fix: preserve reasoning_tokens and cached_tokens in usage conversion

[pjb157]

Problem

When the Responses API adapter converts a Chat Completions response to a Responses API response, it constructs a ResponseUsage from the upstream Usage. The previous code hardcoded both detail fields to zero:

input_tokens_details: InputTokensDetails { cached_tokens: 0 },
output_tokens_details: OutputTokensDetails { reasoning_tokens: 0 },

This silently dropped the reasoning_tokens count reported by thinking models (DeepSeek-R1, Qwen3, etc.) and the cached_tokens count reported by providers that support prompt caching, even when the upstream had populated completion_tokens_details.reasoning_tokens and prompt_tokens_details.cached_tokens correctly.

The bug was duplicated across five call sites:

• adapter.rs — non-streaming to_responses_response_with_usage default usage
• streaming.rs — StreamingState usage capture
• handlers.rs × 3 — usage accumulation across tool-loop iterations, the unhandled-tools response path, and the final response path

The accumulation site was additionally dropping the entire completion_tokens_details and prompt_tokens_details JSON blobs (= None) when summing across iterations, so even the chat-completions passthrough for tool-loop responses lost the detail.

Fix

Adds a single shared helper in src/strict/mod.rs:

pub(crate) fn chat_usage_to_response_usage(
    u: &schemas::chat_completions::Usage,
) -> schemas::responses::ResponseUsage

It extracts reasoning_tokens from completion_tokens_details.reasoning_tokens and cached_tokens from prompt_tokens_details.cached_tokens (both stored as raw serde_json::Value on the chat completions side), and threads them through into the typed Responses API fields. All five call sites now use this helper.

The accumulation site in handlers.rs now preserves the latest detail JSON across iterations using .or() instead of dropping it.

Test plan

• cargo build — clean, no warnings
• cargo test — all 317 tests pass
• cargo clippy — no new warnings introduced

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes loss of token-usage detail when converting Chat Completions responses into Responses API responses in strict-mode, ensuring reasoning_tokens (thinking models) and cached_tokens (prompt caching) are preserved end-to-end.

Changes:

• Added a shared chat_usage_to_response_usage helper to extract cached_tokens / reasoning_tokens from chat usage detail JSON into typed ResponseUsage.
• Replaced hardcoded-zero usage conversions in adapter, streaming, and handler paths to use the shared helper.
• Updated tool-loop usage accumulation to preserve the latest *_tokens_details JSON instead of dropping it entirely.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

File	Description
src/strict/mod.rs	Introduces shared chat→responses usage conversion helper that extracts cached_tokens and reasoning_tokens.
src/strict/adapter.rs	Uses shared helper when mapping ChatCompletionResponse.usage into ResponsesResponse.usage.
src/strict/handlers.rs	Uses shared helper for aggregate usage; preserves *_tokens_details JSON during accumulation.
src/strict/streaming.rs	Uses shared helper to capture usage from final streaming chunk.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated no new comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.