Added two test cases:

• test_anthropic_thinking_history_in_count_tokens : sends Anthropic-format messages with interleaved thinking + tool use through /v1/messages/count_tokens and verifies thinking blocks increase the token count (i.e. they're not silently dropped by convert_anthropic_to_oai).
• test_anthropic_thinking_history_in_template : uses /apply-template with the Qwen3 template to verify the converted reasoning_content renders inside tags in the prompt.

Both use the Qwen3 template which natively handles reasoning_content.

@T0mSIlver just wanted to open a PR for the same issue when I came across yours. I've been struggling with the same problems and qwen35 models. The core fix is basically identical to mine - i compiled it against ROCm and it solves the issue with claude code.

On a side note, it also fixes the autoparser branch where you'd previously get 500 {"error":{"code":500,"message":"Failed to parse input at pos 67: ","type":"server_error"}} (or other low positions within malformed thinking blocks) after some turns.