Enabling LangSmith tracing (LANGCHAIN_TRACING_V2=true) makes every agent run fail: "onChainEnd: Run ID … not found in run map"

[franco-teramot]

What happened?

With LangSmith tracing enabled through the standard LangChain environment variables (LANGCHAIN_TRACING_V2=true, LANGCHAIN_API_KEY, LANGCHAIN_PROJECT), every agent run fails at its first graph step and the user gets an error bubble. No model call is made. The trace that reaches LangSmith is the failed AgentGraph root run carrying only the error below.

This happens with a minimal agent (Anthropic provider, no tools, instructions "Reply with the single word OK."), so it is not related to MCP or tools. It also happens with agents that have MCP tools attached. The moment LANGCHAIN_TRACING_V2 is unset or false, the same agent and the same request succeed.

Also tried: LANGCHAIN_CALLBACKS_BACKGROUND=false — no change.

Title generation (TitleRun) succeeds and is traced normally, so the tracer itself connects; it is the agent graph's streaming path that breaks.

Version Information

Reproduced on both:

• ghcr.io/danny-avila/librechat:v0.8.7
• ghcr.io/danny-avila/librechat:v0.8.8-rc2

Docker Compose on macOS (Docker Desktop), MongoDB 7, no Redis, no Meilisearch, no RAG API. Config schema version: 1.3.15. endpoints.agents.capabilities: [tools, artifacts, context, deferred_tools], toolApproval.enabled: true (also fails with a tool-less agent, where approval never triggers).

Steps to Reproduce

1. Set in .env:

   LANGCHAIN_TRACING_V2=true
   LANGCHAIN_API_KEY=<a LangSmith key>
   LANGCHAIN_PROJECT=<any project>
   LANGCHAIN_ENDPOINT=https://api.smith.langchain.com
   ANTHROPIC_API_KEY=<key>
   ANTHROPIC_MODELS=claude-sonnet-4-6
   

2. Start LibreChat, log in, create an agent: provider anthropic, model claude-sonnet-4-6, instructions Reply with the single word OK., no tools, no capabilities beyond defaults.
3. Send say OK to that agent (UI, or POST /api/agents/chat/agents with the usual message body).
4. Observe the error bubble in the chat and the log lines below. In LangSmith the project shows an AgentGraph run with status error and the same message.
5. Set LANGCHAIN_TRACING_V2=false, restart, repeat step 3: the agent answers OK.

What browsers are you seeing the problem on?

Not browser related (reproduced via API with curl as well). Chrome for the UI runs.

Relevant log output

Error in handler EventStreamCallbackHandler, handleChainEnd: Error: onChainEnd: Run ID 01a08142-968d-730f-a33d-ad84ee93149a not found in run map.
2026-09-08 13:43:39 error: [api/server/controllers/agents/client.js #sendCompletion] Unhandled error type {"userId":"…","requestId":"…","request_method":"POST","request_path":"/api/agents/chat/agents"}

The stored assistant message has a single content part of type error. The LangSmith root run (AgentGraph) has error: "onChainEnd: Run ID … not found in run map."; a sibling TitleRun for the same message succeeds.

Screenshots

LangSmith project after five attempts: one green TitleRun, five red AgentGraph/AgentRun runs, all with the onChainEnd error (available on request).

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

Langsmith is not a first-class agent tracing solution for LibreChat, even though LC Agents SDK does use Langchain primitives

Will look into it soon but moved to troubleshooting due to above.

All recent agent tracing initiatives revolve around Langfuse as a first-class citizen of LibreChat for agentic tracing,

[franco-teramot]

Thanks for the quick answer, that clarifies it. We've moved our LibreChat tracing to Langfuse and it works well for us, so this is low priority on our side — feel free to close it or keep it as a note if the run-map error is still worth a look for people who have LANGCHAIN_TRACING_V2 set for other reasons. Appreciate the work on Langfuse as the first-class path.