AI Engineering from Scratch（英文原版）

Given an agent runtime, wire OTel GenAI semantic conventions.

Produce:

1. invoke_agent span per agent run. Kind CLIENT for remote agent services, INTERNAL for in-process. Name: invoke_agent \{gen_ai.agent.name\}.
2. chat span per LLM call with gen_ai.operation.name=chat, gen_ai.provider.name, gen_ai.request.model, gen_ai.response.model.
3. tool_call span per tool invocation with gen_ai.tool.name and, when applicable, gen_ai.data_source.id (RAG corpus / memory store).
4. Opt-in content capture: default OFF; when ON, store inputs/outputs externally and record *.reference_id on spans.
5. Context propagation: use W3C trace context headers so multi-process runs (Claude Agent SDK CLI subprocess) stitch into one trace.

Hard rejects:

• Capturing full prompts/outputs inline by default. PII and secret leakage risk; also violates the spec.
• Missing gen_ai.provider.name. Multi-provider dashboards break.
• Orphan tool spans. Always set parent-child relation via active context.

Refusal rules:

• If the runtime cannot propagate context across process boundaries, refuse. Multi-process trace stitching is required for Claude Agent SDK + CLI users.
• If the product has regulatory constraints (HIPAA, GDPR), refuse inline content capture. External store with access control only.
• If the backend does not set OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental, warn: attribute names may change on collector upgrade.

Output: tracer.py, attributes.py, content_store.py, README.md explaining span structure, stability opt-in, and content-capture policy. End with "what to read next" pointing to Lesson 24 (backends: Langfuse, Phoenix, Opik) or Lesson 17 for Claude Agent SDK trace-context propagation.