From 82b8a3af7e1a73baf0930d1ae27138edc8aecbfb Mon Sep 17 00:00:00 2001
From: Xubin Ren <xubinrencs@gmail.com>
Date: Sun, 26 Apr 2026 12:47:09 +0000
Subject: [PATCH] fix(provider): handle incomplete DeepSeek reasoning history

---
 nanobot/providers/openai_compat_provider.py | 45 +++++++++++++
 tests/providers/test_litellm_kwargs.py      | 75 +++++++++++++++++++++
 2 files changed, 120 insertions(+)

diff --git a/nanobot/providers/openai_compat_provider.py b/nanobot/providers/openai_compat_provider.py
index 558093822..fdbad585c 100644
--- a/nanobot/providers/openai_compat_provider.py
+++ b/nanobot/providers/openai_compat_provider.py
@@ -384,6 +384,47 @@ class OpenAICompatProvider(LLMProvider):
                 clean["tool_call_id"] = map_id(clean["tool_call_id"])
         return self._enforce_role_alternation(sanitized)
 
+    def _drop_deepseek_incomplete_reasoning_history(
+        self,
+        messages: list[dict[str, Any]],
+        reasoning_effort: str | None,
+    ) -> list[dict[str, Any]]:
+        if (
+            not self._spec
+            or self._spec.name != "deepseek"
+            or not reasoning_effort
+            or reasoning_effort.lower() == "none"
+        ):
+            return messages
+
+        bad_idx = None
+        for idx, msg in enumerate(messages):
+            if (
+                msg.get("role") == "assistant"
+                and msg.get("tool_calls")
+                and not msg.get("reasoning_content")
+            ):
+                bad_idx = idx
+        if bad_idx is None:
+            return messages
+
+        keep_from = None
+        for idx in range(bad_idx + 1, len(messages)):
+            if messages[idx].get("role") == "user":
+                keep_from = idx
+                break
+
+        if keep_from is None:
+            trimmed = messages[:bad_idx]
+        else:
+            prefix = [msg for msg in messages[:keep_from] if msg.get("role") == "system"]
+            trimmed = prefix + messages[keep_from:]
+        logger.warning(
+            "Dropped {} DeepSeek thinking history message(s) with incomplete reasoning_content",
+            len(messages) - len(trimmed),
+        )
+        return trimmed
+
     # ------------------------------------------------------------------
     # Build kwargs
     # ------------------------------------------------------------------
@@ -424,6 +465,10 @@ class OpenAICompatProvider(LLMProvider):
         if spec and spec.strip_model_prefix:
             model_name = model_name.split("/")[-1]
 
+        messages = self._drop_deepseek_incomplete_reasoning_history(
+            messages,
+            reasoning_effort,
+        )
         kwargs: dict[str, Any] = {
             "model": model_name,
             "messages": self._sanitize_messages(self._sanitize_empty_content(messages)),
diff --git a/tests/providers/test_litellm_kwargs.py b/tests/providers/test_litellm_kwargs.py
index dfa0f58ac..1c3cfb851 100644
--- a/tests/providers/test_litellm_kwargs.py
+++ b/tests/providers/test_litellm_kwargs.py
@@ -585,6 +585,81 @@ def test_openai_compat_preserves_message_level_reasoning_fields() -> None:
     assert sanitized[1]["tool_calls"][0]["extra_content"] == {"google": {"thought_signature": "sig"}}
 
 
+def _deepseek_kwargs(messages: list[dict]) -> dict:
+    with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):
+        provider = OpenAICompatProvider(
+            api_key="sk-test",
+            default_model="deepseek-v4-flash",
+            spec=find_by_name("deepseek"),
+        )
+
+    return provider._build_kwargs(
+        messages=messages,
+        tools=None,
+        model="deepseek-v4-flash",
+        max_tokens=1024,
+        temperature=0.7,
+        reasoning_effort="high",
+        tool_choice=None,
+    )
+
+
+def _tool_call(call_id: str) -> dict:
+    return {
+        "id": call_id,
+        "type": "function",
+        "function": {"name": "my", "arguments": "{}"},
+    }
+
+
+def test_deepseek_thinking_drops_tool_history_missing_reasoning_content() -> None:
+    kwargs = _deepseek_kwargs([
+        {"role": "system", "content": "system"},
+        {"role": "user", "content": "can we use wechat?"},
+        {"role": "assistant", "content": "", "tool_calls": [_tool_call("call_bad")]},
+        {"role": "tool", "tool_call_id": "call_bad", "name": "my", "content": "channels"},
+        {"role": "user", "content": "continue"},
+    ])
+
+    assert kwargs["messages"] == [
+        {"role": "system", "content": "system"},
+        {"role": "user", "content": "continue"},
+    ]
+
+
+def test_deepseek_thinking_keeps_tool_history_with_reasoning_content() -> None:
+    kwargs = _deepseek_kwargs([
+        {"role": "user", "content": "can we use wechat?"},
+        {
+            "role": "assistant",
+            "content": "",
+            "reasoning_content": "I should inspect supported channels.",
+            "tool_calls": [_tool_call("call_good")],
+        },
+        {"role": "tool", "tool_call_id": "call_good", "name": "my", "content": "channels"},
+        {"role": "user", "content": "continue"},
+    ])
+
+    assistant = kwargs["messages"][1]
+    assert assistant["role"] == "assistant"
+    assert assistant["reasoning_content"] == "I should inspect supported channels."
+    assert kwargs["messages"][2]["role"] == "tool"
+
+
+def test_deepseek_thinking_drops_current_bad_tool_turn_without_followup_user() -> None:
+    kwargs = _deepseek_kwargs([
+        {"role": "system", "content": "system"},
+        {"role": "user", "content": "can we use wechat?"},
+        {"role": "assistant", "content": "", "tool_calls": [_tool_call("call_bad")]},
+        {"role": "tool", "tool_call_id": "call_bad", "name": "my", "content": "channels"},
+    ])
+
+    assert kwargs["messages"] == [
+        {"role": "system", "content": "system"},
+        {"role": "user", "content": "can we use wechat?"},
+    ]
+
+
 def test_openai_compat_keeps_tool_calls_after_consecutive_assistant_messages() -> None:
     with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):
         provider = OpenAICompatProvider()