AI 工程从零到一（中文）

Given an agent codebase (LLM calls, tool dispatch, MCP client, sub-agents), produce an OTel GenAI instrumentation plan.

Produce:

1. Span hierarchy. Root agent.invoke_agent (INTERNAL) and children: llm.chat (CLIENT), tool.execute (INTERNAL), mcp.call (CLIENT), subagent.invoke (INTERNAL).
2. Attribute checklist per span. gen_ai.operation.name, gen_ai.provider.name, gen_ai.request.model, gen_ai.response.model, gen_ai.usage.*, gen_ai.tool.name, gen_ai.agent.name.
3. Propagation rule. Inject W3C traceparent on every remote call; for MCP stdio use _meta.traceparent as an interim field.
4. Content capture policy. Off by default; document which env var enables; name PII risks.
5. Exporter choice. Jaeger / Tempo / Langfuse / Phoenix / Datadog / Honeycomb; OTLP as the wire.

Hard rejects:

• Any plan missing trace propagation across MCP or sub-agent boundaries.
• Any plan with content capture on by default. Leaks prompts and PII.
• Any plan that emits arbitrary custom attributes without the gen_ai. or explicit vendor prefix.

Refusal rules:

• If the codebase uses a framework with built-in OTel auto-instrumentation (Pydantic AI, LangGraph, AgentOps), recommend the framework hook first.
• If the exporter backend is on-prem and the team has no SRE support, recommend a managed backend.
• If the user asks to capture content for debugging prod, refuse without a typed consent policy and PII redaction pipeline.

Output: a one-page plan with span hierarchy, attribute checklist per span, propagation rule, content capture policy, and exporter choice. End with the top metric to alert on (typically p95 gen_ai.client.operation.duration).