点火Ignition

任务简报Brief

学习目标Learning goals

• 用 Agent / ReActAgent 构造可调用对象，并传入系统提示词。

  Build a callable agent with Agent / ReActAgent and a system prompt.

• 用一次普通调用拿到 Msg，再用流式事件观察增量文本。

  Get a Msg from a normal call, then observe incremental text with streaming events.

• 理解 max_iters 是安全阀，不是提示词。

  Understand that max_iters is a safety valve, not a prompt.

• 跑通 check，拿到第一枚通行码。

  Run check and earn your first pass code.

核心概念Core concepts

双轨实现Build it twice

先完整跑通一条轨道，再切换重来。两条轨道写出同一种产物，方便直接对照行为。Run one track end to end, then switch and repeat. Both tracks write the same artifact so you can compare behavior directly.

from agentscope.agent import Agent, ReActConfig
from agentscope.message import UserMsg

agent = Agent(
    name="ignition",
    system_prompt="You are a precise AgentScope tutor.",
    model=model,
    react_config=ReActConfig(max_iters=3),
)
reply = await agent.reply(UserMsg("user", "What did you use to answer me?"))
print(reply.get_text_content())

from agentscope.event import TextBlockDeltaEvent, ReplyEndEvent

async for event in agent.reply_stream(UserMsg("user", "Explain reply_stream.")):
    if isinstance(event, TextBlockDeltaEvent):
        print(event.delta, end="")
    elif isinstance(event, ReplyEndEvent):
        print("\n[done]")

# 在 lab/ 目录下
python lab.py check p0

# 期望看到
# PASS CODE: AS-P0-7Q4M

python lab.py check p0

ReActAgent agent = ReActAgent.builder()
        .name("ignition")
        .sysPrompt("You are a precise AgentScope tutor.")
        .model(model)
        .maxIters(3)
        .build();

RuntimeContext context = RuntimeContext.builder()
        .sessionId("p0").userId("lab").build();

Msg reply = agent.call("What did you use to answer me?", context).block();
System.out.println(reply.getTextContent());

streaming.streamEvents("Explain streamEvents.", context)
        .doOnNext(event -> System.out.println(event.getType().name()))
        .blockLast();

cd lab/java
mvn -q -o compile exec:java -Dexec.args="check p0"

# 期望看到
# PASS CODE: AS-P0-7Q4M

cd lab/java && mvn -q -o compile exec:java -Dexec.args="check p0"

Python ↔ Java 差异Python ↔ Java differences

• Python 用 reply / reply_stream；Java 用 call / streamEvents(Flux)。语义一致，但 Java 需要自己订阅 Reactor。

  Python uses reply / reply_stream; Java uses call / streamEvents(Flux). Same semantics, but Java subscribes to Reactor explicitly.

• Python 的事件是类（TextBlockDeltaEvent）；Java 的事件类型是枚举（AgentEventType.TEXT_BLOCK_DELTA）。

  Python events are classes (TextBlockDeltaEvent); Java event types are enum constants (AgentEventType.TEXT_BLOCK_DELTA).

• 两边都用 max_iters / maxIters 限制循环；不要用提示词祈祷模型停下。

  Both bound the loop with max_iters / maxIters; never rely on prompt wording to stop a runaway loop.

常见坑Pitfalls

1. 把 max_iters 当成模型参数：它限制的是循环次数，不是 token 数，也不是超时。生产环境还要加超时和成本上限。

   Treating max_iters as a model parameter: it bounds loop iterations, not tokens or time. Add timeouts and cost caps in production.

2. 在事件循环里同步等待：Python 用 await，Java 的 .block() 只适合 CLI/测试；Web 服务应返回 Flux 或异步结果。

   Blocking inside an event loop: Python uses await; Java .block() is for CLI/tests only. Web services should return Flux or async results.

3. 忽略 AGENT_END / ReplyEndEvent：只处理 delta 会让 UI 无法判断何时结束，流式渲染会一直转圈。

   Ignoring AGENT_END / ReplyEndEvent: handling only deltas leaves the UI unable to know when to stop.

领取通行码Claim your pass code

先运行当前轨道的 check。CLI 会打印 PASS CODE、XP 与徽章。粘贴到这里；同一关在每条轨道可各领取一次。Run the check for the current track. The CLI prints PASS CODE plus the exact XP and badge. Paste it here; the same mission can be claimed once on each track.

费曼复述Feynman recall

合上教程，用自己的话讲清机制。关键点覆盖 65%、篇幅 20%、结构 15%；60 分通过并一次性获得 20 XP。Explain the mechanism without looking. Coverage 65%, length 20%, structure 15%; 60 passes and earns +20 XP once.

间隔复习Spaced review

首次领取后会在明天安排复习。按真实感受评分，下一次间隔会自适应。After the first claim a review is scheduled for tomorrow. Rate how it felt; the next interval adapts.