fix(agent): send thinking control to DeepSeek direct API to prevent 400 on tool-call replay

agent/transports/chat_completions.py

@@ -182,6 +182,7 @@ def build_kwargs(
             is_github_models: bool
             is_nvidia_nim: bool
             is_kimi: bool
+            is_deepseek_direct: bool
             is_lmstudio: bool
             is_custom_provider: bool
             ollama_num_ctx: int | None
@@ -353,6 +354,19 @@ def build_kwargs(
                 "type": "enabled" if _kimi_thinking_enabled else "disabled",
             }
 
+        # DeepSeek direct API: extra_body.thinking
+        is_deepseek_direct = params.get("is_deepseek_direct", False)
+        if is_deepseek_direct:
+            _ds_thinking_enabled = True
+            if reasoning_config and isinstance(reasoning_config, dict):
+                if reasoning_config.get("enabled") is False:
+                    _ds_thinking_enabled = False
+                elif (reasoning_config.get("effort") or "").strip().lower() == "none":
+                    _ds_thinking_enabled = False
+            extra_body["thinking"] = {
+                "type": "enabled" if _ds_thinking_enabled else "disabled",
+            }
+
         # Reasoning. LM Studio is handled above via top-level reasoning_effort,
         # so skip emitting extra_body.reasoning for it.
         if params.get("supports_reasoning", False) and not params.get("is_lmstudio", False):

run_agent.py

@@ -8410,6 +8410,10 @@ def _build_api_kwargs(self, api_messages: list) -> dict:
             or base_url_host_matches(self.base_url, "moonshot.ai")
             or base_url_host_matches(self.base_url, "moonshot.cn")
         )
+        _is_deepseek_direct = (
+            (self.provider or "").strip().lower() == "deepseek"
+            or base_url_host_matches(self.base_url, "api.deepseek.com")
+        )
         _is_tokenhub = base_url_host_matches(self._base_url_lower, "tokenhub.tencentmaas.com")
         _is_lmstudio = (self.provider or "").strip().lower() == "lmstudio"
 
@@ -8484,6 +8488,7 @@ def _build_api_kwargs(self, api_messages: list) -> dict:
             is_github_models=_is_gh,
             is_nvidia_nim=_is_nvidia,
             is_kimi=_is_kimi,
+            is_deepseek_direct=_is_deepseek_direct,
             is_tokenhub=_is_tokenhub,
             is_lmstudio=_is_lmstudio,
             is_custom_provider=self.provider == "custom",

tests/agent/transports/test_chat_completions.py

@@ -454,6 +454,49 @@ def test_non_moonshot_tools_are_not_mutated(self, transport):
         assert "type" not in kw["tools"][0]["function"]["parameters"]["properties"]["q"]
 
 
+class TestChatCompletionsDeepSeekDirect:
+    """DeepSeek direct API (api.deepseek.com) thinking control via extra_body."""
+
+    def test_deepseek_thinking_enabled_by_default(self, transport):
+        kw = transport.build_kwargs(
+            model="deepseek-chat", messages=[{"role": "user", "content": "Hi"}],
+            is_deepseek_direct=True,
+        )
+        assert kw["extra_body"]["thinking"] == {"type": "enabled"}
+
+    def test_deepseek_thinking_disabled_when_enabled_false(self, transport):
+        kw = transport.build_kwargs(
+            model="deepseek-chat", messages=[{"role": "user", "content": "Hi"}],
+            is_deepseek_direct=True,
+            reasoning_config={"enabled": False},
+        )
+        assert kw["extra_body"]["thinking"] == {"type": "disabled"}
+
+    def test_deepseek_thinking_disabled_when_effort_none(self, transport):
+        kw = transport.build_kwargs(
+            model="deepseek-chat", messages=[{"role": "user", "content": "Hi"}],
+            is_deepseek_direct=True,
+            reasoning_config={"effort": "none"},
+        )
+        assert kw["extra_body"]["thinking"] == {"type": "disabled"}
+
+    def test_openrouter_deepseek_no_thinking_param(self, transport):
+        kw = transport.build_kwargs(
+            model="deepseek/deepseek-chat",
+            messages=[{"role": "user", "content": "Hi"}],
+            is_openrouter=True,
+        )
+        assert "thinking" not in kw.get("extra_body", {})
+
+    def test_kimi_thinking_unchanged_with_deepseek_present(self, transport):
+        kw = transport.build_kwargs(
+            model="kimi-k2", messages=[{"role": "user", "content": "Hi"}],
+            is_kimi=True,
+            max_tokens_param_fn=lambda n: {"max_tokens": n},
+        )
+        assert kw["extra_body"]["thinking"] == {"type": "enabled"}
+
+
 class TestChatCompletionsLmStudioReasoning:
     """LM Studio publishes per-model reasoning ``allowed_options``. When the
     user requests an effort the model can't honor (e.g. ``high`` on a