Module telemetry 

Observability as an ECS system: watch the components every other system already writes and narrate them into the installed TelemetrySink.

The pipeline’s collect systems leave pure-data ActivityRecords on the agent (an inference landed, a tool batch ran, a compaction finished); observe_lifecycle runs once per schedule pass near the end of the tick, turns those plus the agent’s own state into TelemetryEvents, and emits them. Ordering in the tick chain is load-bearing twice over: the system must run before sync_tool_stages (which consumes the transient StageJustEntered marker) and before dispatch_persistence (which drains StageIoBuffer - this system only reads the buffer, so running first is what makes each log line observed exactly once).

Structs

StageActivity

Buffered ActivityRecords awaiting the observer. Inserted alongside TelemetryState the first time the observer sees an agent, so the collect systems treat it as optional and skip recording until then (an agent’s first inference cannot land before the observer has run once).

Telemetry

The installed telemetry sink. crate::world::PipelineWorld::new installs leviath_core::telemetry::NoopSink; a host that wants export replaces the resource (the same way it installs PolicyGate or TitleSettings).

TelemetryState

The observer’s per-agent memory: what it has already narrated.

Enums

ActivityRecord

One completed piece of stage work, recorded by the collect system that applied it and drained into events by observe_lifecycle. Carries only what the collect site knows; run/stage identity is added at drain time.

Functions

observe_lifecycle

Emit lifecycle, activity, and log events for every agent run.