`instrumentLangGraph` and `createLangChainCallbackHandler` silently drop spans when used together

[priscilawebdev]

Is there an existing issue for this?

• I have checked for existing issues https://github.com/getsentry/sentry-javascript/issues
• I have reviewed the documentation https://docs.sentry.io/
• I am using the latest SDK release https://github.com/getsentry/sentry-javascript/releases

How do you use Sentry?

Sentry Saas (sentry.io)

Which SDK are you using?

@sentry/browser

SDK Version

^10.40.0

Framework Version

No response

Link to Sentry event

https://pri-dogfooding-ai.sentry.io/insights/ai-agents/?project=4510951776059472&statsPeriod=30d&trace=043dd6239e7645d9b70f641fe3ec2f83&span=undefined&trace-timestamp=1772622191

Reproduction Example/SDK Setup

No response

Steps to Reproduce

const graph = new StateGraph(IdeaState)
  .addNode("expand", expandNode)     // calls llm.invoke()
  .addNode("validate", validateNode) // calls llm.invoke()
  .addNode("differentiate", diffNode) // calls llm.invoke()
  .addNode("roadmap", roadmapNode)   // calls llm.invoke()
  .addNode("pitch", pitchNode)       // calls llm.invoke()
  .addEdge(START, "expand")
  .addEdge("expand", "validate")
  .addEdge("validate", "differentiate")
  .addEdge("differentiate", "roadmap")
  .addEdge("roadmap", "pitch")
  .addEdge("pitch", END);

Sentry.instrumentLangGraph(graph, { recordInputs: true, recordOutputs: true });
const compiled = graph.compile({ name: "idea-forge" });

const sentryHandler = Sentry.createLangChainCallbackHandler({
  recordInputs: true,
  recordOutputs: true,
});
const result = await compiled.invoke({ idea: "test" }, { callbacks: [sentryHandler] });

Expected Result

All 5 LLM calls should produce chat spans in the trace.

Actual Result

Setup	chat spans captured
instrumentLangGraph + createLangChainCallbackHandler	1 out of 5
createLangChainCallbackHandler only	5 out of 5
instrumentLangGraph only	0 (expected — no callback handler)

4 out of 5 chat spans are silently dropped with no error or warning.

Additionally, the trace contains multiple spurious nested invoke_agent sub-spans with near-zero durations (0.20ms) that should not exist:

invoke_agent idea-forge        7.43s    ← top-level (expected)
  invoke_agent idea-forge      2.70ms   ← spurious
  invoke_agent idea-forge      1.60ms   ← spurious
  invoke_agent idea-forge      0.20ms   ← spurious
  invoke_agent idea-forge      4.41s
    chat claude-haiku-4-5-20251001  4.41s  ← only 1 of 5 chat spans
  invoke_agent idea-forge      0.20ms   ← spurious
  invoke_agent idea-forge      0.20ms   ← spurious

Additional Context

Possible Root Cause:

instrumentLangGraph wraps invoke() inside Sentry.startSpan(). As async execution passes through sequential nodes, the span context from startSpan is not consistently maintained across all node invocations. Some handleChatModelStart callbacks from the handler fire outside an active span context and their spans are dropped.

Priority

_{React with 👍 to help prioritize this issue. Please use comments to provide useful context, avoiding +1 or me too, to help us triage it.}

[linear]

JS-1859