Telemetry (OpenTelemetry)

Off by default. When switched on, every system_one call sends metrics and events to the OTLP collector you name, and nowhere else. The names follow Claude Code's telemetry, so a team that already collects claude_code.* can put jul.* next to it with the same collector and dashboards.

pip install "jul[otel]"
export JUL_ENABLE_TELEMETRY=1
export OTEL_METRICS_EXPORTER=otlp          # otlp, console, none
export OTEL_LOGS_EXPORTER=otlp             # otlp, console, none
export OTEL_EXPORTER_OTLP_ENDPOINT=http://collector:4318
export OTEL_RESOURCE_ATTRIBUTES="team.id=risk,department=eng"   # optional, lands on every record

The standard OTEL_* variables apply (per-signal endpoints, headers, OTEL_METRIC_EXPORT_INTERVAL). The default protocol is http/protobuf; for grpc, also install opentelemetry-exporter-otlp-proto-grpc.

Metrics, meter com.usejul.jul, resource service.name=jul:

The standard attributes are session.id, app.version, app.entrypoint (sdk-py or cli), model and backend. method is how the question was actually read: vector, letters, cross, contrastive, head (a tuned head from autotune), pointer (a decision model), letter-readout (a letter-readout decision model), laya or remote (a SystemOneHTTP server).

Events (OTel logs), each with event.name (the same value as the record's event_name, e.g. jul.decision), event.timestamp and event.sequence:

• jul.request: one per call. request_id, duration_ms, input_tokens, question_count, state_length, state.
• jul.decision: one per question, joined to its call by request_id. question.index, question.type, method, answer, confidence, option_count.
• jul.request_error: error_type, duration_ms, question_count.

Choosing what leaves

By default only metadata leaves the process: the state is sent as <REDACTED> with its length, and question names and instructions are not sent at all. Each kind of content has its own switch:

The answer is on by default because it is an option key you wrote, and it is what a routing or audit dashboard needs; set JUL_OTEL_LOG_ANSWERS=0 to send <REDACTED> instead. The switches have their own JUL_ names so that a machine configured to log Claude Code prompts does not start sending JuL states without being asked.

A telemetry failure (collector down, OpenTelemetry missing) logs one warning and never fails the decision. With telemetry off, OpenTelemetry is not imported. jul[otel] needs OpenTelemetry 1.38 or later (Logger.emit(event_name=)); CI runs the telemetry tests at that floor.

A session is one client: session.id is drawn when a TypeSafeClient (or a SystemOneHTTP) makes its first call, and jul.session.count counts clients. Under jul serve every HTTP request goes through the server's one client, so a server process is one session, whoever calls it.

Remote deciders report too. A SystemOneHTTP call (Jev via typesafe, Ollama, Clef on Workers AI, a remote jul serve) is recorded like a local one, with model the model the server answered with (e.g. jev-1.13.0), backend and method set to remote, input_tokens from the server's usage, and the same switches and redaction. Under Escalation / --escalate-to, each tier's call is recorded under its own model, so a dashboard can put the local model and the remote one side by side. An HTTP error is a jul.request_error (error_type=RemoteError).

The two histograms have explicit buckets, since OpenTelemetry's default ones (0, 5, 10, 25 … 10000) fit neither:

backend is always set: torch, mlx or onnx for a local model, torch for Laya (its own PyTorch runtime), remote for a SystemOneHTTP server.