From 8e0ce59c0e2d6dd1accc7b9e493ecd5139adf76c Mon Sep 17 00:00:00 2001
From: hanyuanling <ushouldknowr0@gmail.com>
Date: Mon, 27 Apr 2026 00:30:45 +0800
Subject: [PATCH] fix(provider): normalize DeepSeek non-string message content

---
 nanobot/providers/openai_compat_provider.py | 20 ++++++++
 tests/providers/test_litellm_kwargs.py      | 51 +++++++++++++++++++++
 2 files changed, 71 insertions(+)

diff --git a/nanobot/providers/openai_compat_provider.py b/nanobot/providers/openai_compat_provider.py
index fdbad585..6cd1ffb8 100644
--- a/nanobot/providers/openai_compat_provider.py
+++ b/nanobot/providers/openai_compat_provider.py
@@ -345,10 +345,25 @@ class OpenAICompatProvider(LLMProvider):
             return json.dumps(arguments, ensure_ascii=False)
         return "{}"
 
+    @staticmethod
+    def _coerce_content_to_string(content: Any) -> str | None:
+        """Coerce block/list content into plain text for strict string-only APIs."""
+        if content is None or isinstance(content, str):
+            return content
+        text = OpenAICompatProvider._extract_text_content(content)
+        if isinstance(text, str) and text:
+            return text
+        try:
+            dumped = json.dumps(content, ensure_ascii=False)
+        except Exception:
+            dumped = str(content)
+        return dumped or "(empty)"
+
     def _sanitize_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
         """Strip non-standard keys, normalize tool_call IDs."""
         sanitized = LLMProvider._sanitize_request_messages(messages, _ALLOWED_MSG_KEYS)
         id_map: dict[str, str] = {}
+        force_string_content = bool(self._spec and self._spec.name == "deepseek")
 
         def map_id(value: Any) -> Any:
             if not isinstance(value, str):
@@ -382,6 +397,11 @@ class OpenAICompatProvider(LLMProvider):
                     clean["content"] = None
             if "tool_call_id" in clean and clean["tool_call_id"]:
                 clean["tool_call_id"] = map_id(clean["tool_call_id"])
+            if (
+                force_string_content
+                and not (clean.get("role") == "assistant" and clean.get("tool_calls"))
+            ):
+                clean["content"] = self._coerce_content_to_string(clean.get("content"))
         return self._enforce_role_alternation(sanitized)
 
     def _drop_deepseek_incomplete_reasoning_history(
diff --git a/tests/providers/test_litellm_kwargs.py b/tests/providers/test_litellm_kwargs.py
index 1c3cfb85..e91d1cae 100644
--- a/tests/providers/test_litellm_kwargs.py
+++ b/tests/providers/test_litellm_kwargs.py
@@ -929,6 +929,57 @@ def test_backfill_does_not_touch_messages_when_thinking_off() -> None:
                 assert "reasoning_content" not in msg
 
 
+def test_deepseek_coerces_list_content_to_string() -> None:
+    """DeepSeek chat endpoint expects message.content to be a string."""
+    spec = find_by_name("deepseek")
+    with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):
+        p = OpenAICompatProvider(api_key="k", default_model="deepseek-chat", spec=spec)
+
+    kw = p._build_kwargs(
+        messages=[{
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "hello "},
+                {"type": "text", "text": "world"},
+            ],
+        }],
+        tools=None,
+        model="deepseek-chat",
+        max_tokens=1024,
+        temperature=0.7,
+        reasoning_effort=None,
+        tool_choice=None,
+    )
+
+    assert isinstance(kw["messages"][0]["content"], str)
+    assert "hello" in kw["messages"][0]["content"]
+    assert "world" in kw["messages"][0]["content"]
+
+
+def test_non_deepseek_keeps_list_content() -> None:
+    """Only DeepSeek should force string content; OpenAI-compatible providers keep blocks."""
+    spec = find_by_name("openai")
+    with patch("nanobot.providers.openai_compat_provider.AsyncOpenAI"):
+        p = OpenAICompatProvider(api_key="k", default_model="gpt-4o", spec=spec)
+
+    kw = p._build_kwargs(
+        messages=[{
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "hello"},
+            ],
+        }],
+        tools=None,
+        model="gpt-4o",
+        max_tokens=1024,
+        temperature=0.7,
+        reasoning_effort=None,
+        tool_choice=None,
+    )
+
+    assert isinstance(kw["messages"][0]["content"], list)
+
+
 def test_openai_no_thinking_extra_body() -> None:
     """Non-thinking providers should never get extra_body for thinking."""
     kw = _build_kwargs_for("openai", "gpt-4o", reasoning_effort="medium")