skill-graph

Discover useful routes

Run skill-search search 'task' --limit 5, inspect selected entrypoints with load EXACT_ID, then use neighbors, path 'task', related or graph. Choose the smallest applicable set. Relationships never grant execution, delegation or approval.

The packaged profile is offline keyword search. After inventory changes, run skill-search refresh --lexical; every discovery/export checks freshness. Neural map and path --semantic require a separately configured pinned BGE model and llama-embedding. Missing models fail visibly; the declared graph remains usable.

skill-search export --query 'task' --limit 5 emits knowledge-snapshot/v2 JSONL with typed skills/workflows, resource revisions, labeled edges and independent ranks.

knowledge-intake --root harness=/absolute/skills inspects existing bundles as data. Review the report before adding --index: this registers read-only, namespaced, unreviewed nodes without rewriting or activating sources. Conflicts and unresolved links remain visible. Indexing never creates scanner acceptance.

Read references/maintenance.md before editing registered skills. Stage local changes, preserve licenses/provenance, apply construction review, generate backlinks, and verify the graph before promotion. Direct remote registry installs require exact final bundle acceptance: read references/remote-loading.md. Historical receipts cannot bless changed bytes. Resolve paths from AGENT_KNOWLEDGE_HOME.