From 9daa0620a6bdfc3413b7520a7f5194931e8d97da Mon Sep 17 00:00:00 2001
From: codez <codezhujr@gmail.com>
Date: Sun, 26 Apr 2026 02:04:52 +0800
Subject: [PATCH] =?UTF-8?q?fix(agent):=20ordering=20fix=20in=20=5Fcopy=5Fr?=
 =?UTF-8?q?easoning=5Fcontent=5Ffor=5Fapi=20=E2=80=94=20cross-provider=20r?=
 =?UTF-8?q?easoning=20isolation?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Fix logic-ordering bug where normalized_reasoning promotion returns
before the DeepSeek/Kimi needs_empty_reasoning guard, causing
cross-provider reasoning content (MiniMax → DeepSeek) to leak into
reasoning_content and trigger HTTP 400.

Changes:
- Reorder branching: existing reasoning_content check first
- Add 'not has_reasoning' guard so poisoned histories (no reasoning)
  still get '' injected for DeepSeek/Kimi
- Healthy same-provider reasoning promotion path unchanged

Refs: #15250, #15213
---
 run_agent.py | 41 +++++++++++++++++++++++++++++------------
 1 file changed, 29 insertions(+), 12 deletions(-)

diff --git a/run_agent.py b/run_agent.py
index 55cbe04193..a28d04f278 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -7744,25 +7744,42 @@ class AIAgent:
         if source_msg.get("role") != "assistant":
             return
 
-        explicit_reasoning = source_msg.get("reasoning_content")
-        if isinstance(explicit_reasoning, str):
-            api_msg["reasoning_content"] = explicit_reasoning
+        # 1. Explicit reasoning_content already set — preserve it verbatim
+        # (includes DeepSeek/Kimi's own empty-string placeholder written at
+        # creation time, and any valid reasoning content from the same provider).
+        existing = source_msg.get("reasoning_content")
+        if isinstance(existing, str):
+            api_msg["reasoning_content"] = existing
             return
 
+        # 2. DeepSeek / Kimi thinking mode: tool-call turns with neither
+        # reasoning_content nor reasoning are "poisoned history" — a prior
+        # provider (MiniMax, etc.) left them empty. DeepSeek rejects HTTP 400
+        # if reasoning_content is absent on replay. Inject "" to satisfy the
+        # provider's requirement without forwarding cross-provider content.
+        has_reasoning = isinstance(source_msg.get("reasoning"), str) and source_msg.get("reasoning")
+        needs_empty_reasoning = (
+            not has_reasoning
+            and source_msg.get("tool_calls")
+            and (
+                self._needs_kimi_tool_reasoning()
+                or self._needs_deepseek_tool_reasoning()
+            )
+        )
+        if needs_empty_reasoning:
+            api_msg["reasoning_content"] = ""
+            return
+
+        # 3. Healthy session: promote 'reasoning' field to 'reasoning_content'
+        # for providers that use the internal 'reasoning' key.
         normalized_reasoning = source_msg.get("reasoning")
         if isinstance(normalized_reasoning, str) and normalized_reasoning:
             api_msg["reasoning_content"] = normalized_reasoning
             return
 
-        # Providers that require an echoed reasoning_content on every
-        # assistant tool-call turn. Detection logic lives in the per-provider
-        # helpers so both the creation path (_build_assistant_message) and
-        # this replay path stay in sync.
-        if source_msg.get("tool_calls") and (
-            self._needs_kimi_tool_reasoning()
-            or self._needs_deepseek_tool_reasoning()
-        ):
-            api_msg["reasoning_content"] = ""
+        # 4. reasoning_content was present but not a string (e.g. None after
+        # context compaction).  Don't pass null to the API.
+        api_msg.pop("reasoning_content", None)
 
     @staticmethod
     def _sanitize_tool_calls_for_strict_api(api_msg: dict) -> dict: