Local Knowledge Router

Local Knowledge Router

When to use

Use this skill before guessing paths or doing broad file searches when a user asks about local or private operational knowledge, including:

• installed or custom Hermes skills and their support docs;
• generated tool OKFs, runbooks, operational docs, and memory docs;
• helper scripts and automation entry points;
• cron jobs and scheduled maintenance;
• MCP servers and wrapper scripts;
• service workflows in the configured local source tree.

The plugin indexes whole artifacts. It identifies the first artifact to inspect; it does not replace reading that source.

Workflow

1. Search for the user's intent:

   knowledge_search(query="backup runbook", limit=8)
   knowledge_search(query="paperless review automation", limit=8)
   knowledge_search(query="siyuan mcp wrapper", limit=8)
   

2. Fetch the best artifact before acting:

   knowledge_get(artifact_id="skill:example-skill", include_neighbors=true)
   

   Use knowledge_neighbors separately when graph context will help:

   knowledge_neighbors(artifact_id="cron:daily-review", limit=20)
   

3. Inspect the routed source of truth:

   • skill → load it with skill_view.
   • script → read the script and help text before running it.
   • tool_okf → use it as routing context, then inspect the live tool schema/docs before high-impact calls.
   • runbook, memory_doc, doc, or skill_support_doc → read the file before changing systems.
   • cron_job → verify the live cron registry before mutating jobs.
   • mcp_server → inspect the live wrapper/config before troubleshooting.

4. Check freshness when it matters. Managed lookups rebuild an index that is missing, corrupt, older-format, or marked dirty by completed tool-OKF publication. They do not detect ordinary source-file, cron-registry, or MCP-config changes. When those changed recently or results look stale, force a rebuild:

   knowledge_search(query="new helper script", limit=8, rebuild=true)
   

   knowledge_get and knowledge_neighbors also accept rebuild=true. An operator may optionally schedule the plugin's explicit CLI build command to enforce a freshness interval, but no schedule is required.

5. If mixed results look weak, retry once with a shorter core-intent query and the likely artifact_type. Current plugin revisions can perform one such retry automatically when a matching explicit useful route exists and its accepted query is no longer than the current query; the remembered artifact still has to be rediscovered in the live index. A newer matching rejection vetoes an older overlap route. Default-enabled implicit feedback may supply a lower-priority route from mature same-turn evidence under the same safeguards: current-index promotion or one verified typed retry. Matching explicit routes take precedence.

6. Record clear lookup outcomes:

   knowledge_feedback(event_id=<usage_event_id>, rating="useful", artifact_id="skill:example-skill")
   knowledge_feedback(query="missing workflow phrase", rating="missing", note="Expected the deployment runbook")
   

Pitfalls

• Search results are routing hints, not proof; read the artifact before relying on it.
• If a query returns no results, retry once with broader domain synonyms before broad repository search.
• Do not include secrets or private document text in queries or feedback notes; telemetry is local but persistent.
• Treat generated tool_okf files as compact routing hints, not proof of current tool behavior.
• After installing this skill or enabling the plugin, start a fresh session (or reload skills and reset) so the toolset and instructions enter the prompt.