From cb511928e2be4b7ee234582ecba96801fccf94fe Mon Sep 17 00:00:00 2001
From: oobabooga <112222186+oobabooga@users.noreply.github.com>
Date: Mon, 6 Apr 2026 12:06:28 -0700
Subject: [PATCH] Fix GPT-OSS tag leak during streaming between thinking and
 tool calls

---
 modules/reasoning.py    | 13 ++++++++++---
 modules/tool_parsing.py | 12 +++++++++---
 2 files changed, 19 insertions(+), 6 deletions(-)

diff --git a/modules/reasoning.py b/modules/reasoning.py
index 4a7cfa79..2b260818 100644
--- a/modules/reasoning.py
+++ b/modules/reasoning.py
@@ -73,9 +73,16 @@ def extract_reasoning(text, html_escaped=False):
                 if content_pos != -1:
                     content_start = content_pos + len(content_esc)
                 else:
-                    # Content tag not present — fall back to content after
-                    # end_tag (e.g. GPT-OSS tool calls skip the final channel).
-                    content_start = end_pos + len(end_esc)
+                    # Content tag not present yet.  In GPT-OSS the region
+                    # between <|end|> and the content tag contains internal
+                    # markup (<|start|>assistant…) that must not be shown.
+                    # Suppress it to prevent tag leaks during streaming.
+                    remainder = text[end_pos + len(end_esc):].lstrip()
+                    framing_token = esc('<|start|>')
+                    if not remainder or remainder.startswith(framing_token) or framing_token.startswith(remainder):
+                        content_start = len(text)
+                    else:
+                        content_start = end_pos + len(end_esc)
             else:
                 content_start = end_pos + len(end_esc)
 
diff --git a/modules/tool_parsing.py b/modules/tool_parsing.py
index 7fcf58b7..aa3e0e95 100644
--- a/modules/tool_parsing.py
+++ b/modules/tool_parsing.py
@@ -638,9 +638,15 @@ def parse_tool_call(answer: str, tool_names: list[str], return_prefix: bool = Fa
     # Strip thinking blocks so tool-call syntax inside <think> is ignored.
     original_answer = answer
     _, answer = extract_reasoning(answer)
-    # Offset between original and stripped text, used to map start_pos
-    # back to the original string when returning a prefix.
-    reasoning_offset = len(original_answer) - len(answer)
+    # Reasoning extraction returns empty content when GPT-OSS internal
+    # markup (<|start|>assistant…) follows the thinking block without a
+    # content tag.  Fall back to the full text so tool-call markers can
+    # be found.
+    if not answer.strip():
+        answer = original_answer
+        reasoning_offset = 0
+    else:
+        reasoning_offset = len(original_answer) - len(answer)
 
     matches = []
     start_pos = None