feat(langchain): report cached and reasoning token usage in Langchain/Langgraph

[joaquinhuigomez]

Description

The Langchain and Langgraph integrations only recorded input/output/total token counts, dropping cached input tokens and reasoning output tokens — even though SPANDATA defines both and the OpenAI, OpenAI Agents, and Google GenAI integrations already report them. This makes cache-discount and reasoning-token cost calculations inaccurate for Langchain/Langgraph apps.

This extracts cache_read/cached_tokens and reasoning/reasoning_tokens from both the LangChain usage_metadata detail shape (input_token_details/output_token_details) and OpenAI-style details dicts (prompt_tokens_details/completion_tokens_details), and sets gen_ai.usage.input_tokens.cached / gen_ai.usage.output_tokens.reasoning in both integrations. Tests cover both detail shapes, span data emission, and the Langgraph accumulation loop (all three fail on master and pass with this change); ruff clean.

Issues

• resolves: Langchain and Langgraph don't report cached/reasoning token usage #6799
• resolves: PY-2606

[sentry]

Bug: The use of the or operator incorrectly handles legitimate 0 values for token counts, causing them to be evaluated as None and dropped from span data.
_{Severity: MEDIUM}

Suggested Fix

Refactor the logic to avoid using the or operator for fallback when 0 is a valid value. Instead, explicitly check if the first value is None before attempting to retrieve the second value. This ensures that a 0 from the primary key is preserved.

Prompt for AI Agent

Review the code at the location below. A potential bug has been identified by an AI
agent. Verify if this is a real issue. If it is, propose a fix; if not, explain why it's
not valid.

Location: sentry_sdk/integrations/langchain.py#L752-L754

Potential issue: In the `_extract_tokens` function, the use of the `or` operator to
select between two potential keys for token counts (e.g., `cache_read` or
`cached_tokens`) leads to incorrect behavior when a valid token count is `0`. Because
`0` is a falsy value in Python, the expression proceeds to evaluate the right-hand side.
If the alternative key does not exist, the result becomes `None`. This causes legitimate
zero-token counts for cached or reasoning tokens to be silently dropped instead of being
reported in span data. While this has minimal impact on cost calculations, it is a
correctness issue that prevents accurate reporting of API responses.

Also affects:

• sentry_sdk/integrations/langchain.py:761~763
• sentry_sdk/integrations/langchain.py:737~738
• sentry_sdk/integrations/langchain.py:742~743

_{Did we get this right? 👍 / 👎 to inform future reviews.}

[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 cb26c51. Configure here.}

[cursor]

Langgraph misses LangChain token shapes

Medium Severity

_set_usage_data only reads cached_tokens and reasoning_tokens from OpenAI-style prompt_tokens_details / completion_tokens_details. It never checks LangChain’s input_token_details / output_token_details or the cache_read / reasoning keys that _extract_tokens handles in the Langchain integration, so cached and reasoning usage can be omitted on Langgraph spans.

 

^{Reviewed by Cursor Bugbot for commit cb26c51. Configure here.}