LLM Observability Integration Skill

This skill covers creating LLMObs plugins that instrument LLM library operations and emit span events. Supported operations: chat completions (streaming and non-streaming), embeddings, agent runs, orchestration (workflows / graphs), tool calls, retrieval (RAG / vector DB).

Read Upstream Source First

LLM libraries iterate fast — six-month-old assumptions about an SDK's response shape, streaming contract, or tool-call format are usually wrong. Before category detection or any plugin work, read the upstream library's source for the installed version (versions/<lib>@<range>/node_modules/<lib>). The shape checklist below depends on facts the source carries (does this package make HTTP calls? does it orchestrate? does it support multiple providers?). See apm-integrations § Read Upstream Source First for the shallow-clone / npm pack shapes.

Core Concepts

1. LLMObsPlugin Base Class

Leaf plugins extend LLMObsPlugin and implement two methods:

• getLLMObsSpanRegisterOptions(ctx) — returns a required kind plus any available name, model and session fields.
• setLLMObsTags(ctx) — tags the operation's input, output, metrics, and metadata.

A composite root such as ai/index.js extends CompositePlugin and selects leaf plugins.

On the usual promise-backed channel, start(ctx) registers the span and captures context, end(ctx) restores the parent after the wrapped call returns, and asyncEnd(ctx) calls setLLMObsTags() after the operation settles.

See references/plugin-architecture.md for the full implementation surface.

2. Package Shape

Settle each instrumented surface's shape before writing anything — it decides which methods to hook and how the operation gets its response. These are working categories for reasoning, not constants in the codebase, so classify by reading the source rather than looking for an enum.

• LLM client — owns the provider endpoint, transport and authentication (openai, anthropic, genai). Hook the chat / completion methods.
• Multi-provider — accepts provider implementations behind one surface (ai, langchain). The providers may live in separate packages. Hook the provider abstraction layer.
• Orchestration — runs a graph or workflow and holds state, with no provider HTTP of its own (langgraph). Hook the workflow lifecycle (invoke, stream, run).
• Infrastructure — implements a protocol across a client / server split (modelcontextprotocol-sdk). Hook the protocol handlers.

The shape decides the response source and test harness. The instrumented operation decides its span kind and fields. Hybrid packages such as ai and LangChain must be classified per operation. Test strategy per shape lives in llmobs-testing.

See references/category-detection.md for heuristics and worked examples.

3. LLM Span Kinds

SPAN_KINDS in packages/dd-trace/src/llmobs/constants/tags.js lists llm, agent, workflow, task, tool, embedding, retrieval. Chat completions and text generation are llm; graph or chain execution is workflow; agent runs are agent; vector-DB and RAG lookups are retrieval. Only the public SDK validates against that list, so a plugin may register a kind outside it — ai v7 and claude-agent-sdk both use step.

4. Message Extraction

llm operations convert provider-specific messages to the tagger's message shape:

Common shape: [{ content?: string, role: string, toolCalls?: object[], toolResults?: object[] }]

role defaults to an empty string. Tool-call or tool-result-only messages may omit content.

Provider-specific handling:

• OpenAI: Direct format match, handle function_call and tool_calls
• Anthropic: Map role values, flatten nested content arrays
• Google GenAI: Extract from parts arrays, map role names
• Multi-provider: Detect provider and apply appropriate extraction

See references/message-extraction.md for provider-specific patterns.

Implementation Steps

1. Map each surface's response source and operation kind, from the upstream source rather than the package name.
2. Create leaf plugins under packages/dd-trace/src/llmobs/plugins/{integration}/ extending LLMObsPlugin.
3. Implement getLLMObsSpanRegisterOptions(ctx) — span kind plus any available name, model and session fields.
4. Implement setLLMObsTags(ctx) — input, output, metrics and metadata from the fields the instrumentation publishes on ctx, tagged through this._tagger.
5. Cover the edges: streaming, kind-specific error output, non-standard formats, absent metadata.

Export the class itself when the package needs one plugin (openai, anthropic, genai), or an array when several operations each need their own (langchain, langgraph, modelcontextprotocol-sdk, claude-agent-sdk). Use a CompositePlugin root when one integration selects between child implementations, as ai does. The required static fields and the rest of the surface are in references/plugin-architecture.md.