Observability
Bonfire
The failure that brings your agent stack down won't be the one you were monitoring. Bonfire watches the rest.
Your dashboards say green. Bonfire is already watching the thing you haven't checked — the silent accumulation, the single-provider dependency, the agent that hasn't emitted a token event in three hours. Bonfire is the telemetry layer the rest of your reliability stack reads. It runs without your attention so you can focus yours.
What it does
Token event streams
Bonfire writes real-time telemetry to bonfire_tokens.jsonl as every model call completes — counts, efficiency ratios, model used, latency. No dollar values. What you actually know, not an approximation.
Health snapshots
bonfire_health.json is updated continuously. Agent911 reads it directly. Transmission's governor checks it before every model dispatch. The rest of your reliability stack reads Bonfire — not the other way around.
Alert history
bonfire_alerts.log records threshold crossings with timestamps. Not a dashboard. A file. Grep it, tail it, pipe it. Works in any environment.
Governor preflight (live)
Since 2026-03-18, Transmission's router calls Bonfire's token_governor.py before every model dispatch. Budget and lane policy enforcement is real-time, not retrospective. Bonfire stops the expensive call before it happens.
REB consumer
Bonfire receives events from the Resilience Event Bus and routes them to Commander and Quartermaster. Your governance layer sees what happened without you relaying it manually.
What Bonfire is not
- Not a cost estimator. Bonfire reports token counts and efficiency ratios. Dollar values depend on your actual per-token rate, which varies by contract. Use
operator_ratesto configure your own cost view. - Not a dashboard. Bonfire is a file-based telemetry system. The outputs are structured JSONL and log files that other tools read. Agent911, Transmission, and Quartermaster read Bonfire. Not the other way around.
- Not a standalone product. Bonfire ships as part of the Operator Bundle. It's the telemetry backbone that makes the rest of the resilience layer intelligent.
Included in
Bonfire is part of the Operator Bundle — the complete resilience layer. Detection → Readiness → Recovery, fully wired.
bonfire statusConfirm REB events are flowing. Run this after bundle install.