From 813db960458753a9e743ed0d10d7c1fd052c88c6 Mon Sep 17 00:00:00 2001
From: Houyun Sun <18763177732@139.com>
Date: Wed, 28 Oct 2026 16:14:53 +0800
Subject: [PATCH] fix(streaming): report tool_calls when an openai-compatible
 stream ends with stop

Non-streaming rewrites a "stop" to "tool_calls" whenever the response carries tool
calls (convert_dict_to_response). The stream only did it through
finish_reason_handler, and that is reached for providers that end on a separate
empty chunk. For openai/azure compatible chunks the finish chunk goes through the
received_finish_reason branch, which returns the chunk as-is, so the upstream
"stop" reached the caller unchanged.

Two things were off:
- the branch above returned before the tool_call flag could be used
- the flag itself is only set from completion_obj["tool_calls"], which
  openai-compatible chunks never fill; their tool calls sit on the delta

Now the delta is checked as well, and the same stop -> tool_calls rewrite is
applied in the received_finish_reason branch. length and content_filter are left
alone, same as the non-streaming path.

Fixes #48716