observability-tracerlisted

# observability-tracer Emits a **tree of typed spans** for an agent run. Each span carries OpenTelemetry GenAI attribute names (`gen_ai.*`); trace-level attributes (user/session id) propagate to every child. Message content is captured **only when you opt in** — the privacy default is off. ## When to use - Debugging "what did the agent actually do, in what order, and how long did each step take?" - Recording token usage and model per call for cost analysis. - Producing portable JSONL traces you can diff in CI or load into a viewer. ## Run it ``` python -m harness.observability trace # content dropped (default) python -m harness.observability trace --capture-content # opt in to message content python .claude/skills/observability-tracer/scripts/run.py trace --out trace.jsonl ``` ## The privacy choke point Content attributes (`gen_ai.input.messages`, `gen_ai.output.messages`) are dropped at a single choke point — `Span.set_attribute` — so they cannot leak whether passed at span-open or added later. Turning capture on is a deliberate `capture_content=True`. This is the one invariant the skill defends: **traces are safe to ship by default.** ## Span kinds and attributes - **kinds**: `agent`, `chain`, `llm_call`, `tool_call`, `retrieval`. - **non-content attrs** (always kept): `gen_ai.request.model`, `gen_ai.operation.name`, `gen_ai.usage.input_tokens`, `gen_ai.usage.output_tokens`. - **content attrs** (opt-in only): `gen_ai.input.messages`, `gen_