From 5789d78c34c6dae9aaac60ea0e316b1d0d0ae8ad Mon Sep 17 00:00:00 2001
From: "anbei.yuan" <anbei.yuan@nio.com>
Date: Sat, 9 May 2026 10:13:10 +0800
Subject: [PATCH] fix: strip <think> tokens from reasoning model output

Models with thinking/reasoning capabilities (DeepSeek-R1, MiniMax-M2.7,
QwQ, etc.) include <think>...</think> blocks in their response content.
These internal reasoning tokens leak into agent output and downstream
node inputs, corrupting the workflow.

Add _strip_thinking_tokens() classmethod to OpenAIProvider that filters
<think>...</think> blocks via regex. Applied in both:
- _deserialize_chat_response() (Message content)
- _append_chat_response_output() (timeline content)

The fix is zero-cost for models without thinking tokens (fast path
checks for '<think>' substring before regex).

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
---
 .../node/agent/providers/openai_provider.py    | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

diff --git a/runtime/node/agent/providers/openai_provider.py b/runtime/node/agent/providers/openai_provider.py
index 3b64efcc4f..809f6a00a7 100755
--- a/runtime/node/agent/providers/openai_provider.py
+++ b/runtime/node/agent/providers/openai_provider.py
@@ -2,6 +2,7 @@
 
 import base64
 import hashlib
+import re
 
 import binascii
 import os
@@ -383,18 +384,31 @@ def _deserialize_chat_response(self, response: Any) -> Message:
                     type="function"
                 ))
         
+        content = self._get_attr(msg, "content") or ""
+        content = self._strip_thinking_tokens(content)
+
         return Message(
             role=MessageRole.ASSISTANT,
-            content=self._get_attr(msg, "content") or "",
+            content=content,
             tool_calls=tool_calls
         )
 
+    _THINK_PATTERN = re.compile(r"<think>.*?</think>\s*", re.DOTALL)
+
+    @classmethod
+    def _strip_thinking_tokens(cls, text: str) -> str:
+        """Strip <think>...</think> blocks from model output (e.g. DeepSeek-R1, MiniMax-M2.7)."""
+        if "<think>" not in text:
+            return text
+        return cls._THINK_PATTERN.sub("", text).strip()
+
     def _append_chat_response_output(self, timeline: List[Any], response: Any) -> None:
         """Add chat response to timeline, preserving tool_calls (Chat API compatible)."""
         msg = response.choices[0].message
+        content = self._strip_thinking_tokens(msg.content or "")
         assistant_msg = {
             "role": "assistant",
-            "content": msg.content or ""
+            "content": content
         }
 
         if getattr(msg, "tool_calls", None):