feat: bridge Telegram updates to harness agents

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-14 16:22:46 +08:00
parent 4a9a41484e
commit 2b800da381
2 changed files with 587 additions and 0 deletions
@@ -0,0 +1,270 @@
+import type { Context } from '@deepseek-ai/cordis'
+import type { AgentHandle } from '@deepseek-ai/dsh-agent'
+import { createUserMessage } from '@deepseek-ai/dsh-llm/message'
+import type { ContentBlock, TextBlock } from '@deepseek-ai/dsh-llm/types'
+import { SessionId, type SessionEvent } from '@deepseek-ai/dsh-session/types'
+import { isAuthorized } from './auth.ts'
+import { TelegramClient, type TelegramClientLike, type TelegramUpdate } from './client.ts'
+import { MSG, parseCommand } from './commands.ts'
+import { markdownToHtml, splitMessage } from './format.ts'
+
+export interface TelegramBridgeOptions {
+  token: string
+  allowedUserIds: number[]
+  allowAllUsers: boolean
+  client?: TelegramClientLike
+  sleep?: (ms: number) => Promise<void>
+  provider?: string
+  model?: string
+  maxMessageLength?: number
+  cwd?: string
+}
+
+interface ChatState {
+  chatId: number
+  handle: AgentHandle
+  sessionId: ReturnType<typeof SessionId>
+}
+
+interface SessionLike {
+  id: ReturnType<typeof SessionId>
+}
+
+const defaultSleep = (ms: number) => new Promise<void>((resolve) => setTimeout(resolve, ms))
+
+function contentToText(content: readonly ContentBlock[]): string {
+  return content
+    .filter((block): block is TextBlock => block.type === 'text')
+    .map((block) => block.text)
+    .join('')
+}
+
+export class TelegramBridge {
+  private readonly ctx: Context
+  private readonly token: string
+  private readonly allowedUserIds: number[]
+  private readonly allowAllUsers: boolean
+  private readonly client: TelegramClientLike
+  private readonly sleep: (ms: number) => Promise<void>
+  private readonly provider: string
+  private readonly model: string
+  private readonly maxMessageLength: number
+  private readonly cwd: string
+
+  private readonly chats = new Map<string, ChatState>()
+  private polling = false
+  private offset: number | undefined
+  private pollPromise: Promise<void> | undefined
+  private pollAbort: AbortController | undefined
+  private disposeSessionListener: (() => void) | undefined
+
+  constructor(ctx: Context, options: TelegramBridgeOptions) {
+    this.ctx = ctx
+    this.token = options.token
+    this.allowedUserIds = options.allowedUserIds
+    this.allowAllUsers = options.allowAllUsers
+    this.client = options.client ?? new TelegramClient(options.token)
+    this.sleep = options.sleep ?? defaultSleep
+    this.provider = options.provider ?? 'deepseek-official'
+    this.model = options.model ?? 'deepseek-v4-flash'
+    this.maxMessageLength = options.maxMessageLength ?? 4096
+    this.cwd = options.cwd ?? process.cwd()
+  }
+
+  start(): void {
+    this.disposeSessionListener?.()
+    this.disposeSessionListener = this.ctx.on('session/event', (session, event) => {
+      void this.onSessionEvent(session, event)
+    })
+    if (!this.polling) {
+      this.polling = true
+      this.pollAbort = new AbortController()
+      this.pollPromise = this.pollLoop()
+    }
+  }
+
+  async stop(): Promise<void> {
+    this.polling = false
+    this.pollAbort?.abort()
+    this.pollAbort = undefined
+    this.disposeSessionListener?.()
+    this.disposeSessionListener = undefined
+    const disposals = [...this.chats.values()].map((chat) => chat.handle.dispose())
+    this.chats.clear()
+    await Promise.all(disposals)
+    if (this.pollPromise) {
+      await this.pollPromise.catch(() => {})
+      this.pollPromise = undefined
+    }
+  }
+
+  async processUpdate(update: TelegramUpdate): Promise<void> {
+    const message = update.message
+    if (!message?.text) return
+
+    const chatId = message.chat.id
+    const userId = message.from?.id
+
+    if (!isAuthorized({ allowAllUsers: this.allowAllUsers, allowedUserIds: this.allowedUserIds, userId })) {
+      await this.client.sendMessage(chatId, MSG.DENIED)
+      return
+    }
+
+    const parsed = parseCommand(message.text)
+    switch (parsed.type) {
+      case 'start':
+        await this.client.sendMessage(chatId, MSG.WELCOME)
+        return
+      case 'help':
+        await this.client.sendMessage(chatId, MSG.HELP)
+        return
+      case 'new':
+        await this.resetChat(chatId)
+        await this.client.sendMessage(chatId, MSG.NEW_SESSION)
+        return
+      case 'unknown':
+        await this.client.sendMessage(chatId, MSG.unknown(parsed.command))
+        return
+      case 'plain':
+        await this.ensureChat(chatId)
+        await this.followup(chatId, parsed.text)
+        return
+    }
+  }
+
+  private async pollLoop(): Promise<void> {
+    const signal = this.pollAbort?.signal
+    let errorCount = 0
+    while (this.polling) {
+      try {
+        const updates = await this.client.getUpdates(this.offset)
+        if (!this.polling) break
+        errorCount = 0
+        if (updates.length === 0) {
+          await this.interruptibleDelay(50, signal)
+          continue
+        }
+        for (const update of updates) {
+          if (!this.polling) break
+          await this.processUpdate(update)
+          this.offset = update.update_id + 1
+        }
+      } catch (err) {
+        if (!this.polling) break
+        errorCount += 1
+        this.ctx.logger.error(this.redact(err))
+        await this.interruptibleSleep(Math.min(1000 * errorCount, 10_000), signal)
+      }
+    }
+  }
+
+  private interruptibleDelay(ms: number, signal?: AbortSignal): Promise<void> {
+    if (signal?.aborted || !this.polling) return Promise.resolve()
+    return new Promise((resolve) => {
+      const timer = setTimeout(resolve, ms)
+      signal?.addEventListener('abort', () => {
+        clearTimeout(timer)
+        resolve()
+      }, { once: true })
+    })
+  }
+
+  private async interruptibleSleep(ms: number, signal?: AbortSignal): Promise<void> {
+    if (signal?.aborted || !this.polling) return
+    await Promise.race([
+      this.sleep(ms),
+      new Promise<void>((resolve) => {
+        if (signal?.aborted) {
+          resolve()
+          return
+        }
+        signal?.addEventListener('abort', () => resolve(), { once: true })
+      }),
+    ])
+  }
+
+  private async onSessionEvent(session: SessionLike, event: SessionEvent): Promise<void> {
+    const chat = this.findChatBySessionId(session.id)
+    if (!chat) return
+
+    if (event.type === 'turn/start') {
+      await this.client.sendChatAction(chat.chatId, 'typing')
+      return
+    }
+
+    if (event.type === 'assistant/message') {
+      const text = contentToText(event.data.message.content)
+      if (text) {
+        await this.deliver(chat.chatId, text)
+      }
+    }
+  }
+
+  private findChatBySessionId(sessionId: ReturnType<typeof SessionId>): ChatState | undefined {
+    for (const chat of this.chats.values()) {
+      if (chat.sessionId === sessionId) return chat
+    }
+    return undefined
+  }
+
+  private async ensureChat(chatId: number): Promise<ChatState> {
+    const key = String(chatId)
+    const existing = this.chats.get(key)
+    if (existing) return existing
+
+    const sessionId = SessionId(`telegram:${chatId}`)
+    const handle = await this.createAgent(sessionId)
+    const state: ChatState = { chatId, handle, sessionId }
+    this.chats.set(key, state)
+    return state
+  }
+
+  private async resetChat(chatId: number): Promise<void> {
+    const key = String(chatId)
+    const existing = this.chats.get(key)
+    if (existing) {
+      await existing.handle.dispose()
+      this.chats.delete(key)
+    }
+    const sessionId = SessionId(`telegram:${chatId}:${Date.now()}`)
+    const handle = await this.createAgent(sessionId)
+    this.chats.set(key, { chatId, handle, sessionId })
+  }
+
+  private async createAgent(sessionId: ReturnType<typeof SessionId>): Promise<AgentHandle> {
+    return this.ctx.agents.create({
+      sessionId,
+      meta: { cwd: this.cwd },
+      agentOptions: {
+        provider: this.provider,
+        model: this.model,
+      },
+    })
+  }
+
+  private async followup(chatId: number, text: string): Promise<void> {
+    const chat = await this.ensureChat(chatId)
+    const message = createUserMessage({
+      content: [{ type: 'text', text }],
+      source: { kind: 'user' },
+    })
+    chat.handle.agent.followup(message)
+  }
+
+  private async deliver(chatId: number, markdown: string): Promise<void> {
+    const chunks = splitMessage(markdown, this.maxMessageLength)
+    for (const chunk of chunks) {
+      const html = markdownToHtml(chunk)
+      try {
+        await this.client.sendMessage(chatId, html, 'HTML')
+      } catch {
+        await this.client.sendMessage(chatId, chunk)
+      }
+    }
+  }
+
+  private redact(value: unknown): string {
+    const message = value instanceof Error ? value.message : String(value)
+    return message.split(this.token).join('***')
+  }
+}
@@ -0,0 +1,317 @@
+import assert from 'node:assert/strict'
+import test from 'node:test'
+import type { CreateAgentOptions } from '@deepseek-ai/dsh-agent'
+import type { UserMessage } from '@deepseek-ai/dsh-llm/types'
+import { SessionId } from '@deepseek-ai/dsh-session/types'
+import { TelegramBridge } from '../src/bridge.ts'
+import type { TelegramClientLike, TelegramUpdate } from '../src/client.ts'
+import { MSG } from '../src/commands.ts'
+
+type SentMessage = { chatId: number; text: string; parseMode?: string }
+
+function fakeClient(
+  sent: SentMessage[],
+  overrides: Partial<TelegramClientLike> = {},
+): TelegramClientLike {
+  return {
+    getMe: async () => ({ id: 1 }),
+    getUpdates: async () => [],
+    sendMessage: async (chatId, text, parseMode) => {
+      sent.push({ chatId, text, parseMode })
+      return { message_id: 1, date: 0, chat: { id: chatId, type: 'private' }, text }
+    },
+    sendChatAction: async () => true,
+    ...overrides,
+  }
+}
+
+function messageUpdate(chatId: number, userId: number, text: string, updateId = 1): TelegramUpdate {
+  return {
+    update_id: updateId,
+    message: {
+      message_id: updateId,
+      date: 0,
+      chat: { id: chatId, type: 'private' },
+      from: { id: userId },
+      text,
+    },
+  }
+}
+
+function fakeAgents(track: {
+  createCount: number
+  disposeCount: number
+  sessionIds: string[]
+  followups: UserMessage[][]
+}) {
+  return {
+    create: async (options: CreateAgentOptions) => {
+      track.createCount += 1
+      track.sessionIds.push(String(options.sessionId))
+      const followups: UserMessage[] = []
+      track.followups.push(followups)
+      return {
+        agent: {
+          followup(message: UserMessage) {
+            followups.push(message)
+          },
+        },
+        dispose: async () => {
+          track.disposeCount += 1
+        },
+      }
+    },
+  }
+}
+
+function bridgeCtx(
+  agents: ReturnType<typeof fakeAgents>,
+  onSessionEvent?: (session: { id: ReturnType<typeof SessionId> }, event: unknown) => void,
+) {
+  return {
+    logger: { warn() {}, error() {} },
+    agents,
+    on(event: string, listener: (session: { id: ReturnType<typeof SessionId> }, event: unknown) => void) {
+      if (event === 'session/event' && onSessionEvent) {
+        onSessionEvent = listener
+      }
+      return () => {}
+    },
+    _fireSessionEvent(session: { id: ReturnType<typeof SessionId> }, ev: unknown) {
+      onSessionEvent?.(session, ev)
+    },
+  }
+}
+
+test('unauthorized user gets denied and no agent', async () => {
+  const sent: SentMessage[] = []
+  let createCount = 0
+  const ctx = {
+    logger: { warn() {}, error() {} },
+    agents: {
+      create: async () => {
+        createCount += 1
+        throw new Error('should not create')
+      },
+    },
+    on() {
+      return () => {}
+    },
+  }
+  const bridge = new TelegramBridge(ctx as never, {
+    token: 't',
+    allowedUserIds: [1],
+    allowAllUsers: false,
+    client: fakeClient(sent),
+    sleep: async () => {},
+  })
+  await bridge.processUpdate(messageUpdate(99, 99, 'hi'))
+  assert.equal(createCount, 0)
+  assert.equal(sent[0]?.text, MSG.DENIED)
+})
+
+test('/help sends help without creating agent', async () => {
+  const sent: SentMessage[] = []
+  const track = { createCount: 0, disposeCount: 0, sessionIds: [] as string[], followups: [] as UserMessage[][] }
+  const bridge = new TelegramBridge(bridgeCtx(fakeAgents(track)) as never, {
+    token: 't',
+    allowedUserIds: [1],
+    allowAllUsers: false,
+    client: fakeClient(sent),
+    sleep: async () => {},
+  })
+  await bridge.processUpdate(messageUpdate(42, 1, '/help'))
+  assert.equal(track.createCount, 0)
+  assert.equal(sent[0]?.text, MSG.HELP)
+})
+
+test('first plain text creates agent then followups', async () => {
+  const sent: SentMessage[] = []
+  const track = { createCount: 0, disposeCount: 0, sessionIds: [] as string[], followups: [] as UserMessage[][] }
+  const bridge = new TelegramBridge(bridgeCtx(fakeAgents(track)) as never, {
+    token: 't',
+    allowedUserIds: [1],
+    allowAllUsers: false,
+    client: fakeClient(sent),
+    sleep: async () => {},
+  })
+  await bridge.processUpdate(messageUpdate(42, 1, 'hello'))
+  assert.equal(track.createCount, 1)
+  assert.equal(track.sessionIds[0], 'telegram:42')
+  assert.equal(track.followups.length, 1)
+  assert.equal((track.followups[0]![0]!.content[0] as { text: string }).text, 'hello')
+})
+
+test('second plain text in same chat only followups', async () => {
+  const sent: SentMessage[] = []
+  const track = { createCount: 0, disposeCount: 0, sessionIds: [] as string[], followups: [] as UserMessage[][] }
+  const bridge = new TelegramBridge(bridgeCtx(fakeAgents(track)) as never, {
+    token: 't',
+    allowedUserIds: [1],
+    allowAllUsers: false,
+    client: fakeClient(sent),
+    sleep: async () => {},
+  })
+  await bridge.processUpdate(messageUpdate(42, 1, 'hello'))
+  await bridge.processUpdate(messageUpdate(42, 1, 'again', 2))
+  assert.equal(track.createCount, 1)
+  assert.equal(track.followups.length, 1)
+  assert.equal(track.followups[0]!.length, 2)
+  assert.equal((track.followups[0]![1]!.content[0] as { text: string }).text, 'again')
+})
+
+test('/new disposes previous handle and creates fresh session', async () => {
+  const sent: SentMessage[] = []
+  const track = { createCount: 0, disposeCount: 0, sessionIds: [] as string[], followups: [] as UserMessage[][] }
+  const bridge = new TelegramBridge(bridgeCtx(fakeAgents(track)) as never, {
+    token: 't',
+    allowedUserIds: [1],
+    allowAllUsers: false,
+    client: fakeClient(sent),
+    sleep: async () => {},
+  })
+  await bridge.processUpdate(messageUpdate(42, 1, 'hello'))
+  await bridge.processUpdate(messageUpdate(42, 1, '/new', 2))
+  assert.equal(track.disposeCount, 1)
+  assert.equal(track.createCount, 2)
+  assert.equal(track.sessionIds[0], 'telegram:42')
+  assert.match(track.sessionIds[1]!, /^telegram:42:\d+$/)
+  assert.equal(sent.at(-1)?.text, MSG.NEW_SESSION)
+})
+
+test('turn/start sends typing for matching session', async () => {
+  const sent: SentMessage[] = []
+  const actions: Array<{ chatId: number; action: string }> = []
+  const track = { createCount: 0, disposeCount: 0, sessionIds: [] as string[], followups: [] as UserMessage[][] }
+  let fire: ((session: { id: ReturnType<typeof SessionId> }, event: unknown) => void) | undefined
+  const ctx = {
+    logger: { warn() {}, error() {} },
+    agents: fakeAgents(track),
+    on(_event: string, listener: (session: { id: ReturnType<typeof SessionId> }, event: unknown) => void) {
+      fire = listener
+      return () => {}
+    },
+  }
+  const bridge = new TelegramBridge(ctx as never, {
+    token: 't',
+    allowedUserIds: [1],
+    allowAllUsers: false,
+    client: fakeClient(sent, {
+      sendChatAction: async (chatId, action) => {
+        actions.push({ chatId, action })
+        return true
+      },
+    }),
+    sleep: async () => {},
+  })
+  bridge.start()
+  await bridge.processUpdate(messageUpdate(42, 1, 'hello'))
+  fire!({ id: SessionId(track.sessionIds[0]!) }, { type: 'turn/start', data: { turn: 1 } })
+  await new Promise((resolve) => setImmediate(resolve))
+  assert.deepEqual(actions, [{ chatId: 42, action: 'typing' }])
+  await bridge.stop()
+})
+
+test('assistant/message delivers HTML chunks for matching session', async () => {
+  const sent: SentMessage[] = []
+  const track = { createCount: 0, disposeCount: 0, sessionIds: [] as string[], followups: [] as UserMessage[][] }
+  let fire: ((session: { id: ReturnType<typeof SessionId> }, event: unknown) => void) | undefined
+  const ctx = {
+    logger: { warn() {}, error() {} },
+    agents: fakeAgents(track),
+    on(_event: string, listener: (session: { id: ReturnType<typeof SessionId> }, event: unknown) => void) {
+      fire = listener
+      return () => {}
+    },
+  }
+  const bridge = new TelegramBridge(ctx as never, {
+    token: 't',
+    allowedUserIds: [1],
+    allowAllUsers: false,
+    client: fakeClient(sent),
+    sleep: async () => {},
+  })
+  bridge.start()
+  await bridge.processUpdate(messageUpdate(42, 1, 'hello'))
+  fire!(
+    { id: SessionId(track.sessionIds[0]!) },
+    {
+      type: 'assistant/message',
+      data: {
+        turn: 1,
+        step: 1,
+        message: {
+          content: [{ type: 'text', text: '**bold**' }],
+        },
+      },
+    },
+  )
+  await new Promise((resolve) => setImmediate(resolve))
+  const reply = sent.find((m) => m.parseMode === 'HTML')
+  assert.ok(reply)
+  assert.match(reply!.text, /<b>bold<\/b>/)
+  await bridge.stop()
+})
+
+test('HTML send failure retries plain text for that chunk', async () => {
+  const sent: SentMessage[] = []
+  const track = { createCount: 0, disposeCount: 0, sessionIds: [] as string[], followups: [] as UserMessage[][] }
+  let fire: ((session: { id: ReturnType<typeof SessionId> }, event: unknown) => void) | undefined
+  const ctx = {
+    logger: { warn() {}, error() {} },
+    agents: fakeAgents(track),
+    on(_event: string, listener: (session: { id: ReturnType<typeof SessionId> }, event: unknown) => void) {
+      fire = listener
+      return () => {}
+    },
+  }
+  const bridge = new TelegramBridge(ctx as never, {
+    token: 't',
+    allowedUserIds: [1],
+    allowAllUsers: false,
+    client: fakeClient(sent, {
+      sendMessage: async (chatId, text, parseMode) => {
+        if (parseMode === 'HTML') throw new Error('bad html')
+        sent.push({ chatId, text, parseMode })
+        return { message_id: 1, date: 0, chat: { id: chatId, type: 'private' }, text }
+      },
+    }),
+    sleep: async () => {},
+  })
+  bridge.start()
+  await bridge.processUpdate(messageUpdate(42, 1, 'hello'))
+  fire!(
+    { id: SessionId(track.sessionIds[0]!) },
+    {
+      type: 'assistant/message',
+      data: {
+        turn: 1,
+        step: 1,
+        message: {
+          content: [{ type: 'text', text: 'plain chunk' }],
+        },
+      },
+    },
+  )
+  await new Promise((resolve) => setImmediate(resolve))
+  assert.equal(sent.length, 1)
+  assert.equal(sent[0]?.parseMode, undefined)
+  assert.equal(sent[0]?.text, 'plain chunk')
+  await bridge.stop()
+})
+
+test('stop disposes all chat handles', async () => {
+  const sent: SentMessage[] = []
+  const track = { createCount: 0, disposeCount: 0, sessionIds: [] as string[], followups: [] as UserMessage[][] }
+  const bridge = new TelegramBridge(bridgeCtx(fakeAgents(track)) as never, {
+    token: 't',
+    allowedUserIds: [1],
+    allowAllUsers: false,
+    client: fakeClient(sent),
+    sleep: async () => {},
+  })
+  await bridge.processUpdate(messageUpdate(42, 1, 'one'))
+  await bridge.processUpdate(messageUpdate(43, 1, 'two', 2))
+  await bridge.stop()
+  assert.equal(track.disposeCount, 2)
+})