[bot] Anthropic usage extractor reads non-existent `reasoning_tokens` field instead of `output_tokens_details.thinking_tokens`

[braintrust-bot]

<!-- provider-gap-audit: anthropic-thinking-tokens-field-mismatch -->

Summary

extract_anthropic_usage() in src/extractors.rs looks for a field literally named reasoning_tokens — both as a flat usage.reasoning_tokens field and nested under output_tokens_details.reasoning_tokens — to populate reasoning/thinking token counts for Anthropic extended-thinking responses. Neither field exists in Anthropic's real API. The actual field, per Anthropic's own documentation, is usage.output_tokens_details.thinking_tokens. As a result, completion_reasoning_tokens is always None for genuine Anthropic extended-thinking API responses — this is not a partial-detail gap, it's a complete miss caused by a wrong field name.

What is wrong

src/extractors.rs lines 135-138 read a flat top-level usage.reasoning_tokens field that Anthropic never emits.

src/extractors.rs lines 140-156 build completion_tokens_details from usage.get("output_tokens_details"), and at line 150 read details.get("reasoning_tokens") — again the wrong key. The correct key per Anthropic's docs is thinking_tokens.

This mistaken assumption is also baked into the test suite: the extracts_anthropic_cache_metrics test (src/extractors.rs, around line 299) asserts a "reasoning_tokens": 12 fixture value and expects completion_reasoning_tokens() == Some(12) — encoding the incorrect field name as expected behavior rather than testing against Anthropic's real response shape.

Braintrust docs status

unclear — checked https://www.braintrust.dev/docs/integrations/ai-providers/anthropic (2026-09-02). The page lists thinking as a captured request parameter (Python/Go/Ruby sections), but none of the documented usage metrics tables include a thinking_tokens or reasoning_tokens field for any language SDK. Braintrust does not explicitly state whether extended-thinking output token counts are captured as a metric.

Upstream sources

• Anthropic extended thinking guide — exact quote: "monitor the usage.output_tokens_details.thinking_tokens field in the response... When streaming, this breakdown appears only on the final message_delta event." https://platform.claude.com/docs/en/build-with-claude/extended-thinking
• Anthropic Messages API reference (usage object shape): https://platform.claude.com/docs/en/api/messages

Local files inspected

• src/extractors.rs lines 92-188 (extract_anthropic_usage), specifically lines 135-138 and 140-156
• src/extractors.rs around line 299 (extracts_anthropic_cache_metrics test fixture asserting the incorrect field name)
• src/types.rs (CompletionTokensDetails, UsageMetrics structs)
• Full codebase grep for thinking_tokens — zero results

Fix sketch

Read thinking_tokens as the primary (or aliased) key when parsing Anthropic's output_tokens_details, e.g. details.get("thinking_tokens").or_else(|| details.get("reasoning_tokens")).