fix(provider): narrow DeepSeek reasoning history cleanup

Made-with: Cursor

nanobot/providers/openai_compat_provider.py

@@ -452,6 +452,7 @@ class OpenAICompatProvider(LLMProvider):
     def _drop_deepseek_incomplete_reasoning_history(
         self,
         messages: list[dict[str, Any]],
+        model_name: str,
         reasoning_effort: str | None,
     ) -> list[dict[str, Any]]:
         if (
@@ -460,11 +461,17 @@ class OpenAICompatProvider(LLMProvider):
         ):
             return messages
 
-        # For DeepSeek models, always check for incomplete reasoning history
-        # when thinking mode might be active. reasoning_effort may not be set
-        # explicitly but the model could still be using thinking mode by default.
-        # Only skip this check when reasoning_effort is explicitly set to "none".
-        if reasoning_effort is not None and reasoning_effort.lower() == "none":
+        semantic_effort = reasoning_effort.lower() if isinstance(reasoning_effort, str) else None
+        if semantic_effort in {"none", "minimal", "minimum"}:
+            return messages
+
+        # DeepSeek-V4 can require reasoning_content even when the config did
+        # not explicitly request reasoning_effort. Keep that implicit-thinking
+        # cleanup scoped to known thinking-capable DeepSeek models so normal
+        # deepseek-chat history is not trimmed.
+        if semantic_effort is None:
+            model_lower = model_name.lower()
+            if not any(token in model_lower for token in ("deepseek-v4", "deepseek-reasoner")):
                 return messages
 
         bad_idx = None
@@ -537,6 +544,7 @@ class OpenAICompatProvider(LLMProvider):
 
         messages = self._drop_deepseek_incomplete_reasoning_history(
             messages,
+            model_name,
             reasoning_effort,
         )
         kwargs: dict[str, Any] = {

tests/providers/test_litellm_kwargs.py

@@ -913,8 +913,8 @@ def test_deepseek_backfills_reasoning_content_on_legacy_tool_call_messages() ->
             assert msg["reasoning_content"] == ""
 
 
-def test_backfill_does_not_touch_messages_when_thinking_off() -> None:
-    """When reasoning_effort is None or minimal, legacy messages must NOT be altered."""
+def test_backfill_does_not_touch_messages_when_thinking_explicitly_off() -> None:
+    """When thinking is explicitly disabled, legacy messages must NOT be altered."""
     spec = find_by_name("deepseek")
     with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):
         p = OpenAICompatProvider(api_key="k", default_model="deepseek-v4-pro", spec=spec)
@@ -926,7 +926,7 @@ def test_backfill_does_not_touch_messages_when_thinking_off() -> None:
         {"role": "tool", "tool_call_id": "tc1", "content": "result"},
         {"role": "user", "content": "thanks"},
     ]
-    for effort in (None, "minimal"):
+    for effort in ("minimal", "none"):
         kw = p._build_kwargs(
             messages=list(messages), tools=None, model="deepseek-v4-pro",
             max_tokens=1024, temperature=0.7,
@@ -937,6 +937,56 @@ def test_backfill_does_not_touch_messages_when_thinking_off() -> None:
                 assert "reasoning_content" not in msg
 
 
+def test_deepseek_v4_drops_incomplete_reasoning_history_when_effort_implicit() -> None:
+    """DeepSeek-V4 may default to thinking, so incomplete legacy history is trimmed."""
+    spec = find_by_name("deepseek")
+    with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):
+        p = OpenAICompatProvider(api_key="k", default_model="deepseek-v4-pro", spec=spec)
+    messages = [
+        {"role": "system", "content": "system"},
+        {"role": "user", "content": "hi"},
+        {"role": "assistant", "content": "", "tool_calls": [
+            {"id": "tc1", "type": "function", "function": {"name": "web_search", "arguments": "{}"}}
+        ]},
+        {"role": "tool", "tool_call_id": "tc1", "content": "result"},
+        {"role": "user", "content": "thanks"},
+    ]
+
+    kw = p._build_kwargs(
+        messages=list(messages), tools=None, model="deepseek-v4-pro",
+        max_tokens=1024, temperature=0.7,
+        reasoning_effort=None, tool_choice=None,
+    )
+
+    assert [msg["role"] for msg in kw["messages"]] == ["system", "user"]
+    assert kw["messages"][-1]["content"] == "thanks"
+
+
+def test_deepseek_chat_keeps_tool_history_when_effort_implicit() -> None:
+    """Implicit cleanup must not trim non-thinking DeepSeek chat models."""
+    spec = find_by_name("deepseek")
+    with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):
+        p = OpenAICompatProvider(api_key="k", default_model="deepseek-chat", spec=spec)
+    messages = [
+        {"role": "user", "content": "hi"},
+        {"role": "assistant", "content": "", "tool_calls": [
+            {"id": "tc1", "type": "function", "function": {"name": "web_search", "arguments": "{}"}}
+        ]},
+        {"role": "tool", "tool_call_id": "tc1", "content": "result"},
+        {"role": "user", "content": "thanks"},
+    ]
+
+    kw = p._build_kwargs(
+        messages=list(messages), tools=None, model="deepseek-chat",
+        max_tokens=1024, temperature=0.7,
+        reasoning_effort=None, tool_choice=None,
+    )
+
+    roles = [msg["role"] for msg in kw["messages"]]
+    assert roles == ["user", "assistant", "tool", "user"]
+    assert kw["messages"][1]["tool_calls"]
+
+
 def test_deepseek_coerces_list_content_to_string() -> None:
     """DeepSeek chat endpoint expects message.content to be a string."""
     spec = find_by_name("deepseek")