Can´t use DeepSeek API

[raffaeler]

Describe the bug

When I run the following command:

export COPILOT_PROVIDER_BASE_URL="https://api.deepseek.com" \
export COPILOT_PROVIDER_API_KEY="...." \
export COPILOT_PROVIDER_TYPE="openai" \
export COPILOT_MODEL="deepseek-v4-pro" \
copilot

When I make a question to the copilot, I get this answer back:

● Request failed (transient_bad_request). Retrying...
● Request failed (transient_bad_request). Retrying...
✗ 400 The `reasoning_content` in the thinking mode must be passed back to the API. 

I confirm that the DeepSeek API needs the entire reasoning_content field to be returned back to the API every time. In my own code I had the same problem and could rapidly fix the issue by sending back the the entire conversation.

Affected version

No response

Steps to reproduce the behavior

I run the following command:

export COPILOT_PROVIDER_BASE_URL="https://api.deepseek.com" \
export COPILOT_PROVIDER_API_KEY="...." \
export COPILOT_PROVIDER_TYPE="openai" \
export COPILOT_MODEL="deepseek-v4-pro" \
copilot

Ensure to have a valid api key for the DeepSeek endpoint

Expected behavior

I expect the request to be fulfilled.

Additional context

No response

[raffaeler]

I created this Proxy as a workaround for this bug:
https://github.com/raffaeler/mindthegap

If you like it, please give it a star.

[abchatra]

SxS team evaluating DeepSeek is also hitting the same issue https://github.com/github/cli-side-by-side-launcher

[Aquaakuma]

Using DeepSeek's Anthropic-compatible API can resolve this issue, but it may still be triggered under certain circumstances. For example, during a long conversation, if you run commands like copilot CLI or /plugins to check some information, then wait for a while before continuing the conversation, the issue might occur. Additionally, the DeepSeek API included in the OpenCode Go subscription plan also has usability problems, as it uses an OpenAI-compatible API.

[raffaeler]

@Aquaakuma I tried but unfortunately it does not resolve.
I could fully resolve using my own proxy (see my comment above) which also has the bonus to show the thinking tokens.

I would expect the Copilot CLI to give a better user experience in adding, editing and switching providers/models. At the moment I am using a tool called cpm.

[Aquaakuma]

@Aquaakuma I tried but unfortunately it does not resolve. I could fully resolve using my own proxy (see my comment above) which also has the bonus to show the thinking tokens.

I would expect the Copilot CLI to give a better user experience in adding, editing and switching providers/models. At the moment I am using a tool called cpm.

If you are using the DeepSeek Official API, you can try this method: [接入 GitHub Copilot CLI]（https://api-docs.deepseek.com/zh-cn/quick_start/agent_integrations/copilot_cli）. but it may still be triggered in some times.

[raffaeler]

If you are using the DeepSeek Official API, you can try this method: [接入 GitHub Copilot CLI]（https://api-docs.deepseek.com/zh-cn/quick_start/agent_integrations/copilot_cli）. but it may still be triggered in some times.

@Aquaakuma No, it does not help as I wrote in my previous comment.
I tried the anthropic mode but GitHub Copilot CLI still complains about the reasoning_content field.

The DeepSeek API expects the reasoning_content field is always returned to the API after the first turn.
My proxy moves the reasoning_content into the main content and then puts it back in the reasoning_content at the following turn.
This works with any client, not just with the CLI.

[Aquaakuma]

If you are using the DeepSeek Official API, you can try this method: [接入 GitHub Copilot CLI]（https://api-docs.deepseek.com/zh-cn/quick_start/agent_integrations/copilot_cli）. but it may still be triggered in some times.

@Aquaakuma No, it does not help as I wrote in my previous comment. I tried the anthropic mode but GitHub Copilot CLI still complains about the reasoning_content field.

The DeepSeek API expects the reasoning_content field is always returned to the API after the first turn. My proxy moves the reasoning_content into the main content and then puts it back in the reasoning_content at the following turn. This works with any client, not just with the CLI.

This implementation approach is not very user-friendly. The ideal solution is to fix it in the Copilot CLI source code, or perhaps to execute scripts during the conversation via hooks, ensuring that it is invoked for every conversation.

[raffaeler]

This implementation approach is not very user-friendly. The ideal solution is to fix it in the Copilot CLI source code, or perhaps to execute scripts during the conversation via hooks, ensuring that it is invoked for every conversation.

@Aquaakuma I opened this issue to solicit the change in GitHub Copilot CLI, but the same problem also affects Visual Studio Code and Visual Studio 2026. This means that currently the proxy is the best workaround.
I am not sure the hooks can get to the reasoning_content. Most probably they do not have access to it. Give a try ;-)

[copilot-cli-bot]

This appears to be resolved as of GitHub Copilot CLI v1.0.88. Please update to that release or a newer stable version. We're closing this as completed. If it still happens, please comment with your version and reproduction steps.