OpenClaw Operations: Seven Commands for the First 60 Seconds, and 'It Feels Dumber' Is Usually Not the Model

🌏 中文版

This is the article in the series most worth bookmarking — not because it is interesting, but because you will need it at 2 AM.

The first 60 seconds: run these in order

The docs call it the "triage front door: 2 minutes to a diagnosis, then jump to the deep page."

openclaw status
openclaw status --all
openclaw gateway probe
openclaw gateway status
openclaw doctor
openclaw channels status --probe
openclaw logs --follow

What each should show:

One reading is worth flagging: Read probe: limited - missing scope: operator.read is degraded diagnostics, not a connect failure. Seeing it is no reason to start suspecting the network.

The most common symptom: it feels dumber

Assistant feels limited or missing tools

Easily misdiagnosed as "the model is bad," but the docs give it the first dedicated section and the answer is usually the tool profile:

And per-agent agents.entries.*.tools overrides narrow or expand the root profile — so "the other agent on the same Gateway can do it" does not mean your config is fine.

Diagnose with openclaw status / --all / doctor, then after changing the profile, restart or reload the Gateway and recheck with openclaw status --all.

Local OpenAI-compatible backend: works directly, fails through OpenClaw

A concrete escalation worth memorizing. The symptom: your self-hosted /v1 backend answers direct /v1/chat/completions probes but fails on openclaw infer model run or normal agent turns.

In order:

1. The error mentions messages[].content expecting a string → set models.providers.<id>.models[].compat.requiresStringContent: true
2. Still fails only on OpenClaw agent turns → set models.providers.<id>.models[].compat.supportsTools: false and retry
3. Tiny direct calls work but larger OpenClaw prompts crash the backend → that is an upstream model/server limit, not an OpenClaw bug

The wording on the third step is refreshingly honest. Documentation willing to say "this is not our problem" instead of hedging actually helps the person debugging — it stops you looking in the wrong place.

Three ways plugins get stuck

1. package.json missing openclaw.extensions

The plugin package uses a shape OpenClaw no longer accepts. Fix it in the plugin: add openclaw.extensions pointing at built runtime files (usually ./dist/index.js), republish, and reinstall.

2. Install policy failing closed

The symptom: an update finishes but plugins are stale, disabled, or show blocked by install policy, install policy failed closed, or Disabled "<plugin>" after plugin update failure.

The cause is usually security.installPolicy rules written too rigidly. The docs list policy shapes to avoid:

• Freezing OpenClaw-owned plugins to one exact old version (allowing only @openclaw/*@2026.5.3, say)
• Blocking by source kind alone (every npm, network, or request.mode: "update" request)
• Treating the policy command as optional — with security.installPolicy enabled, a missing, slow, unreadable, or permission-blocked policy executable fails closed
• Approving versions without checking the request's openclawVersion against plugin candidate metadata

The reason behind it: @openclaw/* plugin versions normally move with the OpenClaw release, so an OpenClaw update can require a matching plugin update during post-update sync.

Recovery:

openclaw doctor --deep
openclaw plugins update --all
openclaw status --all

If the policy is intentionally strict, relax it for the trusted upgrade window, rerun the update, then restore it. If an update failure disabled a plugin, inspect before re-enabling:

openclaw plugins inspect <plugin-id> --runtime --json
openclaw plugins enable <plugin-id>

3. Suspicious file ownership

blocked plugin candidate: suspicious ownership (... uid=1000, expected uid=0 or root)
plugin present but blocked

The plugin files are owned by a different Unix user than the loading process. Do not remove the plugin config — fix the ownership, or run OpenClaw as the user that owns the state directory.

Docker installs run as node (uid 1000), so repair the host bind mounts:

sudo chown -R 1000:1000 /path/to/openclaw-config /path/to/openclaw-workspace
openclaw doctor --fix

If you intentionally run as root, repair the managed plugin root instead:

sudo chown -R root:root /path/to/openclaw-config/npm
openclaw doctor --fix

One specific model-side error

HTTP 429: rate_limit_error: Extra usage is required for long context requests — this is Anthropic requiring an extra-usage plan for long-context requests, with its own dedicated section upstream. It is not an ordinary rate limit, and changing your retry strategy will not help.

The big picture

How this runbook is organized is itself worth learning from: symptom first, not component first. At 2 AM what you know is "it stopped answering," not "the Gateway's channel module has a problem."

The most useful entry is the tool-profile section — when an agent "feels dumber," check what tools it is allowed to use before suspecting the model. That lesson holds in any agent system.

Changelog

• 2026-08-18: Substantially revised against the current official docs. Added: the seven-command first-60-seconds ladder with what each line should show (including Read probe: limited being degraded diagnostics rather than a connect failure, and gateway status --require-rpc), the tool-profile diagnosis for "it feels dumber" (minimal allowing only session_status, coding as the new local default, per-agent overrides changing the result), the three-step compatibility escalation for local OpenAI-compatible backends (with the third step explicitly identified as an upstream limit), the three plugin failure modes (missing openclaw.extensions, the four security.installPolicy shapes to avoid plus recovery steps, and ownership-blocked plugins with Docker and root fixes), and the Anthropic long-context 429.

References

This article draws on the following official OpenClaw documentation:

• General troubleshooting — the symptom-first triage front door
• Gateway troubleshooting — deeper Gateway and model runbooks
• Tool profiles — the full profile and group table
• Plugins — install policy and ownership issues
• Channel troubleshooting — per-channel diagnosis