goose

goose is an aaif open-source AI coding agent. It emits structured OpenTelemetry data for every session: agent reply spans, model provider calls, tool dispatches, session token counts, and duration. Routing this to Last9 lets you measure adoption, model latency, cost, and tool reliability — within your existing observability stack.

goose exports all three OpenTelemetry signal types:

• Traces — spans covering reply, reply_stream, complete, stream_response_from_provider, dispatch_tool_call, plus events like WAITING_LLM_STREAM_START/END
• Metrics — derived from monotonic_counter.* attributes on log events; instrumented via tracing-opentelemetry
• Logs — session events with token counts, duration, exit type, message count

Default temporality is cumulative, so metrics ingest cleanly into Last9 without conversion.

What gets exported

Spans

Session counters (emitted as log attributes)

These counters propagate as native OTel metrics when tracing-opentelemetry’s MetricsLayer is active.

Prerequisites

1. Last9 account — Sign up at app.last9.io
2. goose installed — brew install block-goose-cli or follow install docs
3. OTLP credentials — Get your endpoint and auth header from Integrations → OpenTelemetry

Setup

goose honors the standard OpenTelemetry SDK environment variables — no custom prefix.

1. Get your Last9 OTLP credentials

   Navigate to Integrations → OpenTelemetry in your Last9 dashboard. Copy:

   • OTLP Endpoint (e.g., https://otlp.last9.io or https://otlp-aps1.last9.io:443)
   • Authorization header value (e.g., Basic <base64-token>)

2. Add environment variables to your shell profile

   Append the following to ~/.zshrc, ~/.bashrc, or equivalent:

   • HTTP (recommended)
   • gRPC

   export OTEL_EXPORTER_OTLP_ENDPOINT="https://<your-last9-otlp-endpoint>"
   export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Basic <your-last9-auth-token>"
   export OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf
   export OTEL_SERVICE_NAME=goose
   export OTEL_RESOURCE_ATTRIBUTES="deployment.environment=local"

3. Reload your shell

   source ~/.zshrc

4. Run goose

   goose session

   Or non-interactively:

   goose run -t "explain this file"

5. Verify data is arriving

   • Traces — open Traces in Last9, filter by service.name = goose. Look for reply, complete, stream_response_from_provider span names
   • Metrics — open Metrics, search for goose_session_tokens or goose_session_duration_ms
   • Logs — open Logs, filter by service.name = goose. Session events include total_tokens, duration_ms, exit_type

Configuration reference

Core

Per-signal control

Sampling and tuning

Anonymous usage data (separate from OTel)

Configuration via goose config file

Alternatively, configure via ~/.config/goose/config.yaml:

otel_exporter_otlp_endpoint: "https://<your-last9-otlp-endpoint>"
otel_exporter_otlp_timeout: 30

Environment variables take precedence over config file values.

What you can do in Last9

Model latency tracking (traces)

stream_response_from_provider span duration is end-to-end LLM call latency. Filter by gen_ai.request.model to compare providers/models. p95/p99 over time catches degradation:

histogram_quantile(0.95,
  sum by (le, gen_ai_request_model) (rate(traces_span_metrics_duration_milliseconds_bucket{ServiceName="goose",SpanName="stream_response_from_provider"}[5m]))
)

Token consumption (logs / metrics)

Per-session token counters are emitted as log attributes (monotonic_counter.goose.session_tokens). Aggregate via Last9 logs panel or convert to a metric. Sum over a time window for daily/weekly totals.

Session duration distribution (logs / metrics)

monotonic_counter.goose.session_duration_ms plus the structured log fields (exit_type, message_count, total_tokens) let you find long-running sessions or sessions that exited abnormally.

Tool reliability (traces)

dispatch_tool_call spans carry input, output, and session.id. Filter for failed spans (StatusCode=ERROR) to spot tool errors. Group by tool name from span attributes.

Session replay via session.id (traces + logs)

Every span and log carries session.id. Filter by it to reconstruct a single goose session — agent loop, model calls, tool dispatches — in trace order.

Sampling tip

For high-volume environments, sample 10% of traces to reduce volume:

export OTEL_TRACES_SAMPLER="parentbased_traceidratio"
export OTEL_TRACES_SAMPLER_ARG="0.1"

Counters and logs continue at full rate; only trace volume reduces.

Troubleshooting

• No data in Last9

  • Confirm echo $OTEL_EXPORTER_OTLP_ENDPOINT in the same shell where you run goose. Restart the shell after editing ~/.zshrc.
  • Verify OTEL_SDK_DISABLED is not set to true.

• 401 / authentication errors

  • Header format must be Authorization=Basic <token> (key=value, not HTTP colon syntax)
  • Regenerate the token from Integrations → OpenTelemetry if it has expired

• Service name appears as unknown_service

  • Set OTEL_SERVICE_NAME=goose explicitly. goose’s resource builder defaults to goose but env vars take precedence.

• High trace volume

  • Set OTEL_TRACES_SAMPLER=parentbased_traceidratio and OTEL_TRACES_SAMPLER_ARG=0.1 to sample 10% of traces

• goose anonymous telemetry vs OpenTelemetry export

  • These are separate. GOOSE_TELEMETRY_ENABLED controls goose’s built-in anonymous usage data collection. The OTEL_* vars control the OpenTelemetry export to your collector. You can run both, neither, or just one.

Please get in touch with us on Discord or Email if you have any questions.