fix:timinggate非法工具会重试
This commit is contained in:
@@ -3,7 +3,7 @@ from types import SimpleNamespace
|
||||
|
||||
import pytest
|
||||
|
||||
from src.core.tooling import ToolExecutionResult
|
||||
from src.core.tooling import ToolExecutionResult, ToolInvocation
|
||||
from src.llm_models.payload_content.tool_option import ToolCall
|
||||
from src.maisaka.chat_loop_service import ChatResponse, MaisakaChatLoopService
|
||||
from src.maisaka.context_messages import AssistantMessage, TIMING_GATE_INVALID_TOOL_HINT_SOURCE
|
||||
@@ -45,8 +45,12 @@ async def test_timing_gate_invalid_tool_defaults_to_no_reply(monkeypatch: pytest
|
||||
runtime._enter_stop_state = _enter_stop_state
|
||||
engine = MaisakaReasoningEngine(runtime) # type: ignore[arg-type]
|
||||
|
||||
call_count = 0
|
||||
|
||||
async def _fake_timing_gate_sub_agent(**kwargs: object) -> ChatResponse:
|
||||
nonlocal call_count
|
||||
del kwargs
|
||||
call_count += 1
|
||||
return _build_chat_response([
|
||||
ToolCall(call_id="invalid-timing-tool", func_name="finish", args={}),
|
||||
])
|
||||
@@ -61,6 +65,7 @@ async def test_timing_gate_invalid_tool_defaults_to_no_reply(monkeypatch: pytest
|
||||
action, response, tool_results, tool_monitor_results = await engine._run_timing_gate(object()) # type: ignore[arg-type]
|
||||
|
||||
assert action == "no_reply"
|
||||
assert call_count == 3
|
||||
assert response.tool_calls[0].func_name == "finish"
|
||||
assert runtime.stopped is True
|
||||
assert tool_monitor_results == []
|
||||
@@ -68,10 +73,72 @@ async def test_timing_gate_invalid_tool_defaults_to_no_reply(monkeypatch: pytest
|
||||
assert runtime._chat_history[0].source == TIMING_GATE_INVALID_TOOL_HINT_SOURCE
|
||||
assert "finish" in runtime._chat_history[0].processed_plain_text
|
||||
assert tool_results == [
|
||||
"- retry [非法 Timing 工具]: 返回了 finish,将重试 (1/3)",
|
||||
"- retry [非法 Timing 工具]: 返回了 finish,将重试 (2/3)",
|
||||
"- no_reply [非法 Timing 工具]: 返回了 finish,已停止本轮并等待新消息",
|
||||
]
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_timing_gate_invalid_tool_retries_until_valid(monkeypatch: pytest.MonkeyPatch) -> None:
|
||||
runtime = SimpleNamespace(
|
||||
_force_next_timing_continue=False,
|
||||
_chat_history=[],
|
||||
log_prefix="[test]",
|
||||
stopped=False,
|
||||
)
|
||||
|
||||
def _enter_stop_state() -> None:
|
||||
runtime.stopped = True
|
||||
|
||||
runtime._enter_stop_state = _enter_stop_state
|
||||
engine = MaisakaReasoningEngine(runtime) # type: ignore[arg-type]
|
||||
responses = [
|
||||
_build_chat_response([ToolCall(call_id="invalid-timing-tool", func_name="finish", args={})]),
|
||||
_build_chat_response([ToolCall(call_id="valid-timing-tool", func_name="continue", args={})]),
|
||||
]
|
||||
|
||||
async def _fake_timing_gate_sub_agent(**kwargs: object) -> ChatResponse:
|
||||
del kwargs
|
||||
return responses.pop(0)
|
||||
|
||||
async def _fake_invoke_tool_call(
|
||||
tool_call: ToolCall,
|
||||
latest_thought: str,
|
||||
anchor_message: object,
|
||||
*,
|
||||
append_history: bool = True,
|
||||
store_record: bool = True,
|
||||
) -> tuple[ToolInvocation, ToolExecutionResult, None]:
|
||||
del latest_thought, anchor_message, append_history, store_record
|
||||
return (
|
||||
ToolInvocation(tool_name=tool_call.func_name, call_id=tool_call.call_id),
|
||||
ToolExecutionResult(
|
||||
tool_name=tool_call.func_name,
|
||||
success=True,
|
||||
content="继续执行主流程",
|
||||
metadata={"timing_action": "continue"},
|
||||
),
|
||||
None,
|
||||
)
|
||||
|
||||
monkeypatch.setattr(engine, "_run_timing_gate_sub_agent", _fake_timing_gate_sub_agent)
|
||||
monkeypatch.setattr(engine, "_invoke_tool_call", _fake_invoke_tool_call)
|
||||
|
||||
action, response, tool_results, tool_monitor_results = await engine._run_timing_gate(object()) # type: ignore[arg-type]
|
||||
|
||||
assert action == "continue"
|
||||
assert response.tool_calls[0].func_name == "continue"
|
||||
assert runtime.stopped is False
|
||||
assert len(runtime._chat_history) == 2
|
||||
assert all(message.source != TIMING_GATE_INVALID_TOOL_HINT_SOURCE for message in runtime._chat_history)
|
||||
assert tool_results == [
|
||||
"- retry [非法 Timing 工具]: 返回了 finish,将重试 (1/3)",
|
||||
"- continue [成功]: 继续执行主流程",
|
||||
]
|
||||
assert tool_monitor_results[0]["tool_name"] == "continue"
|
||||
|
||||
|
||||
def test_timing_gate_invalid_tool_hint_keeps_only_latest() -> None:
|
||||
old_hint = SimpleNamespace(source=TIMING_GATE_INVALID_TOOL_HINT_SOURCE)
|
||||
runtime = SimpleNamespace(_chat_history=[old_hint])
|
||||
|
||||
Reference in New Issue
Block a user