fix: improve OpenAI forced tool-call fallback parsing

app/main.py

@@ -44,6 +44,7 @@ from .http.tool_bridge import (
     _anthropic_tool_result_block,
     _anthropic_tool_use_block,
     _extract_function_call_event_from_text,
+    _forced_tool_fallback_event,
     _json_string,
     _openai_forced_tool_name,
     _openai_tool_call,
@@ -660,10 +661,17 @@ async def v1_chat_completions(req: ChatCompletionsRequest, request: Request):
                         yield _text_payload(text)
 
                     if buffered_text_parts and forced_tool_name and not saw_tool_call:
+                        merged_text = "".join(buffered_text_parts)
                         inferred = _extract_function_call_event_from_text(
-                            "".join(buffered_text_parts),
+                            merged_text,
                             forced_tool_name=forced_tool_name,
                         )
+                        if inferred is None:
+                            inferred = _forced_tool_fallback_event(
+                                merged_text,
+                                forced_tool_name=forced_tool_name,
+                                tools=req.tools,
+                            )
                         if inferred is not None:
                             tool_id = "call_inferred_0"
                             tool_call_indexes[tool_id] = 0
@@ -805,6 +813,12 @@ async def v1_chat_completions(req: ChatCompletionsRequest, request: Request):
                 message_content,
                 forced_tool_name=forced_tool_name,
             )
+            if inferred is None:
+                inferred = _forced_tool_fallback_event(
+                    message_content,
+                    forced_tool_name=forced_tool_name,
+                    tools=req.tools,
+                )
             if inferred is not None:
                 tool_calls.append(
                     _openai_tool_call(inferred, forced_id="call_inferred_0")