Deprecation Notice: LangSmith Tracing Changes

[Mason Daugherty (mdrxy)]

Summary

Starting with langchain-core 1.2.4, we've made important changes to how runs are traced to LangSmith. If you're using an older version of LangChain, you may experience missing token counts, cost information, or input data in your LangSmith traces.

What's Changing

1. Token Counts and Cost Metrics

Issue: Token counts and cost information may not appear in LangSmith traces when using older LangChain versions.

What happened: We've updated where token usage metadata is stored in traces. The LangSmith platform now expects this information in a specific location that older LangChain versions don't populate.

Action Required: Upgrade to langchain-core 1.2.4 or later to ensure token counts and cost metrics continue to appear in your traces. Information on backports below.

2. Streaming Runnable Inputs

Issue: When using the .transform() or .atransform() methods on any runnable, input data may not appear in LangSmith traces on older LangChain versions.

What happened: Previously, these methods would create a trace before inputs were fully materialized, then attempt to update the inputs later. LangSmith now requires all inputs to be provided when the trace is initially created, not as a later update.

Action Required: Upgrade to langchain-core 1.2.4 or later to ensure inputs from .transform() and .atransform() are properly captured in traces. Information on backports below.

Recommended Action

Upgrade to the latest version of LangChain:

uv pip install --upgrade langchain-core

Affected Versions

• langchain-core 1.2.4: Versions prior to 1.2.4
• Backport Support: We've backported these fixes langchain-core 0.3.82

Timeline

• Now: Deprecation warnings will begin appearing in affected versions
• Soon: LangSmith will fully enforce these requirements - older versions will experience missing data in traces

Questions?

If you have questions or need assistance upgrading, please:

• Check our LangChain v1 migration guide or versioning documentation
• Open an issue on GitHub
• Contact us at support@langchain.dev

---

Note: This change only affects tracing to LangSmith. Core LangChain functionality is not impacted. However, upgrading ensures you maintain full observability of your LLM applications.

[Ahmed abdullah (D-engahmed)]

Is there a recommended way to verify that token usage and streaming inputs are being traced correctly after upgrading? For example, is there a minimal example or checklist we can use in CI to confirm traces contain the expected metadata?

[Mason Daugherty (mdrxy)]

For CI, use a dedicated LangSmith project and run a small end-to-end smoke test with explicit root run IDs. After calling wait_for_all_tracers(), fetch each run with Client.read_run(run_id, load_child_runs=True)—with a short retry for ingestion.

• For token usage, locate the child llm run and assert that run.metadata["usage_metadata"] contains positive input_tokens and output_tokens, and that run.total_tokens is populated.
• For transform()/atransform(), send a known value split across multiple input chunks and assert that the runnable’s inputs contains the fully materialized value rather than {"input": ""}.

Don’t use nonzero cost as the baseline assertion: cost also depends on model/provider metadata and workspace pricing configuration. We recommend the latest langchain-core; if remaining on the 1.2 line, use 1.2.17 or later.
Relevant docs: submitting traces before exit and querying runs through the SDK.