skill-graph

Use this skill when a task could benefit from specialized agent skills, especially work involving frontend polish, testing, documentation, deployment, security review, performance, or framework-specific implementation.

Runtime Loop

1. Resolve the task before specialized work:

   skill-graph resolve "<task>" --format markdown

2. Load the selected shallow context first.
3. Inspect the task and repository.

4. Expand only nodes justified by new evidence:

   skill-graph expand <node-id> --depth summary

5. Expand to full context only when the summary is insufficient:

   skill-graph expand <node-id> --depth full

6. Ask before installing any remote skill.
7. At the end, report which nodes were loaded and which frontier nodes were skipped.

Commands

• skill-graph index: scan local skill roots and write .skill-graph/index.json.
• skill-graph index --skills-sh-query "<query>": include not-installed skills.sh candidates in the graph for the current domain.
• skill-graph remote-cache "<query>": cache skills.sh candidates and show approval-required install commands.
• skill-graph embeddings index: build a local semantic embedding index. Use the default qwen3-local provider only when local Python dependencies and model storage are acceptable; use --provider deterministic for tests and demos.
• skill-graph embeddings index --trust-remote-code: allow model repository code execution only after explicit human review of the model source.
• skill-graph embeddings info: show the saved local embedding provider, model, dimensions, and vector count.
• skill-graph edges suggest: propose inferred edges from embedding similarity for human review; do not treat proposed edges as canonical.
• skill-graph search "<query>": search the local graph with BM25 by default; use --strategy lexical for baseline comparison, --strategy semantic after embeddings exist, or --strategy hybrid for BM25 plus lexical plus semantic fusion.
• skill-graph resolve "<task>": return selected nodes, context depths, frontier nodes, conflicts, missing nodes, scoring provider provenance, and reasons. Hybrid resolution includes semantic results only when a local embedding index already exists.
• skill-graph expand <node-id> --depth <depth>: load deeper context for one node.
• skill-graph context: show context layers loaded in the current workspace.
• skill-graph explain --last: explain the last resolution path.
• skill-graph install <node-id>: show the exact dry-run install command for a remote node when available.

Operating Rules

• Prefer installed local skills when they are good enough.
• Use ancestors at shallow depth.
• Use direct matches at the deepest useful depth allowed by budget.
• Keep complements on the frontier until the task requires them.
• Treat conflicts as warnings that need human review before loading both skills at full depth.
• Treat inferred edge suggestions as review items, not active resolver policy.
• Never install remote skills without explicit user approval.
• Treat remote-cache results as metadata only until the user approves an install.
• Do not enable any embedding provider that uploads local task, repository, or private skill text without explicit human approval.
• Do not pass --trust-remote-code unless the user explicitly accepts local execution of model repository code.