feat: add DeepSeek V4 model support and fix thinking mode detection

[roomote-v0]

This PR attempts to address Issue #12174. Feedback and guidance are welcome.

Changes

1. Add DeepSeek V4 model (deepseek-v4-0324)

• Added to packages/types/src/providers/deepseek.ts with:
  • 128K context window, 16K max output tokens
  • Vision support enabled
  • Thinking mode (preserveReasoning: true)
  • Pricing from https://api-docs.deepseek.com/quick_start/pricing

2. Fix thinking mode detection (root cause of the 400 error)

• Changed src/api/providers/deepseek.ts to detect thinking mode via the preserveReasoning flag on ModelInfo instead of hardcoding modelId.includes("deepseek-reasoner")
• This means any model with preserveReasoning: true automatically gets thinking mode enabled, including v4 and any future models
• This fixes the reported 400 error ("The reasoning_content in the thinking mode must be passed back to the API") for v4 models

3. Tests

• Added model info tests for deepseek-v4-0324
• Added thinking mode behavior tests for v4 model
• Updated mock to recognize v4 models as reasoning models
• All 27 tests pass

Notes

• The pricing values for v4 may need verification against the latest DeepSeek API docs
• If additional v4 model variants exist (e.g. flash), they can be added following the same pattern

Interactively review PR in Roo Code Cloud

[designgears]

model names should be
deepseek-v4-pro
deepseek-v4-flash