Summary

• Fix HTTP 400 errors when using DeepSeek /anthropic endpoint with thinking mode enabled
• Widen DeepSeek V-series model regex to accept bracket suffixes like [1m] (e.g. deepseek-v4-flash[1m] for 1M context)

Problem

When replaying multi-turn conversations through DeepSeek's /anthropic endpoint, the API returns:

HTTP 400: The `content[].thinking` in the thinking mode must be passed back to the API.

Root cause: _extract_preserved_thinking_blocks() extracts signed thinking blocks from reasoning_details first, setting _already_has_thinking=True. This prevents the reasoning_content path from creating unsigned blocks. The signed blocks are then stripped by the thinking-block management pass (since DeepSeek cannot validate Anthropic signatures), resulting in no thinking blocks at all.

Fix

In convert_messages_to_anthropic(), skip extracting signed thinking blocks from reasoning_details when the target endpoint is DeepSeek /anthropic. This allows the reasoning_content path to create unsigned blocks that DeepSeek accepts.

Also widen the _DEEPSEEK_V_SERIES_RE regex from ^deepseek-v\d+([-.].+)?$ to ^deepseek-v\d+.*$ so model names with bracket suffixes (like deepseek-v4-flash[1m]) are correctly recognized as V-series models instead of being normalized to deepseek-chat.

Test plan

• Verified with DeepSeek deepseek-v4-flash[1m] via /anthropic endpoint — multi-turn conversations now work without HTTP 400
• Verified model normalization: deepseek-v4-flash[1m] → deepseek-v4-flash[1m] (not deepseek-chat)