OpenAI Chat Completions: tool rounds drop the model's reasoning #326

OpenAILanguageModel's Chat Completions tool loop drops the model's reasoning between tool rounds. Thinking models return it alongside their tool calls, in reasoning_content (DeepSeek's API) or reasoning and reasoning_details (OpenRouter, Vercel AI Gateway). The next request in the same response rebuilds the assistant message from the content and tool calls only, so the reasoning that planned those calls is lost.

Why it matters

• DeepSeek requires it. Its thinking mode docs say reasoning_content must be passed back on tool-call turns, and its API returns 400 when it isn't. Other clients have hit this: NousResearch/hermes-agent#17400, langchain-ai/langchain#39370, Kilo-Org/kilocode#9501.
• Through a gateway it fails silently. The Vercel AI Gateway accepts the requests, so nothing errors. But each round, the model sees its earlier tool calls without the reasoning behind them.

Reproduction

Run a session with tools against deepseek/deepseek-v4.1-flash-fast or zai/glm-5.3-fast on the Vercel AI Gateway (https://ai-gateway.vercel.sh/v1), and record the request bodies. Each response's choices[0].message has reasoning and reasoning_details next to tool_calls. In the next request, the matching assistant message has only content and tool_calls. Across one 8-round response, 0 of the 8 assistant messages carried the reasoning back.

PR #290 added reasoning transcript entries and replay for Anthropic. Chat Completions doesn't read or send these fields yet (checked on main at 80d50b0).

A possible fix

In the tool loop, decode the three fields from the response message and send them back unchanged on that round's assistant message. Keeping them within the one response matches DeepSeek's rule, so the transcript doesn't need to change. I did this for the nonstreaming loop in a fork, with a test: gtokman/AnyLanguageModel@daf184b. The commit is on 0.16.0. The streaming loop would need the same change, gathering the reasoning deltas.

A fuller version would store the reasoning as Transcript.Reasoning entries, as #290 does for Anthropic, so it could also be replayed after a transcript is restored. I'm happy to open a PR for either approach.