Skip to content

observability.md

Latest commit

 

History

History
29 lines (17 loc) · 2.34 KB

observability.md

File metadata and controls

29 lines (17 loc) · 2.34 KB
Raw

Flue observability retention

Bobsled treats observability as durable evidence first and a UI decision later. Flue's observe() API is process-global and live-only, so the application installs one subscriber from the provider bootstrap shared by the server and standalone agent entry points.

What is retained

Every FlueObservation delivered after process startup is stored in bobsled.db. This includes all event families and the live-only observation detail that Flue does not replay: full model-visible turn requests, prompts and messages, model responses, reasoning and text deltas, tool arguments and effective results, task prompts/results, compaction, token/cost usage, structured logs, and classified errors with live stacks.

Each payload is stored twice:

  1. payload_blob is the authoritative lossless Node V8 serialization.
  2. payload_json is a human/query-tool-friendly projection with explicit markers for cycles, undefined, big integers, non-finite numbers, binary data, errors, maps, and sets.

The row also projects stable envelope and correlation fields into indexed columns. Token accounting must use one semantic level only: turn is the leaf, while operation and compaction usage are rollups that already include their turns.

Deliberate exclusions

  • Flue replaces raw base64 image bytes with its omission sentinel before observers run.
  • Bobsled does not persist FlueEventContext.env.
  • Request headers, cookies, authorization data, and URL query strings are not persisted.
  • There is no API for raw telemetry content yet.

The retained content remains highly sensitive because it can include source code, prompts, reasoning, tool inputs/results, filesystem paths, and error stacks. The database is mode 0600 and belongs in protected durable storage and backups.

Delivery behavior

The synchronous observer performs only cheap queueing. A microtask batches SQLite writes in one transaction; failed batches remain queued for a later retry and never affect agent execution. A clean process exit flushes the outstanding batch. Abrupt host or process failure can still lose the final in-memory batch because Flue's source stream itself is live-only and best-effort.

Current retention is indefinite. A later policy may add rotation, encryption, export, or tiered redaction, but it must be versioned and must not silently change historical evidence.