Telemetry

Forge telemetry is the structured evidence stream for an agent run. It exists so operators, tests, and auditors can reason about behavior without reading unstructured logs.

What gets reported

Telemetry should cover:

  • agent DID and lineage metadata where identity is bound
  • lifecycle transitions
  • model provider and model reference
  • token usage and finish reason
  • tool invocation id, tier, status, and duration
  • approval and authorization outcomes
  • health status changes
  • errors with stable error categories

Event sources

Source Examples
Core model runtime request started, chunk emitted, usage reported, finish reason
Agent loop step started, tool call requested, final answer emitted
Tool layer approval requested, tool executed, tool failed
Identity and auth DID attached, ACT verified, delegation narrowed
Health layer state transition, degraded threshold crossed

OpenTelemetry compatibility

The Rust reference includes a forge-telemetry crate for tracing and OpenTelemetry initialization. Other SDKs should preserve the same semantic fields even when they use language-native telemetry libraries.

Release discipline

Telemetry claims must be verifiable. If a page says a runtime emits a field, the release gate should be able to prove that field exists in a scenario or unit test.

See Release Gate and Security Model.