程序包 io.agentscope.core.middleware

类 TaskReminderMiddleware

java.lang.Object

io.agentscope.core.middleware.TaskReminderMiddleware

所有已实现的接口:

MiddlewareBase

public class TaskReminderMiddleware
extends Object
implements MiddlewareBase

Keeps the agent's todo list in front of the model so long-running tasks don't drift.

Two hooks:

• onSystemPrompt(io.agentscope.core.agent.Agent, io.agentscope.core.agent.RuntimeContext, java.lang.String) appends a one-time, static explanation of how to use the todo_write tool (grounding).
• onReasoning(io.agentscope.core.agent.Agent, io.agentscope.core.agent.RuntimeContext, io.agentscope.core.middleware.ReasoningInput, java.util.function.Function<io.agentscope.core.middleware.ReasoningInput, reactor.core.publisher.Flux<io.agentscope.core.event.AgentEvent>>) appends a fresh <system-reminder> rendering the current AgentState.tasksContext before every reasoning step, so the model always sees the latest state regardless of how many tool calls happened since.

The reminder is appended transiently to the reasoning input only; it is never written into AgentState.context, so it is never persisted, compacted, or recalled. It is additionally tagged with Msg.METADATA_SYNTHETIC so any consumer that does observe it can skip it.

Unlike opencode (which leaves the latest list in the most recent tool output), agentscope re-injects the list every turn: after conversation compaction the tool output may be gone, but tasksContext survives, so this middleware guarantees the list is always visible.

构造器概要

构造器

构造器	说明
TaskReminderMiddleware()

方法概要

修饰符和类型	方法	说明
reactor.core.publisher.Flux<AgentEvent>	onReasoning(Agent agent, RuntimeContext ctx, ReasoningInput input, Function<ReasoningInput,reactor.core.publisher.Flux<AgentEvent>> next)	Intercept the reasoning phase (LLM call + streaming output parsing).
reactor.core.publisher.Mono<String>	onSystemPrompt(Agent agent, RuntimeContext ctx, String currentPrompt)	Transform the system prompt string (pipeline pattern).

从类继承的方法 java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

从接口继承的方法 io.agentscope.core.middleware.MiddlewareBase

onActing, onAgent, onModelCall

构造器详细资料

TaskReminderMiddleware

public TaskReminderMiddleware()

方法详细资料

onSystemPrompt

public reactor.core.publisher.Mono<String> onSystemPrompt(Agent agent,
 RuntimeContext ctx,
 String currentPrompt)

从接口复制的说明: MiddlewareBase

Transform the system prompt string (pipeline pattern).

Multiple middlewares are applied sequentially; each receives the output of the previous one.

指定者:

onSystemPrompt 在接口中 MiddlewareBase

参数:

agent - the agent instance

ctx - per-call runtime context (session, user, attributes)

currentPrompt - the current system prompt

返回:

the (possibly transformed) system prompt

onReasoning

public reactor.core.publisher.Flux<AgentEvent> onReasoning(Agent agent,
 RuntimeContext ctx,
 ReasoningInput input,
 Function<ReasoningInput,reactor.core.publisher.Flux<AgentEvent>> next)

从接口复制的说明: MiddlewareBase

Intercept the reasoning phase (LLM call + streaming output parsing).

指定者:

onReasoning 在接口中 MiddlewareBase

参数:

agent - the agent instance

ctx - per-call runtime context (session, user, attributes)

input - reasoning input (messages, tools, options)

next - calls the next middleware or the core reasoning logic

返回:

event stream from reasoning