perf: detect repeated tool calls by name and arguments (#9311)

* fix: detect repeated tool calls by name and arguments

* Update tests/test_tool_loop_agent_runner.py

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

---------

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

astrbot/core/agent/runners/tool_loop_agent_runner.py

@@ -148,21 +148,23 @@ class ToolLoopAgentRunner(BaseAgentRunner[TContext]):
     MALFORMED_TOOL_NAME_PLACEHOLDER = "__malformed_tool_name__"
     REPEATED_TOOL_NOTICE_L1_TEMPLATE = (
         "\n\n[SYSTEM NOTICE] By the way, you have executed the same tool "
-        "`{tool_name}` {streak} times consecutively. Double-check whether another "
-        "tool, different arguments, or a summary would move the task forward better."
+        "`{tool_name}` with the same arguments {streak} times consecutively. "
+        "Double-check whether another tool, different arguments, or a summary would "
+        "move the task forward better."
     )
     REPEATED_TOOL_NOTICE_L2_TEMPLATE = (
         "\n\n[SYSTEM NOTICE] Important: you have executed the same tool "
-        "`{tool_name}` {streak} times consecutively. Unless this repetition is "
-        "clearly necessary, stop repeating the same action and either switch "
-        "tools, refine parameters, or summarize what is still missing."
+        "`{tool_name}` with the same arguments {streak} times consecutively. "
+        "Unless this repetition is clearly necessary, stop repeating the same action "
+        "and either switch tools, refine parameters, or summarize what is still "
+        "missing."
     )
     REPEATED_TOOL_NOTICE_L3_TEMPLATE = (
         "\n\n[SYSTEM NOTICE] Important: you have executed the same tool "
-        "`{tool_name}` {streak} times consecutively. Repetition is now very "
-        "high. Continue only if each call is clearly producing new information. "
-        "Otherwise, change strategy, adjust arguments, or explain the limitation "
-        "to the user."
+        "`{tool_name}` with the same arguments {streak} times consecutively. "
+        "Repetition is now very high. Continue only if each call is clearly producing "
+        "new information. Otherwise, change strategy, adjust arguments, or explain "
+        "the limitation to the user."
     )
     TOOL_RESULT_OVERFLOW_NOTICE_TEMPLATE = (
         "Truncated tool output preview shown above. "
@@ -281,6 +283,7 @@ class ToolLoopAgentRunner(BaseAgentRunner[TContext]):
         self._pending_follow_ups: list[FollowUpTicket] = []
         self._follow_up_seq = 0
         self._last_tool_name: str | None = None
+        self._last_tool_args: dict[str, T.Any] | None = None
         self._same_tool_streak = 0
 
         # These two are used for tool schema mode handling
@@ -662,11 +665,29 @@ class ToolLoopAgentRunner(BaseAgentRunner[TContext]):
             return content
         return f"{content}{notice}"
 
-    def _track_tool_call_streak(self, tool_name: str) -> int:
-        if tool_name == self._last_tool_name:
+    def _track_tool_call_streak(
+        self,
+        tool_name: str,
+        tool_args: dict[str, T.Any] | None,
+    ) -> int:
+        """Track consecutive tool calls with the same name and arguments.
+
+        Args:
+            tool_name: Name of the called tool.
+            tool_args: Arguments passed to the tool.
+
+        Returns:
+            Number of consecutive calls with the same name and arguments.
+        """
+        normalized_args = {} if tool_args is None else tool_args
+        if (
+            tool_name == self._last_tool_name
+            and normalized_args == self._last_tool_args
+        ):
             self._same_tool_streak += 1
         else:
             self._last_tool_name = tool_name
+            self._last_tool_args = copy.deepcopy(normalized_args)
             self._same_tool_streak = 1
         return self._same_tool_streak
 
@@ -1032,7 +1053,10 @@ class ToolLoopAgentRunner(BaseAgentRunner[TContext]):
             llm_response.tools_call_ids,
         ):
             tool_result_blocks_start = len(tool_call_result_blocks)
-            tool_call_streak = self._track_tool_call_streak(func_tool_name)
+            tool_call_streak = self._track_tool_call_streak(
+                func_tool_name,
+                func_tool_args,
+            )
             yield _HandleFunctionToolsResult.from_message_chain(
                 MessageChain(
                     type="tool_call",

tests/test_tool_loop_agent_runner.py

@@ -323,9 +323,14 @@ class CapturingToolLoopProvider(MockProvider):
 
 
 class SequentialToolProvider(MockProvider):
-    def __init__(self, tool_sequence: list[str]):
+    def __init__(
+        self,
+        tool_sequence: list[str],
+        tool_args_sequence: list[dict[str, Any]] | None = None,
+    ):
         super().__init__()
         self.tool_sequence = tool_sequence
+        self.tool_args_sequence = tool_args_sequence
 
     async def text_chat(self, **kwargs) -> LLMResponse:
         self.call_count += 1
@@ -338,11 +343,16 @@ class SequentialToolProvider(MockProvider):
             )
 
         tool_name = self.tool_sequence[self.call_count - 1]
+        tool_args = (
+            self.tool_args_sequence[self.call_count - 1]
+            if self.tool_args_sequence is not None
+            else {"query": f"step-{self.call_count}"}
+        )
         return LLMResponse(
             role="assistant",
             completion_text="",
             tools_call_name=[tool_name],
-            tools_call_args=[{"query": f"step-{self.call_count}"}],
+            tools_call_args=[tool_args],
             tools_call_ids=[f"call_{self.call_count}"],
             usage=TokenUsage(input_other=10, output=5),
         )
@@ -968,7 +978,10 @@ async def test_same_tool_consecutive_results_include_escalating_guidance(
 ):
     runner_cls = type(runner)
     total_calls = runner_cls.REPEATED_TOOL_NOTICE_L3_THRESHOLD
-    provider = SequentialToolProvider(["test_tool"] * total_calls)
+    provider = SequentialToolProvider(
+        ["test_tool"] * total_calls,
+        [{"query": "same"}] * total_calls,
+    )
     tool = FunctionTool(
         name="test_tool",
         description="测试工具",
@@ -1031,6 +1044,46 @@ async def test_same_tool_consecutive_results_include_escalating_guidance(
             assert level_3_notice in content
 
 
+@pytest.mark.asyncio
+async def test_same_tool_with_different_args_does_not_include_repeated_guidance(
+    runner, mock_tool_executor, mock_hooks
+):
+    runner_cls = type(runner)
+    total_calls = runner_cls.REPEATED_TOOL_NOTICE_L3_THRESHOLD
+    provider = SequentialToolProvider(["test_tool"] * total_calls)
+    tool = FunctionTool(
+        name="test_tool",
+        description="测试工具",
+        parameters={"type": "object", "properties": {"query": {"type": "string"}}},
+        handler=AsyncMock(),
+    )
+    request = ProviderRequest(
+        prompt="使用不同参数连续执行工具",
+        func_tool=ToolSet(tools=[tool]),
+        contexts=[],
+    )
+
+    await runner.reset(
+        provider=provider,
+        request=request,
+        run_context=ContextWrapper(context=None),
+        tool_executor=mock_tool_executor,
+        agent_hooks=mock_hooks,
+        streaming=False,
+    )
+
+    async for _ in runner.step_until_done(total_calls + 1):
+        pass
+
+    tool_messages = [
+        m for m in runner.run_context.messages if getattr(m, "role", None) == "tool"
+    ]
+    assert len(tool_messages) == total_calls
+    assert all(
+        "[SYSTEM NOTICE]" not in str(message.content) for message in tool_messages
+    )
+
+
 @pytest.mark.asyncio
 async def test_same_tool_streak_resets_after_switching_tools(
     runner, mock_tool_executor, mock_hooks
@@ -1038,7 +1091,8 @@ async def test_same_tool_streak_resets_after_switching_tools(
     runner_cls = type(runner)
     repeated_after_reset = runner_cls.REPEATED_TOOL_NOTICE_L1_THRESHOLD
     provider = SequentialToolProvider(
-        ["test_tool", "other_tool", *(["test_tool"] * repeated_after_reset)]
+        ["test_tool", "other_tool", *(["test_tool"] * repeated_after_reset)],
+        [{"query": "same"}] * (repeated_after_reset + 2),
     )
     tool_a = FunctionTool(
         name="test_tool",