Fix streaming tool calls when reasoning parser is active

[janhilgard]

Summary

• Fix streaming tool calls being emitted as raw text when --reasoning-parser is enabled
• The reasoning parser branch in stream_chat_completion() was skipping tool call parsing entirely, causing <tool_call> markup to leak into content instead of being parsed into structured tool_calls objects
• Adds tool call parsing to the reasoning branch, mirroring the existing logic in the standard path

Problem

When both --reasoning-parser deepseek_r1 and --tool-call-parser hermes are enabled, streaming responses emit raw tool call markup as plain text:

{"delta": {"content": "<tool_call>\n<function=get_weather>\n<parameter=location>\nPrague\n</parameter>\n</function>\n</tool_call>", "tool_calls": null}, "finish_reason": "stop"}

Fix

After this change, tool calls are correctly parsed and emitted:

{"delta": {"tool_calls": [{"index": 0, "id": "call_4b99a63b", "type": "function", "function": {"name": "get_weather", "arguments": "{\"location\": \"Prague\"}"}}]}, "finish_reason": "tool_calls"}

Test plan

• Streaming tool call with --reasoning-parser deepseek_r1 --tool-call-parser hermes returns structured tool_calls
• finish_reason is "tool_calls" (not "stop")
• Reasoning-only responses (no tool calls) still work correctly
• Non-streaming tool calls unaffected
• black formatting passes

🤖 Generated with Claude Code