feat: add DeepSeek V4 Pro model support

[roomote-v0]

Related GitHub Issue

Closes: #12177

Description

This PR attempts to address Issue #12177 where deepseek-v4-pro fails with error "The reasoning_content in the thinking mode must be passed back to the API".

The root cause is that deepseek-v4-pro was not recognized as a thinking/reasoning model. Three changes were made:

1. Added deepseek-v4-pro to the model registry (packages/types/src/providers/deepseek.ts) with:

   • 164K context window, 16K max output tokens
   • Vision support enabled
   • preserveReasoning: true so reasoning_content is passed back to the API during multi-turn tool call conversations
   • Pricing based on info shared in the issue

2. Updated the isThinkingModel check (src/api/providers/deepseek.ts) to match deepseek-v4 prefix in addition to deepseek-reasoner. This ensures:

   • The thinking: { type: "enabled" } parameter is sent
   • mergeToolResultText is enabled to prevent reasoning_content from being dropped

3. Added tests for the new model in src/api/providers/__tests__/deepseek.spec.ts

Feedback and guidance are welcome, especially regarding the exact pricing and context window values.

Test Procedure

• All 27 existing DeepSeek tests pass (1 pre-existing skip)
• Added 4 new tests for deepseek-v4-pro:
  • Streaming reasoning_content handling
  • Thinking parameter inclusion in API calls
  • preserveReasoning flag verification
  • Model info correctness (maxTokens, contextWindow, supportsImages)
• Run: cd src && npx vitest run api/providers/__tests__/deepseek.spec.ts

Pre-Submission Checklist

• Issue Linked: This PR is linked to an approved GitHub Issue.
• Scope: Changes are focused on the linked issue.
• Self-Review: Performed self-review.
• Testing: New tests added for deepseek-v4-pro model.
• Documentation Impact: No documentation updates required.
• Contribution Guidelines: Read and agree.

Documentation Updates

• No documentation updates are required.

Additional Notes

The pricing values used ($2.19/M input, $8.87/M output, $0.55/M cache read) were estimated based on the pricing screenshot shared in the issue. These may need adjustment once official API documentation is confirmed.

Interactively review PR in Roo Code Cloud

[hunterschuler]

164K context window, 16K max output tokens

That context window is incorrect. DeepSeek recommends a minimum context window of 384K and supports up to a 1M context window, per https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro

16K output tokens also may be too strict. DeepSeek API supports 384K max output tokens, per https://api-docs.deepseek.com/quick_start/pricing