OpenClaw documentation

Use the synchronized official documentation as the source of truth. Keep context small by loading only the files and sections needed for the current request.

Establish the version boundary

1. Read references/SOURCE.json when provenance or freshness matters.
2. For operational work, obtain openclaw --version when it is available and relevant.
3. Treat the references as documentation for the recorded upstream main revision. Flag a possible version mismatch when the installed OpenClaw version is older.
4. Never invent a command, flag, configuration key, default, or migration step. Search the references and state uncertainty when the documentation does not establish the answer.

Do not update the installed skill during ordinary questions. Use the local snapshot without blocking the user. Refresh only when the user asks to install or update the skill.

Find documentation progressively

Choose the narrowest route that fits the request.

Exact error, command, or configuration key

Search references/ directly for:

• the exact error message first;
• the full CLI command or subcommand;
• the complete configuration key;
• distinctive log text, provider name, or channel name.

When shell search is available, prefer:

rg -n -i --glob '*.md' --glob '*.mdx' '<exact text>' references
rg -l -i --glob '*.md' --glob '*.mdx' '<keyword>' references

Use the host's equivalent file-search tool when rg is unavailable. Do not load the global router before an exact search unless the search produces no useful candidate.

Broad topic

1. Read references/SKILL_INDEX.md.
2. Open exactly one matching catalog under references/_catalog/.
3. If that catalog links to alphabetical sections, open exactly one matching section.
4. Select at most three candidate documents using their summary and read_when metadata.
5. Inspect headings or search within those documents before reading long sections.
6. Expand to another document, section, or catalog only when the first candidates are insufficient.

Topic routing

User intent	Start with
Installation, update, migration, deployment	install catalog
Gateway configuration, service, networking	gateway catalog
Telegram, WhatsApp, Discord, Slack, LINE, or other messaging	channels catalog
Anthropic, OpenAI, Gemini, Ollama, or other models	providers catalog
Exact openclaw command or flag	CLI catalog
Exec, browser, web, skills, permissions, or agent tools	tools catalog
Cron, hooks, tasks, or webhooks	automation catalog
Agents, sessions, memory, routing, or architecture	concepts catalog
Nodes or OS-specific behavior	nodes catalog and platforms catalog
Control UI, WebChat, dashboard, or TUI	web catalog
Security, exposure, or incident response	security catalog and gateway catalog
Unclear symptom or general troubleshooting	help catalog

If a catalog does not exist in an older installed snapshot, search the corresponding references/<topic>/ directory directly.

Diagnose from evidence

Match diagnostics to the symptom instead of running a universal command ladder.

• For Gateway reachability, inspect status, service state, and relevant Gateway logs.
• For a single channel, inspect Gateway reachability and that channel's status, policy, and channel-specific troubleshooting page.
• For model failures, inspect provider authentication, model resolution, and the exact provider error without probing unrelated channels.
• For configuration failures, identify the active config path, exact rejected key, and matching configuration reference.
• For update or migration failures, establish the installed version, install method, and target version before recommending changes.

Begin with read-only observations. Preserve exact errors and command output when searching the documentation.

Control state-changing actions

Treat edits, installation, updates, --fix, --force, restarts, uninstalls, credential changes, pairing approvals, and message sends as state-changing actions.

1. Explain what will change and the likely impact.
2. Confirm the target and scope from available evidence.
3. Obtain the user's authorization when it is not already explicit.
4. Prefer previews, validation, backups, and reversible operations.
5. Re-check status after the change.

Never expose tokens, passwords, session data, auth profiles, or secrets in the response. Redact them from copied output.

Produce the answer

• Lead with the diagnosis or requested outcome.
• Give commands only after confirming them in the synchronized references.
• Separate documented facts from inferences.
• Mention relevant version assumptions or mismatch risks.
• Cite the local reference paths used so the user can verify the answer.
• Keep unexplored alternatives out of context unless the primary route fails.

Maintain this skill

Run maintenance only when the user explicitly asks to refresh this repository:

sh scripts/sync-docs.sh
python3 scripts/generate_index.py --check
python3 -m unittest -v tests/test_repo.py

For an installed Git checkout, install or update with:

bash <skill-directory>/scripts/install-skill.sh <skill-directory>

Do not run scripts/sync-docs.sh or regenerate indexes during normal OpenClaw assistance.