OTLP exporter

TokenOps ships envelopes to any OTLP/HTTP/JSON-compatible collector (OpenTelemetry Collector, Honeycomb, Grafana Agent, …) when otel.enabled is set.

Each envelope becomes one OTLP LogRecord. Numeric fields (tokens, latency, cost) become attributes keyed by the GenAI semantic conventions (gen_ai.usage.input_tokens, gen_ai.system, …) plus TokenOps-specific keys under tokenops.*.

Enable

otel:
  enabled: true
  endpoint: http://localhost:4318
  service_name: tokenops
  redact: true        # default: redact secrets before exporting
  headers:            # forwarded on every export request
    x-honeycomb-team: ...

Or via env:

export TOKENOPS_OTEL_ENABLED=true
export TOKENOPS_OTEL_ENDPOINT=http://localhost:4318

Redaction

When redact: true (the default), the redaction pipeline runs on every envelope before encoding so secrets never leave the daemon. The same pattern + entropy detectors used for the local event store apply.

Attribute reference

GenAI semantic conventions:

• gen_ai.system (openai / anthropic / gcp.gemini)
• gen_ai.request.model
• gen_ai.response.model
• gen_ai.usage.input_tokens
• gen_ai.usage.output_tokens
• gen_ai.usage.total_tokens

TokenOps-specific:

• tokenops.event.type
• tokenops.prompt.hash
• tokenops.cost_usd
• tokenops.workflow.id / tokenops.agent.id / tokenops.session.id
• tokenops.cache.hit
• tokenops.optimization.{type,mode,decision,estimated_savings_tokens}

The full key set lives in pkg/eventschema/otel.go.