The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in embed.py and backfill_multimodal.py to manage local Docker containers (embry-embedding, embry-embedding-mm). This is used for service orchestration and health recovery (restarting containers on failure).
[EXTERNAL_DOWNLOADS]: The skill fetches dependencies from vendor-owned repositories during installation, specifically graph-memory and scillm from github.com/grahama1970. These are used for Knowledge Graph operations and LLM proxying.
[INDIRECT_PROMPT_INJECTION]: The edge-verifier skill processes untrusted source text through an LLM to determine relationships for the Knowledge Graph. While it uses a system prompt to define output format, it lacks explicit sanitization for the ingested content, creating a surface for indirect injection.
Ingestion points: verify_edges.py via the --text argument.
Boundary markers: System prompt defined within the verification loop.
Capability inventory: Performs db.aql.execute to UPSERT verified edges into ArangoDB.
Sanitization: Relies on system prompt and JSON parsing; no specific content filtering for the source text.
[CREDENTIALS_UNSAFE]: The file edge-verifier/verify_edges.py contains a hardcoded default value sk-dev-proxy-123 for the SCILLM_PROXY_KEY. While clearly intended as a local development placeholder for the localhost:4001 proxy, it matches a credential pattern.

embedding