[BUG] Switching deepseek providers during a conversation will result in an error

[anyiba]

Problem (one or two sentences)

DeepSeek completion error: 400 Missing reasoning_content field in the assistant message at message index 3. For more information, please refer to https://api-docs.deepseek.com/guides/thinking_mode#tool-calls

Context (who is affected and when)

An error occurred when switching suppliers to deepseek during the conversation

Reproduction steps

1. Start the conversation with another provider, or start the conversation with the deepseek provider first
2. Switching to another supplier other than deepseek during the process
3. An error occurred when switching back to the deepseek provider:
   DeepSeek completion error: 400 Missing reasoning_content field in the assistant message at message index 3. For more information, please refer to https://api-docs.deepseek.com/guides/thinking_mode#tool-calls

Expected result

It won't report errors and has context memory

Actual result

error：DeepSeek completion error: 400 Missing reasoning_content field in the assistant message at message index 3. For more information, please refer to https://api-docs.deepseek.com/guides/thinking_mode#tool-calls

Variations tried (optional)

No response

App Version

v3.36.11

API Provider (optional)

DeepSeek

Model Used (optional)

deepseek-reansoner

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

DeepSeek completion error: 400 Missing `reasoning_content` field in the assistant message at message index 3. For more information, please refer to https://api-docs.deepseek.com/guides/thinking_mode#tool-calls

[roomote]

I've analyzed this issue and here's my proposed implementation plan:

Problem Analysis

When switching providers during a conversation (e.g., from OpenAI to DeepSeek reasoner), the DeepSeek API returns a 400 error: "Missing reasoning_content field in the assistant message at message index 3."

This happens because:

1. Messages from other providers in the conversation history do not have a reasoning_content field
2. The DeepSeek API in thinking mode requires ALL assistant messages to have a reasoning_content field (even if empty)
3. The current convertToR1Format() function preserves existing reasoning_content but does not add it to messages that lack it

Proposed Solution

Modify the convertToR1Format() function in src/api/transform/r1-format.ts to accept an option for adding empty reasoning_content to assistant messages that do not have reasoning content. This ensures compatibility with DeepSeek's thinking mode when switching providers mid-conversation.

Changes:

1. Add a new option to convertToR1Format() - Add addEmptyReasoning?: boolean to the options object
2. Apply empty reasoning_content - When this option is true, add reasoning_content: "" to all assistant messages that do not already have reasoning content
3. Enable the option in DeepSeek handler - In src/api/providers/deepseek.ts, pass addEmptyReasoning: true when converting messages for the deepseek-reasoner model
4. Add tests - Add test cases to verify the fix works correctly

Code Changes:

src/api/transform/r1-format.ts:

• Add addEmptyReasoning?: boolean option
• When building assistant messages, if addEmptyReasoning is true and no reasoning exists, add reasoning_content: ""

src/api/providers/deepseek.ts:

• Pass addEmptyReasoning: isThinkingModel to convertToR1Format()

This approach ensures backward compatibility since the option defaults to false, and only affects DeepSeek reasoner models that require this field.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[anyiba]

@roomote approve

[roomote]

Thanks for the approval! I've seen your comment and will start implementing the fix now. I'll open a draft PR to share my progress.

[roomote]

Opened a draft PR with my attempt at addressing this: #10172

The fix adds a new addEmptyReasoning option to convertToR1Format() that adds empty reasoning_content to assistant messages that lack it. This ensures compatibility with DeepSeek's thinking mode when switching providers mid-conversation.

Feedback welcome!