feat(langchain): add the langchain and langgraph callback handler - #217

LangChain.js and LangGraph.js apps can now send traces to telemetry.dev through our own package. At this time the LangChain guide covers only Python, through OpenLLMetry, and a TypeScript app must use a third-party instrumentation.

import { init } from "@telemetry-dev/sdk";
import { TelemetryDevCallbackHandler } from "@telemetry-dev/langchain";

init({ serviceName: "support-agent" });

const agent = graph.compile().withConfig({ callbacks: [new TelemetryDevCallbackHandler()] });

The handler is a LangChain callback handler on @telemetry-dev/sdk, as the provider wrappers are. Thus the init() options (capture, mask, batching), propagateAttributes(), and the spans of observe() operate on the LangChain spans too. A LangGraph run becomes an invoke_agent span, and each node becomes a span under it. Each model call becomes a chat <model> span with usage, finish reason, and the type and name of the bound tools in gen_ai.tool.definitions. Each tool becomes an execute_tool span. LangGraph copies configurable.thread_id into the metadata of each run, and the handler uses it as the session.

The handler does not record runs that LangChain tags langsmith:hidden, such as __start__. Their child runs go under the nearest recorded run.

The handler sets awaitHandlers. Without it, LangChain calls handlers from a shared background queue. There the active context is not the context of the caller, and concurrent runs go under the wrong span of the app.

The declaration of handleChainStart in @langchain/core lists the parameters in a different order from the call that LangChain makes. The handler uses the order of the call, and a comment in handler.ts tells this.

If the state of a graph has a messages list, the output of the graph span is the last message. The trace list shows the first 4,000 characters of the root output, and a full message state puts the final answer after that limit.

Each span has langchain.run.id. With telemetry-dev/telemetry.dev#1113, telemetry.dev shows these spans as framework langchain. An older ingest shows them as otel.

The lockfile adds @langchain/core 1.3.4 and @langchain/langgraph 1.7.0 as dev dependencies. pnpm then shares their @cfworker/json-schema and ws 8.21.3 with the @modelcontextprotocol/sdk and @google/genai entries, and the peer suffixes of those entries change.

The LangGraph quickstart agent from the LangChain docs, with Claude Sonnet 4.6, in telemetry.dev:

Related: telemetry-dev/telemetry.dev#1113

Follow-up for the docs page and the setup flow: https://github.com/telemetry-dev/telemetry.dev/issues/1114