Is this a bug? Positron Assistant error with DeepSeek reasoning models

[Vara326]

Hello👋, the DeepSeek models have recently been updated with "deepseek-v4-pro" and "deepseek-v4-flash" models. Additionally, they enable the thinking mode by default. I typically use DeepSeek's official API under the "OpenAI-compatible Custom providers" option to call them, which results in the following error:

If I did not upload the image correctly, the error message is roughly "The reasoning_content in the thinking mode must be passed back to the API."

I personally understand that DeepSeek requires me to pass back all the content generated by the model in the previous round (including the thought process).

I tried using a small tool to convert reasoning_content into tags and splice them into the content, but it still couldn't resolve the error:

(In the image, the deepseek-v4-flash-max is still deepseek-v4-flash—the max suffix simply means it converts reasoning_content into tags and appends them to the content.)

I haven't been able to solve this problem. I'm wondering if this is a bug?

Of course, after I disable the thinking mode through some settings, this error no longer appears.

This issue didn't just appear recently. A few months ago, DeepSeek’s older models were divided into deepseek-chat and deepseek-reasoner. If I used deepseek-reasoner, it would give the same error, but that was tolerable because I could use deepseek-chat without much setup. However, since DeepSeek’s recent update, both models it offers now have the thinking mode enabled by default, which is very inconvenient. I have to manually disable the thinking mode for both models to use them. The problem is, when I use Alibaba Cloud’s Kimi K2.5 and Qwen 3.6 Plus, they don't report errors, even though these two models also enable thinking mode by default.

I’d like to know if these issues are due to my improper settings or if Positron Assistant hasn’t adapted to DeepSeek’s official API. If it hasn’t, I hope Positron could consider adding support for it. DeepSeek is an excellent model, and for Chinese users, it’s the easiest to use and the most cost-effective model. With all of DeepSeek’s models now defaulting to thinking mode, a user who is not tech-savvy may never figure out how to disable it to resolve the errors. I think the best solution would be for Positron to simply support DeepSeek’s thinking mode entirely.

I don’t use GitHub very often, and I’m not very familiar with programming. If I’ve said something wrong or posted this in the wrong place, please let me know. Thank you all.

This information was translated into English by AI.

[kv9898]

Same problem here - besides, one cannot use DeepSeek through the anthropic API as its API end point does not support model fetching. The agent mode of DeepSeek also used to run into the problem of mismatched lines (e.g. it was trying to modify line 7 but ended up modifying line 9).

[juliasilge]

@Vara326 and @kv9898, do you think the problems you are experiencing are captured in #13211? Would you be able to add the info you have on the error messages you are seeing?

[Vara326]

@juliasilge，hello, thank you for your reply!

I noticed that there was no upload error in #13211 , so I'd like to provide more details:

When using DeepSeek models (e.g., deepseek-v4-flash and deepseek-v4-pro) through a custom provider compatible with OpenAI, asking them questions will fail and always produce the following error.
[deepseek-v4-flash] Error in chat response: {"error":{"message":"The reasoning_content in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":"","code":"invalid_request_error"}}

I ran another test. Strangely enough, when I said simple things to him like "hello," he answered correctly. But when I asked him to "check my project library," he executed a small step and then displayed [deepseek-v4-flash] Error in chat response: {"error":{"message":"The reasoning_content in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":"","code":"invalid_request_error"}}.

(In the figure, my second sentence means "Check my project library." This conversation took place during my self-study project of R4DS, but the same situation occurs regardless of which project I switch to, so I don't think it is a settings issue with any particular project. The model used is deepseek-v4-flash, and it has thinking mode enabled by default.)

By contrast, as shown in the image below, when I turned off the thinking mode of DeepSeek-v4-flash, it was able to respond normally (I added thinking={"type":"disabled"} via parameter overrides using new-api)

(The deepseek-v4-flash-chat in the figure refers to the DeepSeek-v4-flash with the thinking mode turned off using the method described above.)

This plugin is the positron assistant, and the AI mode is agent mode.

For another new plugin that seems to be in beta, called posit assistant (I'm not quite sure), the exact same situation occurs:

(In the image, when I said "hello" to it, it responded normally, but when I asked it to "check my project library," it reported an error: [An error occurred: [deepseek-v4-flash] Error in chat response: {"error":{"message":"The reasoning_content in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":"","code":"invalid_request_error"}}])

Similarly, when I switch to "deepseek-v4-flash-chat", it doesn't report an error:

The above content uses DeepSeek-V4-Flash as an example, and the same situation occurs when using DeepSeek-V4-Pro.

My version information is:

Version: Positron Version: 2026.04.1 (system setup) build 10
Code - OSS Version: 1.109.2
Commit: 5dfc92424b4381ca7139f3f0a122a65abc2a9383
Date: 2026-04-14T01:26:22.333Z
Submission: 39.3.0
Date: 142.0.7444.265
Electron: 22.21.1
ElectronBuildId: 14.2.231.22-electron.0
Chromium: Windows_NT x64 10.0.26200
Node.js: undefined
V8: undefined
OS: undefined

The version of Posit Assistant is 0.3.3. I couldn't find any version information for Positron Assistant. Sorry.

That's roughly the situation. If there's anything else you need me to do, please feel free to ask. I will get back to you as soon as possible! Thank you!

[juliasilge]

Thank you!

[kv9898]

@Vara326 and @kv9898, do you think the problems you are experiencing are captured in #13211? Would you be able to add the info you have on the error messages you are seeing?

Yes. It's the same as #13211

[kv9898]

Created a PR to address this: #13289

[richardchen874-sys]

This looks like the same provider-state issue that other DeepSeek reasoning-model integrations are running into.

For DeepSeek V4 Flash / Pro with thinking mode enabled, reasoning_content is not just display-only metadata. In multi-turn or agent-mode workflows, the assistant state may need to preserve it and send it back on the next request. That explains why a simple “hello” works, but “check my project library” fails after the assistant starts doing agent-style steps.

A safer implementation might need to:

• preserve reasoning_content on assistant messages when DeepSeek returns it
• keep it hidden from normal visible chat content
• send it back only for DeepSeek reasoning / thinking models that require it
• avoid applying this behavior to other OpenAI-compatible providers
• test both normal chat and agent-mode tool workflows
• test thinking-disabled DeepSeek models to avoid regressions

The line-mismatch issue mentioned above may be a separate agent-editing problem, but the reasoning_content must be passed back error looks specifically like conversation history serialization for DeepSeek thinking models.

This is a good example of why OpenAI-compatible providers need workflow-level testing, not only a single chat-completion test.

[melissa-barca]

We've added a dedicated DeepSeek provider for Posit Assistant in our 2026.07 daily builds. You can enabled it by setting assistant.provider.deepseek.enabled in your settings.json. This should handle reasoning_content for you unlike the custom provider.

[richardchen874-sys]

Thanks for the update — that sounds like the right direction.

A dedicated DeepSeek provider makes much more sense here than relying on the generic OpenAI-compatible custom provider, especially for reasoning models where reasoning_content needs provider-specific handling.

That should also reduce confusion for users who can get normal chat working, but then hit errors once agent mode starts preserving multi-turn state or tool workflow context.

I’ll keep an eye on the 2026.07 daily builds. This is a good example of why reasoning-model support often needs provider-level integration, not only a custom base URL.

[wangbinzjcc]

I’m thrilled to see that Posit Assistant now supports direct integration with DeepSeek. This is so convenient – I’m absolutely delighted!