feat: add DeepSeek V4 model support with thinking/reasoning_content handling

[roomote-v0]

This PR attempts to address Issue #12203.

Problem

When using new DeepSeek V4 model IDs (deepseek-v4-flash, deepseek-v4-pro) via the DeepSeek provider or OpenAI-compatible mode, the system fails to handle the reasoning_content field, resulting in a 400 Bad Request error because thinking mode is not enabled and reasoning_content is not passed back to the API.

Changes

1. packages/types/src/providers/deepseek.ts - Added deepseek-v4-flash and deepseek-v4-pro model definitions with preserveReasoning: true
2. src/api/providers/deepseek.ts - Changed isThinkingModel check to use the model info preserveReasoning flag instead of hardcoding deepseek-reasoner, making it extensible for any future thinking models
3. src/api/providers/openai.ts - Updated the OpenAI-compatible provider to also recognize deepseek-v4 model IDs for R1 format handling
4. src/api/providers/__tests__/deepseek.spec.ts - Added tests for V4 model preserveReasoning flags and thinking parameter passing

Feedback and guidance are welcome.

Interactively review PR in Roo Code Cloud

[lazyupdate]

Thanks for this PR! The refactored preserveReasoning flag approach is clean and extensible. A few things I noticed with the V4 model definitions:

Pricing

The current pricing (inputPrice: 0.28, outputPrice: 0.42) looks like it might be from the legacy deepseek-chat pricing. Per the official V4 pricing page:

Model	Input (cache miss)	Output	Cache Hit
deepseek-v4-flash	$0.14/M	$0.28/M	$0.0028/M
deepseek-v4-pro	$0.435/M ($1.74 disc)	$0.87/M ($3.48 disc)	$0.0145/M (disc)

Might be worth updating inputPrice, outputPrice, cacheWritesPrice, and cacheReadsPrice to match.

Context window

- contextWindow: 128_000,
+ contextWindow: 1_048_576,

V4 supports 1M context — could bump this from 128K.

maxTokens

- maxTokens: 8192,
+ maxTokens: 65536,

V4 supports up to 384K max output. 64K would be a reasonable lower bound for thinking models — 8K tends to be tight for reasoning chains plus tool call sequences. Thoughts?

Nice work on the preserveReasoning flag refactor btw, makes it easy to add future thinking models 👍