Fix Responses API request handling (#25881)

Co-authored-by: Kai-Hsun Chen <kaihsun@apache.org>
Co-authored-by: Kristin Cowalcijk <kristincowalcijk@gmail.com>
Co-authored-by: aerosta <63026763+aerosta@users.noreply.github.com>
Co-authored-by: glaziermag <glaziermag@users.noreply.github.com>
Co-authored-by: Blake Ledden <blake.ledden@gmail.com>
Co-authored-by: PanJason <pyyjason@gmail.com>
Co-authored-by: Leoyzen <leoyzen@gmail.com>
Co-authored-by: kennyu <966806+kennyu@users.noreply.github.com>

python/sglang/srt/parser/jinja_template_utils.py

@@ -157,14 +157,16 @@ def process_content_for_template_format(
             if isinstance(chunk, dict):
                 chunk_type = chunk.get("type")
 
-                if chunk_type == "image_url":
+                if chunk_type in ("image_url", "input_image"):
                     image_obj = chunk.get("image_url") or {}
+                    if isinstance(image_obj, str):
+                        image_obj = {"url": image_obj, "detail": chunk.get("detail")}
                     mdp = image_obj.get("max_dynamic_patch", None)
                     # Also allow flat style: chunk["max_dynamic_patch"]
                     image_data.append(
                         ImageData(
                             url=image_obj["url"],
-                            detail=image_obj.get("detail", "auto"),
+                            detail=image_obj.get("detail") or "auto",
                             max_dynamic_patch=mdp,
                         )
                     )
@@ -194,13 +196,15 @@ def process_content_for_template_format(
                     audio_data.append(chunk["audio_url"]["url"])
                     # Normalize to simple 'audio' type
                     processed_content_parts.append({"type": "audio"})
-                elif chunk_type == "text":
+                elif chunk_type in ("text", "input_text"):
                     # For v32 encoding, collect text parts separately
                     if use_dpsk_v32_encoding:
                         text_parts.append(chunk["text"])
                     else:
                         # Keep text content as-is for openai format
-                        processed_content_parts.append(chunk)
+                        processed_content_parts.append(
+                            {"type": "text", "text": chunk["text"]}
+                        )
                 elif chunk_type == "tool_reference":
                     # GLM-specific extension: pass through so the chat template
                     # can match tool_reference.name against tools[*].function.name
@@ -220,7 +224,7 @@ def process_content_for_template_format(
         # String format: flatten to text only (for templates like DeepSeek)
         text_parts = []
         for chunk in msg_dict["content"]:
-            if isinstance(chunk, dict) and chunk.get("type") == "text":
+            if isinstance(chunk, dict) and chunk.get("type") in ("text", "input_text"):
                 text_parts.append(chunk["text"])
             # Note: For string format, we ignore images/audio since the template
             # doesn't expect structured content - multimodal placeholders would

python/sglang/srt/parser/reasoning_parser.py

@@ -257,6 +257,9 @@ class Qwen3Detector(BaseReasoningFormatDetector):
             think_excluded_tokens=think_excluded_tokens,
             force_reasoning=force_reasoning,
             stream_reasoning=stream_reasoning,
+            # Qwen3.5 sometimes opens ``<tool_call>`` without closing
+            # ``</think>``; treat it as an implicit reasoning close.
+            tool_start_token="<tool_call>",
             continue_final_message=continue_final_message,
             previous_content=previous_content,
             thinks_internally=True,