Files
deepseek-harness/website/zh-CN/develop/framework/events.md
T
lintianle 6ce9f16030 website: wire the site into the repo gates; make every tutorial example compile
- website joins the pnpm workspace; root scripts website:dev/website:build;
  run-gates gains a website-build gate (ci-primary + ci-static) — the
  VitePress build doubles as the site's dead-link check; AGENTS.md documents
  the commands.
- doc-typecheck + verify-type-equiv now scan website/zh-CN/**/*.md; every
  ```typescript fence converted to ```ts and made standalone-compilable
  (55 compiled, 1 ignore-check). Phantom APIs the compiler caught are fixed:
  invented event names (agent/turn-end, tool/call, llm/pre-request, ready,
  dispose) replaced with real catalog events or per-plugin declare-module
  merges; presentCall/inject/Config claims corrected to the real shapes.
- guide/config.md entry-fields table completed against loader EntryOptions;
  its coding-agent example brought in line with examples/coding-agent.
2026-07-16 18:12:22 +08:00

5.5 KiB
Raw Blame History

事件系统

事件是 Cordis 插件间通信的核心机制。Harness 大量使用事件来实现松耦合的扩展点。

基本用法

监听事件

import type { Context } from 'cordis'

declare module 'cordis' {
  interface Events {
    'event-name'(payload: string): void
  }
}

declare const ctx: Context

ctx.on('event-name', (payload) => {
  // 处理事件
})

触发事件

import type { Context } from 'cordis'

declare module 'cordis' {
  interface Events {
    'event-name'(payload: string): void
  }
}

declare const ctx: Context
declare const payload: string

ctx.emit('event-name', payload)

事件模式

Cordis 提供多种事件触发模式,适用于不同场景:

emit — 广播

所有监听器并行执行,不关心返回值:

import type { Context } from 'cordis'

declare module 'cordis' {
  interface Events {
    'my-plugin/turn-end'(agentId: string, turnIndex: number): void
  }
}

declare const ctx: Context
declare const agentId: string
declare const turnIndex: number

// 触发
ctx.emit('my-plugin/turn-end', agentId, turnIndex)

// 监听
ctx.on('my-plugin/turn-end', (agentId, turnIndex) => {
  console.log(`Turn ${turnIndex} ended`)
})

bail — 短路

依次调用监听器,第一个返回非 undefined 值的结果作为最终值:

import type { Context } from 'cordis'

declare module 'cordis' {
  interface Events {
    'some-check'(input: string): string | undefined
  }
}

declare const ctx: Context
declare const input: string
declare function shouldBlock(input: string): boolean

// 触发
const result = ctx.bail('some-check', input)

// 监听(返回值阻止后续监听器)
ctx.on('some-check', (input) => {
  if (shouldBlock(input)) return 'blocked'
  // 返回 undefined 继续传递给下一个监听器
  return undefined
})

serial — 顺序执行

所有监听器按注册顺序依次执行(异步安全):

import type { Context } from 'cordis'

declare module 'cordis' {
  interface Events {
    'setup-phase'(context: object): Promise<void> | void
  }
}

declare const ctx: Context
declare const context: object

await ctx.serial('setup-phase', context)

waterfall — 管道

监听器围绕默认实现层层包裹,形成数据管道。必须调用 next() 委托给下游,不调用即为否决:

import type { Context } from 'cordis'
import type { Message } from '@deepseek-ai/dsh-llm'

declare module 'cordis' {
  interface Events {
    'my-plugin/messages'(messages: Message[], next: () => Promise<Message[]>): Promise<Message[]>
  }
}

declare const ctx: Context
declare const messages: Message[]
declare const extraMessage: Message

// 触发:最后一个参数是默认实现(所有监听器都调用 next 时的最终值)
const finalMessages = await ctx.waterfall('my-plugin/messages', messages, async () => messages)

// 监听(必须调用 next
ctx.on('my-plugin/messages', async (messages, next) => {
  // next() 委托给下游监听器(最终到达默认实现),返回值可以被加工
  const result = await next()
  return [...result, extraMessage]
})

::: warning Waterfall 监听器必须调用 next()。不调用 next 等于否决整个管道,这是故意为之的设计——用于实现拦截/网关逻辑。 :::

Typed Events

Harness 使用 TypeScript 声明合并来为事件提供类型安全:

import type {} from 'cordis'

declare module 'cordis' {
  interface Events {
    'my-plugin/ready'(payload: { id: string }): void
    'my-plugin/check'(input: string): boolean | undefined
  }
}

// 现在 ctx.on('my-plugin/ready', ...) 和 ctx.emit('my-plugin/ready', ...)
// 都有正确的类型推导

命名约定

Harness 事件遵循 namespace/action 命名:

agent/pre-step       — 每个 step 开始前的检查点(serial
agent/step-result    — step 的 assistant 消息组装完成(waterfall
tools/pre-execute    — tool 执行前的允许/拒绝门(waterfall
tools/post-execute   — tool 执行后的检查/改写缝(waterfall
llm/stream           — 每次流式模型调用的环绕点(waterfall
session/event        — 会话事件被记录(emit
session/flush        — 会话持久化检查点(parallel)

完整的事件列表(含每个事件的签名与派发模式)见仓库中的 docs/cordis-catalog/events.md

事件也是效果

通过 ctx.on() 注册的监听器会在插件卸载时自动移除:

import type { Context } from 'cordis'
import type { Agent, AgentStatus } from '@deepseek-ai/dsh-agent'

declare function handler(agent: Agent, status: AgentStatus): void

export function apply(ctx: Context) {
  // 这个监听器在插件 dispose 时自动清理
  ctx.on('agent/status', handler)
}

实战示例:日志插件

一个记录所有 tool 调用的简单插件:

import type { Context } from 'cordis'
import type {} from '@deepseek-ai/dsh-tools'

export const name = 'tool-logger'

export function apply(ctx: Context) {
  ctx.on('tools/execute', async (exec, next) => {
    console.log(`[tool] ${exec.name}(${JSON.stringify(exec.arguments)})`)
    const result = await next()
    const text = result.content
      .map(b => b.type === 'text' ? b.text : '')
      .join('')
    console.log(`[tool result] ${text.slice(0, 100)}`)
    return result
  })
}

下一步