Feature hasn't been suggested before.
Describe the enhancement you want to request
vLLM renamed reasoning_content to reasoning on their chat completions API (vllm-project/vllm#27755) and dropped reasoning_content on input entirely. The interleaved config only allows "reasoning_content" and "reasoning_details" as field values, so there's no way to target the field name vLLM reads.
Result: on self-hosted vLLM with reasoning models (MiniMax-M2, DeepSeek, etc.), reasoning is sent back as reasoning_content on assistant messages but vLLM silently ignores it. The model never sees its prior thinking in multi-turn/tool-call chains.
Fix: add "reasoning" to the zod enum in models.ts:32 and provider.ts:807. The transform already handles it generically via [field].
Happy to open a PR for this.
Feature hasn't been suggested before.
Describe the enhancement you want to request
vLLM renamed
reasoning_contenttoreasoningon their chat completions API (vllm-project/vllm#27755) and droppedreasoning_contenton input entirely. The interleaved config only allows"reasoning_content"and"reasoning_details"as field values, so there's no way to target the field name vLLM reads.Result: on self-hosted vLLM with reasoning models (MiniMax-M2, DeepSeek, etc.), reasoning is sent back as
reasoning_contenton assistant messages but vLLM silently ignores it. The model never sees its prior thinking in multi-turn/tool-call chains.Fix: add
"reasoning"to the zod enum inmodels.ts:32andprovider.ts:807. The transform already handles it generically via[field].Happy to open a PR for this.