Tracing Conventions

Tracing conventions

tracing is the single logging surface and a user-facing product feature. Libraries emit spans and events always-on; consumers (CLIs, services, xberg-enterprise) install the subscriber and export. This skill is the canonical contract referenced by the logging-tracing and tracing-product-surface rules.

Level semantics (authoritative)

Hoisted from xberg-enterprise/crates/observability/src/telemetry.rs:

RUST_LOG (env-filter) always takes precedence over a caller-supplied default level.

Emit vs. install

• Libraries emit only. Never call tracing_subscriber::* or install a global subscriber in a library — that steals the choice from the consumer and breaks composition. Add tracing as a non-optional dependency; spans compile to near-zero cost when no subscriber is attached.
• CLIs and services install. Initialize tracing-subscriber (env-filtered, stderr) once in main. The OTLP/OpenTelemetry export layer lives behind the otel feature (see the feature-flag-conventions skill), on top of the always-on tracing.
• xberg-enterprise is the reference consumer: init_telemetry wires the fmt layer, the OTLP tracer/metrics providers (when OTEL_EXPORTER_OTLP_ENDPOINT is set), a GoldenSignalsLayer, HTTP context propagation, and a panic hook that routes panics through tracing::error! with a forced backtrace. Library spans surface through this stack automatically once the library emits them.

Span and field naming (semver-relevant)

Span names, targets, and field keys are part of the public observability API. Changing them can break dashboards, alerts, and log queries — treat them like any other public symbol and keep them stable.

• Span names: crate_name::operation (e.g. html_to_markdown::convert). Set an explicit level and stable name.
• #[tracing::instrument] on public entry points. skip/skip_all large or non-Debug arguments (input buffers, big option structs); record cheap, useful fields instead — input_len, resolved option booleans, selected strategy.
• Field keys: prefer stable, lowercase, snake_case keys. Reuse conventional keys across crates (input_len, output_len, node_count, error, reason, elapsed_ms). Avoid the reserved message field name for structured data — it collides with the event body; use a descriptive key like detail.
• Events at stage boundaries: debug! with small structured fields at parse/walk/render (or equivalent) boundaries; warn! at every silent-recovery / fallback / skip site; error! before returning a hard failure.

Instrumentation checklist

1. tracing is a non-optional dependency of the library.
2. Top-level public operations carry an #[instrument] span with recorded fields, not raw args.
3. Every place the code already recovers or silently drops something emits a WARN.
4. No per-token/per-node INFO/DEBUG spam — that detail is TRACE.
5. No subscriber installed in library code.
6. Raw println!/eprintln!/dbg! are gone (see logging-tracing); the only stdout writes left are a CLI's result output behind #[expect(clippy::print_stdout)].