fix: harden Hermes stream recovery around tool boundaries (#189)

2026-04-24 15:42:42 +02:00
parent edd41e6eb7
commit 009acc1c28
6 changed files with 496 additions and 114 deletions
@@ -0,0 +1,59 @@
+// @vitest-environment jsdom
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+import { streamRunEvents, type RunEvent } from '@/api/hermes/chat'
+
+class MockEventSource {
+  static instances: MockEventSource[] = []
+
+  url: string
+  onmessage: ((event: { data: string }) => void) | null = null
+  onerror: (() => void) | null = null
+  close = vi.fn()
+
+  constructor(url: string) {
+    this.url = url
+    MockEventSource.instances.push(this)
+  }
+
+  emit(data: string) {
+    this.onmessage?.({ data })
+  }
+
+  fail() {
+    this.onerror?.()
+  }
+}
+
+describe('streamRunEvents', () => {
+  beforeEach(() => {
+    window.localStorage.clear()
+    MockEventSource.instances = []
+    vi.stubGlobal('EventSource', MockEventSource)
+  })
+
+  it('maps non-JSON EventSource data to message.delta so raw text is rendered', () => {
+    const events: RunEvent[] = []
+
+    streamRunEvents('run-raw', event => events.push(event), vi.fn(), vi.fn())
+    MockEventSource.instances[0].emit('原因：raw fallback')
+
+    expect(events).toEqual([{ event: 'message.delta', delta: '原因：raw fallback' }])
+  })
+
+  it('parses colon-containing JSON deltas and closes on completion', () => {
+    const events: RunEvent[] = []
+    const onDone = vi.fn()
+
+    streamRunEvents('run-json', event => events.push(event), onDone, vi.fn())
+    const source = MockEventSource.instances[0]
+    source.emit(JSON.stringify({ event: 'message.delta', delta: '让我直接读文件：A: B' }))
+    source.emit(JSON.stringify({ event: 'run.completed' }))
+
+    expect(events).toEqual([
+      { event: 'message.delta', delta: '让我直接读文件：A: B' },
+      { event: 'run.completed' },
+    ])
+    expect(source.close).toHaveBeenCalledTimes(1)
+    expect(onDone).toHaveBeenCalledTimes(1)
+  })
+})
@@ -41,13 +41,36 @@ function makeSummary(id: string, title = 'Session') {
  }
 }

-function makeDetail(id: string, messages: Array<Record<string, any>>) {
+function makeDetail(id: string, messages: Array<Record<string, any>>, overrides: Record<string, any> = {}) {
  return {
    ...makeSummary(id),
+    ...overrides,
    messages,
  }
 }

+function makeHermesMessage(
+  id: number,
+  role: 'user' | 'assistant' | 'system' | 'tool',
+  content: string,
+  overrides: Record<string, any> = {},
+) {
+  return {
+    id,
+    session_id: overrides.session_id || 'sess-1',
+    role,
+    content,
+    tool_call_id: null,
+    tool_calls: null,
+    tool_name: null,
+    timestamp: 1710000000 + id,
+    token_count: null,
+    finish_reason: null,
+    reasoning: null,
+    ...overrides,
+  }
+}
+
 async function flushPromises() {
  await Promise.resolve()
  await Promise.resolve()
@@ -294,4 +317,186 @@ describe('Chat Store', () => {
    expect(store.isRunActive).toBe(false)
    expect(window.localStorage.getItem(inFlightKey('sess-1'))).toBeNull()
  })
+
+  it('keeps colon deltas before and after a tool boundary', async () => {
+    mockChatApi.streamRunEvents.mockImplementation((
+      _runId: string,
+      onEvent: (event: any) => void,
+    ) => {
+      onEvent({ event: 'message.delta', delta: '让我直接读文件：' })
+      onEvent({ event: 'tool.started', tool: 'read_file', preview: 'notes.md' })
+      onEvent({ event: 'tool.completed' })
+      onEvent({ event: 'message.delta', delta: '读取后结论: final' })
+      onEvent({ event: 'run.completed' })
+      return { abort: vi.fn() }
+    })
+
+    const store = useChatStore()
+    await flushPromises()
+    await store.sendMessage('check file')
+    await flushPromises()
+
+    const assistantText = store.messages
+      .filter(m => m.role === 'assistant')
+      .map(m => m.content)
+      .join('')
+    expect(assistantText).toBe('让我直接读文件：读取后结论: final')
+    expect(store.messages.some(m => m.role === 'tool' && m.toolName === 'read_file' && m.toolStatus === 'done')).toBe(true)
+  })
+
+  it('renders raw SSE fallback message events as assistant deltas', async () => {
+    mockChatApi.streamRunEvents.mockImplementation((
+      _runId: string,
+      onEvent: (event: any) => void,
+    ) => {
+      onEvent({ event: 'message', delta: '原因：raw fallback' })
+      onEvent({ event: 'run.completed' })
+      return { abort: vi.fn() }
+    })
+
+    const store = useChatStore()
+    await flushPromises()
+    await store.sendMessage('raw stream')
+    await flushPromises()
+
+    expect(store.messages.some(m => m.role === 'assistant' && m.content === '原因：raw fallback')).toBe(true)
+  })
+
+  it('does not stop polling when server messages are stable but the session is still active', async () => {
+    vi.useFakeTimers()
+
+    let fetchSessionCalls = 0
+    mockSessionsApi.fetchSession.mockImplementation(async () => {
+      fetchSessionCalls += 1
+      if (fetchSessionCalls === 1) return null
+      return makeDetail('sess-1', [
+        makeHermesMessage(1, 'user', 'tool gap prompt'),
+        makeHermesMessage(2, 'assistant', '让我直接读文件：'),
+      ], { ended_at: null })
+    })
+
+    mockChatApi.streamRunEvents.mockImplementation((
+      _runId: string,
+      onEvent: (event: any) => void,
+      _onDone: () => void,
+      onError: (err: Error) => void,
+    ) => {
+      onEvent({ event: 'message.delta', delta: '让我直接读文件：' })
+      setTimeout(() => onError(new Error('SSE connection error')), 0)
+      return { abort: vi.fn() }
+    })
+
+    const store = useChatStore()
+    await flushPromises()
+    await store.sendMessage('tool gap prompt')
+    const sid = store.activeSessionId!
+
+    await vi.advanceTimersByTimeAsync(0)
+    await flushPromises()
+    await vi.advanceTimersByTimeAsync(9000)
+    await flushPromises()
+
+    expect(window.localStorage.getItem(inFlightKey(sid))).toBeTruthy()
+    expect(store.isRunActive).toBe(true)
+  })
+
+  it('reconciles the final session after run.completed to recover missed deltas', async () => {
+    let fetchSessionCalls = 0
+    mockSessionsApi.fetchSession.mockImplementation(async () => {
+      fetchSessionCalls += 1
+      if (fetchSessionCalls === 1) return null
+      return makeDetail('sess-1', [
+        makeHermesMessage(1, 'user', 'finish prompt'),
+        makeHermesMessage(2, 'assistant', '让我直接读文件：读取后结论：完整回答'),
+      ], { ended_at: 1710000010 })
+    })
+
+    mockChatApi.streamRunEvents.mockImplementation((
+      _runId: string,
+      onEvent: (event: any) => void,
+    ) => {
+      onEvent({ event: 'message.delta', delta: '让我直接读文件：' })
+      onEvent({ event: 'run.completed' })
+      return { abort: vi.fn() }
+    })
+
+    const store = useChatStore()
+    await flushPromises()
+    await store.sendMessage('finish prompt')
+    await flushPromises()
+
+    expect(store.messages.some(m => m.role === 'assistant' && m.content === '让我直接读文件：读取后结论：完整回答')).toBe(true)
+  })
+
+  it('does not replace longer local tool-boundary text with a stale shorter final fetch', async () => {
+    let fetchSessionCalls = 0
+    const stalePrefix = '让我直接读文件：较长的工具前说明'
+    mockSessionsApi.fetchSession.mockImplementation(async () => {
+      fetchSessionCalls += 1
+      if (fetchSessionCalls === 1) return null
+      return makeDetail('sess-1', [
+        makeHermesMessage(1, 'user', 'stale prompt'),
+        makeHermesMessage(2, 'assistant', stalePrefix),
+      ], { ended_at: 1710000010 })
+    })
+
+    mockChatApi.streamRunEvents.mockImplementation((
+      _runId: string,
+      onEvent: (event: any) => void,
+    ) => {
+      onEvent({ event: 'message.delta', delta: stalePrefix })
+      onEvent({ event: 'tool.started', tool: 'read_file', preview: 'notes.md' })
+      onEvent({ event: 'tool.completed' })
+      onEvent({ event: 'message.delta', delta: 'OK' })
+      onEvent({ event: 'run.completed' })
+      return { abort: vi.fn() }
+    })
+
+    const store = useChatStore()
+    await flushPromises()
+    await store.sendMessage('stale prompt')
+    await flushPromises()
+
+    const assistantText = store.messages
+      .filter(m => m.role === 'assistant')
+      .map(m => m.content)
+      .join('')
+    expect(assistantText).toBe(`${stalePrefix}OK`)
+    expect(store.messages.some(m => m.role === 'tool' && m.toolStatus === 'done')).toBe(true)
+  })
+
+  it('does not let delayed completion reconciliation clear a newer in-flight run', async () => {
+    let resolveReconcile: ((detail: any) => void) | null = null
+    const reconcilePromise = new Promise<any>(resolve => { resolveReconcile = resolve })
+    mockSessionsApi.fetchSession.mockImplementation(() => reconcilePromise)
+    mockChatApi.startRun
+      .mockResolvedValueOnce({ run_id: 'run-1', status: 'queued' })
+      .mockResolvedValueOnce({ run_id: 'run-2', status: 'queued' })
+    let firstRunEvent: ((event: any) => void) | null = null
+    mockChatApi.streamRunEvents.mockImplementation((
+      runId: string,
+      onEvent: (event: any) => void,
+    ) => {
+      if (runId === 'run-1') firstRunEvent = onEvent
+      return { abort: vi.fn() }
+    })
+
+    const store = useChatStore()
+    await flushPromises()
+    await store.sendMessage('first')
+    firstRunEvent!({ event: 'run.completed' })
+    await flushPromises()
+    const sid = store.activeSessionId!
+    await store.sendMessage('second')
+
+    expect(JSON.parse(window.localStorage.getItem(inFlightKey(sid)) || '{}').runId).toBe('run-2')
+
+    resolveReconcile!(makeDetail(sid, [
+      makeHermesMessage(1, 'user', 'first', { session_id: sid }),
+      makeHermesMessage(2, 'assistant', 'first done', { session_id: sid }),
+    ], { ended_at: 1710000010 }))
+    await flushPromises()
+
+    expect(JSON.parse(window.localStorage.getItem(inFlightKey(sid)) || '{}').runId).toBe('run-2')
+  })
 })