400 "reasoning_content in the thinking mode must be passed back to the API" — DeepSeek thinking models (deepseek-flash) fail on multi-turn (v1.0.87-0)

[jeremkief]

🏷️ Discussion Type

Bug

💬 Feature/Topic Area

Copilot Agent Mode

Body

Describe the bug

When using a DeepSeek thinking model (deepseek-flash) against the DeepSeek API, the first request in a conversation can succeed, but every subsequent request in the same conversation fails with:

400 The reasoning_content in the thinking mode must be passed back to the API.

The DeepSeek API requires the reasoning_content field returned in assistant messages to be echoed back verbatim in the assistant message of subsequent requests once thinking mode has been used. Copilot's provider adapter appears to drop this field when persisting assistant turns, so the follow-up request is rejected with HTTP 400. It is triggered more easily by agent flows that issue several internal/tool requests within one turn.

Affected version

• Copilot CLI: 1.0.87-0 (GitHub Copilot desktop app 1.1.23)
• OS: Windows 11 Pro (10.0.26200)
• Model: deepseek-flash
• Provider: DeepSeek API (OpenAI-compatible, https://api.deepseek.com)
• Thinking/reasoning mode: enabled

Steps to reproduce the behavior

1. Configure Copilot to use the DeepSeek API with deepseek-flash (e.g. COPILOT_PROVIDER_BASE_URL=https://api.deepseek.com, COPILOT_PROVIDER_TYPE=openai, COPILOT_MODEL=deepseek-flash).
2. Enable thinking/reasoning mode.
3. Start a conversation and send a prompt that triggers thinking mode.
4. Continue the same conversation with a follow-up message.
5. Copilot sends the next request and it fails with the 400 error above.

Expected behavior

Copilot preserves the reasoning_content returned by DeepSeek and includes it in the assistant message of subsequent requests, as the DeepSeek API requires.

Actual behavior

reasoning_content is omitted from the conversation history sent upstream, so the DeepSeek API rejects subsequent turns with HTTP 400 and the conversation cannot proceed.

Additional context

• Related: BUG: Pinned discussions has scroll bars #2995 ("Can't use DeepSeek API") — same 400 with DeepSeek + OpenAI-compatible provider; closed as resolved in v1.0.88. This report is from v1.0.87-0.
• Same root cause on another Copilot surface: community discussion Copilot Chat + OpenRouter DeepSeek-V4-Flash: reasoning_content not passed back in multi-turn conversations (400 error) #193953 (VS Code Copilot Chat + OpenRouter deepseek/deepseek-v4-flash).
• Possibly related open issues: Mobile version of GitHub.com doesn't allow us to see who reacted to a comment #4196, creating a metal cluster input file in lanterns #3195, Large organization member searching issue #4946.
• DeepSeek's docs state that when thinking mode is used with tool calls, reasoning_content from previous assistant turns must be passed back in subsequent requests.
• Suggested fix: preserve extra/unknown fields on assistant messages when persisting conversation history, and include reasoning_content in the assistant message objects sent upstream for DeepSeek thinking models.

Thank you.

[github-actions[bot]]

💬 Your Product Feedback Has Been Submitted 🎉

Thank you for taking the time to share your insights with us! Your feedback is invaluable as we build a better GitHub experience for all our users.

Here's what you can expect moving forward ⏩

• Your input will be carefully reviewed and cataloged by members of our product teams. 
  • Due to the high volume of submissions, we may not always be able to provide individual responses.
  • Rest assured, your feedback will help chart our course for product improvements.
• Other users may engage with your post, sharing their own perspectives or experiences. 
• GitHub staff may reach out for further clarification or insight. 
  • We may 'Answer' your discussion if there is a current solution, workaround, or roadmap/changelog post related to the feedback.

Where to look to see what's shipping 👀

• Read the Changelog for real-time updates on the latest GitHub features, enhancements, and calls for feedback.
• Explore our Product Roadmap, which details upcoming major releases and initiatives.

What you can do in the meantime 💻

• Upvote and comment on other user feedback Discussions that resonate with you.
• Add more information at any point! Useful details include: use cases, relevant labels, desired outcomes, and any accompanying screenshots.

As a member of the GitHub community, your participation is essential. While we can't promise that every suggestion will be implemented, we want to emphasize that your feedback is instrumental in guiding our decisions and priorities.

Thank you once again for your contribution to making GitHub even better! We're grateful for your ongoing support and collaboration in shaping the future of our platform. ⭐

[KasraK2K]

Hit the same wall with deepseek-flash through an OpenAI-compatible endpoint, not on Copilot specifically but the error is identical, so this is worth confirming as a general pattern rather than something unique to your setup.

What's happening is that DeepSeek's reasoning models (same story with most "thinking mode" models, OpenRouter's deepseek-v4-flash included) return reasoning_content as a separate field alongside the normal content in the assistant message. Most OpenAI-compatible client code only persists role and content when it builds conversation history, because that's all a non-reasoning chat completion ever had. The reasoning field gets silently dropped the moment the SDK serializes the assistant turn back into history, and the next request comes back with that 400 because the API can tell the trailing reasoning context is missing.

Until the fix lands in Copilot's provider adapter, a couple of things that work as a stopgap if you're scripting around this or building your own adapter:

• If you're not actually relying on the model seeing its own prior reasoning for context (most people aren't, since the point of thinking mode is usually just quality of the final answer), you can sidestep the whole problem by starting a fresh conversation per turn instead of carrying history forward, obviously only workable for stateless single-turn tasks.
• • If you do need multi-turn with thinking preserved, the fix in your own adapter code is literally just copying reasoning_content onto the stored assistant message object alongside content when you persist it, then echoing it back unmodified in the next request's messages array. DeepSeek's API doesn't want you to regenerate or summarize it, just pass through the exact string you got back.
• • Worth checking whether disabling thinking mode mid-conversation on a retry is a usable fallback for your use case, since the requirement only kicks in once reasoning has been used in that conversation.
    Given the linked discussion Copilot Chat + OpenRouter DeepSeek-V4-Flash: reasoning_content not passed back in multi-turn conversations (400 error) #193953 shows the same bug on a completely different Copilot surface (VS Code Chat + OpenRouter) and BUG: Pinned discussions has scroll bars #2995 was supposedly fixed in v1.0.88, it sounds like the fix didn't cover every provider adapter path, just whichever one was tested for that release. Might be worth linking both of those in your report explicitly if you haven't, since it strengthens the case that this is a shared serialization bug rather than three unrelated ones.