antinet-provenance

Provenance Tracing Skill (Tàishǐgé)

How to Use

• Called by each government office at key operation nodes (or pushed asynchronously via the event bus), writing to the chain of evidence.
• Provides retrieval interfaces by trace_id / time window / government office dimension, supporting the audit panel and rollback localization.

Input

• event: operation event, containing actor(government office name),action、target、timestamp、trace_id
• payload: (optional) structured artifact reference associated with the event

Output

• evidence_chain: traceable chain of evidence linked by trace_id
• vector_index: vector index entries for semantic retrieval
• query_api: interface description for retrieving historical evidence by condition

Dependencies

• Qdrant (vector storage, accessed via MCP)
• SQLite (structured event persistence, local fallback)
• Event reporting protocol of each government office (unified schema)

Failure Handling

• Vector store write failure → cache event in local SQLite, asynchronously backfill after recovery, without blocking the main flow.
• Single event with illegal schema → record and discard that entry, without affecting the write of the entire chain.
• Retrieval timeout → return the most recent successful snapshot and mark stale。

Reuse Value

• Observability foundation: any multi-Agent system can mount it directly to obtain out-of-the-box auditing and replay capabilities.
• Aligned with evaluation: a natural scoring point for the Agent Infra track's "Engineering Implementation and Runtime Verification and Security Auditing (20%)" dimension.

Final-round code package execution (runnable package)

• Real entry point:scripts/run_provenance.py
• Execution is equivalent to core.runtime.AgentSession.run_stage("provenance"), calling memory.taishige.TaiShiGeAgent.writeback(return full-chain events and four-color cards to the chain of evidence, pure Python, can be offline).
• Run:python skills/provenance/scripts/run_provenance.py
• Artifacts:examples/snse_survey/provenance/(trace.jsonl + trace_summary.json, containing every Junjichu dispatch and government office execution event).