What problem does it solve?
Multi-agent systems often lack end-to-end observability, making it impossible to audit who did what, when, and why. This Skill records every dispatch, scan, parse, and card-generation event into a traceable evidence chain so agent runs become auditable and replayable.
Core Features & Use Cases
- Evidence Chain Generation: Links events by trace_id into a queryable provenance chain stored as trace.jsonl and trace_summary.json.
- Vector Indexing: Writes semantic index entries to Qdrant for retrieval, with SQLite as a local fallback when the vector store fails.
- Resilient Failure Handling: Caches events locally on vector-store failure, drops malformed events without breaking the chain, and returns stale snapshots on query timeout.
- Use Case: Run the provenance stage after a multi-agent pipeline to produce a full audit trail of every agent dispatch and execution event for security review or rollback analysis.
Quick Start
Run the provenance skill with python scripts/run_provenance.py to execute the full agent pipeline and write the trace evidence chain to the examples output directory.