fix(rollout): handle invalid tool arguments

miles/rollout/generate_utils/tool_call_utils.py

@@ -16,6 +16,7 @@
 from miles.utils.types import Sample
 
 _DUMMY_USER = {"role": "user", "content": "dummy"}
+_INVALID_TOOL_ARGUMENTS_MESSAGE = "Error: Tool arguments must be a valid JSON object."
 
 
 def create_tool_call_parser(tool_specs, tool_call_parser):
@@ -40,21 +41,36 @@ async def _execute_tool_call(
 ) -> dict[str, Any]:
     if isinstance(call, ChatCompletionMessageToolCall):
         name = call.function.name
-        params = json.loads(call.function.arguments) if call.function.arguments else {}
+        raw_arguments = call.function.arguments
         tool_call_id = call.id
     elif isinstance(call, ToolCallItem):
         name = call.name
-        params = json.loads(call.parameters) if call.parameters else {}
+        raw_arguments = call.parameters
         tool_call_id = f"call_{uuid.uuid4().hex[:24]}"
     else:
         raise TypeError(f"Unsupported tool call type: {type(call)}")
 
-    result = await execute_one(name, params)
-    assert isinstance(result, str)
+    params = _parse_tool_arguments(raw_arguments)
+    if params is None:
+        result = _INVALID_TOOL_ARGUMENTS_MESSAGE
+    else:
+        result = await execute_one(name, params)
+        assert isinstance(result, str)
 
     return {"role": "tool", "tool_call_id": tool_call_id, "content": result, "name": name}
 
 
+def _parse_tool_arguments(raw_arguments: str | None) -> dict[str, Any] | None:
+    """Return executor kwargs, or None when recognized arguments are invalid."""
+    if not raw_arguments:
+        return {}
+    try:
+        params = json.loads(raw_arguments)
+    except json.JSONDecodeError:
+        return None
+    return params if isinstance(params, dict) else None
+
+
 def update_sample_with_tool_responses(sample: Sample, tool_messages: list[dict[str, Any]], tokenizer):
     next_obs_tokens_ids: list[int] = tokenize_tool_responses(tool_messages, tokenizer=tokenizer)
     sample.response += tokenizer.decode(next_obs_tokens_ids)

tests/fast/rollout/generate_hub/test_tool_call_utils.py

@@ -1,6 +1,15 @@
+from unittest.mock import AsyncMock
+
 import pytest
+from openai.types.chat import ChatCompletionMessageToolCall
+from sglang.srt.function_call.core_types import ToolCallItem
 
-from miles.rollout.generate_utils.tool_call_utils import _DUMMY_USER, _build_dummy_assistant, tokenize_tool_responses
+from miles.rollout.generate_utils.tool_call_utils import (
+    _DUMMY_USER,
+    _build_dummy_assistant,
+    _execute_tool_call,
+    tokenize_tool_responses,
+)
 from miles.utils.processing_utils import load_tokenizer
 
 TOOL_CALL_TEST_MODELS = [
@@ -59,6 +68,59 @@
 ]
 
 
+def _make_tool_call(call_type: str, arguments: str) -> ToolCallItem | ChatCompletionMessageToolCall:
+    if call_type == "sglang":
+        return ToolCallItem(tool_index=0, name="bash", parameters=arguments)
+    return ChatCompletionMessageToolCall(
+        id="call_test",
+        type="function",
+        function={"name": "bash", "arguments": arguments},
+    )
+
+
+class TestExecuteToolCall:
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("call_type", ["sglang", "openai"])
+    @pytest.mark.parametrize(
+        ("arguments", "expected_params"),
+        [
+            ('{"cmd": "pytest"}', {"cmd": "pytest"}),
+            ("", {}),
+        ],
+    )
+    async def test_valid_object_arguments_reach_executor(self, call_type, arguments, expected_params):
+        execute_one = AsyncMock(return_value="ok")
+
+        message = await _execute_tool_call(_make_tool_call(call_type, arguments), execute_one)
+
+        execute_one.assert_awaited_once_with("bash", expected_params)
+        assert message["role"] == "tool"
+        assert message["content"] == "ok"
+        assert message["name"] == "bash"
+        assert message["tool_call_id"]
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("call_type", ["sglang", "openai"])
+    @pytest.mark.parametrize(
+        "arguments",
+        [
+            pytest.param('{"cmd": "pytest"', id="malformed"),
+            pytest.param("[1, 2]", id="list"),
+            pytest.param("42", id="number"),
+        ],
+    )
+    async def test_invalid_arguments_return_tool_error_without_execution(self, call_type, arguments):
+        execute_one = AsyncMock(return_value="must not run")
+
+        message = await _execute_tool_call(_make_tool_call(call_type, arguments), execute_one)
+
+        execute_one.assert_not_awaited()
+        assert message["role"] == "tool"
+        assert message["content"] == "Error: Tool arguments must be a valid JSON object."
+        assert message["name"] == "bash"
+        assert message["tool_call_id"]
+
+
 class TestTokenizeToolResponses:
     @pytest.mark.parametrize("model_name", ["Qwen/Qwen3-0.6B"])
     def test_snapshot(self, model_name):