Ensure tool-call JSON stays deterministic for local models

Author: latent-variable

mini_agent/llm/openai_client.py

@@ -65,8 +65,12 @@ async def _make_api_request(
         params = {
             "model": self.model,
             "messages": api_messages,
+            # Add consistent parameters for better KV cache stability
+            "temperature": 0.7,  # Consistent temperature
+            "seed": 42,  # Fixed seed for deterministic behavior and better caching
             # Enable reasoning_split to separate thinking content
-            "extra_body": {"reasoning_split": True},
+            # NOTE: Commenting out for local models - may cause cache invalidation
+            # "extra_body": {"reasoning_split": True},
         }
 
         if tools:
@@ -144,13 +148,19 @@ def _convert_messages(self, messages: list[Message]) -> tuple[str | None, list[d
                 if msg.tool_calls:
                     tool_calls_list = []
                     for tool_call in msg.tool_calls:
+                        arguments_json = tool_call.function.arguments_json
+                        if arguments_json is None:
+                            # Fall back to deterministic dump if raw string missing
+                            arguments_json = json.dumps(
+                                tool_call.function.arguments, separators=(",", ":"), sort_keys=True
+                            )
                         tool_calls_list.append(
                             {
                                 "id": tool_call.id,
                                 "type": "function",
                                 "function": {
                                     "name": tool_call.function.name,
-                                    "arguments": json.dumps(tool_call.function.arguments),
+                                    "arguments": arguments_json,
                                 },
                             }
                         )
@@ -223,8 +233,9 @@ def _parse_response(self, response: Any) -> LLMResponse:
         tool_calls = []
         if response.tool_calls:
             for tool_call in response.tool_calls:
-                # Parse arguments from JSON string
-                arguments = json.loads(tool_call.function.arguments)
+                # Parse arguments from JSON string while preserving the raw text
+                raw_arguments = tool_call.function.arguments
+                arguments = json.loads(raw_arguments) if raw_arguments else {}
 
                 tool_calls.append(
                     ToolCall(
@@ -233,6 +244,7 @@ def _parse_response(self, response: Any) -> LLMResponse:
                         function=FunctionCall(
                             name=tool_call.function.name,
                             arguments=arguments,
+                            arguments_json=raw_arguments,
                         ),
                     )
                 )

mini_agent/schema/schema.py

@@ -15,7 +15,8 @@ class FunctionCall(BaseModel):
     """Function call details."""
 
     name: str
-    arguments: dict[str, Any]  # Function arguments as dict
+    arguments: dict[str, Any]  # Parsed function arguments
+    arguments_json: str | None = None  # Raw JSON string (for deterministic replay)
 
 
 class ToolCall(BaseModel):