Custom inference DeepSeek thinking models need reasoning_content round-trip

[thesammykins]

Problem

Using DeepSeek thinking models through an OpenAI-compatible Custom Inference endpoint can fail on follow-up requests because reasoning_content is not being round-tripped in conversation history.

Preview-build error observed with the OpenCode endpoint:

Response stream finished unexpectedly with internal error: POST "https://opencode.ai/zen/go/v1/chat/completions": 400 Bad Request {"message":"Error from provider (DeepSeek): The `reasoning_content` in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}

Local endpoint evidence

Tested against https://opencode.ai/zen/go/v1/ without including credentials in logs or this issue.

• GET /models lists DeepSeek thinking-capable models including deepseek-v4-pro and deepseek-v4-flash.
• Direct POST /chat/completions probes against listed DeepSeek models worked.
• DeepSeek responses included reasoning_content in assistant messages / stream deltas.

Provider docs

DeepSeek Thinking Mode docs describe that thinking-mode tool-call turns must pass reasoning_content back to the API on the next request:

https://api-docs.deepseek.com/guides/thinking_mode

Likely scope

This looks like provider conversation/history handling in the backend/provider adapter path, not a client model-list or Custom Inference modal contract issue.

The model autofill proposal can import DeepSeek model IDs correctly, but successful multi-turn use of these models likely requires preserving and replaying reasoning_content when constructing follow-up chat-completion requests for DeepSeek-compatible providers.

Acceptance criteria

• DeepSeek thinking-mode responses preserve reasoning_content in conversation state when required by the provider.
• Follow-up requests for DeepSeek-compatible models include the required reasoning_content fields in prior assistant messages.
• The fix is scoped so providers that do not support or require reasoning_content are unaffected.
• Add regression coverage for a multi-turn DeepSeek-compatible response that includes reasoning_content followed by a request that must replay it.

Related issue

Model autofill for Custom Inference endpoints is tracked separately in #11586.

[warp-for-oss]

@thesammykins

Here's what I found while triaging this issue:

DeepSeek's docs confirm that thinking-mode tool-call turns must replay reasoning_content on later requests. This looks distinct from #11586's model-import work and likely needs provider-history handling rather than Custom Inference UI changes.

Maintainer details

I concluded that custom Inference follow-up requests to DeepSeek thinking models appear to fail because provider-specific reasoning_content is not preserved and replayed in chat history..

Triage summary

This appears to be a provider-adapter bug for OpenAI-compatible Custom Inference models that expose DeepSeek thinking-mode responses. The report includes an exact 400 from the OpenCode endpoint showing DeepSeek rejected a follow-up request because reasoning_content from prior assistant/tool-call history was not replayed.

Evidence

• DeepSeek's thinking-mode docs state that tool-call turns must pass the assistant message's reasoning_content back on subsequent requests.
• The current client-side custom model path sends endpoint/model metadata through custom_model_providers and has generic reasoning display support, but this checkout has no reasoning_content field or DeepSeek-specific round-trip handling.
• That makes the most likely failing layer the server/provider adapter that reconstructs OpenAI-compatible chat-completion history for custom endpoints, rather than the Custom Inference modal/model-list UI.

Reproducibility

Medium. The report has a concrete endpoint, model family, provider error, and documentation backing; it still depends on a configured Custom Inference endpoint and a follow-up conversation shape that triggers DeepSeek thinking-mode tool-call history.

Duplicate check

I searched all open issues and recently closed issues. #11586 and #11514 are related to model discovery/import for Custom Inference endpoints, but they do not cover the reasoning_content round-trip failure, and fewer than two likely duplicates were found.

This is my automated analysis and may be incorrect. A maintainer will verify the details.

Powered by Oz

[RajeevRKC]

Confirming this is blocking multi-turn DeepSeek conversations with Thinking Mode. The
easoning_content field must be preserved in conversation history for subsequent API requests to work. This is a documented requirement in the DeepSeek Thinking Mode API docs.

[DennisZhangOiler]

I met same issue and I wanna fix it, but I am not familiar with the repo, can anyone guide me to the place where I should begin?