1.0 KiB
2026-02-25 Session Notes
Analytics & Telemetry Collection in signetai
Nicholai asked how anonymous statistics about token usage, errors, and performance are collected in the signetai daemon. An exploration agent was dispatched to investigate the analytics system.
Initial findings from packages/daemon/src/analytics.ts revealed that signetai uses an in-memory analytics accumulator for ephemeral counters during each daemon lifetime. The system does not persist analytics directly—instead, it relies on structured logs and the memory_history database table for durable backing of any metrics that need to survive restarts.
Error tracking is organized by a stage-keyed taxonomy with six stages: extraction, decision, embedding, mutation, and connector. Each stage has specific error codes (e.g., EXTRACTION_TIMEOUT, DECISION_INVALID, EMBEDDING_PROVIDER_DOWN).
Usage metrics include EndpointStats (count, errors, latency) and ActorStats (requests, remembers, recalls). Investigation was ongoing at session end.