Build and run agents you can see, understand and trust.
Python agentchatbotlarge-language-modelsllmllm-agentmcpmulti-agentmulti-modalreact-agentrealtime-agentvoice-agentvoice-assistant
fix(middleware): retire the budget hint at a new reply boundary (#3599)
Felix Shi9aeab115940b
diff --git a/src/agentscope/middleware/_budget.py b/src/agentscope/middleware/_budget.pyindex a34ae5049..6115f7cbf 100644--- a/src/agentscope/middleware/_budget.py+++ b/src/agentscope/middleware/_budget.py@@ -25,6 +25,43 @@ _STRUCTURED_OUTPUT_HINT_MESSAGE = ("tools.</system-reminder>")+# Marks the wrap-up hints this middleware injects. They live in the persisted+# context, so a new reply -- which starts with a fresh budget and no forced+# tool_choice -- has to be able to find and retire the previous reply's hints.+_BUDGET_HINT_SOURCE = (+ '{"label": "System", "sublabel": "Reply Budget Exhausted"}'+)+++def _retire_budget_hints(agent: "Agent") -> None:+ """Remove wrap-up hints this middleware injected for an earlier reply.++ Only hints carrying :data:`_BUDGET_HINT_SOURCE` are removed, so hints from+ other middleware, and every message that still has real content, stay put.+ A message left with no content at all was created solely to carry the hint+ and is dropped with it.+ """+ context = agent.state.context+ kept = []+ for message in context:+ content = getattr(message, "content", None)+ if isinstance(content, list):+ remaining = [+ block+ for block in content+ if not (+ isinstance(block, HintBlock)+ and block.source == _BUDGET_HINT_SOURCE+ )+ ]+ if len(remaining) != len(content):+ if not remaining:+ continue+ message.content = remaining+ kept.append(message)+ if len(kept) != len(context):+ context[:] = kept+class ReplyBudgetControlMiddleware(MiddlewareBase):"""Middleware that enforces a weighted token budget per reply.@@ -132,6 +169,9 @@ class ReplyBudgetControlMiddleware(MiddlewareBase):async for event in next_handler(**input_kwargs):if isinstance(event, ReplyStartEvent):+ # A new reply starts with a fresh budget, so the previous+ # reply's wrap-up instruction must not leak into it.+ _retire_budget_hints(agent)# Start counting for this reply, dropping the previous one.# Not cleaned on ReplyEndEvent: an outer middleware may# swallow it and the same reply then keeps spending tokens@@ -201,6 +241,7 @@ class ReplyBudgetControlMiddleware(MiddlewareBase):if structured_output_pendingelse self.hint_message),+ source=_BUDGET_HINT_SOURCE,)if (len(agent.state.context) > 0diff --git a/tests/middleware_budget_test.py b/tests/middleware_budget_test.pyindex 5f1d470f6..6f0f9d8a8 100644--- a/tests/middleware_budget_test.py+++ b/tests/middleware_budget_test.py@@ -408,6 +408,63 @@ class TestBudgetControlMiddleware(IsolatedAsyncioTestCase):{agent.state.reply_id: 28},)+ async def test_new_reply_does_not_inherit_the_previous_budget_hint(+ self,+ ) -> None:+ """A wrap-up hint belongs to one reply; the next reply starts clean.++ Reply 1 spends its whole 300-token budget on the first step, so the+ second step runs under the hint. Reply 2 starts with a fresh counter+ and must not see that wrap-up instruction again.+ """+ toolkit = Toolkit(tools=[DummyTool()])++ model = MockModel()+ model.set_responses(+ [+ ChatResponse(+ content=[+ ToolCallBlock(id="tc_1", name="dummy", input="{}"),+ ],+ is_last=True,+ usage=ChatUsage(+ input_tokens=200,+ output_tokens=100,+ time=0.0,+ ),+ ),+ _response("wrapped up", input_tokens=5, output_tokens=5),+ _response("fresh reply", input_tokens=5, output_tokens=5),+ ],+ )++ middleware = ReplyBudgetControlMiddleware(token_budget=300)+ agent = Agent(+ name="test_agent",+ system_prompt="you are helpful",+ model=model,+ toolkit=toolkit,+ middlewares=[middleware],+ )++ await agent.reply(UserMsg("user", "first"))+ self.assertTrue(+ any(+ _has_hint_block(message, middleware.hint_message)+ for message in agent.state.context+ ),+ "the exhausted reply should have injected the wrap-up hint",+ )++ await agent.reply(UserMsg("user", "second"))+ self.assertFalse(+ any(+ _has_hint_block(message, middleware.hint_message)+ for message in agent.state.context+ ),+ "a fresh reply must not carry the previous reply's budget hint",+ )+async def test_swallowed_reply_end_keeps_counting(self) -> None:"""A redo round forced by swallowing the ReplyEndEvent spends thebudget of the same reply."""