fix: streaming OpenAI→Anthropic block index collisions (qwen3.7-max, deepseek-v4-flash)

[yinnho]

Problem

Claude Code reported Content block not found when proxying streaming responses from qwen3.7-max and deepseek-v4-flash (OpenAI-format providers).

Root Cause

convert_openai_stream_to_anthropic used a single shared block_index that only incremented on close, causing index collisions when multiple blocks were open simultaneously:

1. qwen3.7-max / DeepSeek send reasoning_content: "" to mark end of thinking. The empty string was skipped by !is_empty(), leaving the thinking block open → next text block reused the same index.

2. Parallel tool calls (deepseek-v4-flash, common in Claude Code) each opened a tool_use block at the current block_index without incrementing, so two tools shared index 0 and their argument deltas were misrouted.

Fix

Refactored StreamState so each block records its own uniquely allocated index:

Before	After
block_index: u32 (shared, increment on close)	next_block_index: u32 (alloc on open)
thinking_block_open: bool	thinking_block: Option<u32>
text_block_open: bool	text_block: Option<u32>
tool_blocks_open: Vec<bool>	tool_blocks: Vec<Option<u32>>

Every content_block_start now calls alloc_index() which assigns and increments immediately, guaranteeing unique indices regardless of how many blocks are open or the order of transitions.

Also closes the thinking block on empty/absent reasoning_content and adds a safety net before opening text/tool blocks.

Tests

6 new regression tests, all 91 tests pass:

• test_qwen37_max_empty_reasoning_content_closes_thinking_block
• test_reasoning_content_disappearing_closes_thinking_block
• test_safety_net_closes_thinking_before_text
• test_deepseek_parallel_tool_calls_unique_indices ← the key one
• test_deepseek_reasoning_then_tool_calls
• test_deepseek_flash_reasoning_content_tools

🤖 Generated with Claude Code