 需要经过批准
disabled	渠道完全关闭	无

检查当前安全状态

cat ~/.openclaw/openclaw.json | jq '{
  whatsapp: {policy: .channels.whatsapp.dmPolicy, from: .channels.whatsapp.allowFrom, groups: .channels.whatsapp.groupPolicy, selfChat: .channels.whatsapp.selfChatMode},
  signal: {policy: .channels.signal.dmPolicy, from: .channels.signal.allowFrom, groups: .channels.signal.groupPolicy},
  telegram: {policy: .channels.telegram.dmPolicy, groups: .channels.telegram.groupPolicy, bots: [.channels.telegram.accounts | to_entries[] | "\(.key)=\(.value.enabled)"]},
  imessage: {enabled: .channels.imessage.enabled, policy: .channels.imessage.dmPolicy}
}'

---

工作区文件

文件	内容	编辑时机
SOUL.md	个性：语气、风格（“不用长破折号，随意使用小写”）	更改机器人的说话方式时
IDENTITY.md	名称（Jarvis）、生物类型、表情符号	更改品牌形象时
USER.md	所有者信息、偏好	用户背景信息发生变化时
AGENTS.md	操作规则：记忆协议、安全、群聊行为、心跳指令	更改机器人行为时
BOOT.md	启动指令（自动驾驶通知协议：WA → Signal）	更改启动时的行为时
HEARTBEAT.md	定期检查清单（为空 = 不调用心跳 API）	添加／移除定期任务时
MEMORY.md	精选的长期记忆（仅在主会话／直接会话中加载）	由机器人自行管理
TOOLS.md	联系人、SSH 主机、设备昵称	添加本地工具备注时
memory/*.md	每日原始日志、特定主题的聊天日志	由机器人自动写入

---

会话 JSONL 格式

每个 .jsonl 文件每行包含一个 JSON 对象。类型：

类型	内容
session	会话标头：id、时间戳、cwd
message	对话轮次：角色（user/assistant/toolResult）、内容、模型、用量
custom	元数据：model-snapshot、openclaw.cache-ttl
compaction	上下文窗口已裁剪（丢弃了较早的消息）
model_change	会话中途切换了模型
thinking_level_change	调整了思考级别

会话索引（sessions.json）中的键：

• 模式：agent:main:{channel}:{contact} 或 agent:main:cron:{job-uuid}
• 字段：sessionId（UUID = 文件名）、lastChannel、origin.label（人类可读名称）、origin.from（规范地址）、updatedAt（epoch 毫秒）、chatType（direct/group）

---

Gateway 启动顺序

正常启动大约需要 3 秒：

[heartbeat] started
[gateway] listening on ws://127.0.0.1:18789
[browser/service] Browser control service ready
[hooks] loaded 3 internal hook handlers (boot-md, command-logger, session-memory)
[whatsapp] [default] starting provider
[signal] [default] starting provider (http://127.0.0.1:8080)
[telegram] [coder] starting provider
[telegram] [sales] starting provider
[whatsapp] Listening for personal WhatsApp inbound messages.
[signal] signal-cli: Started HTTP server on /127.0.0.1:8080

如果缺少任何一行，说明对应组件未能启动。请检查 gateway.err.log。

---

常见错误模式

错误	含义	解决方法
Web connection closed (status 408)	WhatsApp Web 超时，自动重试最多 12 次	通常会自动恢复。如果达到 12/12，请重启网关
Signal RPC -1: Failed to send message	signal-cli 守护进程连接丢失	重启网关
Signal RPC -5: Failed to send message due to rate limiting	Signal 速率限制	等待后重试，降低消息发送频率
No profile name set（signal-cli WARN）	日志会被大量刷屏，但无害	signal-cli -a +ACCOUNT updateProfile --given-name "Name"
Cross-context messaging denied	Agent 尝试跨频道发送消息	这不是 bug，而是安全防护措施。消息必须来自正确频道的会话
getUpdates timed out after 500 seconds	Telegram 机器人丢失轮询连接	重启网关
Unrecognized keys: "token", "username"	Telegram 机器人配置键错误	在 openclaw.json 中使用 botToken，而不是 token
RESOURCE_EXHAUSTED（Gemini 429）	嵌入速率限制	减少工作区文件变动，或升级 Gemini 配额
lane wait exceeded	Agent 被较长的大模型调用阻塞	等待；如果卡住超过 2 分钟，则重启
embedded run timeout: timeoutMs=600000	Agent 响应超过 10 分钟	将任务拆分成更小的部分
gateway timeout after 10000ms	重启期间网关无法访问	网关关闭时触发了 Cron——这是暂时现象

---

扩展 OpenClaw

OpenClaw 有 4 个扩展层，每一层解决不同的问题：

层	内容	位置	添加方式
技能	Agent 按需加载的知识和工作流	/opt/homebrew/lib/node_modules/openclaw/skills/ 或 ~/.openclaw/workspace/skills/	clawdhub install <slug> 或 npx add-skill <repo>
扩展	自定义频道插件（TypeScript）	~/.openclaw/extensions/{name}/	创建 openclaw.plugin.json 和 TypeScript 源文件
频道	消息平台（内置）	openclaw.json → channels.* + plugins.entries.*	在 openclaw.json 中配置并添加凭据
Cron 任务	定时自主任务	~/.openclaw/cron/jobs.json	Agent 通过工具创建，或直接编辑 jobs.json

技能：ClawdHub 生态系统

技能是扩展 Agent 知识和能力的主要方式。它们是 Markdown 文件，可包含可选脚本或资源，并会在相关时加载到上下文中。

# Search for skills (vector search across the registry)
clawdhub search "postgres optimization"
clawdhub search "image generation"

# Browse latest skills
clawdhub explore

# Install a skill
clawdhub install supabase-postgres-best-practices
clawdhub install nano-banana-pro

# Install a specific version
clawdhub install my-skill --version 1.2.3

# List what's installed
clawdhub list

# Update all installed skills
clawdhub update --all

# Update a specific skill
clawdhub update my-skill
clawdhub update my-skill --force  # overwrite local changes

当前已安装的技能（与 OpenClaw 捆绑）：

类别	技能
消息	discord、slack、imsg、wacli、voice-call
社交/网页	bird（X/Twitter）、blogwatcher、github、trello、notion
Google	gog、google-workspace-mcp、goplaces、local-places
媒体	nano-banana-pro（Gemini 图像生成）、openai-image-gen、video-frames、gifgrep、pixelation、sag（TTS）、openai-whisper、sherpa-onnx-tts、songsee、camsnap
编码 Agent	coding-agent（Codex/Claude Code/Pi）、ccbg（后台运行器）、tmux
效率工具	apple-notes、apple-reminders、bear-notes、things-mac、obsidian、himalaya（电子邮件）
智能家居	openhue（Philips Hue）、eightctl（Eight Sleep）、sonoscli、blucli（BluOS）
开发工具	github、worktree、starter、desktop、supabase-postgres-best-practices、superdesign
内容	remotion-best-practices、remotion-fastest-tech-stack、humanizer、summarize、market、buildinpublic
元工具	skill-creator、clawdhub、find-skills、add-skill、model-usage、session-logs、recentplans、canvas

创建自己的技能

技能只是一个包含 SKILL.md 的文件夹：

my-skill/
├── SKILL.md              # Required: YAML frontmatter + markdown instructions
├── scripts/              # Optional: executable scripts
├── references/           # Optional: docs loaded on demand
└── assets/               # Optional: templates, images

SKILL.md 格式：

---
name: my-skill
description: What this does and WHEN to trigger it. The description is the primary
  trigger — the agent reads this to decide whether to load the full skill.
---

# My Skill

Instructions go here. Only loaded AFTER the skill triggers.
Keep under 500 lines. Split large content into references/ files.

核心原则：上下文窗口是一种共享资源。 只包含 Agent 尚不了解的内容。相比冗长的说明，优先使用简洁的示例。

# Publish to ClawdHub
clawdhub login
clawdhub publish ./my-skill --slug my-skill --name "My Skill" --version 1.0.0

# Or publish to GitHub for npx add-skill
# (see add-skill skill for details)

多 Agent 编排

OpenClaw 可以将其他 AI Agent（Codex、Claude Code、Pi）作为后台工作进程启动。借助这一功能，你可以并行执行编码任务、代码审查，或任何适合由多个 Agent 协作完成的工作。

模式： bash pty:true background:true workdir:/path command:"agent 'task'"

# Spawn Codex to build something (background, auto-approve)
bash pty:true workdir:~/project background:true command:"codex exec --full-auto 'Build a REST API for todos'"
# Returns sessionId for tracking

# Spawn Claude Code on a different task
bash pty:true workdir:~/other-project background:true command:"claude 'Refactor the auth module'"

# Monitor all running agents
process action:list

# Check output of a specific agent
process action:log sessionId:XXX

# Send input if agent asks a question
process action:submit sessionId:XXX data:"yes"

# Kill a stuck agent
process action:kill sessionId:XXX

并行审查 PR：

# Fetch all PR refs
git fetch origin '+refs/pull/*/head:refs/remotes/origin/pr/*'

# Launch one agent per PR
bash pty:true workdir:~/project background:true command:"codex exec 'Review PR #86. git diff origin/main...origin/pr/86'"
bash pty:true workdir:~/project background:true command:"codex exec 'Review PR #87. git diff origin/main...origin/pr/87'"

使用 git worktrees 并行修复问题：

git worktree add -b fix/issue-78 /tmp/issue-78 main
git worktree add -b fix/issue-99 /tmp/issue-99 main

bash pty:true workdir:/tmp/issue-78 background:true command:"codex --yolo 'Fix issue #78: description. Commit and push.'"
bash pty:true workdir:/tmp/issue-99 background:true command:"codex --yolo 'Fix issue #99: description. Commit and push.'"

Agent 完成时自动通知：

bash pty:true workdir:~/project background:true command:"codex --yolo exec 'Your task.

When completely finished, run: openclaw gateway wake --text \"Done: summary\" --mode now'"

添加新频道

频道是 Agent 可以通过其进行通信的平台。内置频道包括：WhatsApp、Signal、Telegram、iMessage、Discord、Slack。

启用内置频道：

# 1. Enable the plugin
jq '.plugins.entries.discord.enabled = true' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

# 2. Add channel config
jq '.channels.discord = {enabled: true, dmPolicy: "pairing", groupPolicy: "disabled"}' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

# 3. Add credentials (channel-specific)
# 4. Restart gateway
openclaw gateway restart

构建自定义频道扩展：

~/.openclaw/extensions/{name}/
├── openclaw.plugin.json    # {"id": "name", "channels": ["name"], "configSchema": {...}}
├── package.json            # Standard npm package
├── index.ts                # Entry point
└── src/
    ├── channel.ts          # Inbound message handling + outbound send
    ├── actions.ts          # Tool actions the agent can invoke
    ├── runtime.ts          # Lifecycle: start, stop, health check
    ├── config-schema.ts    # JSON schema for plugin config
    └── types.ts            # TypeScript types

# List installed extensions
ls ~/.openclaw/extensions/

# View extension manifests
cat ~/.openclaw/extensions/*/openclaw.plugin.json | jq .

# Check extension source files
find ~/.openclaw/extensions/ -name "*.ts" -not -path "*/node_modules/*"

跨频道通信

Agent 可以在一个频道接收消息，并在另一个频道发送消息，但需遵守以下防护措施：

# Check cross-context settings
cat ~/.openclaw/openclaw.json | jq '.tools.message.crossContext'
# allowAcrossProviders: true = agent CAN send across channels
# marker.enabled: false = no "[via Signal]" prefix on cross-channel messages

# If you see "Cross-context messaging denied" errors:
# The agent tried to send from a session bound to channel A to channel B.
# This is a security feature. To allow it:
jq '.tools.message.crossContext.allowAcrossProviders = true' ~/.openclaw/openclaw.json > /tmp/oc.json && mv /tmp/oc.json ~/.openclaw/openclaw.json

BOOT.md 通知协议（已配置）： Agent 接收 WhatsApp 消息，在 WhatsApp 上回复，然后向 Signal 发送通知摘要。这是主要的跨频道模式——在一个频道上自动运行，在另一个频道上进行控制。

Canvas：已连接设备的 Web UI

向已连接的 Mac/iOS/Android 节点推送 HTML/游戏/仪表板：

# Check canvas config
cat ~/.openclaw/openclaw.json | jq '.canvasHost // "not configured"'

# List connected nodes
openclaw nodes list

# Present HTML content on a node
# canvas action:present node:<node-id> target:http://localhost:18793/__moltbot__/canvas/my-page.html

# Canvas files live in:
ls ~/.openclaw/canvas/

语音通话

通过 Twilio/Telnyx/Plivo 发起电话呼叫：

# Check if voice-call plugin is enabled
cat ~/.openclaw/openclaw.json | jq '.plugins.entries["voice-call"] // "not configured"'

# CLI usage
openclaw voicecall call --to "+15555550123" --message "Hello"
openclaw voicecall status --call-id <id>

Cron：计划的自主任务

# View all jobs
cat ~/.openclaw/cron/jobs.json | jq '.jobs[] | {name, enabled, schedule: .schedule, channel: .payload.channel, to: .payload.to}'

# Job schedule types:
# "kind": "at", "atMs": <epoch>          — one-shot at specific time
# "kind": "every", "everyMs": <ms>       — recurring interval

# Job delivery targets:
# channel + to fields determine where results go (signal, whatsapp, telegram)
# sessionTarget: "isolated" = fresh context each run (no memory of previous runs)

# To add a job, the agent creates it via tool, or edit jobs.json:
# See existing jobs as templates in ~/.openclaw/cron/jobs.json

---

许可证

MIT