ClawMonitor (OpenClaw monitoring)
Use this skill when the user asks questions like:
- - “Did my agent finish? Why no feedback?”
- “Which session/thread received the last message, and when?”
- “Is the agent working, interrupted, or stuck? Any delivery failures?”
- “Which model is failing, and is the problem the provider or OpenClaw itself?”
- “Which session burned the most tokens recently?”
- “Do we have zombie/helper processes or a dirty gateway cgroup?”
Preconditions
- - This skill runs on a machine that has OpenClaw state at
~/.openclaw/. - INLINECODE1 is installed on that same machine.
Preflight (recommended)
Before using the commands below, verify the binary exists and can read local OpenClaw state:
CODEBLOCK0
If clawmonitor is missing, or status fails, install it first.
Install:
CODEBLOCK1
Alternative installs:
CODEBLOCK2
CODEBLOCK3
Notes:
- - Some OpenClaw environments intentionally disallow installing packages at runtime. If installs are blocked, ask the user to install
clawmonitor on the host first. - INLINECODE6 writes config under
~/.config/clawmonitor/config.toml and is safe to re-run.
Core commands
0) Tree view (who owns which sessions)
If you suspect ACP/subagent routing issues (for example Telegram thread bindings), start with:
CODEBLOCK4
1) Status (Markdown)
Show the core status table:
CODEBLOCK5
For a more verbose table including task/message previews:
CODEBLOCK6
2) Model health
Probe configured models directly and/or through OpenClaw:
CODEBLOCK7
Useful variants:
CODEBLOCK8
3) Drill down on one session
Export a redacted report for a single session key:
CODEBLOCK9
4) Nudge (ask the session to report progress)
Send a progress request into the session:
CODEBLOCK10
5) Full TUI for interactive inspection
Run the TUI when you need interactive triage:
CODEBLOCK11
Important TUI keys:
- -
v: cycle Sessions / Models / System - INLINECODE9 : jump directly to System
- INLINECODE10 : toggle right-side Status / History
- INLINECODE11 : cycle token windows (
now / 1d / 7d / 30d) - INLINECODE16 : refresh current active surface
- INLINECODE17 : cycle pane widths
- INLINECODE18 : fullscreen detail
- INLINECODE19 : open the English operator note in System view
Troubleshooting quick wins
- - If
clawmonitor status shows DELIVERY_FAILED or NO_FEEDBACK, export a report next:
CODEBLOCK12
- - If model checks disagree, separate
direct provider path from OpenClaw path. - If Telegram looks bound to the wrong session key, run
clawmonitor tree and follow the bound session instead of assuming agent:main:... is the active path. - If the problem is service-level, use the TUI
System view and summarize risk, reclaimable memory estimate, and whether zombies/orphans were detected. - If the TUI is unavailable, fall back to
clawmonitor status --format md --detail plus clawmonitor models --format md.
Reply guidelines
- - Prefer
--format md outputs for IM replies. - If status shows
DELIVERY_FAILED or NO_FEEDBACK, include the relevant sessionKey and recommend a report export next. - If token questions are time-windowed, mention whether the number is a current session snapshot or a
1d / 7d / 30d Gateway usage range. - Avoid pasting raw gateway logs unless the user asks; use
clawmonitor report which redacts common secrets.
ClawMonitor(OpenClaw监控)
当用户提出以下问题时使用此技能:
- - “我的代理完成了吗?为什么没有反馈?”
- “哪个会话/线程收到了最后一条消息,以及何时收到的?”
- “代理正在运行、被中断还是卡住了?有没有投递失败?”
- “哪个模型出了问题,是提供商的问题还是OpenClaw本身的问题?”
- “最近哪个会话消耗的令牌最多?”
- “我们有没有僵尸/辅助进程或脏的网关cgroup?”
前置条件
- - 此技能在具有 ~/.openclaw/ 下OpenClaw状态的机器上运行。
- 同一台机器上安装了 clawmonitor。
预检(推荐)
在使用以下命令之前,验证二进制文件存在并能读取本地OpenClaw状态:
bash
clawmonitor --version
clawmonitor init --non-interactive || true
clawmonitor status --format md
如果 clawmonitor 缺失,或 status 失败,请先安装它。
安装:
bash
pip install -U clawmonitor
其他安装方式:
bash
pipx install clawmonitor
bash
python3 -m venv .venv
. .venv/bin/activate
pip install -U pip
pip install clawmonitor
注意:
- - 某些OpenClaw环境有意禁止在运行时安装包。如果安装被阻止,请要求用户先在主机上安装 clawmonitor。
- clawmonitor init 在 ~/.config/clawmonitor/config.toml 下写入配置,可以安全地重新运行。
核心命令
0)树状视图(谁拥有哪些会话)
如果您怀疑ACP/子代理路由问题(例如Telegram线程绑定),请从以下命令开始:
bash
clawmonitor tree
1)状态(Markdown格式)
显示核心状态表:
bash
clawmonitor status --format md
要获取包含任务/消息预览的更详细表格:
bash
clawmonitor status --format md --detail
2)模型健康检查
直接和/或通过OpenClaw探测配置的模型:
bash
clawmonitor models --format md
有用的变体:
bash
clawmonitor models --mode direct --format json
clawmonitor models --mode openclaw --timeout 20
3)深入分析单个会话
为单个会话密钥导出经过编辑的报告:
bash
clawmonitor report --session-key agent:main:main --format md
4)提示(要求会话报告进度)
向会话发送进度请求:
bash
clawmonitor nudge --session-key agent:main:main --template progress
5)用于交互式检查的完整TUI
当需要交互式分类时运行TUI:
bash
clawmonitor tui
重要的TUI按键:
- - v:循环切换 会话 / 模型 / 系统
- s:直接跳转到系统
- h:切换右侧 状态 / 历史记录
- u:循环切换令牌窗口(now / 1d / 7d / 30d)
- r:刷新当前活动界面
- z:循环切换面板宽度
- Z:全屏详情
- o:在系统视图中打开英文操作员说明
快速故障排除
- - 如果 clawmonitor status 显示 DELIVERYFAILED 或 NOFEEDBACK,接下来导出报告:
bash
clawmonitor report --session-key agent:main:main --format md
- - 如果模型检查不一致,请将 直接提供商路径 与 OpenClaw路径 分开。
- 如果Telegram看起来绑定到了错误的会话密钥,运行 clawmonitor tree 并跟随绑定的会话,而不是假设 agent:main:... 是活动路径。
- 如果问题是服务级别的,使用TUI的 系统 视图,并总结 风险、可回收内存估计 以及是否检测到僵尸/孤儿进程。
- 如果TUI不可用,回退到 clawmonitor status --format md --detail 加上 clawmonitor models --format md。
回复指南
- - 对于即时消息回复,优先使用 --format md 输出。
- 如果状态显示 DELIVERYFAILED 或 NOFEEDBACK,包含相关的sessionKey并建议下一步导出 report。
- 如果令牌问题有时间窗口限制,请说明数字是当前会话快照还是 1d / 7d / 30d 的网关使用范围。
- 除非用户要求,否则避免粘贴原始网关日志;使用 clawmonitor report,它会编辑常见的秘密信息。