[Lucas Gomide (lucasgomide)]

Note

Medium Risk
Updates core token-usage accounting and the UsageMetrics schema across multiple LLM providers, which could affect downstream reporting/analytics and any consumers expecting the previous metrics shape. No auth/data-path behavior changes beyond usage extraction and aggregation.

Overview
Extends LLM usage tracking to record reasoning_tokens (thinking/reasoning models) and cache_creation_tokens (cache write costs), and propagates these through BaseLLM._token_usage and UsageMetrics aggregation.

Enhances provider-specific usage extraction to populate these fields and improve cached-token detection (OpenAI Chat + Responses API, Azure, Gemini, Anthropic cache writes, and Bedrock cache-read keys), and adds/updates tests to cover the new metrics and default/absence behavior.

^{Reviewed by Cursor Bugbot for commit 4fb3b2b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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.}

^{Reviewed by Cursor Bugbot for commit 981b1ba8290518b7553c2983ebaf6b68cf85f754. Configure here.}