Observability

autosolve has OpenTelemetry baked in at every level. Disabled by default — zero overhead until you need it.

Enable OTEL

Set otel.sdkDisabled: false in config.yml and point to your collector:

otel:
  sdkDisabled: false
  serviceName: autosolve
  exporter:
    endpoint: "localhost:4317"
    protocol: grpc

That's it. Traces, metrics, and logs will flow via OTLP/gRPC.

What Gets Instrumented

Traces

Every task invocation in longrun is automatically wrapped in a span. You get a trace for each polling cycle without writing any OTEL code:

[longrun/task: "worker:issue-poller:thumbrise/autosolve"]
  └─ [GitHub API call]
       └─ [SQL INSERT]

Errors are recorded on spans automatically: span.RecordError(err) + span.SetStatus(codes.Error, ...).

Metrics

Baseline retries emit counters and histograms:

Metric	Type	What it tells you
longrun_baseline_retry_total	Counter	Retries per task and error category
longrun_degraded_total	Counter	How often a task enters degraded mode
longrun_degraded_duration_seconds	Histogram	Time spent in degraded state

Useful alert: "task X in degraded for 10 minutes".

Logs

All logging uses slog with context. When OTEL is configured, every log line carries trace_id and span_id — automatic correlation with traces, zero boilerplate in business code.

Compatible Backends

Anything that speaks OTLP: Grafana Tempo, Jaeger, Uptrace, Datadog, Honeycomb, SigNoz, etc.

OTEL Environment Variables

All config fields can be overridden via env vars:

export AUTOSOLVE_OTEL_SDKDISABLED=false
export AUTOSOLVE_OTEL_EXPORTER_ENDPOINT=localhost:4317

See the OpenTelemetry SDK env var spec for the full list.