feat(langchain): Support cache/reasoning token metrics based on Langchain UsageMetadata

[flippingbitss]

Summary

Cache read/Cache write tokens and output reasoning tokens weren't being picked up by the Langchain integration.

• We now include these additional usage metrics and get the usages from Langchain's provider-agnostic UsageMetadata shape instead.
• Ran the scripts/generate-test-files.sh to update Langchain version as the older latest version wasn't returning cache_write metric.

Description

As described by my initial approach here, this PR adds a few new token usage metrics cache_read/cache_creation/reasoning/audio and populates the following additional span tags.

• gen_ai.usage.input_tokens.cached
• gen_ai.usage.input_tokens.cache_write
• gen_ai.usage.output_tokens.reasoning

Note: audio tokens are available and read from UsageMetadata but Sentry doesn't have any span tags for these yet, so no span tags are populated for these.

Testing

Add new asserts in Langchain v1+ only tests (test_langchain_create_agent, test_tool_execution_span) as suggested by @alexander-alderman-webb

Ran unit tests with TESTPATH=tests/integrations/langchain/test_langchain.py uv run tox -e <ENV> for following envs:

• py3.14-langchain-base-v0.1.20
• py3.14-langchain-base-latest

Issues

Fixes #6799
Fixes PY-2606

Related

OTel's Langchain Integration's set_chat_response_usage impl

Test Evidence

1. py3.14-langchain-base-v0.1.20

2. py3.14-langchain-base-latest

[alexander-alderman-webb]

Re-generating the test matrix has pulled in some test failures.
Could you revert changes to tox.ini, scripts/populate_tox/releases.jsonl, and scripts/populate_tox/package_dependencies.jsonl?

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want reviews to match your repository better? Bugbot Learning can learn team-specific rules from PR activity. A team admin can enable Learning in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 12997a4. Configure here.}

[cursor]

Token counts inflated for multiple choices

Medium Severity

_record_token_usage now always prefers summing usage from response.generations before considering response-level llm_output usage. Providers typically attach the same request-level usage_metadata to every choice, so calls with n > 1 can report multiplied input, output, total, cache, and reasoning tokens. Previously, present llm_output token usage was used once and avoided that inflation.

Additional Locations (1)

• sentry_sdk/integrations/langchain.py#L786-L793

 

^{Reviewed by Cursor Bugbot for commit 12997a4. Configure here.}

[alexander-alderman-webb]

Going to review this now.

[alexander-alderman-webb]

Just an update that I found a bug handling multi-candidate responses. We'll want to fix that first.
I'll take over your work and add you as a co-author.