feat: add DeepSeek extreme reasoning tier

agent_platform/runtime/loop.py

@@ -599,6 +599,10 @@ class AgentLoop:
                 model=spec.provider_model,
                 messages=messages,
                 tools=available_specs,
+                provider=spec.provider,
+                thinking_enabled=spec.thinking_enabled,
+                reasoning_effort=spec.reasoning_effort,
+                max_output_tokens=spec.max_output_tokens,
             )
             choice = response["choices"][0]
             message = choice.get("message") or {}
@@ -648,9 +652,12 @@ class AgentLoop:
                         "completion.rejected",
                         {"reason": failure, "iteration": iteration + 1, "depth": depth},
                     )
+                    assistant_checkpoint = {"role": "assistant", "content": candidate}
+                    if isinstance(message.get("reasoning_content"), str):
+                        assistant_checkpoint["reasoning_content"] = message["reasoning_content"]
                     messages.extend(
                         [
-                            {"role": "assistant", "content": candidate},
+                            assistant_checkpoint,
                             {"role": "user", "content": recovery_message(failure)},
                         ]
                     )
@@ -683,6 +690,8 @@ class AgentLoop:
                 "content": message.get("content"),
                 "tool_calls": tool_calls,
             }
+            if isinstance(message.get("reasoning_content"), str):
+                assistant_message["reasoning_content"] = message["reasoning_content"]
             messages.append(assistant_message)
             blocked_call_reasons: dict[str, str] = {}
             batch_mutation_epoch = 0
@@ -801,7 +810,15 @@ class AgentLoop:
                 ),
             }
         )
-        response = await self.provider.complete(model=spec.provider_model, messages=messages, tools=None)
+        response = await self.provider.complete(
+            model=spec.provider_model,
+            messages=messages,
+            tools=None,
+            provider=spec.provider,
+            thinking_enabled=spec.thinking_enabled,
+            reasoning_effort=spec.reasoning_effort,
+            max_output_tokens=spec.max_output_tokens,
+        )
         answer = str(response["choices"][0].get("message", {}).get("content") or "").strip()
         if completion_failure(
             artifact_required,