feat: preserve pages and compress context

2026-08-15 20:44:47 +08:00
parent 1031ea618b
commit c2b9ce4ff0
23 changed files with 1972 additions and 507 deletions
@@ -72,19 +72,37 @@ describe('context compression planning', () => {

  it('uses an optional model context limit as an earlier trigger', () => {
    const history = [
-      { role: 'user' as const, content: 'a'.repeat(14_000) },
-      { role: 'assistant' as const, content: 'b'.repeat(14_000) },
-      { role: 'user' as const, content: 'c'.repeat(14_000) },
-      { role: 'assistant' as const, content: 'd'.repeat(14_000) }
+      { role: 'user' as const, content: 'a'.repeat(16_000) },
+      { role: 'assistant' as const, content: 'b'.repeat(16_000) },
+      { role: 'user' as const, content: 'c'.repeat(16_000) },
+      { role: 'assistant' as const, content: 'd'.repeat(16_000) }
    ]
    const plan = planContextCompression({
      history,
      prompt: 'Continue',
      settings: compressionSettings(),
-      contextWindowTokens: 30_000
+      contextWindowTokens: 32_000
    })

-    expect(plan?.effectiveTriggerTokens).toBe(18_000)
+    expect(plan?.effectiveTriggerTokens).toBe(20_000)
+    expect(plan?.earlierMessages.length).toBeGreaterThan(0)
+  })
+
+  it('defensively clamps legacy undersized context limits', () => {
+    const history = [
+      { role: 'user' as const, content: 'a'.repeat(40_000) },
+      { role: 'assistant' as const, content: 'b'.repeat(40_000) },
+      { role: 'user' as const, content: 'c'.repeat(40_000) },
+      { role: 'assistant' as const, content: 'd'.repeat(40_000) }
+    ]
+    const plan = planContextCompression({
+      history,
+      prompt: 'Continue',
+      settings: compressionSettings(),
+      contextWindowTokens: 10_000
+    })
+
+    expect(plan?.effectiveTriggerTokens).toBe(20_000)
    expect(plan?.earlierMessages.length).toBeGreaterThan(0)
  })
 })
@@ -1,4 +1,14 @@
 import type { ContextCompressionSettings } from '../../shared/contracts'
+import {
+  estimateContextInputTokens,
+  estimateMessagesTokens,
+  getEffectiveContextTriggerTokens
+} from '../../shared/context-window'
+
+export {
+  estimateMessagesTokens,
+  estimateTextTokens
+} from '../../shared/context-window'

 export type CompressibleConversationMessage = {
  role: 'user' | 'assistant'
@@ -12,34 +22,6 @@ export type ContextCompressionPlan = {
  effectiveTriggerTokens: number
 }

-const reservedOutputAndSafetyTokens = 12_000
-const estimatedRequestOverheadTokens = 4_000
-
-export function estimateTextTokens(value: string): number {
-  let asciiCharacters = 0
-  let nonAsciiCharacters = 0
-  for (const character of value) {
-    if (character.codePointAt(0)! <= 0x7f) {
-      asciiCharacters += 1
-    } else {
-      nonAsciiCharacters += 1
-    }
-  }
-  return Math.max(
-    1,
-    Math.ceil(asciiCharacters / 4 + nonAsciiCharacters)
-  )
-}
-
-export function estimateMessagesTokens(
-  messages: readonly CompressibleConversationMessage[]
-): number {
-  return messages.reduce(
-    (total, message) => total + estimateTextTokens(message.content) + 4,
-    0
-  )
-}
-
 function groupConversationTurns(
  messages: readonly CompressibleConversationMessage[]
 ): CompressibleConversationMessage[][] {
@@ -61,24 +43,19 @@ function groupConversationTurns(
 export function planContextCompression(input: {
  history: readonly CompressibleConversationMessage[]
  prompt: string
+  summaryTokens?: number
  settings: ContextCompressionSettings
  contextWindowTokens?: number
 }): ContextCompressionPlan | undefined {
-  const estimatedInputTokens =
-    estimateMessagesTokens(input.history) +
-    estimateTextTokens(input.prompt) +
-    estimatedRequestOverheadTokens
-  const contextLimitedTrigger =
-    input.contextWindowTokens === undefined
-      ? input.settings.triggerTokens
-      : Math.max(
-          8_000,
-          input.contextWindowTokens - reservedOutputAndSafetyTokens
-        )
-  const effectiveTriggerTokens = Math.min(
-    input.settings.triggerTokens,
-    contextLimitedTrigger
-  )
+  const estimatedInputTokens = estimateContextInputTokens({
+    history: input.history,
+    prompt: input.prompt,
+    summaryTokens: input.summaryTokens
+  })
+  const effectiveTriggerTokens = getEffectiveContextTriggerTokens({
+    triggerTokens: input.settings.triggerTokens,
+    contextWindowTokens: input.contextWindowTokens
+  })
  if (estimatedInputTokens < effectiveTriggerTokens) {
    return undefined
  }
@@ -379,9 +379,30 @@ describe('ModelAgentRuntime', () => {
    expect(answerMessages).toContain('new-user-')
    expect(answerMessages).not.toContain('old-user-')
    expect(events).toContainEqual(
+      expect.objectContaining({
+        type: 'context-compression',
+        state: 'started',
+        estimatedBeforeTokens: expect.any(Number)
+      })
+    )
+    expect(events).toContainEqual(
+      expect.objectContaining({
+        type: 'context-compression',
+        state: 'completed',
+        estimatedAfterTokens: expect.any(Number)
+      })
+    )
+    expect(events).toContainEqual(
+      expect.objectContaining({
+        type: 'context-metrics',
+        coveredMessageCount: 4,
+        summaryTokens: expect.any(Number)
+      })
+    )
+    expect(events).not.toContainEqual(
      expect.objectContaining({
        type: 'status',
-        message: '较早的对话已压缩，正在生成回答'
+        message: '正在准备直连模型上下文'
      })
    )
    expect(events).toContainEqual(
@@ -42,8 +42,13 @@ import {
 import { readBoundedResponseText } from './bounded-response'
 import {
  formatConversationForSummary,
-  planContextCompression
+  planContextCompression,
+  estimateMessagesTokens
 } from './context-compression'
+import {
+  estimateContextInputTokens,
+  getEffectiveContextTriggerTokens
+} from '../../shared/context-window'

 type ConversationMessage = {
  role: 'user' | 'assistant'
@@ -1784,23 +1789,19 @@ export class ModelAgentRuntime implements AgentRuntime {
    return { summary: summary.trim(), usageEvents }
  }

-  private async prepareCompressedRequest(
+  private async *prepareCompressedRequest(
    request: AgentExecutionRequest,
    signal: AbortSignal
-  ): Promise<{
+  ): AsyncGenerator<RuntimeEvent, {
    request: AgentExecutionRequest
    compressed: boolean
-    usageEvents: RuntimeModelUsageEvent[]
-  }> {
+  }, void> {
    const compression = this.options.contextCompression
-    if (
-      !compression?.settings.enabled ||
-      !request.history?.length
-    ) {
-      return { request, compressed: false, usageEvents: [] }
+    if (!compression) {
+      return { request, compressed: false }
    }

-    const history = request.history
+    const history = request.history ?? []
    let state = this.conversationSummaries.get(request.conversationId)
    if (
      state &&
@@ -1815,13 +1816,42 @@ export class ModelAgentRuntime implements AgentRuntime {
    const remainingHistory = history.slice(
      state?.coveredMessageCount ?? 0
    )
+    const requestPrompt = [
+      request.trustedInstructions ?? '',
+      request.prompt
+    ].join('\n')
+    const currentSummaryTokens = state
+      ? estimateMessagesTokens(this.summaryHistory(state.summary))
+      : 0
+    const effectiveTriggerTokens =
+      getEffectiveContextTriggerTokens({
+        triggerTokens: compression.settings.triggerTokens,
+        contextWindowTokens: compression.contextWindowTokens
+      })
+    const estimatedInputTokens = estimateContextInputTokens({
+      history: remainingHistory,
+      prompt: requestPrompt,
+      summaryTokens: currentSummaryTokens
+    })
+    yield {
+      requestId: request.requestId,
+      type: 'context-metrics',
+      estimatedInputTokens,
+      effectiveTriggerTokens,
+      contextWindowTokens: compression.contextWindowTokens,
+      compressionEnabled: compression.settings.enabled,
+      recentRawTokens: compression.settings.recentRawTokens,
+      coveredMessageCount: state?.coveredMessageCount ?? 0,
+      summaryTokens: currentSummaryTokens
+    }
+    if (!compression.settings.enabled || history.length === 0) {
+      return { request, compressed: false }
+    }
+
    const plan = planContextCompression({
      history: remainingHistory,
-      prompt: [
-        state?.summary ?? '',
-        request.trustedInstructions ?? '',
-        request.prompt
-      ].join('\n'),
+      prompt: requestPrompt,
+      summaryTokens: currentSummaryTokens,
      settings: compression.settings,
      contextWindowTokens: compression.contextWindowTokens
    })
@@ -1836,20 +1866,29 @@ export class ModelAgentRuntime implements AgentRuntime {
              ]
            },
            compressed: false,
-            usageEvents: []
          }
-        : { request, compressed: false, usageEvents: [] }
+        : { request, compressed: false }
    }

+    const coveredMessageCount =
+      (state?.coveredMessageCount ?? 0) +
+      plan.earlierMessages.length
+    yield {
+      requestId: request.requestId,
+      type: 'context-compression',
+      state: 'started',
+      estimatedBeforeTokens: plan.estimatedInputTokens,
+      effectiveTriggerTokens: plan.effectiveTriggerTokens,
+      contextWindowTokens: compression.contextWindowTokens,
+      recentRawTokens: compression.settings.recentRawTokens,
+      coveredMessageCount
+    }
    const summarized = await this.summarizeEarlierHistory(
      request,
      plan.earlierMessages,
      state?.summary,
      signal
    )
-    const coveredMessageCount =
-      (state?.coveredMessageCount ?? 0) +
-      plan.earlierMessages.length
    state = {
      coveredMessageCount,
      coveredHistoryDigest: this.historyDigest(
@@ -1858,6 +1897,40 @@ export class ModelAgentRuntime implements AgentRuntime {
      summary: summarized.summary
    }
    this.conversationSummaries.set(request.conversationId, state)
+    for (const usageEvent of summarized.usageEvents) {
+      yield usageEvent
+    }
+    const summaryTokens = estimateMessagesTokens(
+      this.summaryHistory(state.summary)
+    )
+    const estimatedAfterTokens = estimateContextInputTokens({
+      history: plan.recentMessages,
+      prompt: requestPrompt,
+      summaryTokens
+    })
+    yield {
+      requestId: request.requestId,
+      type: 'context-compression',
+      state: 'completed',
+      estimatedBeforeTokens: plan.estimatedInputTokens,
+      estimatedAfterTokens,
+      effectiveTriggerTokens: plan.effectiveTriggerTokens,
+      contextWindowTokens: compression.contextWindowTokens,
+      recentRawTokens: compression.settings.recentRawTokens,
+      coveredMessageCount,
+      summaryTokens
+    }
+    yield {
+      requestId: request.requestId,
+      type: 'context-metrics',
+      estimatedInputTokens: estimatedAfterTokens,
+      effectiveTriggerTokens: plan.effectiveTriggerTokens,
+      contextWindowTokens: compression.contextWindowTokens,
+      compressionEnabled: true,
+      recentRawTokens: compression.settings.recentRawTokens,
+      coveredMessageCount,
+      summaryTokens
+    }
    return {
      request: {
        ...request,
@@ -1866,8 +1939,7 @@ export class ModelAgentRuntime implements AgentRuntime {
          ...plan.recentMessages
        ]
      },
-      compressed: true,
-      usageEvents: summarized.usageEvents
+      compressed: true
    }
  }

@@ -2791,29 +2863,21 @@ export class ModelAgentRuntime implements AgentRuntime {
      throw new Error('当前模型连接未启用图像输入')
    }

-    if (
-      this.options.contextCompression?.settings.enabled &&
-      request.history?.length
-    ) {
-      yield {
-        requestId: request.requestId,
-        type: 'status',
-        message: '正在准备直连模型上下文'
-      }
-    }
-    const prepared = await this.prepareCompressedRequest(
+    const preparation = this.prepareCompressedRequest(
      request,
      signal
    )
-    for (const usageEvent of prepared.usageEvents) {
-      yield usageEvent
+    let prepared: {
+      request: AgentExecutionRequest
+      compressed: boolean
    }
-    if (prepared.compressed) {
-      yield {
-        requestId: request.requestId,
-        type: 'status',
-        message: '较早的对话已压缩，正在生成回答'
+    while (true) {
+      const result = await preparation.next()
+      if (result.done) {
+        prepared = result.value
+        break
      }
+      yield result.value
    }
    const executionRequest = prepared.request

@@ -226,10 +226,9 @@ describe.runIf(enabled)('runtime end-to-end', () => {
        contextCompression: {
          settings: {
            ...defaultContextCompressionSettings,
-            enabled: true,
-            triggerTokens: 8_000,
-            recentRawTokens: 4_000
-          }
+            enabled: true
+          },
+          contextWindowTokens: 32_000
        }
      })
      const events: RuntimeEvent[] = []
@@ -248,7 +247,7 @@ describe.runIf(enabled)('runtime end-to-end', () => {
                content: [
                  'The project codename is ORBIT-739.',
                  'Background notes:',
-                  'alpha '.repeat(1_200)
+                  'alpha '.repeat(5_000)
                ].join('\n')
              },
              {
@@ -256,7 +255,7 @@ describe.runIf(enabled)('runtime end-to-end', () => {
                content: [
                  'I will remember the project codename.',
                  'Acknowledgement notes:',
-                  'gamma '.repeat(1_000)
+                  'gamma '.repeat(4_000)
                ].join('\n')
              },
              {
@@ -264,7 +263,7 @@ describe.runIf(enabled)('runtime end-to-end', () => {
                content: [
                  'The deploy region is AP-SOUTH-7.',
                  'Recent notes:',
-                  'beta '.repeat(900)
+                  'beta '.repeat(3_000)
                ].join('\n')
              },
              {
@@ -289,8 +288,15 @@ describe.runIf(enabled)('runtime end-to-end', () => {
        .join('')
      expect(events).toContainEqual(
        expect.objectContaining({
-          type: 'status',
-          message: '较早的对话已压缩，正在生成回答'
+          type: 'context-compression',
+          state: 'started'
+        })
+      )
+      expect(events).toContainEqual(
+        expect.objectContaining({
+          type: 'context-compression',
+          state: 'completed',
+          estimatedAfterTokens: expect.any(Number)
        })
      )
      expect(events).toContainEqual(
@@ -156,6 +156,61 @@ describe('RuntimeSettingsStore', () => {
    })
  })

+  it('rejects undersized model context windows and repairs legacy values', async () => {
+    const { filePath, store } = await createStore()
+    const profileId = '00000000-0000-4000-8000-000000000062'
+    const profile = {
+      id: profileId,
+      name: 'Small context',
+      baseUrl: 'https://model.example/v1',
+      modelName: 'small-model',
+      protocol: 'openai-responses' as const,
+      authentication: 'api-key' as const,
+      supportsImageInput: false,
+      contextWindowTokens: 10_000,
+      imageGenerationQuality: 'auto' as const,
+      apiKey: { action: 'keep' as const }
+    }
+
+    expect(() =>
+      runtimeSettingsInputSchema.parse(
+        settings({
+          modelProfiles: [profile],
+          defaultModelProfileId: profileId
+        })
+      )
+    ).toThrow()
+
+    await store.update(
+      settings({
+        modelProfiles: [
+          {
+            ...profile,
+            contextWindowTokens: 32_000
+          }
+        ],
+        defaultModelProfileId: profileId
+      })
+    )
+    const persisted = JSON.parse(
+      await readFile(filePath, 'utf8')
+    ) as {
+      modelProfiles: Array<{ contextWindowTokens?: number }>
+    }
+    persisted.modelProfiles[0]!.contextWindowTokens = 10_000
+    await writeFile(filePath, JSON.stringify(persisted), 'utf8')
+
+    const migrated = new RuntimeSettingsStore(filePath, cipher, {})
+    await expect(migrated.getPublicSettings()).resolves.toMatchObject({
+      modelProfiles: [
+        expect.objectContaining({
+          id: profileId,
+          contextWindowTokens: undefined
+        })
+      ]
+    })
+  })
+
  it('configures bundled runtimes from the default model profile', async () => {
    const { store } = await createStore()

@@ -14,6 +14,7 @@ import {
  imageGenerationQualitySchema,
  isAgentRuntimeModelProtocol,
  isDeepSeekHarnessModelProfile,
+  minimumModelContextWindowTokens,
  modelAuthenticationSchema,
  modelProtocolSchema,
  runtimeModelSourceSchema,
@@ -555,7 +556,12 @@ function migrateVersion10(
 function normalizeStoredSettings(settings: StoredSettings): StoredSettings {
  const modelProfiles = settings.modelProfiles.map((profile) => ({
    ...profile,
-    baseUrl: normalizeModelBaseUrl(profile.baseUrl)
+    baseUrl: normalizeModelBaseUrl(profile.baseUrl),
+    contextWindowTokens:
+      profile.contextWindowTokens === undefined ||
+      profile.contextWindowTokens >= minimumModelContextWindowTokens
+        ? profile.contextWindowTokens
+        : undefined
  }))
  const fallbackProfileId = compatibleTextProfileId({
    modelProfiles,
@@ -902,6 +902,32 @@ describe('App', () => {
    }
  })

+  it('keeps a recently visited workspace page mounted', async () => {
+    render(<App />)
+    fireEvent.click(
+      await screen.findByRole('button', { name: '知识库' })
+    )
+    const heading = await screen.findByRole('heading', {
+      level: 1,
+      name: '知识库'
+    })
+    const route = heading.closest<HTMLElement>('[data-route="knowledge"]')
+    expect(route).not.toHaveAttribute('hidden')
+
+    fireEvent.click(screen.getByRole('button', { name: '对话' }))
+    expect(heading).toBeInTheDocument()
+    expect(route).toHaveAttribute('hidden')
+
+    fireEvent.click(screen.getByRole('button', { name: '知识库' }))
+    expect(
+      await screen.findByRole('heading', {
+        level: 1,
+        name: '知识库'
+      })
+    ).toBe(heading)
+    expect(route).not.toHaveAttribute('hidden')
+  })
+
  it('preserves title, message, and project filtering with deferred search', async () => {
    vi.mocked(api.conversations.list).mockResolvedValueOnce([
      {
@@ -1585,7 +1611,8 @@ describe('App', () => {
    fireEvent.scroll(chat)

    fireEvent.click(screen.getByRole('button', { name: '知识库' }))
-    expect(container.querySelector('.chat')).not.toBeInTheDocument()
+    expect(chat).toBeInTheDocument()
+    expect(chat.closest('[data-route="chat"]')).toHaveAttribute('hidden')
    act(() => {
      agentListener?.({
        requestId: request.requestId,
@@ -1597,7 +1624,7 @@ describe('App', () => {
    fireEvent.click(screen.getByRole('button', { name: '对话' }))
    expect(await screen.findByText('后台新增的回复内容')).toBeInTheDocument()
    const restoredChat = container.querySelector<HTMLElement>('.chat')
-    expect(restoredChat).not.toBe(chat)
+    expect(restoredChat).toBe(chat)
    expect(restoredChat?.scrollTop).toBe(175)
    expect(
      screen.getByRole('button', { name: '到底部' })
@@ -1609,6 +1636,16 @@ describe('App', () => {
      '00000000-0000-4000-8000-000000000461'
    const secondConversationId =
      '00000000-0000-4000-8000-000000000462'
+    const draftAttachment = {
+      id: '00000000-0000-4000-8000-000000000463',
+      name: '第一段草稿附件.md',
+      size: 1_024,
+      preview: '会话级草稿附件',
+      kind: 'text' as const
+    }
+    vi.mocked(api.context.selectFiles).mockResolvedValueOnce([
+      draftAttachment
+    ])
    vi.mocked(api.conversations.list).mockResolvedValueOnce([
      {
        id: firstConversationId,
@@ -1619,6 +1656,7 @@ describe('App', () => {
          id: `00000000-0000-4000-8100-${String(index).padStart(12, '0')}`,
          role: index % 2 === 0 ? ('user' as const) : ('assistant' as const),
          content: `第一段历史 ${String(index).padStart(3, '0')}`,
+          reasoning: index === 159 ? '需要保留展开状态' : undefined,
          createdAt: 1_775_000_000_000 + index,
          state: 'complete' as const
        }))
@@ -1647,31 +1685,65 @@ describe('App', () => {
        name: '加载更早的消息（还剩 81 条）'
      })
    )
-    expect(container.querySelectorAll('.message')).toHaveLength(160)
-    const firstChat = container.querySelector<HTMLElement>('.chat')
+    const firstPane = container.querySelector<HTMLElement>(
+      `[data-conversation-id="${firstConversationId}"]`
+    )
+    expect(firstPane?.querySelectorAll('.message')).toHaveLength(160)
+    const firstChat = firstPane?.querySelector<HTMLElement>('.chat')
    if (!firstChat) {
      throw new Error('Missing first chat scroll container')
    }
+    const reasoningDetails = firstPane?.querySelector<HTMLDetailsElement>(
+      '.message-reasoning'
+    )
+    if (!reasoningDetails) {
+      throw new Error('Missing reasoning details')
+    }
+    reasoningDetails.open = true
    Object.defineProperties(firstChat, {
      clientHeight: { configurable: true, value: 400 },
      scrollHeight: { configurable: true, value: 1_600 },
      scrollTop: { configurable: true, writable: true, value: 225 }
    })
    fireEvent.scroll(firstChat)
+    fireEvent.change(screen.getByLabelText('向 GoodBuddy 提问'), {
+      target: { value: '第一段会话草稿' }
+    })
+    fireEvent.click(screen.getByLabelText('添加附件'))
+    expect(
+      await screen.findByText(draftAttachment.name)
+    ).toBeInTheDocument()

    fireEvent.click(
      screen.getByText('第二段会话').closest('button')!
    )
    expect(await screen.findByText('第二段会话内容')).toBeInTheDocument()
+    expect(screen.getByLabelText('向 GoodBuddy 提问')).toHaveValue('')
+    expect(
+      screen.queryByText(draftAttachment.name)
+    ).not.toBeInTheDocument()
+    fireEvent.change(screen.getByLabelText('向 GoodBuddy 提问'), {
+      target: { value: '第二段会话草稿' }
+    })
    fireEvent.click(
      screen.getByText('第一段长会话').closest('button')!
    )

    expect(await screen.findByText('第一段历史 001')).toBeInTheDocument()
-    expect(container.querySelectorAll('.message')).toHaveLength(160)
-    expect(container.querySelector<HTMLElement>('.chat')?.scrollTop).toBe(
-      225
+    const restoredFirstPane = container.querySelector<HTMLElement>(
+      `[data-conversation-id="${firstConversationId}"]`
    )
+    expect(restoredFirstPane).toBe(firstPane)
+    expect(restoredFirstPane?.querySelectorAll('.message')).toHaveLength(
+      160
+    )
+    expect(restoredFirstPane?.querySelector('.chat')).toBe(firstChat)
+    expect(firstChat.scrollTop).toBe(225)
+    expect(reasoningDetails).toHaveAttribute('open')
+    expect(screen.getByLabelText('向 GoodBuddy 提问')).toHaveValue(
+      '第一段会话草稿'
+    )
+    expect(screen.getByText(draftAttachment.name)).toBeInTheDocument()
  })

  it('requires an accessible confirmation before permanently deleting a conversation', async () => {
@@ -2227,6 +2299,124 @@ describe('App', () => {
    expect(screen.getByText('正在分析真实推理内容')).toBeVisible()
  })

+  it('shows live context usage and keeps explicit compression status', async () => {
+    const settings = await api.settings.getRuntime()
+    vi.mocked(api.settings.getRuntime).mockResolvedValueOnce({
+      ...settings,
+      provider: 'model',
+      modelProfiles: settings.modelProfiles.map((profile) => ({
+        ...profile,
+        contextWindowTokens: 32_000
+      })),
+      contextCompression: {
+        enabled: true,
+        triggerTokens: 200_000,
+        recentRawTokens: 32_000,
+        modelSource: { kind: 'current' },
+        summaryPrompt: 'Preserve important facts.'
+      }
+    })
+    render(<App />)
+
+    expect(
+      await screen.findByText(/上下文 ≈.+ \/ 32\.0K · \d+%/u)
+    ).toBeInTheDocument()
+    expect(
+      screen.getByRole('progressbar', {
+        name: '当前上下文使用量'
+      })
+    ).toBeInTheDocument()
+    expect(
+      screen.queryByText('只读问答，不修改文件')
+    ).not.toBeInTheDocument()
+    expect(
+      await screen.findByText('快捷唤起：', { exact: false })
+    ).toHaveTextContent('快捷唤起：Ctrl+Shift+Space')
+
+    fireEvent.change(screen.getByLabelText('向 GoodBuddy 提问'), {
+      target: { value: '中'.repeat(1_000) }
+    })
+    expect(
+      screen.getByText(/上下文 ≈5\.\dK \/ 32\.0K/u)
+    ).toBeInTheDocument()
+    fireEvent.click(screen.getByLabelText('发送'))
+    await waitFor(() => expect(run).toHaveBeenCalledOnce())
+    const request = run.mock.calls[0]?.[0]
+    if (!request) {
+      throw new Error('Missing request')
+    }
+
+    act(() => {
+      agentListener?.({
+        requestId: request.requestId,
+        type: 'context-compression',
+        state: 'started',
+        estimatedBeforeTokens: 22_000,
+        effectiveTriggerTokens: 20_000,
+        contextWindowTokens: 32_000,
+        recentRawTokens: 32_000,
+        coveredMessageCount: 2
+      })
+    })
+    expect(
+      screen.getByText('正在压缩较早对话…')
+    ).toBeInTheDocument()
+
+    act(() => {
+      agentListener?.({
+        requestId: request.requestId,
+        type: 'status',
+        message: 'sonnet-5 正在思考'
+      })
+    })
+    expect(
+      screen.getByText('正在压缩较早对话…')
+    ).toBeInTheDocument()
+    expect(screen.getByText('sonnet-5 正在思考')).toBeInTheDocument()
+
+    act(() => {
+      agentListener?.({
+        requestId: request.requestId,
+        type: 'context-compression',
+        state: 'completed',
+        estimatedBeforeTokens: 22_000,
+        estimatedAfterTokens: 9_000,
+        effectiveTriggerTokens: 20_000,
+        contextWindowTokens: 32_000,
+        recentRawTokens: 32_000,
+        coveredMessageCount: 2,
+        summaryTokens: 1_000
+      })
+      agentListener?.({
+        requestId: request.requestId,
+        type: 'context-metrics',
+        estimatedInputTokens: 9_000,
+        effectiveTriggerTokens: 20_000,
+        contextWindowTokens: 32_000,
+        compressionEnabled: true,
+        recentRawTokens: 32_000,
+        coveredMessageCount: 2,
+        summaryTokens: 1_000
+      })
+    })
+
+    expect(
+      screen.getByText('已压缩较早对话 · ≈22.0K → ≈9.0K')
+    ).toBeInTheDocument()
+    expect(
+      screen.getByText('上下文 ≈9.0K / 32.0K · 28%')
+    ).toBeInTheDocument()
+    act(() => {
+      agentListener?.({
+        requestId: request.requestId,
+        type: 'done'
+      })
+    })
+    expect(
+      screen.getByText('已压缩较早对话 · ≈22.0K → ≈9.0K')
+    ).toBeInTheDocument()
+  })
+
  it('keeps a tool failure in details and hides retry after continuing', async () => {
    render(<App />)

@@ -2871,9 +3061,7 @@ describe('App', () => {
    expect(
      screen.queryByRole('heading', { name: '设置中心' })
    ).not.toBeInTheDocument()
-    expect(
-      await screen.findByText(/请先配置可用的模型或 Agent Runtime/u)
-    ).toBeInTheDocument()
+    expect(screen.getByLabelText('发送')).toBeDisabled()
    expect(run).not.toHaveBeenCalled()
  })

@@ -3794,9 +3982,7 @@ describe('App', () => {
    expect(mode).toBeEnabled()
    expect(mode.closest('.composer')).not.toBeNull()
    expect(
-      await screen.findByText(
-        new RegExp(`${label} Ask 模式.*只允许搜索当前启用的知识库`)
-      )
+      await screen.findByText('快捷唤起：', { exact: false })
    ).toBeInTheDocument()
    selectComposerOption('工作模式', 'Execute · 受控执行')
    expect(mode).toHaveAccessibleName(
@@ -4076,7 +4262,11 @@ describe('App', () => {
      expect(notification).toHaveTextContent(
        '当前对话已切换到 OpenCode · 默认模型'
      )
-      expect(screen.getByText(/Ask 模式：只读问答/)).toBeInTheDocument()
+      expect(
+        screen.getByRole('button', {
+          name: '工作模式：Ask · 只读问答'
+        })
+      ).toBeInTheDocument()

      fireEvent.click(
        screen.getByRole('button', {
@@ -4102,7 +4292,11 @@ describe('App', () => {
          '当前对话已切换到 Continue · 默认模型'
        )
      ).not.toBeInTheDocument()
-      expect(screen.getByText(/Ask 模式：只读问答/)).toBeInTheDocument()
+      expect(
+        screen.getByRole('button', {
+          name: '工作模式：Ask · 只读问答'
+        })
+      ).toBeInTheDocument()
    } finally {
      vi.useRealTimers()
    }
@@ -1,4 +1,4 @@
-import { cleanup, render } from '@testing-library/react'
+import { cleanup, render, screen } from '@testing-library/react'
 import { afterEach, describe, expect, it, vi } from 'vitest'
 import {
  ChatTimeline,
@@ -85,4 +85,54 @@ describe('ChatTimeline', () => {
    )
    expect(unchangedDetails).toHaveAttribute('open')
  })
+
+  it('places compression progress between the user and assistant messages', () => {
+    const messages: Message[] = [
+      {
+        id: 'user-message',
+        role: 'user',
+        content: 'Continue',
+        createdAt: 1_775_000_000_000,
+        state: 'complete'
+      },
+      {
+        id: 'assistant-message',
+        role: 'assistant',
+        content: 'Answer',
+        contextCompression: {
+          state: 'completed',
+          estimatedBeforeTokens: 22_000,
+          estimatedAfterTokens: 9_000
+        },
+        createdAt: 1_775_000_001_000,
+        state: 'complete'
+      }
+    ]
+    const { container } = render(
+      <ChatTimeline
+        artifactById={new Map()}
+        conversationId="conversation-1"
+        hiddenMessageCount={0}
+        isUnusedConversation={false}
+        locale="zh-CN"
+        messageStartIndex={0}
+        messages={messages}
+        {...callbacks}
+        retryContent=""
+        totalMessageCount={messages.length}
+      />
+    )
+
+    const children = Array.from(
+      container.querySelector('.message-list')?.children ?? []
+    )
+    expect(children.map((element) => element.className)).toEqual([
+      'message message--user',
+      'context-compression-event context-compression-event--completed',
+      'message message--assistant'
+    ])
+    expect(
+      screen.getByRole('status')
+    ).toHaveTextContent('已压缩较早对话 · ≈22.0K → ≈9.0K')
+  })
 })
@@ -51,6 +51,11 @@ export type Message = {
  createdAt: number
  state: 'streaming' | 'complete' | 'error'
  status?: string
+  contextCompression?: {
+    state: 'compressing' | 'completed' | 'failed'
+    estimatedBeforeTokens: number
+    estimatedAfterTokens?: number
+  }
  tools?: ToolActivity[]
  subagents?: SubagentActivity[]
  approval?: {
@@ -74,6 +79,14 @@ export type ImageViewerItem = {
  title: string
 }

+function formatCompactTokens(tokens: number): string {
+  if (tokens < 1_000) {
+    return tokens.toLocaleString()
+  }
+  const value = tokens / 1_000
+  return `${value >= 100 ? Math.round(value) : value.toFixed(1)}K`
+}
+
 type MessageBlockRenderItem =
  | {
      kind: 'block'
@@ -270,6 +283,41 @@ function ChatMessageRowView({
  const { t } = useTranslation('app')

  return (
+    <>
+      {message.role === 'assistant' && message.contextCompression && (
+        <div
+          aria-live="polite"
+          className={`context-compression-event context-compression-event--${message.contextCompression.state}`}
+          role="status"
+        >
+          <span className="context-compression-event__line" />
+          <span className="context-compression-event__label">
+            <span
+              aria-hidden="true"
+              className={
+                message.contextCompression.state === 'compressing'
+                  ? 'message__status-dot message__status-dot--active'
+                  : 'message__status-dot'
+              }
+            />
+            {message.contextCompression.state === 'compressing'
+              ? t('chat.contextCompression.compressing')
+              : message.contextCompression.state === 'completed' &&
+                  message.contextCompression.estimatedAfterTokens !==
+                    undefined
+                ? t('chat.contextCompression.completed', {
+                    before: formatCompactTokens(
+                      message.contextCompression.estimatedBeforeTokens
+                    ),
+                    after: formatCompactTokens(
+                      message.contextCompression.estimatedAfterTokens
+                    )
+                  })
+                : t('chat.contextCompression.failed')}
+          </span>
+          <span className="context-compression-event__line" />
+        </div>
+      )}
    <article
      className={`message message--${message.role}`}
      ref={(element) => onArticleRef(message.id, element)}
@@ -755,6 +803,7 @@ function ChatMessageRowView({
        )}
      </div>
    </article>
+    </>
  )
 }

@@ -604,6 +604,7 @@ describe('SettingsPanel runtime files', () => {
      '上下文上限（可选）'
    )
    expect(contextWindow).toHaveValue(null)
+    expect(contextWindow).toHaveAttribute('min', '32')
    fireEvent.change(contextWindow, { target: { value: '256' } })
    fireEvent.click(screen.getByRole('button', { name: '保存设置' }))

@@ -2418,7 +2418,7 @@ export function SettingsPanel({
                          aria-label={t('model.profile.contextWindow')}
                          inputMode="numeric"
                          max={10_000}
-                          min={8}
+                          min={32}
                          onChange={(event) => {
                            const value = event.target.valueAsNumber
                            updateModelProfile(profile.id, {
@@ -187,6 +187,12 @@ export const app = {
      streaming: 'Reasoning',
      complete: 'Reasoning process'
    },
+    contextCompression: {
+      compressing: 'Compressing earlier conversation…',
+      completed:
+        'Earlier conversation compressed · ≈{{before}} → ≈{{after}}',
+      failed: 'Earlier conversation compression failed'
+    },
    sources: 'Sources: {{sources}}',
    citations: {
      view: 'View {{count}} evidence references',
@@ -314,6 +320,14 @@ export const app = {
    send: 'Send',
    sendTitle: 'Send message',
    shortcut: 'Quick access: ',
+    context: {
+      tokenCount: 'Context ≈{{used}}',
+      windowUsage: 'Context ≈{{used}} / {{total}} · {{percentage}}%',
+      thresholdUsage:
+        'Compression threshold ≈{{used}} / {{total}} · {{percentage}}%',
+      progressLabel: 'Current context usage',
+      compressionTrigger: 'Automatic compression at ≈{{tokens}}'
+    },
    experts: {
      general: 'General assistant',
      generalDescription: 'Default single assistant',
@@ -495,7 +495,7 @@ export const settings = {
        'When enabled, GoodBuddy can send image context to this model connection.',
      contextWindow: 'Context window (optional)',
      contextWindowDescription:
-        'Enter K tokens. Leave blank when unknown. This value is used only for GoodBuddy local budget calculations.',
+        'Enter 32K–10000K tokens. Leave blank when unknown. This value is used only for GoodBuddy local budget calculations.',
      imageQuality: 'Image quality',
      imageQualityAriaLabel: 'Image quality for {{name}}',
      quality: {
@@ -182,6 +182,11 @@ export const app = {
      streaming: '正在推理',
      complete: '推理过程'
    },
+    contextCompression: {
+      compressing: '正在压缩较早对话…',
+      completed: '已压缩较早对话 · ≈{{before}} → ≈{{after}}',
+      failed: '较早对话压缩失败'
+    },
    sources: '来源：{{sources}}',
    citations: {
      view: '查看 {{count}} 条证据引用',
@@ -306,6 +311,14 @@ export const app = {
    send: '发送',
    sendTitle: '发送消息',
    shortcut: '快捷唤起：',
+    context: {
+      tokenCount: '上下文 ≈{{used}}',
+      windowUsage: '上下文 ≈{{used}} / {{total}} · {{percentage}}%',
+      thresholdUsage:
+        '距压缩阈值 ≈{{used}} / {{total}} · {{percentage}}%',
+      progressLabel: '当前上下文使用量',
+      compressionTrigger: '自动压缩线：≈{{tokens}}'
+    },
    experts: {
      general: '通用助手',
      generalDescription: '默认单助手',
@@ -452,7 +452,7 @@ export const settings = {
        '启用后，GoodBuddy 可将图片上下文发送给此模型连接。',
      contextWindow: '上下文上限（可选）',
      contextWindowDescription:
-        '以 K tokens 填写。留空表示未知；此值仅用于 GoodBuddy 本地预算计算。',
+        '以 K tokens 填写，范围为 32K–10000K。留空表示未知；此值仅用于 GoodBuddy 本地预算计算。',
      imageQuality: '图片质量',
      imageQualityAriaLabel: '图片质量 {{name}}',
      quality: {
@@ -0,0 +1,96 @@
+import { describe, expect, it } from 'vitest'
+import {
+  pruneKeepAliveEntries,
+  touchKeepAliveEntry
+} from './keep-alive-cache'
+
+describe('keep-alive cache', () => {
+  it('updates a visited entry without duplicating it', () => {
+    expect(
+      touchKeepAliveEntry(
+        [
+          { key: 'chat', lastVisitedAt: 10 },
+          { key: 'knowledge', lastVisitedAt: 20 }
+        ],
+        'chat',
+        30
+      )
+    ).toEqual([
+      { key: 'knowledge', lastVisitedAt: 20 },
+      { key: 'chat', lastVisitedAt: 30 }
+    ])
+  })
+
+  it('keeps recent and protected entries while expiring inactive ones', () => {
+    expect(
+      pruneKeepAliveEntries(
+        [
+          { key: 'one', lastVisitedAt: 10 },
+          { key: 'two', lastVisitedAt: 20 },
+          { key: 'three', lastVisitedAt: 30 },
+          { key: 'four', lastVisitedAt: 40 }
+        ],
+        {
+          currentKey: 'four',
+          expiresAfterMs: 50,
+          maximumEntries: 4,
+          now: 100,
+          protectedKeys: new Set(['one']),
+          recentEntries: 2
+        }
+      )
+    ).toEqual([
+      { key: 'four', lastVisitedAt: 40 },
+      { key: 'three', lastVisitedAt: 30 },
+      { key: 'one', lastVisitedAt: 10 }
+    ])
+  })
+
+  it('enforces the hard limit with least-recently-used eviction', () => {
+    expect(
+      pruneKeepAliveEntries(
+        Array.from({ length: 8 }, (_, index) => ({
+          key: `conversation-${index}`,
+          lastVisitedAt: index
+        })),
+        {
+          currentKey: 'conversation-7',
+          expiresAfterMs: 1_000,
+          maximumEntries: 5,
+          now: 10,
+          protectedKeys: new Set(['conversation-0']),
+          recentEntries: 2
+        }
+      ).map((entry) => entry.key)
+    ).toEqual([
+      'conversation-7',
+      'conversation-6',
+      'conversation-5',
+      'conversation-4',
+      'conversation-0'
+    ])
+  })
+
+  it('expires an unprotected entry after one hour', () => {
+    const entries = [{ key: 'knowledge', lastVisitedAt: 1_000 }]
+    const options = {
+      expiresAfterMs: 60 * 60 * 1_000,
+      maximumEntries: 4,
+      protectedKeys: new Set<string>(),
+      recentEntries: 0
+    }
+
+    expect(
+      pruneKeepAliveEntries(entries, {
+        ...options,
+        now: 1_000 + 60 * 60 * 1_000 - 1
+      })
+    ).toEqual(entries)
+    expect(
+      pruneKeepAliveEntries(entries, {
+        ...options,
+        now: 1_000 + 60 * 60 * 1_000
+      })
+    ).toEqual([])
+  })
+})
@@ -0,0 +1,69 @@
+export type KeepAliveCacheEntry<Key extends string> = {
+  key: Key
+  lastVisitedAt: number
+}
+
+export function touchKeepAliveEntry<Key extends string>(
+  entries: readonly KeepAliveCacheEntry<Key>[],
+  key: Key,
+  visitedAt: number
+): KeepAliveCacheEntry<Key>[] {
+  const existing = entries.find((entry) => entry.key === key)
+  if (existing?.lastVisitedAt === visitedAt) {
+    return [...entries]
+  }
+  return [
+    ...entries.filter((entry) => entry.key !== key),
+    { key, lastVisitedAt: visitedAt }
+  ]
+}
+
+export function pruneKeepAliveEntries<Key extends string>(
+  entries: readonly KeepAliveCacheEntry<Key>[],
+  {
+    currentKey,
+    expiresAfterMs,
+    maximumEntries,
+    now,
+    protectedKeys,
+    recentEntries
+  }: {
+    currentKey?: Key
+    expiresAfterMs: number
+    maximumEntries: number
+    now: number
+    protectedKeys?: ReadonlySet<Key>
+    recentEntries: number
+  }
+): KeepAliveCacheEntry<Key>[] {
+  const newestFirst = [...entries].sort(
+    (left, right) => right.lastVisitedAt - left.lastVisitedAt
+  )
+  const alwaysKeep = new Set(
+    newestFirst.slice(0, recentEntries).map((entry) => entry.key)
+  )
+  if (currentKey) {
+    alwaysKeep.add(currentKey)
+  }
+  protectedKeys?.forEach((key) => alwaysKeep.add(key))
+
+  const retained = newestFirst.filter(
+    (entry) =>
+      alwaysKeep.has(entry.key) ||
+      now - entry.lastVisitedAt < expiresAfterMs
+  )
+  if (retained.length <= maximumEntries) {
+    return retained
+  }
+
+  const removableOldestFirst = retained
+    .filter((entry) => !alwaysKeep.has(entry.key))
+    .sort((left, right) => left.lastVisitedAt - right.lastVisitedAt)
+  const removeCount = retained.length - maximumEntries
+  const removedKeys = new Set(
+    removableOldestFirst
+      .slice(0, removeCount)
+      .map((entry) => entry.key)
+  )
+  return retained.filter((entry) => !removedKeys.has(entry.key))
+}
@@ -1963,6 +1963,17 @@ textarea:focus-visible {
  grid-template-rows: 58px minmax(0, 1fr) auto;
 }

+.workspace-route-cache {
+  display: grid;
+  min-width: 0;
+  min-height: 0;
+  overflow: hidden;
+}
+
+.workspace-route-cache[hidden] {
+  display: none;
+}
+
 .topbar {
  display: flex;
  min-width: 0;
@@ -2975,6 +2986,15 @@ button > svg {
  background: var(--surface-raised);
 }

+.chat-history-pane {
+  height: 100%;
+  min-height: 0;
+}
+
+.chat-history-pane[hidden] {
+  display: none;
+}
+
 .chat {
  height: 100%;
  min-height: 0;
@@ -3608,6 +3628,55 @@ button > svg {
  color: var(--danger);
 }

+.context-compression-event {
+  display: flex;
+  width: 100%;
+  align-items: center;
+  color: var(--text-muted);
+  gap: var(--space-2);
+}
+
+.context-compression-event__line {
+  height: 1px;
+  min-width: var(--space-6);
+  flex: 1;
+  background: var(--border-subtle);
+}
+
+.context-compression-event__label {
+  display: inline-flex;
+  align-items: center;
+  padding: var(--space-1) var(--space-2);
+  border: 1px solid var(--border-default);
+  border-radius: var(--radius-control);
+  background: var(--surface-subtle);
+  color: var(--text-secondary);
+  font-size: var(--font-caption);
+  gap: var(--space-2);
+  white-space: nowrap;
+}
+
+.context-compression-event--compressing
+  .context-compression-event__label {
+  border-color: var(--accent-selected);
+  background: var(--accent-subtle);
+  color: var(--accent);
+}
+
+.context-compression-event--completed
+  .context-compression-event__label {
+  border-color: color-mix(in srgb, var(--success) 35%, var(--border-default));
+  background: var(--success-subtle);
+  color: var(--success);
+}
+
+.context-compression-event--failed
+  .context-compression-event__label {
+  border-color: var(--danger-border);
+  background: var(--danger-subtle);
+  color: var(--danger);
+}
+
 .message__status-dot {
  flex: 0 0 auto;
  width: 6px;
@@ -4409,7 +4478,7 @@ button > svg {
 }

 .composer-picker--mode > .model-button {
-  width: 96px;
+  width: 108px;
 }

 .composer-picker--ask svg {
@@ -4595,29 +4664,31 @@ button > svg {
  background: var(--danger-solid);
 }

-.composer-hint {
+.composer-meta {
  display: flex;
  margin: var(--space-2) 0 0;
  align-items: center;
-  justify-content: center;
-  color: var(--text-secondary);
+  justify-content: flex-end;
  font-size: var(--font-caption);
  flex-wrap: wrap;
-  gap: var(--space-1) var(--space-3);
+  gap: var(--space-2) var(--space-3);
  line-height: 1.45;
-  text-align: center;
 }

-.composer-hint--error {
+.composer-meta__error {
+  margin-right: auto;
+}
+
+.composer-meta__error {
  color: var(--danger);
 }

-.composer-hint__shortcut {
+.composer-meta__shortcut {
  color: var(--text-muted);
  white-space: nowrap;
 }

-.composer-hint kbd {
+.composer-meta kbd {
  padding: 1px var(--space-1);
  border: 1px solid var(--border-default);
  border-radius: 4px;
@@ -4627,6 +4698,52 @@ button > svg {
  font-size: 10px;
 }

+.composer-context-meter {
+  display: flex;
+  min-width: 0;
+  align-items: center;
+  margin-right: auto;
+  color: var(--text-muted);
+  gap: var(--space-2);
+}
+
+.composer-context-meter--warning {
+  color: var(--warning);
+}
+
+.composer-context-meter__summary {
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+
+.composer-context-meter__track {
+  position: relative;
+  width: 112px;
+  height: 4px;
+  flex: 0 0 112px;
+  border-radius: 999px;
+  overflow: visible;
+  background: var(--surface-muted);
+}
+
+.composer-context-meter__fill {
+  position: absolute;
+  border-radius: inherit;
+  background: currentColor;
+  inset: 0 auto 0 0;
+  transition: width var(--motion-fast) ease-out;
+}
+
+.composer-context-meter__trigger {
+  position: absolute;
+  top: -2px;
+  width: 1px;
+  height: 8px;
+  background: var(--text-secondary);
+  transform: translateX(-50%);
+}
+
@container (max-width: 700px) {
  .composer__controls {
    flex-wrap: wrap;
@@ -4649,6 +4766,15 @@ button > svg {
  .model-button {
    width: 100%;
  }
+
+  .composer-context-meter {
+    width: 100%;
+  }
+
+  .composer-context-meter__track {
+    min-width: 64px;
+    flex: 1;
+  }
 }

 .settings-backdrop {
@@ -147,6 +147,14 @@ export const conversationMessageSchema = z
    createdAt: z.number().int().nonnegative(),
    state: z.enum(['streaming', 'complete', 'error']),
    status: z.string().max(4_000).optional(),
+    contextCompression: z
+      .object({
+        state: z.enum(['compressing', 'completed', 'failed']),
+        estimatedBeforeTokens: z.number().int().nonnegative(),
+        estimatedAfterTokens: z.number().int().nonnegative().optional()
+      })
+      .strict()
+      .optional(),
    tools: z.array(conversationToolActivitySchema).max(100).optional(),
    sources: z.array(z.string().max(8_192)).max(100).optional(),
    sourceReferences: z
@@ -0,0 +1,63 @@
+export const minimumModelContextWindowTokens = 32_000
+export const maximumModelContextWindowTokens = 10_000_000
+export const contextOutputAndSafetyTokens = 12_000
+export const estimatedContextRequestOverheadTokens = 4_000
+
+export type ContextWindowMessage = {
+  role: 'user' | 'assistant'
+  content: string
+}
+
+export function estimateTextTokens(value: string): number {
+  let asciiCharacters = 0
+  let nonAsciiCharacters = 0
+  for (const character of value) {
+    if (character.codePointAt(0)! <= 0x7f) {
+      asciiCharacters += 1
+    } else {
+      nonAsciiCharacters += 1
+    }
+  }
+  return Math.max(
+    1,
+    Math.ceil(asciiCharacters / 4 + nonAsciiCharacters)
+  )
+}
+
+export function estimateMessagesTokens(
+  messages: readonly ContextWindowMessage[]
+): number {
+  return messages.reduce(
+    (total, message) => total + estimateTextTokens(message.content) + 4,
+    0
+  )
+}
+
+export function estimateContextInputTokens(input: {
+  history: readonly ContextWindowMessage[]
+  prompt: string
+  summaryTokens?: number
+}): number {
+  return (
+    estimateMessagesTokens(input.history) +
+    estimateTextTokens(input.prompt) +
+    (input.summaryTokens ?? 0) +
+    estimatedContextRequestOverheadTokens
+  )
+}
+
+export function getEffectiveContextTriggerTokens(input: {
+  triggerTokens: number
+  contextWindowTokens?: number
+}): number {
+  if (input.contextWindowTokens === undefined) {
+    return input.triggerTokens
+  }
+  return Math.min(
+    input.triggerTokens,
+    Math.max(
+      input.contextWindowTokens,
+      minimumModelContextWindowTokens
+    ) - contextOutputAndSafetyTokens
+  )
+}
@@ -1,4 +1,8 @@
 import { z } from 'zod'
+import {
+  maximumModelContextWindowTokens,
+  minimumModelContextWindowTokens
+} from './context-window'
 import type {
  BrowserProfileCreateInput,
  BrowserProfileRenameInput,
@@ -416,6 +420,11 @@ const modelApiKeyUpdateSchema = z.discriminatedUnion('action', [
  z.object({ action: z.literal('clear') }).strict()
 ])

+export {
+  maximumModelContextWindowTokens,
+  minimumModelContextWindowTokens
+} from './context-window'
+
 const modelProfileInputSchema = z
  .object({
    id: modelProfileIdSchema,
@@ -433,8 +442,8 @@ const modelProfileInputSchema = z
    contextWindowTokens: z
      .number()
      .int()
-      .min(8_000)
-      .max(10_000_000)
+      .min(minimumModelContextWindowTokens)
+      .max(maximumModelContextWindowTokens)
      .optional(),
    imageGenerationQuality: imageGenerationQualitySchema,
    apiKey: modelApiKeyUpdateSchema
@@ -907,6 +916,29 @@ export type AgentEvent =
      type: 'reasoning'
      delta: string
    }
+  | {
+      requestId: string
+      type: 'context-metrics'
+      estimatedInputTokens: number
+      effectiveTriggerTokens: number
+      contextWindowTokens?: number
+      compressionEnabled: boolean
+      recentRawTokens: number
+      coveredMessageCount: number
+      summaryTokens: number
+    }
+  | {
+      requestId: string
+      type: 'context-compression'
+      state: 'started' | 'completed'
+      estimatedBeforeTokens: number
+      estimatedAfterTokens?: number
+      effectiveTriggerTokens: number
+      contextWindowTokens?: number
+      recentRawTokens: number
+      coveredMessageCount: number
+      summaryTokens?: number
+    }
  | {
      requestId: string
      type: 'tool'