goodbuddy/src/main/agent/model-runtime.test.ts

import { createHash } from 'node:crypto'
import { mkdir, mkdtemp, readFile, rm, writeFile } from 'node:fs/promises'
import { tmpdir } from 'node:os'
import { join, relative } from 'node:path'
import { rgPath } from '@vscode/ripgrep'
import { describe, expect, it, vi } from 'vitest'
import {
  ModelToolProvider,
  RecoverableModelToolError,
  type ModelToolDefinition,
  type ModelToolProviderLike,
  type ModelToolResult
} from './model-tool-provider'
import { ModelAgentRuntime } from './model-runtime'
import type { RuntimeEvent } from './runtime'
import { SubagentScheduler } from '../assistant/subagent-scheduler'
import { LocalWorkspaceAccess } from '../workspace'
import { LocalDirectModelProcessService } from './direct-model-process-service'
import { browserTabIdSchema } from '../../shared/contracts'
import type { BrowserToolService } from '../browser/browser-model-tools'

const toolPng = Buffer.from([
  0x89, 0x50, 0x4e, 0x47,
  0x0d, 0x0a, 0x1a, 0x0a
]).toString('base64')

it('preserves full supplied and cached history beyond message and byte caps', async () => {
  const fetcher = vi.fn<typeof fetch>(async () => new Response(
    createEventStream('reply'),
    { headers: { 'content-type': 'text/event-stream' } }
  ))
  const runtime = new ModelAgentRuntime({
    apiKey: 'test-key',
    baseUrl: 'https://example.test',
    model: 'test-model',
    protocol: 'anthropic-messages',
    authentication: 'api-key',
    fetcher
  })
  const history = Array.from({ length: 502 }, (_, index) => ({
    role: index % 2 ? 'assistant' as const : 'user' as const,
    content: `${index}: ${'x'.repeat(4_200)}`
  }))
  for (const suppliedHistory of [history, undefined]) {
    for await (const event of runtime.run({
      requestId: crypto.randomUUID(),
      conversationId: 'full-history',
      prompt: 'next',
      history: suppliedHistory
    }, new AbortController().signal)) {
      expect(event).toBeDefined()
    }
  }
  for (const [, init] of fetcher.mock.calls) {
    const body = JSON.parse(String(init?.body))
    expect(body.messages.slice(0, history.length)).toEqual(history)
  }
  expect(fetcher).toHaveBeenCalledTimes(2)
})

function createTextToolResult(text: string): ModelToolResult {
  return {
    parts: [{ type: 'text', text }],
    contextBytes: Buffer.byteLength(text)
  }
}

function createMultimodalToolResult(): ModelToolResult {
  return {
    parts: [
      { type: 'text', text: 'tool result' },
      {
        type: 'image',
        mimeType: 'image/png',
        data: toolPng
      }
    ],
    contextBytes:
      Buffer.byteLength('tool result') + Buffer.byteLength(toolPng)
  }
}

function createEventStream(
  text: string,
  thinking?: string,
  usage: {
    inputTokens?: number
    outputTokens?: number
    cacheReadTokens?: number
    cacheWriteTokens?: number
  } = {}
): string {
  const {
    inputTokens = 23,
    outputTokens = 11,
    cacheReadTokens = 7,
    cacheWriteTokens = 5
  } = usage
  return [
    'event: message_start',
    `data: ${JSON.stringify({
      type: 'message_start',
      message: {
        id: 'message-1',
        model: 'claude-sonnet-provider',
        usage: {
          input_tokens: inputTokens,
          cache_creation_input_tokens: cacheWriteTokens,
          cache_read_input_tokens: cacheReadTokens
        }
      }
    })}`,
    '',
    ...(thinking
      ? [
          'event: content_block_delta',
          `data: ${JSON.stringify({
            type: 'content_block_delta',
            delta: { type: 'thinking_delta', thinking }
          })}`,
          ''
        ]
      : []),
    'event: content_block_delta',
    `data: ${JSON.stringify({
      type: 'content_block_delta',
      delta: { type: 'text_delta', text }
    })}`,
    '',
    'event: message_delta',
    `data: ${JSON.stringify({
      type: 'message_delta',
      usage: { output_tokens: outputTokens }
    })}`,
    '',
    'event: message_stop',
    'data: {"type":"message_stop"}',
    '',
    ''
  ].join('\n')
}

function createResponsesEventStream(
  text: string,
  reasoning?: string
): string {
  return [
    ...(reasoning
      ? [
          'event: response.reasoning_summary_text.delta',
          `data: ${JSON.stringify({
            type: 'response.reasoning_summary_text.delta',
            delta: reasoning
          })}`,
          ''
        ]
      : []),
    'event: response.output_text.delta',
    `data: ${JSON.stringify({
      type: 'response.output_text.delta',
      delta: text
    })}`,
    '',
    'event: response.completed',
    `data: ${JSON.stringify({
      type: 'response.completed',
      response: {
        id: 'resp-provider-1',
        model: 'gpt-5-provider',
        usage: {
          input_tokens: 29,
          output_tokens: 8,
          total_tokens: 37,
          input_tokens_details: { cached_tokens: 11 }
        }
      }
    })}`,
    '',
    ''
  ].join('\n')
}

function createSseEventStream(
  events: ReadonlyArray<Record<string, unknown>>
): string {
  return [
    ...events.flatMap((event) => [
      `event: ${event.type as string}`,
      `data: ${JSON.stringify(event)}`,
      ''
    ]),
    ''
  ].join('\n')
}

function createToolProvider(
  overrides: Partial<ModelToolProviderLike> = {}
): ModelToolProviderLike {
  const tool: ModelToolDefinition = {
    name: 'workspace_read_text',
    displayName: '读取工作区文本',
    description: 'Read text',
    inputSchema: {
      type: 'object',
      properties: { path: { type: 'string' } },
      required: ['path']
    },
    source: 'builtin'
  }
  return {
    listTools: vi.fn(async () => [tool]),
    getApproval: vi.fn((_definition, _arguments, summary) => ({
      scopeKey: 'model:builtin:workspace_read_text',
      title: '允许读取工作区文本？',
      description: '读取文件',
      toolName: '读取工作区文本',
      argumentSummary: summary
    })),
    callTool: vi.fn(async () => createTextToolResult('tool result')),
    releaseConversation: vi.fn(async () => {}),
    dispose: vi.fn(async () => {}),
    ...overrides
  }
}

describe('ModelAgentRuntime', () => {
  it('rejects images when the model connection disables image input', async () => {
    const fetcher = vi.fn<typeof fetch>()
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      supportsImageInput: false,
      fetcher
    })
    const stream = runtime.run(
      {
        requestId: '3f496642-f47d-4e0a-8944-a32c77b0d6ef',
        conversationId: 'conversation-1',
        prompt: 'describe',
        images: [
          {
            name: 'screenshot.png',
            mediaType: 'image/png',
            data: 'aW1hZ2U='
          }
        ]
      },
      new AbortController().signal
    )

    await expect(stream.next()).rejects.toThrow(
      '当前模型连接未启用图像输入'
    )
    expect(fetcher).not.toHaveBeenCalled()
  })

  it('keeps provider-reported image usage after the response completes', async () => {
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(createResponsesEventStream('image described'), {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.openai.com/v1',
      model: 'gpt-5',
      protocol: 'openai-responses',
      authentication: 'api-key',
      supportsImageInput: true,
      fetcher,
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 20_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        },
        contextWindowTokens: 32_000
      }
    })
    const events: RuntimeEvent[] = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: 'describe',
        images: [
          {
            name: 'screenshot.png',
            mediaType: 'image/png',
            data: toolPng
          }
        ]
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    const body = JSON.parse(
      fetcher.mock.calls[0]?.[1]?.body as string
    ) as { input: unknown[] }
    expect(JSON.stringify(body.input)).toContain('input_image')
    expect(
      events.filter((event) => event.type === 'context-metrics')
    ).toEqual([
      expect.objectContaining({
        contextTokens: 37,
        source: 'provider'
      })
    ])
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it.each([
    {
      protocol: 'anthropic-messages' as const,
      payloadKey: 'messages' as const,
      response: () =>
        new Response(createEventStream('history accepted'), {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
    },
    {
      protocol: 'openai-chat-completions' as const,
      payloadKey: 'messages' as const,
      response: () =>
        new Response(
          [
            'data: {"choices":[{"delta":{"content":"history accepted"}}]}',
            '',
            'data: [DONE]',
            '',
            ''
          ].join('\n'),
          {
            status: 200,
            headers: { 'content-type': 'text/event-stream' }
          }
        )
    },
    {
      protocol: 'openai-responses' as const,
      payloadKey: 'input' as const,
      response: () =>
        new Response(createResponsesEventStream('history accepted'), {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
    }
  ])(
    'omits local message IDs from $protocol history',
    async ({ protocol, payloadKey, response }) => {
      const localMessageIds = [
        '00000000-0000-4000-8000-000000000011',
        '00000000-0000-4000-8000-000000000012'
      ]
      const fetcher = vi.fn<typeof fetch>(async () => response())
      const runtime = new ModelAgentRuntime({
        apiKey: 'test-key',
        baseUrl: 'https://model.example/v1',
        model: 'test-model',
        protocol,
        authentication: 'api-key',
        fetcher
      })

      try {
        for await (const _event of runtime.run(
          {
            requestId: '00000000-0000-4000-8000-000000000010',
            conversationId: 'conversation-local-message-ids',
            prompt: 'current question',
            history: [
              { role: 'user', content: 'earlier question' },
              { role: 'assistant', content: 'earlier answer' }
            ],
            historyMessageIds: localMessageIds
          },
          new AbortController().signal
        )) {
          void _event
        }
      } finally {
        await runtime.dispose()
      }

      const body = JSON.parse(
        fetcher.mock.calls[0]?.[1]?.body as string
      ) as Record<string, unknown>
      const messages = body[payloadKey] as Array<
        Record<string, unknown>
      >
      expect(
        messages.filter(
          (message) =>
            message.role === 'user' ||
            message.role === 'assistant'
        )
      ).toEqual([
        { role: 'user', content: 'earlier question' },
        { role: 'assistant', content: 'earlier answer' },
        { role: 'user', content: 'current question' }
      ])
      expect(JSON.stringify(body)).not.toContain(localMessageIds[0])
      expect(JSON.stringify(body)).not.toContain(localMessageIds[1])
    }
  )

  it('requires generated text from a real minimal Anthropic request when testing the connection', async () => {
    const fetcher = vi.fn<typeof fetch>(async (_input, init) => {
      const body = JSON.parse(init?.body as string) as {
        messages: Array<{ content: string }>
      }
      const marker =
        /GOODBUDDY_MODEL_TEST_[A-F0-9]+/u.exec(
          body.messages[0]?.content ?? ''
        )?.[0]
      if (!marker) {
        throw new Error('missing model test marker')
      }
      return Response.json({
        content: [{ type: 'text', text: marker }]
      })
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher
    })

    await expect(runtime.testConnection()).resolves.toMatchObject({
      available: true,
      id: 'model',
      detail: expect.stringContaining('真实模型生成测试')
    })
    const body = JSON.parse(
      fetcher.mock.calls[0]?.[1]?.body as string
    ) as { max_tokens: number; stream: boolean }
    expect(body).toMatchObject({ max_tokens: 64, stream: false })
  })

  it('rejects a successful HTTP response that does not contain generated test text', async () => {
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher: vi.fn(async () =>
        Response.json({
          content: [{ type: 'text', text: 'generic health check' }]
        })
      )
    })

    await expect(runtime.testConnection()).rejects.toThrow(
      '未完成真实生成测试'
    )
  })

  it('validates generated test text from an OpenAI Chat Completions response', async () => {
    const fetcher = vi.fn<typeof fetch>(async (_input, init) => {
      const body = JSON.parse(init?.body as string) as {
        messages: Array<{ content: string }>
      }
      const marker =
        /GOODBUDDY_MODEL_TEST_[A-F0-9]+/u.exec(
          body.messages[0]?.content ?? ''
        )?.[0]
      if (!marker) {
        throw new Error('missing model test marker')
      }
      return Response.json({
        choices: [
          {
            message: {
              role: 'assistant',
              content: marker
            }
          }
        ]
      })
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://model.example/v1',
      model: 'chat-model',
      protocol: 'openai-chat-completions',
      authentication: 'api-key',
      fetcher
    })

    await expect(runtime.testConnection()).resolves.toMatchObject({
      available: true,
      detail: expect.stringContaining('真实模型生成测试')
    })
    expect(fetcher).toHaveBeenCalledOnce()
  })

  it('merges custom headers and body fields into a direct model request without overriding runtime fields', async () => {
    const fetcher = vi.fn<typeof fetch>(async (_input, init) => {
      const body = JSON.parse(init?.body as string) as {
        messages: Array<{ content: string }>
      }
      const marker =
        /GOODBUDDY_MODEL_TEST_[A-F0-9]+/u.exec(
          body.messages[0]?.content ?? ''
        )?.[0]
      return Response.json({
        choices: [
          {
            message: {
              role: 'assistant',
              content: marker
            }
          }
        ]
      })
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'runtime-key',
      baseUrl: 'https://model.example/v1',
      model: 'runtime-model',
      protocol: 'openai-chat-completions',
      authentication: 'api-key',
      requestHeaders: {
        'x-tenant-id': 'tenant-a',
        authorization: 'Bearer custom-key'
      },
      requestBody: {
        temperature: 0.2,
        model: 'custom-model',
        stream: true
      },
      fetcher
    })

    await expect(runtime.testConnection()).resolves.toMatchObject({
      available: true
    })
    const init = fetcher.mock.calls[0]?.[1]
    const headers = new Headers(init?.headers)
    const body = JSON.parse(String(init?.body)) as Record<string, unknown>
    expect(headers.get('x-tenant-id')).toBe('tenant-a')
    expect(headers.get('authorization')).toBe('Bearer runtime-key')
    expect(body).toMatchObject({
      model: 'runtime-model',
      stream: false,
      temperature: 0.2
    })
  })

  it('uses the Anthropic messages endpoint and streams text deltas', async () => {
    const fetcher = vi.fn<typeof fetch>(async () => {
      return new Response(createEventStream('真实模型回答', '先分析问题'), {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      maximumOutputTokens: 48_000,
      skillInstructions: '# 文档写作',
      fetcher
    })
    const events: Array<{ type: string; state?: string }> = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed125',
        conversationId: 'conversation-1',
        prompt: '你好',
        trustedInstructions: 'Trusted specialist system instruction.'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledOnce()
    const [input, init] = fetcher.mock.calls[0] ?? []
    expect(input?.toString()).toBe('https://bigtoken.ai/v1/messages')
    expect(init?.method).toBe('POST')

    const body = JSON.parse(init?.body as string) as {
      model: string
      max_tokens: number
      stream: boolean
      system: string
    }
    expect(body).toMatchObject({
      model: 'sonnet-5',
      max_tokens: 48_000,
      stream: true
    })
    expect(body.system).toMatch(
      /Current system time: \d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2}\./u
    )
    expect(body.system).toContain('# 文档写作')
    expect(body.system).toContain('Trusted specialist system instruction.')
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'reasoning',
        delta: '先分析问题'
      })
    )
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'text',
        delta: '真实模型回答'
      })
    )
    expect(events.filter((event) => event.type === 'model-usage')).toEqual([
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed125',
        type: 'model-usage',
        callId: expect.stringMatching(
          /^model-call:[0-9a-f]{32}:message-1$/u
        ),
        runtime: 'model',
        provider: 'anthropic',
        model: 'claude-sonnet-provider',
        inputTokens: 23,
        outputTokens: 11,
        cacheReadTokens: 7,
        cacheWriteTokens: 5
      }
    ])
    expect(events.at(-2)).toMatchObject({ type: 'model-usage' })
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('uses a hard-limit preflight guard while preserving recent raw turns', async () => {
    const fetcher = vi
      .fn<typeof fetch>()
      .mockResolvedValueOnce(
        new Response(createEventStream('压缩后的摘要'), {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
      .mockResolvedValueOnce(
        new Response(createEventStream('继续回答'), {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      requestHeaders: { 'x-model-route': 'summary-test' },
      requestBody: { temperature: 0.1 },
      fetcher,
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 15_000,
          recentRawTokens: 5_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Summarize earlier history.'
        },
        contextWindowTokens: 32_000
      }
    })
    const history = [
      { role: 'user' as const, content: `old-user-${'a'.repeat(16_000)}` },
      {
        role: 'assistant' as const,
        content: `old-assistant-${'b'.repeat(16_000)}`
      },
      { role: 'user' as const, content: `mid-user-${'c'.repeat(16_000)}` },
      {
        role: 'assistant' as const,
        content: `mid-assistant-${'d'.repeat(16_000)}`
      },
      { role: 'user' as const, content: `new-user-${'e'.repeat(16_000)}` },
      {
        role: 'assistant' as const,
        content: `new-assistant-${'f'.repeat(16_000)}`
      }
    ]
    const historyMessageIds = [
      '00000000-0000-4000-8000-000000000221',
      '00000000-0000-4000-8000-000000000222',
      '00000000-0000-4000-8000-000000000223',
      '00000000-0000-4000-8000-000000000224',
      '00000000-0000-4000-8000-000000000225',
      '00000000-0000-4000-8000-000000000226'
    ]
    const events: RuntimeEvent[] = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed222',
        conversationId: 'conversation-compressed',
        prompt: '继续',
        history,
        historyMessageIds
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledTimes(2)
    for (const [, init] of fetcher.mock.calls) {
      expect(new Headers(init?.headers).get('x-model-route')).toBe(
        'summary-test'
      )
      expect(JSON.parse(String(init?.body))).toMatchObject({
        temperature: 0.1
      })
    }
    const summaryBody = JSON.parse(
      fetcher.mock.calls[0]![1]!.body as string
    ) as { max_tokens: number; system: string; messages: unknown[] }
    expect(summaryBody.max_tokens).toBe(32_000)
    expect(summaryBody.system).toContain('Summarize earlier history.')
    expect(JSON.stringify(summaryBody.messages)).toContain('old-user-')
    expect(JSON.stringify(summaryBody.messages)).not.toContain('new-user-')
    expect(JSON.stringify(summaryBody)).not.toContain(
      historyMessageIds[0]
    )

    const answerBody = JSON.parse(
      fetcher.mock.calls[1]![1]!.body as string
    ) as { messages: unknown[] }
    const answerMessages = JSON.stringify(answerBody.messages)
    expect(answerMessages).toContain('压缩后的摘要')
    expect(answerMessages).toContain('new-user-')
    expect(answerMessages).not.toContain('old-user-')
    for (const id of historyMessageIds) {
      expect(answerMessages).not.toContain(id)
    }
    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),
        conversationState: expect.objectContaining({
          coveredMessageCount: 4,
          coveredFromMessageId: historyMessageIds[0],
          coveredThroughMessageId: historyMessageIds[3],
          coveredHistoryDigest: expect.stringMatching(/^[0-9a-f]{64}$/u),
          summary: '压缩后的摘要'
        })
      })
    )
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'context-metrics',
        contextTokens: 46,
        source: 'provider'
      })
    )
    expect(events).not.toContainEqual(
      expect.objectContaining({
        type: 'status',
        message: '正在准备直连模型上下文'
      })
    )
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'model-usage',
        callId: expect.stringMatching(
          /^context-summary:model-call:[0-9a-f]{32}:message-1$/u
        )
      })
    )
    expect(events).toContainEqual(
      expect.objectContaining({ type: 'text', delta: '继续回答' })
    )
  })

  it('manually compacts Continue history even when automatic compression is disabled', async () => {
    const fetcher = vi.fn<typeof fetch>().mockResolvedValueOnce(
      new Response(createEventStream('手动压缩摘要'), {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      contextCompression: {
        settings: {
          enabled: false,
          triggerTokens: 20_000,
          recentRawTokens: 5_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Summarize earlier history.'
        },
        contextWindowTokens: 32_000
      }
    })
    const history = [
      { role: 'user' as const, content: 'earlier question' },
      { role: 'assistant' as const, content: 'earlier answer' },
      { role: 'user' as const, content: 'recent question' },
      { role: 'assistant' as const, content: 'recent answer' }
    ]
    const outcome = await runtime.compactConversation(
      {
        requestId: '00000000-0000-4000-8000-000000000081',
        conversationId: '00000000-0000-4000-8000-000000000082',
        runtimeSelection: { provider: 'continue' },
        history,
        historyMessageIds: [
          '00000000-0000-4000-8000-000000000083',
          '00000000-0000-4000-8000-000000000084',
          '00000000-0000-4000-8000-000000000085',
          '00000000-0000-4000-8000-000000000086'
        ]
      },
      new AbortController().signal
    )

    expect(fetcher).toHaveBeenCalledOnce()
    const body = JSON.parse(
      fetcher.mock.calls[0]![1]!.body as string
    ) as { messages: unknown[] }
    expect(JSON.stringify(body.messages)).toContain('earlier question')
    expect(JSON.stringify(body.messages)).not.toContain('recent question')
    expect(outcome.result).toMatchObject({
      provider: 'continue',
      strategy: 'goodbuddy-summary',
      compacted: true,
      contextCompressionState: {
        coveredMessageCount: 2,
        coveredFromMessageId:
          '00000000-0000-4000-8000-000000000083',
        coveredThroughMessageId:
          '00000000-0000-4000-8000-000000000084',
        summary: '手动压缩摘要'
      }
    })
    await runtime.dispose()
  })

  it('waits for completed provider usage before normal threshold compression', async () => {
    const fetcher = vi
      .fn<typeof fetch>()
      .mockResolvedValueOnce(
        new Response(
          createEventStream('正常回答', undefined, {
            inputTokens: 15_500,
            outputTokens: 500,
            cacheReadTokens: 0,
            cacheWriteTokens: 0
          }),
          {
            status: 200,
            headers: { 'content-type': 'text/event-stream' }
          }
        )
      )
      .mockResolvedValueOnce(
        new Response(createEventStream('回复后摘要'), {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
    const history = [
      { role: 'user' as const, content: 'a'.repeat(12_000) },
      { role: 'assistant' as const, content: 'b'.repeat(12_000) },
      { role: 'user' as const, content: 'c'.repeat(12_000) },
      { role: 'assistant' as const, content: 'd'.repeat(12_000) }
    ]
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 15_000,
          recentRawTokens: 5_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        },
        contextWindowTokens: 32_000
      }
    })
    const events: RuntimeEvent[] = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: '继续',
        history
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledTimes(2)
    const firstBody = JSON.parse(
      fetcher.mock.calls[0]?.[1]?.body as string
    ) as { system: string; messages: unknown[] }
    expect(firstBody.system).not.toContain(
      'Conversation history and any existing summary'
    )
    expect(JSON.stringify(firstBody.messages)).toContain(
      'a'.repeat(1_000)
    )
    expect(
      events.findIndex(
        (event) =>
          event.type === 'context-compression' &&
          event.state === 'started'
      )
    ).toBeGreaterThan(
      events.findIndex(
        (event) =>
          event.type === 'text' && event.delta === '正常回答'
      )
    )
    expect(
      events.filter((event) => event.type === 'context-metrics')
    ).toEqual([
      expect.objectContaining({
        contextTokens: 16_000,
        source: 'provider'
      })
    ])
  })

  it('reuses persisted conversation summary state after Runtime restart', async () => {
    const history = [
      { role: 'user' as const, content: 'old question' },
      { role: 'assistant' as const, content: 'old answer' },
      { role: 'user' as const, content: 'recent question' },
      { role: 'assistant' as const, content: 'recent answer' }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(createEventStream('continued answer'), {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 200_000,
          recentRawTokens: 32_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        }
      }
    })

    for await (const _event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: 'continue',
        history,
        contextCompressionState: {
          coveredHistoryDigest: createHash('sha256')
            .update(JSON.stringify(history.slice(0, 2)))
            .digest('hex'),
          coveredMessageCount: 2,
          summary: 'persisted summary'
        }
      },
      new AbortController().signal
    )) {
      void _event
    }

    expect(fetcher).toHaveBeenCalledOnce()
    const body = JSON.parse(
      fetcher.mock.calls[0]?.[1]?.body as string
    ) as { messages: unknown[] }
    const messages = JSON.stringify(body.messages)
    expect(messages).toContain('persisted summary')
    expect(messages).toContain('recent question')
    expect(messages).not.toContain('old question')
  })

  it('keeps a persisted summary after its covered prefix rolls out of bounded history', async () => {
    const history = [
      { role: 'user' as const, content: 'recent question' },
      { role: 'assistant' as const, content: 'recent answer' }
    ]
    const historyMessageIds = [
      '00000000-0000-4000-8000-000000000603',
      '00000000-0000-4000-8000-000000000604'
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(createEventStream('continued answer'), {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 200_000,
          recentRawTokens: 32_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        }
      }
    })

    for await (const _event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: 'continue',
        history,
        historyMessageIds,
        contextCompressionState: {
          coveredHistoryDigest: createHash('sha256')
            .update(
              JSON.stringify([
                { role: 'user', content: 'evicted question' },
                { role: 'assistant', content: 'evicted answer' }
              ])
            )
            .digest('hex'),
          coveredMessageCount: 2,
          coveredFromMessageId:
            '00000000-0000-4000-8000-000000000601',
          coveredThroughMessageId:
            '00000000-0000-4000-8000-000000000602',
          summary: 'persisted evicted summary'
        }
      },
      new AbortController().signal
    )) {
      void _event
    }

    const body = JSON.parse(
      fetcher.mock.calls[0]?.[1]?.body as string
    ) as { messages: unknown[] }
    const messages = JSON.stringify(body.messages)
    expect(messages).toContain('persisted evicted summary')
    expect(messages).toContain('recent question')
    expect(messages).not.toContain('evicted question')
  })

  it('compresses a completed oversized response before reporting done', async () => {
    const longAnswer = `最终长回复-${'终'.repeat(12_000)}`
    const fetcher = vi
      .fn<typeof fetch>()
      .mockResolvedValueOnce(
        new Response(
          createEventStream(longAnswer, undefined, {
            inputTokens: 11_500,
            outputTokens: 1_000,
            cacheReadTokens: 0,
            cacheWriteTokens: 0
          }),
          {
            status: 200,
            headers: { 'content-type': 'text/event-stream' }
          }
        )
      )
      .mockResolvedValueOnce(
        new Response(createEventStream('最终回合摘要'), {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 12_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        }
      }
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: '生成长回复'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledTimes(2)
    const compressionEvents = events.filter(
      (event) =>
        event.type === 'context-compression' &&
        event.scope === 'conversation'
    )
    expect(compressionEvents).toEqual([
      expect.objectContaining({
        state: 'started',
        estimatedBeforeTokens: expect.any(Number)
      }),
      expect.objectContaining({
        state: 'completed',
        estimatedAfterTokens: expect.any(Number)
      })
    ])
    expect(
      events.findIndex(
        (event) =>
          event.type === 'context-compression' &&
          event.state === 'started'
      )
    ).toBeGreaterThan(
      events.findIndex(
        (event) =>
          event.type === 'text' && event.delta === longAnswer
      )
    )
    expect(
      events.filter((event) => event.type === 'context-metrics')
    ).toEqual([
      expect.objectContaining({
        contextTokens: 12_500,
        source: 'provider'
      })
    ])
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('keeps a completed answer when post-response compression fails', async () => {
    const longAnswer = `最终长回复-${'终'.repeat(12_000)}`
    const fetcher = vi
      .fn<typeof fetch>()
      .mockResolvedValueOnce(
        new Response(
          createEventStream(longAnswer, undefined, {
            inputTokens: 11_500,
            outputTokens: 1_000,
            cacheReadTokens: 0,
            cacheWriteTokens: 0
          }),
          {
            status: 200,
            headers: { 'content-type': 'text/event-stream' }
          }
        )
      )
      .mockResolvedValueOnce(
        Response.json(
          { error: { message: 'summary unavailable' } },
          { status: 503 }
        )
      )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 12_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        }
      }
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: '生成长回复'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'context-compression',
        scope: 'conversation',
        state: 'failed'
      })
    )
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('does not publish partial usage when a stream ends without message_stop', async () => {
    const completeStream = createEventStream('partial')
    const stream = completeStream.slice(
      0,
      completeStream.indexOf('event: message_stop')
    )
    const fetcher = vi.fn<typeof fetch>(async () => {
      return new Response(stream, {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 20_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        },
        contextWindowTokens: 32_000
      }
    })
    const events: RuntimeEvent[] = []

    const consume = async (): Promise<void> => {
      for await (const event of runtime.run(
        {
          requestId: 'a431666e-5ec8-45e6-beb4-654132eed126',
          conversationId: 'conversation-2',
          prompt: '你好'
        },
        new AbortController().signal
      )) {
        events.push(event)
      }
    }

    await expect(consume()).rejects.toThrow('意外中断')
    expect(
      events.filter((event) => event.type === 'context-metrics')
    ).toEqual([])
  })

  it('estimates completed usage only when the provider omits usage', async () => {
    const stream = [
      'event: content_block_delta',
      'data: {"type":"content_block_delta","delta":{"type":"text_delta","text":"fallback answer"}}',
      '',
      'event: message_stop',
      'data: {"type":"message_stop"}',
      '',
      ''
    ].join('\n')
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async () =>
        new Response(stream, {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      ),
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 20_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        },
        contextWindowTokens: 32_000
      }
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: 'provider omits usage'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    expect(
      events.filter((event) => event.type === 'context-metrics')
    ).toEqual([
      expect.objectContaining({
        contextTokens: expect.any(Number),
        source: 'estimated'
      })
    ])
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('rejects malformed SSE JSON instead of silently skipping it', async () => {
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async () =>
        new Response('data: {invalid}\n\n', {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
    })
    const consume = async (): Promise<void> => {
      for await (const _event of runtime.run(
        {
          requestId: crypto.randomUUID(),
          conversationId: crypto.randomUUID(),
          prompt: 'test'
        },
        new AbortController().signal
      )) {
        void _event
      }
    }

    await expect(consume()).rejects.toThrow('无效的流式 JSON')
  })

  it('keeps the latest confirmed context usage when a model request is cancelled', async () => {
    const controller = new AbortController()
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async (_input, init) => {
        const requestSignal = init?.signal
        if (!requestSignal) {
          throw new Error('Missing request signal')
        }
        return await new Promise<Response>((_resolve, reject) => {
          requestSignal.addEventListener(
            'abort',
            () => reject(requestSignal.reason),
            { once: true }
          )
        })
      }),
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 20_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        },
        contextWindowTokens: 32_000
      }
    })
    const stream = runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: '等待取消'
      },
      controller.signal
    )

    await expect(stream.next()).resolves.toMatchObject({
      value: { type: 'status' }
    })
    const pending = stream.next()
    controller.abort(new Error('用户取消'))

    await expect(pending).rejects.toThrow('用户取消')
  })

  it('keeps the latest confirmed context usage when the model API fails', async () => {
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async () =>
        Response.json(
          { error: { message: 'upstream unavailable' } },
          { status: 503 }
        )
      ),
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 20_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve important facts.'
        },
        contextWindowTokens: 32_000
      }
    })
    const stream = runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: '触发失败'
      },
      new AbortController().signal
    )

    await expect(stream.next()).resolves.toMatchObject({
      value: { type: 'status' }
    })
    await expect(stream.next()).rejects.toThrow('upstream unavailable')
  })

  it('retries a nested ECONNRESET before streaming output', async () => {
    vi.useFakeTimers()
    try {
      const connectionReset = Object.assign(
        new Error(
          'Client network socket disconnected before secure TLS connection was established'
        ),
        { code: 'ECONNRESET', port: 443 }
      )
      const fetcher = vi
        .fn<typeof fetch>()
        .mockRejectedValueOnce(
          new TypeError('fetch failed', { cause: connectionReset })
        )
        .mockResolvedValueOnce(
          new Response(createEventStream('重试后成功'), {
            status: 200,
            headers: { 'content-type': 'text/event-stream' }
          })
        )
      const runtime = new ModelAgentRuntime({
        apiKey: 'test-key',
        baseUrl: 'https://bigtoken.ai',
        model: 'sonnet-5',
        protocol: 'anthropic-messages',
        authentication: 'api-key',
        fetcher
      })
      const events: RuntimeEvent[] = []
      const consume = async (): Promise<void> => {
        for await (const event of runtime.run(
          {
            requestId: crypto.randomUUID(),
            conversationId: crypto.randomUUID(),
            prompt: 'test'
          },
          new AbortController().signal
        )) {
          events.push(event)
        }
      }
      const completed = consume()

      await vi.runAllTimersAsync()
      await completed

      expect(fetcher).toHaveBeenCalledTimes(2)
      expect(events).toContainEqual(
        expect.objectContaining({
          type: 'status',
          message: '模型网络请求失败，等待重试（1/3，0.5 秒后）'
        })
      )
      const retryWaiting = events.findIndex((event) =>
        event.type === 'status' && event.message.includes('等待重试'))
      const retryStarted = events.findIndex((event) =>
        event.type === 'status' && event.message.includes('正在重试请求'))
      const responseStarted = events.findIndex((event) => event.type === 'text')
      expect(retryStarted).toBeGreaterThan(retryWaiting)
      expect(responseStarted).toBeGreaterThan(retryStarted)
      expect(events).toContainEqual(
        expect.objectContaining({
          type: 'text',
          delta: '重试后成功'
        })
      )
      expect(events.at(-1)).toMatchObject({ type: 'done' })
    } finally {
      vi.useRealTimers()
    }
  })

  it('does not retry a transient stream failure after publishing output', async () => {
    const connectionReset = Object.assign(
      new Error('socket disconnected'),
      { code: 'ECONNRESET' }
    )
    let sentOutput = false
    const body = new ReadableStream<Uint8Array>({
      pull(controller) {
        if (sentOutput) {
          controller.error(connectionReset)
          return
        }
        sentOutput = true
        controller.enqueue(
          new TextEncoder().encode(
            [
              'event: content_block_delta',
              'data: {"type":"content_block_delta","delta":{"type":"text_delta","text":"已经显示"}}',
              '',
              ''
            ].join('\n')
          )
        )
      }
    })
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(body, {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher
    })
    const events: RuntimeEvent[] = []
    const consume = async (): Promise<void> => {
      for await (const event of runtime.run(
        {
          requestId: crypto.randomUUID(),
          conversationId: crypto.randomUUID(),
          prompt: 'test'
        },
        new AbortController().signal
      )) {
        events.push(event)
      }
    }

    await expect(consume()).rejects.toThrow('socket disconnected')
    expect(fetcher).toHaveBeenCalledOnce()
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'text',
        delta: '已经显示'
      })
    )
    expect(events).not.toContainEqual(
      expect.objectContaining({
        message: expect.stringContaining('等待重试')
      })
    )
  })

  it('does not retry a reset whose dispatch outcome is unknown', async () => {
    const fetcher = vi.fn<typeof fetch>(async () => {
      throw Object.assign(new Error('socket disconnected'), {
        code: 'ECONNRESET'
      })
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher
    })
    const events: RuntimeEvent[] = []
    const consume = async (): Promise<void> => {
      for await (const event of runtime.run(
        {
          requestId: crypto.randomUUID(),
          conversationId: crypto.randomUUID(),
          prompt: 'test'
        },
        new AbortController().signal
      )) {
        events.push(event)
      }
    }

    await expect(consume()).rejects.toThrow('socket disconnected')
    expect(fetcher).toHaveBeenCalledOnce()
    expect(events).not.toContainEqual(
      expect.objectContaining({
        message: expect.stringContaining('等待重试')
      })
    )
  })

  it('cancels immediately during model retry backoff', async () => {
    const controller = new AbortController()
    const fetcher = vi.fn<typeof fetch>(async () => {
      throw Object.assign(new Error('connect timeout'), {
        code: 'UND_ERR_CONNECT_TIMEOUT'
      })
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher
    })
    const stream = runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: 'test'
      },
      controller.signal
    )

    await expect(stream.next()).resolves.toMatchObject({
      value: { type: 'status', message: 'sonnet-5 请求处理中' }
    })
    await expect(stream.next()).resolves.toMatchObject({
      value: {
        type: 'status',
        message: '模型网络请求失败，等待重试（1/3，0.5 秒后）'
      }
    })
    const pendingRetry = stream.next()
    controller.abort(new Error('用户取消'))

    await expect(pendingRetry).rejects.toThrow('用户取消')
    expect(fetcher).toHaveBeenCalledOnce()
  })

  it('parses CRLF event separators split across response chunks', async () => {
    const payload = createEventStream('split CRLF').replaceAll('\n', '\r\n')
    const splitAt = payload.indexOf('\r\n\r\n') + 3
    const body = new ReadableStream<Uint8Array>({
      start(controller) {
        controller.enqueue(
          new TextEncoder().encode(payload.slice(0, splitAt))
        )
        controller.enqueue(
          new TextEncoder().encode(payload.slice(splitAt))
        )
        controller.close()
      }
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async () =>
        new Response(body, {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: 'test'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'text',
        delta: 'split CRLF'
      })
    )
  })

  it('aborts a model request that exceeds the runtime timeout', async () => {
    vi.useFakeTimers()
    try {
      const fetcher = vi.fn<typeof fetch>(
        async (_input, init) =>
          new Promise<Response>((_resolve, reject) => {
            init?.signal?.addEventListener(
              'abort',
              () => reject(init.signal?.reason),
              { once: true }
            )
          })
      )
      const runtime = new ModelAgentRuntime({
        apiKey: 'test-key',
        baseUrl: 'https://bigtoken.ai',
        model: 'sonnet-5',
        protocol: 'anthropic-messages',
        authentication: 'api-key',
        requestTimeoutMs: 50,
        fetcher
      })
      const consume = async (): Promise<void> => {
        for await (const _event of runtime.run(
          {
            requestId: crypto.randomUUID(),
            conversationId: crypto.randomUUID(),
            prompt: 'test'
          },
          new AbortController().signal
        )) {
          void _event
        }
      }
      const assertion = expect(consume()).rejects.toThrow(
        '模型接口请求超时'
      )

      await vi.runAllTimersAsync()
      await assertion
      expect(fetcher).toHaveBeenCalledOnce()
    } finally {
      vi.useRealTimers()
    }
  })

  it('aborts a stalled response body after headers arrive', async () => {
    vi.useFakeTimers()
    try {
      let responseSignal: AbortSignal | null | undefined
      const runtime = new ModelAgentRuntime({
        apiKey: 'test-key',
        baseUrl: 'https://bigtoken.ai',
        model: 'sonnet-5',
        protocol: 'anthropic-messages',
        authentication: 'api-key',
        requestTimeoutMs: 50,
        fetcher: vi.fn<typeof fetch>(async (_input, init) => {
          responseSignal = init?.signal
          return new Response(
            new ReadableStream<Uint8Array>({
              start(controller) {
                init?.signal?.addEventListener(
                  'abort',
                  () => controller.error(init.signal?.reason),
                  { once: true }
                )
              }
            }),
            {
              status: 200,
              headers: { 'content-type': 'text/event-stream' }
            }
          )
        })
      })
      const consume = async (): Promise<void> => {
        for await (const _event of runtime.run(
          {
            requestId: crypto.randomUUID(),
            conversationId: crypto.randomUUID(),
            prompt: 'test'
          },
          new AbortController().signal
        )) {
          void _event
        }
      }
      const assertion = expect(consume()).rejects.toThrow(
        '模型接口请求超时'
      )

      await vi.runAllTimersAsync()
      await assertion
      expect(responseSignal?.aborted).toBe(true)
    } finally {
      vi.useRealTimers()
    }
  })

  it('streams long ordinary responses without a size limit', async () => {
    const encoder = new TextEncoder()
    const deltaChunks = 48
    const deltaText = 'x'.repeat(65_000)
    const body = new ReadableStream<Uint8Array>({
      start(controller) {
        controller.enqueue(
          encoder.encode(
            `data: ${JSON.stringify({
              type: 'message_start',
              message: {
                id: 'message-long',
                model: 'sonnet-5',
                usage: { input_tokens: 5 }
              }
            })}\n\n`
          )
        )
        for (let index = 0; index < deltaChunks; index += 1) {
          controller.enqueue(
            encoder.encode(
              `data: ${JSON.stringify({
                type: 'content_block_delta',
                delta: {
                  type: 'text_delta',
                  text: deltaText
                }
              })}\n\n`
            )
          )
        }
        controller.enqueue(
          encoder.encode('data: {"type":"message_stop"}\n\n')
        )
        controller.close()
      }
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async () =>
        new Response(body, {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
    })
    let streamedText = ''
    let done = false
    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: 'test'
      },
      new AbortController().signal
    )) {
      if (event.type === 'text') {
        streamedText += event.delta
      }
      if (event.type === 'done') {
        done = true
      }
    }

    expect(done).toBe(true)
    expect(streamedText.length).toBe(deltaChunks * deltaText.length)
  })

  it('preserves bounded provider error messages', async () => {
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'claude-sonnet-5',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async () =>
        Response.json(
          {
            error: {
              message:
                'upstream failed Authorization: Bearer secret-token'
            }
          },
          { status: 502 }
        )
      )
    })
    const consume = async (): Promise<void> => {
      for await (const _event of runtime.run(
        {
          requestId: crypto.randomUUID(),
          conversationId: crypto.randomUUID(),
          prompt: 'test'
        },
        new AbortController().signal
      )) {
        void _event
      }
    }

    await expect(consume()).rejects.toThrow(
      'upstream failed Authorization: Bearer secret-token'
    )
  })

  it('uses OpenAI Chat Completions SSE and omits auth for Ollama', async () => {
    const stream = [
      `data: ${JSON.stringify({
        choices: [{ delta: { content: '本机回答' } }]
      })}`,
      '',
      `data: ${JSON.stringify({
        id: 'chatcmpl-provider-1',
        model: 'qwen3-provider',
        choices: [],
        usage: {
          prompt_tokens: 31,
          completion_tokens: 9,
          total_tokens: 40,
          prompt_tokens_details: { cached_tokens: 13 },
          cache_write_tokens: 4
        }
      })}`,
      '',
      'data: [DONE]',
      '',
      ''
    ].join('\n')
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(stream, {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed127',
        conversationId: 'conversation-3',
        prompt: '你好'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    const [input, init] = fetcher.mock.calls[0] ?? []
    expect(input?.toString()).toBe(
      'http://127.0.0.1:11434/v1/chat/completions'
    )
    expect(init?.headers).toEqual({
      'content-type': 'application/json'
    })
    expect(JSON.parse(init?.body as string)).toMatchObject({
      model: 'qwen3',
      stream: true,
      stream_options: {
        include_usage: true
      },
      messages: [
        expect.objectContaining({ role: 'system' }),
        expect.objectContaining({ role: 'user', content: '你好' })
      ]
    })
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'text',
        delta: '本机回答'
      })
    )
    expect(events.filter((event) => event.type === 'model-usage')).toEqual([
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed127',
        type: 'model-usage',
        callId: expect.stringMatching(
          /^model-call:[0-9a-f]{32}:chatcmpl-provider-1$/u
        ),
        runtime: 'model',
        provider: 'openai',
        model: 'qwen3-provider',
        inputTokens: 31,
        outputTokens: 9,
        cacheReadTokens: 13,
        cacheWriteTokens: 4,
        reportedTotalTokens: 40
      }
    ])
    expect(events.at(-2)).toMatchObject({ type: 'model-usage' })
    expect(events.at(-1)).toMatchObject({ type: 'done' })
    expect(toolProvider.listTools).not.toHaveBeenCalled()
  })

  it('retries transient network failures for direct-model tool rounds', async () => {
    vi.useFakeTimers()
    try {
      const fetcher = vi
        .fn<typeof fetch>()
        .mockRejectedValueOnce(
          Object.assign(new Error('connect timeout'), {
            code: 'UND_ERR_CONNECT_TIMEOUT'
          })
        )
        .mockResolvedValueOnce(
          Response.json({
            choices: [
              {
                message: {
                  role: 'assistant',
                  content: '工具轮次重试成功'
                }
              }
            ]
          })
        )
      const runtime = new ModelAgentRuntime({
        baseUrl: 'http://127.0.0.1:11434/v1',
        model: 'qwen3',
        protocol: 'openai-chat-completions',
        authentication: 'none',
        fetcher,
        toolProvider: createToolProvider()
      })
      const events: RuntimeEvent[] = []
      const consume = async (): Promise<void> => {
        for await (const event of runtime.run(
          {
            requestId: crypto.randomUUID(),
            conversationId: crypto.randomUUID(),
            prompt: '使用工具能力回答',
            workMode: 'execute'
          },
          new AbortController().signal,
          async () => 'once'
        )) {
          events.push(event)
        }
      }
      const completed = consume()

      await vi.runAllTimersAsync()
      await completed

      expect(fetcher).toHaveBeenCalledTimes(2)
      expect(events).toContainEqual(
        expect.objectContaining({
          type: 'status',
          message: '模型网络请求失败，等待重试（1/3，0.5 秒后）'
        })
      )
      expect(events).toContainEqual(
        expect.objectContaining({
          type: 'text',
          delta: '工具轮次重试成功'
        })
      )
      expect(events.at(-1)).toMatchObject({ type: 'done' })
    } finally {
      vi.useRealTimers()
    }
  })

  it('streams OpenAI-compatible reasoning deltas before the answer', async () => {
    const stream = [
      `data: ${JSON.stringify({
        choices: [
          {
            delta: {
              reasoning_content: '先分析'
            }
          }
        ]
      })}`,
      '',
      `data: ${JSON.stringify({
        choices: [
          {
            delta: {
              reasoning_content: '，再验证'
            }
          }
        ]
      })}`,
      '',
      `data: ${JSON.stringify({
        choices: [
          {
            delta: {
              content: '最终回答'
            }
          }
        ]
      })}`,
      '',
      'data: [DONE]',
      '',
      ''
    ].join('\n')
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.deepseek.com',
      model: 'deepseek-reasoner',
      protocol: 'openai-chat-completions',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async () =>
        new Response(stream, {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed128',
        conversationId: 'conversation-deepseek-reasoning',
        prompt: '分析这个问题'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    expect(
      events.filter(
        (event) => event.type === 'reasoning' || event.type === 'text'
      )
    ).toEqual([
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed128',
        type: 'reasoning',
        delta: '先分析'
      },
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed128',
        type: 'reasoning',
        delta: '，再验证'
      },
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed128',
        type: 'text',
        delta: '最终回答'
      }
    ])
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('runs Ask workspace search, text reads and output paging with the production deny authorizer and no optional capabilities', async () => {
    const workspace = await mkdtemp(join(tmpdir(), 'goodbuddy-ask-tools-'))
    const workspaceAccess = new LocalWorkspaceAccess(workspace)
    const processService = new LocalDirectModelProcessService()
    const toolProvider = new ModelToolProvider(
      workspaceAccess, [], undefined, undefined, false,
      { processService, ripgrepExecutablePath: rgPath }
    )
    const signal = new AbortController().signal
    const conversationId = 'conversation-ask'
    const responses: unknown[] = []
    const fetcher = vi.fn<typeof fetch>(async () => Response.json(responses.shift()))
    const runtime = new ModelAgentRuntime({
      baseUrl: 'https://example.test/v1',
      model: 'test-model',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider,
      workspaceAccess
    })
    const authorize = vi.fn(async () => 'deny' as const)
    try {
      await writeFile(join(workspace, 'README.md'), 'readonly-marker\n')
      await writeFile(join(workspace, 'output.cjs'),
        "process.stdout.write('x'.repeat(110000) + 'retained-tail')")
      // Seed output as a preceding Execute operation in the same conversation.
      const previous = await processService.execute(
        { command: 'node output.cjs', timeoutMs: 10_000 },
        { conversationId, workspace: workspaceAccess, signal }
      )
      expect(previous.exitCode).toBe(0)
      const reference = previous.stdoutReference!
      expect(reference).toBeDefined()
      const calls = [
        { name: 'workspace_rg', arguments: { pattern: 'readonly-marker', glob: ['README.md'] } },
        { name: 'workspace_read_text', arguments: { path: 'README.md' } },
        { name: 'output_read', arguments: { handle: reference.handle, cursor: reference.nextCursor } }
      ]
      responses.push({
        choices: [{ message: {
          role: 'assistant', content: null,
          tool_calls: calls.map((call, index) => ({
            id: `ask-${index}`, type: 'function',
            function: { name: call.name, arguments: JSON.stringify(call.arguments) }
          }))
        } }]
      }, { choices: [{ message: { role: 'assistant', content: '只读完成' } }] })
      const events: RuntimeEvent[] = []
      for await (const event of runtime.run({
        requestId: crypto.randomUUID(), conversationId,
        prompt: '搜索并读取文件，续读已有输出', workMode: 'ask'
      }, signal, authorize)) events.push(event)

      const firstBody = JSON.parse(String(fetcher.mock.calls[0]![1]!.body))
      expect(firstBody.tools.map((tool: { function: { name: string } }) =>
        tool.function.name)).toEqual(['workspace_rg', 'workspace_read_text', 'output_read'])
      const secondBody = JSON.parse(String(fetcher.mock.calls[1]![1]!.body))
      const results = secondBody.messages.filter(
        (message: { role: string }) => message.role === 'tool'
      )
      expect(results).toHaveLength(3)
      expect(results[0].content).toContain('README.md:1:1:readonly-marker')
      expect(results[1].content).toContain('readonly-marker')
      expect(JSON.parse(results[2].content)).toMatchObject({
        content: expect.stringContaining('retained-tail'), eof: true
      })
      expect(authorize).not.toHaveBeenCalled()
      expect(events.filter((event) =>
        event.type === 'tool' && event.state === 'completed')).toHaveLength(3)
      expect(events.at(-1)).toMatchObject({ type: 'done' })
      expect(await readFile(join(workspace, 'README.md'), 'utf8')).toBe('readonly-marker\n')
    } finally {
      await runtime.dispose()
      await rm(workspace, { recursive: true, force: true })
    }
  })

  it.each(['workspace_apply_patch', 'workspace_write_text', 'process_execute', 'browser_snapshot'])(
    'rejects unlisted %s in Ask at the runtime boundary',
    async (name) => {
      const workspace = await mkdtemp(join(tmpdir(), 'goodbuddy-ask-denial-'))
      const fetcher = vi.fn<typeof fetch>(async () => Response.json({
        choices: [{ message: {
          role: 'assistant', content: null,
          tool_calls: [{ id: 'spoof', type: 'function', function: {
            name, arguments: JSON.stringify({
              path: 'forbidden.txt', content: 'forbidden',
              patch: '*** Begin Patch\n*** Add File: forbidden.txt\n+forbidden\n*** End Patch',
              command: 'echo forbidden'
            })
          } }]
        } }]
      }))
      const runtime = new ModelAgentRuntime({
        baseUrl: 'https://example.test/v1', model: 'test-model',
        protocol: 'openai-chat-completions', authentication: 'none',
        fetcher, defaultWorkspace: workspace
      })
      const authorize = vi.fn(async () => 'deny' as const)
      try {
        const consume = async () => {
          for await (const event of runtime.run({
            requestId: crypto.randomUUID(), conversationId: 'ask-denied',
            prompt: '只读', workMode: 'ask'
          }, new AbortController().signal, authorize)) void event
        }
        await expect(consume()).rejects.toThrow('未知工具')
        expect(authorize).not.toHaveBeenCalled()
        await expect(readFile(join(workspace, 'forbidden.txt'))).rejects.toMatchObject({ code: 'ENOENT' })
      } finally {
        await runtime.dispose()
        await rm(workspace, { recursive: true, force: true })
      }
    }
  )

  it.each(['openai-chat-completions', 'openai-responses', 'anthropic-messages'] as const)(
    'lists only workspace reads and retained output in plain Ask via %s',
    async (protocol) => {
      const fetcher = vi.fn<typeof fetch>(async () => Response.json(
        protocol === 'anthropic-messages'
          ? { content: [{ type: 'text', text: '只读回答' }], stop_reason: 'end_turn' }
          : protocol === 'openai-responses'
            ? { id: 'resp-ask', output: [{ type: 'message', content: [{ type: 'output_text', text: '只读回答' }] }] }
            : { choices: [{ message: { role: 'assistant', content: '只读回答' } }] }
      ))
      const runtime = new ModelAgentRuntime({
        baseUrl: 'https://example.test/v1', model: 'test-model',
        protocol, authentication: 'none', fetcher
      })
      try {
        for await (const event of runtime.run({
          requestId: crypto.randomUUID(), conversationId: 'plain-ask',
          prompt: '只读', workMode: 'ask'
        }, new AbortController().signal, async () => 'deny')) void event
        const body = JSON.parse(String(fetcher.mock.calls[0]![1]!.body))
        expect(body.tools.map((tool: { name?: string; function?: { name: string } }) =>
          tool.function?.name ?? tool.name)).toEqual(['workspace_read_text', 'output_read'])
      } finally {
        await runtime.dispose()
      }
    }
  )

  it('uses the OpenAI Responses endpoint and streams output text', async () => {
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(
        createResponsesEventStream('Responses 回答', 'Responses 推理'),
        {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
        }
      )
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.openai.com/v1',
      model: 'gpt-5',
      protocol: 'openai-responses',
      authentication: 'api-key',
      fetcher
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed133',
        conversationId: 'conversation-responses',
        prompt: '你好'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    const [input, init] = fetcher.mock.calls[0] ?? []
    expect(input?.toString()).toBe('https://api.openai.com/v1/responses')
    expect(init?.headers).toEqual({
      authorization: 'Bearer test-key',
      'content-type': 'application/json'
    })
    expect(JSON.parse(init?.body as string)).toMatchObject({
      model: 'gpt-5',
      stream: true,
      instructions: expect.stringContaining('GoodBuddy'),
      input: [
        expect.objectContaining({ role: 'user', content: '你好' })
      ]
    })
    expect(JSON.parse(init?.body as string)).not.toHaveProperty(
      'max_output_tokens'
    )
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'reasoning',
        delta: 'Responses 推理'
      })
    )
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'text',
        delta: 'Responses 回答'
      })
    )
    expect(events.filter((event) => event.type === 'model-usage')).toEqual([
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed133',
        type: 'model-usage',
        callId: expect.stringMatching(
          /^model-call:[0-9a-f]{32}:resp-provider-1$/u
        ),
        runtime: 'model',
        provider: 'openai',
        model: 'gpt-5-provider',
        inputTokens: 29,
        outputTokens: 8,
        cacheReadTokens: 11,
        cacheWriteTokens: 0,
        reportedTotalTokens: 37
      }
    ])
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('tests an OpenAI Responses connection with Responses request fields', async () => {
    const fetcher = vi.fn<typeof fetch>(async (_input, init) => {
      const body = JSON.parse(init?.body as string) as {
        input: string
      }
      const marker =
        /GOODBUDDY_MODEL_TEST_[A-F0-9]+/u.exec(body.input)?.[0]
      if (!marker) {
        throw new Error('missing model test marker')
      }
      return Response.json({
        id: 'resp-test',
        status: 'completed',
        output: [
          {
            type: 'message',
            content: [{ type: 'output_text', text: marker }]
          }
        ]
      })
    })
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.openai.com/v1/',
      model: 'gpt-5',
      protocol: 'openai-responses',
      authentication: 'api-key',
      fetcher
    })

    await expect(runtime.testConnection()).resolves.toMatchObject({
      available: true,
      detail: expect.stringContaining('真实模型生成测试')
    })
    expect(fetcher.mock.calls[0]?.[0]?.toString()).toBe(
      'https://api.openai.com/v1/responses'
    )
    expect(
      JSON.parse(fetcher.mock.calls[0]?.[1]?.body as string)
    ).toMatchObject({
      model: 'gpt-5',
      max_output_tokens: 64,
      stream: false
    })
    expect(
      JSON.parse(fetcher.mock.calls[0]?.[1]?.body as string).input
    ).toMatch(/GOODBUDDY_MODEL_TEST_[A-F0-9]+/u)
  })

  it('runs approved direct-model tools and returns their results to OpenAI', async () => {
    const responses = [
      {
        model: 'qwen3',
        choices: [
          {
            message: {
              role: 'assistant',
              content: null,
              tool_calls: [
                {
                  id: 'call-1',
                  type: 'function',
                  function: {
                    name: 'workspace_read_text',
                    arguments: '{"path":"README.md"}'
                  }
                }
              ]
            }
          }
        ],
        usage: { prompt_tokens: 10, completion_tokens: 4 }
      },
      {
        model: 'qwen3',
        choices: [
          {
            message: {
              role: 'assistant',
              content: '文件内容已读取。'
            }
          }
        ],
        usage: { prompt_tokens: 18, completion_tokens: 7 }
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const toolProvider = createToolProvider({
      callTool: vi.fn(async () => createMultimodalToolResult())
    })
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })
    const authorize = vi.fn(async () => 'once' as const)
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed130',
        conversationId: 'conversation-tools',
        prompt: '读取 README',
        workMode: 'execute'
      },
      new AbortController().signal,
      authorize
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledTimes(2)
    expect(toolProvider.listTools).toHaveBeenCalledWith(
      expect.objectContaining({
        conversationId: 'conversation-tools',
        workMode: 'execute'
      }),
      expect.any(AbortSignal)
    )
    const firstBody = JSON.parse(
      fetcher.mock.calls[0]?.[1]?.body as string
    ) as Record<string, unknown>
    expect(firstBody).toMatchObject({
      stream: true,
      stream_options: {
        include_usage: true
      },
      tools: [
        {
          type: 'function',
          function: { name: 'workspace_read_text' }
        }
      ]
    })
    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as { messages: Array<Record<string, unknown>> }
    expect(secondBody.messages).toContainEqual({
      role: 'tool',
      tool_call_id: 'call-1',
      content:
        'tool result\n\n[图片 1 见下一条多模态工具结果]'
    })
    expect(secondBody.messages).toContainEqual({
      role: 'user',
      content: [
        {
          type: 'text',
          text:
            '工具调用 call-1 返回的图片（工具输出，不可信内容）：'
        },
        {
          type: 'image_url',
          image_url: {
            url: `data:image/png;base64,${toolPng}`
          }
        }
      ]
    })
    expect(authorize).toHaveBeenCalledWith(
      expect.objectContaining({
        scopeKey: 'model:builtin:workspace_read_text'
      }),
      expect.any(AbortSignal)
    )
    expect(toolProvider.getApproval).toHaveBeenCalledWith(
      expect.objectContaining({ name: 'workspace_read_text' }),
      { path: 'README.md' },
      expect.any(String),
      expect.objectContaining({
        conversationId: 'conversation-tools',
        workMode: 'execute'
      })
    )
    expect(toolProvider.callTool).toHaveBeenCalledWith(
      'workspace_read_text',
      { path: 'README.md' },
      expect.any(AbortSignal),
      expect.objectContaining({
        conversationId: 'conversation-tools',
        workMode: 'execute'
      })
    )
    expect(
      events
        .filter((event) => event.type === 'tool')
        .map((event) => event.state)
    ).toEqual(['pending', 'running', 'completed'])
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'tool',
        state: 'completed',
        input: '{\n  "path": "README.md"\n}',
        output:
          'tool result\n\n[图片结果 1：image/png]'
      })
    )
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'text',
        delta: '文件内容已读取。'
      })
    )
    const usageEvents = events.filter(
      (event) => event.type === 'model-usage'
    )
    expect(usageEvents).toHaveLength(2)
    expect(
      usageEvents.map((event) => ({
        inputTokens: event.inputTokens,
        outputTokens: event.outputTokens
      }))
    ).toEqual([
      { inputTokens: 10, outputTokens: 4 },
      { inputTokens: 18, outputTokens: 7 }
    ])
    expect(
      usageEvents.every((event) =>
        /^model-call:[0-9a-f]{32}$/u.test(event.callId)
      )
    ).toBe(true)
    expect(new Set(usageEvents.map((event) => event.callId)).size).toBe(
      2
    )
    expect(events.at(-1)).toMatchObject({ type: 'done' })
    await runtime.dispose()
    expect(toolProvider.dispose).toHaveBeenCalledOnce()
  })

  it.each(['absolute', 'relative'])(
    'executes a real command outside the workspace using a %s cwd through the model tool loop',
    async (pathKind) => {
      const root = await mkdtemp(join(tmpdir(), 'goodbuddy-model-cwd-'))
      const workspace = join(root, 'workspace')
      const target = join(root, 'target')
      await mkdir(workspace)
      await mkdir(target)
      await writeFile(join(target, 'input.txt'), 'outside-workspace')
      const command = process.platform === 'win32'
        ? "Copy-Item 'input.txt' 'output.txt'"
        : 'cp input.txt output.txt'
      const responses = [
        {
          choices: [{
            message: {
              role: 'assistant', content: null,
              tool_calls: [{
                id: 'call-process', type: 'function',
                function: {
                  name: 'process_execute',
                  arguments: JSON.stringify({
                    command,
                    cwd: pathKind === 'absolute' ? target : relative(workspace, target)
                  })
                }
              }]
            }
          }]
        },
        { choices: [{ message: { role: 'assistant', content: 'done' } }] }
      ]
      const fetcher = vi.fn<typeof fetch>(async () => Response.json(responses.shift()))
      const runtime = new ModelAgentRuntime({
        baseUrl: 'http://127.0.0.1:11434/v1',
        model: 'test-model',
        protocol: 'openai-chat-completions',
        authentication: 'none',
        fetcher,
        defaultWorkspace: workspace
      })
      try {
        const events: RuntimeEvent[] = []
        for await (const event of runtime.run({
          requestId: crypto.randomUUID(),
          conversationId: crypto.randomUUID(),
          prompt: 'Copy the file in the requested directory.',
          workMode: 'execute'
        }, new AbortController().signal, async () => 'once')) {
          events.push(event)
        }
        expect(await readFile(join(target, 'output.txt'), 'utf8')).toBe('outside-workspace')
        expect(events).toContainEqual(expect.objectContaining({
          type: 'tool', callId: 'call-process', state: 'completed'
        }))
        expect(fetcher).toHaveBeenCalledTimes(2)
      } finally {
        await runtime.dispose()
        await rm(root, { recursive: true, force: true })
      }
    }
  )

  it('lets an Execute Subagent borrow the parent browser tab without changing tab ownership', async () => {
    const conversationId = 'parent-browser-owner'
    const browserTabId = browserTabIdSchema.parse('00000000-0000-4000-8000-000000000203')
    let parentReleased = false
    const browserService: BrowserToolService = {
      getOrigin: vi.fn(() => 'https://example.test'),
      navigate: vi.fn(), click: vi.fn(), type: vi.fn(),
      select: vi.fn(), back: vi.fn(), screenshot: vi.fn(),
      snapshot: vi.fn(async (owner, _signal, tabId) => {
        // BrowserService resolves tabs under their owning conversation, not
        // merely by tab ID; a child's temporary conversation cannot own it.
        expect(owner).toBe(conversationId)
        expect(tabId).toBe(browserTabId)
        expect(parentReleased).toBe(false)
        return { url: 'https://example.test/', title: 'Parent page', nodes: [], truncated: false }
      }),
      releaseConversation: vi.fn(async (owner) => {
        if (owner === conversationId) parentReleased = true
      })
    }
    const toolResponse = (id: string, name: string, args: Record<string, unknown>) => ({
      choices: [{ message: {
        role: 'assistant', content: null,
        tool_calls: [{ id, type: 'function', function: {
          name, arguments: JSON.stringify(args)
        } }]
      } }]
    })
    const responses = [
      toolResponse('delegate', 'subagent_delegate', { task: '读取父请求浏览器页面' }),
      toolResponse('child-browser', 'browser_snapshot', {}),
      { choices: [{ message: { role: 'assistant', content: '子级已读页面' } }] },
      toolResponse('parent-browser', 'browser_snapshot', {}),
      { choices: [{ message: { role: 'assistant', content: '父级继续使用同一页面' } }] }
    ]
    const fetcher = vi.fn<typeof fetch>(async () => Response.json(responses.shift()))
    const scheduler = new SubagentScheduler({ concurrency: 1, queueLimit: 1, timeoutMs: 10_000 })
    const runtime = new ModelAgentRuntime({
      baseUrl: 'https://example.test/v1', model: 'test-model',
      protocol: 'openai-chat-completions', authentication: 'none',
      fetcher, browserService, directModelSubagentScheduler: scheduler
    })
    const listTools = vi.spyOn(ModelToolProvider.prototype, 'listTools')
    const callTool = vi.spyOn(ModelToolProvider.prototype, 'callTool')
    try {
      const events: RuntimeEvent[] = []
      for await (const event of runtime.run({
        requestId: crypto.randomUUID(), conversationId, browserTabId,
        prompt: '委派读取页面后继续使用页面', workMode: 'execute'
      }, new AbortController().signal, async () => 'once')) events.push(event)
      expect(fetcher).toHaveBeenCalledTimes(5)
      const childContext = listTools.mock.calls.find(
        ([context]) => context.delegationDepth === 1
      )![0]
      expect(childContext).toMatchObject({
        workMode: 'execute', browserTabId,
        browserConversationId: conversationId
      })
      expect(childContext.conversationId).not.toBe(conversationId)
      const childBody = JSON.parse(String(fetcher.mock.calls[1]![1]!.body))
      const childTools = childBody.tools.map(
        (tool: { function: { name: string } }) => tool.function.name
      )
      expect(childTools).toEqual(expect.arrayContaining(['browser_navigate', 'browser_snapshot', 'browser_click']))
      expect(childTools).not.toContain('subagent_delegate')
      expect(callTool).toHaveBeenCalledWith(
        'browser_snapshot', {}, expect.any(AbortSignal),
        expect.objectContaining({
          conversationId: childContext.conversationId,
          browserConversationId: conversationId, browserTabId
        })
      )
      expect(browserService.snapshot).toHaveBeenCalledTimes(2)
      expect(browserService.releaseConversation).toHaveBeenCalledWith(childContext.conversationId)
      expect(parentReleased).toBe(false)
      expect(events.filter((event) => event.type === 'subagent').at(-1))
        .toMatchObject({ state: 'completed' })
      await runtime.releaseConversation(conversationId)
      expect(parentReleased).toBe(true)
    } finally {
      listTools.mockRestore()
      callTool.mockRestore()
      await runtime.dispose()
      scheduler.dispose()
      await scheduler.waitForIdle()
    }
  })

  it('delegates one programming Subagent that can run a real project command without recursion', async () => {
    const command =
      process.platform === 'win32'
        ? "[Console]::Write('child-ok')"
        : 'printf child-ok'
    const responses = [
      {
        model: 'qwen3',
        choices: [{
          message: {
            role: 'assistant',
            content: null,
            tool_calls: [{
              id: 'call-delegate',
              type: 'function',
              function: {
                name: 'subagent_delegate',
                arguments: JSON.stringify({
                  task: '运行一个命令确认子代理执行能力'
                })
              }
            }]
          }
        }],
        usage: { prompt_tokens: 10, completion_tokens: 4 }
      },
      {
        model: 'qwen3',
        choices: [{
          message: {
            role: 'assistant',
            content: null,
            tool_calls: [{
              id: 'call-child-process',
              type: 'function',
              function: {
                name: 'process_execute',
                arguments: JSON.stringify({ command })
              }
            }]
          }
        }],
        usage: { prompt_tokens: 12, completion_tokens: 5 }
      },
      {
        model: 'qwen3',
        choices: [{
          message: {
            role: 'assistant',
            content: '子代理已验证 child-ok。'
          }
        }],
        usage: { prompt_tokens: 20, completion_tokens: 6 }
      },
      {
        model: 'qwen3',
        choices: [{
          message: {
            role: 'assistant',
            content: '父模型已综合子代理结果。'
          }
        }],
        usage: { prompt_tokens: 25, completion_tokens: 7 }
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const scheduler = new SubagentScheduler({
      concurrency: 3,
      queueLimit: 20,
      timeoutMs: 10_000
    })
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      defaultWorkspace: process.cwd(),
      directModelSubagentScheduler: scheduler
    })
    const events: RuntimeEvent[] = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed177',
        conversationId: 'conversation-direct-subagent',
        prompt: '委派子代理验证进程能力',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledTimes(4)
    const parentTools = (
      JSON.parse(
        fetcher.mock.calls[0]?.[1]?.body as string
      ) as {
        tools: Array<{ function: { name: string } }>
      }
    ).tools.map((tool) => tool.function.name)
    const childTools = (
      JSON.parse(
        fetcher.mock.calls[1]?.[1]?.body as string
      ) as {
        tools: Array<{ function: { name: string } }>
      }
    ).tools.map((tool) => tool.function.name)
    expect(parentTools).toContain('subagent_delegate')
    expect(parentTools).toContain('process_execute')
    expect(childTools).toContain('process_execute')
    expect(childTools).not.toContain('subagent_delegate')
    expect(
      events
        .filter((event) => event.type === 'subagent')
        .map((event) => event.state)
    ).toEqual(['queued', 'running', 'completed'])
    expect(
      events
        .flatMap(
          (event) =>
            event.type === 'tool' &&
            event.callId.includes('call-child-process')
              ? [event.state]
              : []
        )
    ).toEqual(['pending', 'running', 'completed'])
    expect(
      events.some(
        (event) =>
          event.type === 'text' &&
          event.delta.includes('父模型已综合')
      )
    ).toBe(true)

    await runtime.dispose()
    scheduler.dispose()
    await scheduler.waitForIdle()
  })

  it('compresses complete earlier tool rounds before a later Agent model call', async () => {
    const firstToolResult = `first-result-${'旧'.repeat(5_000)}`
    const secondToolResult = `second-result-${'新'.repeat(5_000)}`
    const thirdToolResult = `third-result-${'终'.repeat(5_000)}`
    const fetcher = vi
      .fn<typeof fetch>()
      .mockResolvedValueOnce(
        Response.json({
          id: 'message-agent-round-1',
          model: 'claude',
          content: [
            {
              type: 'tool_use',
              id: 'toolu-agent-1',
              name: 'workspace_read_text',
              input: { path: 'first.txt' }
            }
          ],
          stop_reason: 'tool_use',
          usage: { input_tokens: 4_000, output_tokens: 20 }
        })
      )
      .mockResolvedValueOnce(
        Response.json({
          id: 'message-agent-round-2',
          model: 'claude',
          content: [
            {
              type: 'tool_use',
              id: 'toolu-agent-2',
              name: 'workspace_read_text',
              input: { path: 'second.txt' }
            }
          ],
          stop_reason: 'tool_use',
          usage: { input_tokens: 12_500, output_tokens: 20 }
        })
      )
      .mockResolvedValueOnce(
        new Response(createEventStream('已完成前两轮读取'), {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
      .mockResolvedValueOnce(
        Response.json({
          id: 'message-agent-round-3',
          model: 'claude',
          content: [
            {
              type: 'tool_use',
              id: 'toolu-agent-3',
              name: 'workspace_read_text',
              input: { path: 'third.txt' }
            }
          ],
          stop_reason: 'tool_use',
          usage: { input_tokens: 12_500, output_tokens: 20 }
        })
      )
      .mockResolvedValueOnce(
        new Response(createEventStream('已完成全部三轮读取'), {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
      .mockResolvedValueOnce(
        Response.json({
          id: 'message-agent-final',
          model: 'claude',
          content: [{ type: 'text', text: '长任务已经完成。' }],
          stop_reason: 'end_turn',
          usage: { input_tokens: 8_000, output_tokens: 30 }
        })
      )
    const callTool = vi
      .fn()
      .mockResolvedValueOnce(createTextToolResult(firstToolResult))
      .mockResolvedValueOnce(createTextToolResult(secondToolResult))
      .mockResolvedValueOnce(createTextToolResult(thirdToolResult))
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'claude',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      toolProvider: createToolProvider({ callTool }),
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 12_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve execution progress.'
        }
      }
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed131',
        conversationId: 'conversation-long-agent',
        prompt: '连续读取三个文件',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledTimes(6)
    const summaryBody = JSON.parse(
      fetcher.mock.calls[2]?.[1]?.body as string
    ) as { system: string; messages: unknown[] }
    expect(summaryBody.system).toContain(
      'Summarize the earlier execution rounds'
    )
    expect(JSON.stringify(summaryBody.messages)).toContain('first.txt')
    expect(JSON.stringify(summaryBody.messages)).toContain('second.txt')
    expect(JSON.stringify(summaryBody.messages)).not.toContain(
      'third-result-'
    )

    const secondSummaryBody = JSON.parse(
      fetcher.mock.calls[4]?.[1]?.body as string
    ) as { messages: unknown[] }
    const secondSummaryMessages = JSON.stringify(
      secondSummaryBody.messages
    )
    expect(secondSummaryMessages).toContain('已完成前两轮读取')
    expect(secondSummaryMessages).toContain('third.txt')

    const finalBody = JSON.parse(
      fetcher.mock.calls[5]?.[1]?.body as string
    ) as { messages: unknown[] }
    const finalMessages = JSON.stringify(finalBody.messages)
    expect(finalMessages).toContain('已完成全部三轮读取')
    expect(finalMessages).not.toContain('first-result-')
    expect(finalMessages).not.toContain('second-result-')
    expect(finalMessages).not.toContain('third-result-')
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'context-compression',
        scope: 'agent-run',
        state: 'started',
        compressionCount: 1
      })
    )
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'context-compression',
        scope: 'agent-run',
        state: 'completed',
        compressionCount: 1,
        estimatedAfterTokens: expect.any(Number)
      })
    )
    expect(
      events
        .flatMap((event) =>
          event.type === 'context-compression' &&
          event.scope === 'agent-run' &&
          event.state === 'completed'
            ? [event.compressionCount]
            : []
        )
    ).toEqual([1, 2])
    expect(
      events
        .flatMap((event) =>
          event.type === 'context-compression' &&
          event.scope === 'agent-run' &&
          event.state === 'completed'
            ? [event]
            : []
        )
        .every(
          (event) =>
            event.estimatedAfterTokens !== undefined &&
            event.estimatedAfterTokens < event.effectiveTriggerTokens
        )
    ).toBe(true)
    const contextMetrics = events.filter(
      (event) => event.type === 'context-metrics'
    )
    expect(contextMetrics).toHaveLength(4)
    expect(
      contextMetrics.map((event) => event.contextTokens)
    ).toEqual([4_020, 12_520, 12_520, 8_030])
    expect(
      contextMetrics.every((event) => event.source === 'provider')
    ).toBe(true)
    const agentSummaryUsageEvents = events.flatMap((event) =>
      event.type === 'model-usage' &&
      event.callId.startsWith('agent-context-summary:')
        ? [event]
        : []
    )
    expect(agentSummaryUsageEvents).toHaveLength(2)
    expect(
      agentSummaryUsageEvents.every((event) =>
        /^agent-context-summary:model-call:[0-9a-f]{32}:message-1$/u.test(
          event.callId
        )
      )
    ).toBe(true)
    expect(
      new Set(
        agentSummaryUsageEvents.map((event) => event.callId)
      ).size
    ).toBe(2)
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('compresses one oversized completed Agent round before the next call', async () => {
    const toolResult = `single-result-${'巨'.repeat(10_000)}`
    const fetcher = vi
      .fn<typeof fetch>()
      .mockResolvedValueOnce(
        Response.json({
          id: 'message-agent-single-round',
          model: 'claude',
          content: [
            {
              type: 'tool_use',
              id: 'toolu-agent-single',
              name: 'workspace_read_text',
              input: { path: 'single.txt' }
            }
          ],
          stop_reason: 'tool_use',
          usage: { input_tokens: 12_500, output_tokens: 20 }
        })
      )
      .mockResolvedValueOnce(
        new Response(createEventStream('已摘要单轮工具结果'), {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        })
      )
      .mockResolvedValueOnce(
        Response.json({
          id: 'message-agent-single-final',
          model: 'claude',
          content: [{ type: 'text', text: '单轮长任务已完成。' }],
          stop_reason: 'end_turn',
          usage: { input_tokens: 8_000, output_tokens: 30 }
        })
      )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'claude',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      toolProvider: createToolProvider({
        callTool: vi.fn(async () => createTextToolResult(toolResult))
      }),
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 12_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve execution progress.'
        }
      }
    })
    const events: RuntimeEvent[] = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed133',
        conversationId: 'conversation-single-agent-round',
        prompt: '读取单个大文件',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledTimes(3)
    const finalBody = JSON.parse(
      fetcher.mock.calls[2]?.[1]?.body as string
    ) as { messages: unknown[] }
    const finalMessages = JSON.stringify(finalBody.messages)
    expect(finalMessages).toContain('已摘要单轮工具结果')
    expect(finalMessages).not.toContain('single-result-')
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'context-compression',
        scope: 'agent-run',
        state: 'completed',
        compressionCount: 1
      })
    )
  })

  it('marks Agent context compression as failed when summarization fails', async () => {
    const fetcher = vi
      .fn<typeof fetch>()
      .mockResolvedValueOnce(
        Response.json({
          id: 'message-agent-failure-1',
          model: 'claude',
          content: [
            {
              type: 'tool_use',
              id: 'toolu-agent-failure-1',
              name: 'workspace_read_text',
              input: { path: 'first.txt' }
            }
          ],
          stop_reason: 'tool_use',
          usage: { input_tokens: 4_000, output_tokens: 20 }
        })
      )
      .mockResolvedValueOnce(
        Response.json({
          id: 'message-agent-failure-2',
          model: 'claude',
          content: [
            {
              type: 'tool_use',
              id: 'toolu-agent-failure-2',
              name: 'workspace_read_text',
              input: { path: 'second.txt' }
            }
          ],
          stop_reason: 'tool_use',
          usage: { input_tokens: 12_500, output_tokens: 20 }
        })
      )
      .mockResolvedValueOnce(
        Response.json(
          { error: { message: 'summary unavailable' } },
          { status: 503 }
        )
      )
    const callTool = vi
      .fn()
      .mockResolvedValueOnce(
        createTextToolResult(`first-result-${'旧'.repeat(5_000)}`)
      )
      .mockResolvedValueOnce(
        createTextToolResult(`second-result-${'新'.repeat(5_000)}`)
      )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'claude',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      toolProvider: createToolProvider({ callTool }),
      contextCompression: {
        settings: {
          enabled: true,
          triggerTokens: 12_000,
          recentRawTokens: 4_000,
          modelSource: { kind: 'current' },
          summaryPrompt: 'Preserve execution progress.'
        }
      }
    })
    const events: RuntimeEvent[] = []
    const consume = async (): Promise<void> => {
      for await (const event of runtime.run(
        {
          requestId: 'a431666e-5ec8-45e6-beb4-654132eed132',
          conversationId: 'conversation-agent-summary-failure',
          prompt: '连续读取两个文件',
          workMode: 'execute'
        },
        new AbortController().signal,
        async () => 'once'
      )) {
        events.push(event)
      }
    }

    await expect(consume()).rejects.toThrow('summary unavailable')
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'context-compression',
        scope: 'agent-run',
        state: 'started',
        compressionCount: 1
      })
    )
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'context-compression',
        scope: 'agent-run',
        state: 'failed',
        compressionCount: 1
      })
    )
  })

  it('streams reasoning while using OpenAI-compatible tools', async () => {
    const streams = [
      [
        `data: ${JSON.stringify({
          choices: [
            {
              delta: {
                reasoning_content: '先读取文件',
                tool_calls: [
                  {
                    index: 0,
                    id: 'call-streamed',
                    type: 'function',
                    function: {
                      name: 'workspace_read_text',
                      arguments: '{"path":'
                    }
                  }
                ]
              }
            }
          ]
        })}`,
        '',
        `data: ${JSON.stringify({
          choices: [
            {
              delta: {
                tool_calls: [
                  {
                    index: 0,
                    id: '',
                    function: {
                      name: '',
                      arguments: '"README.md"}'
                    }
                  }
                ]
              }
            }
          ]
        })}`,
        '',
        'data: [DONE]',
        '',
        ''
      ].join('\n'),
      [
        `data: ${JSON.stringify({
          choices: [
            {
              delta: {
                reasoning_content: '再整理结果'
              }
            }
          ]
        })}`,
        '',
        `data: ${JSON.stringify({
          choices: [
            {
              delta: {
                content: '文件内容已读取。'
              }
            }
          ]
        })}`,
        '',
        'data: [DONE]',
        '',
        ''
      ].join('\n')
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(streams.shift(), {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.deepseek.com',
      model: 'deepseek-v4-flash',
      protocol: 'openai-chat-completions',
      authentication: 'api-key',
      fetcher,
      toolProvider
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed140',
        conversationId: 'conversation-streamed-tools',
        prompt: '读取 README',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    expect(
      events
        .filter(
          (event) =>
            event.type === 'reasoning' ||
            event.type === 'tool' ||
            event.type === 'text'
        )
        .map((event) =>
          event.type === 'tool'
            ? `${event.type}:${event.state}`
            : `${event.type}:${event.delta}`
        )
    ).toEqual([
      'reasoning:先读取文件',
      'tool:pending',
      'tool:running',
      'tool:completed',
      'reasoning:再整理结果',
      'text:文件内容已读取。'
    ])
    expect(toolProvider.callTool).toHaveBeenCalledWith(
      'workspace_read_text',
      { path: 'README.md' },
      expect.any(AbortSignal),
      expect.objectContaining({
        conversationId: 'conversation-streamed-tools',
        workMode: 'execute'
      })
    )
    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as { messages: Array<Record<string, unknown>> }
    expect(secondBody.messages).toContainEqual(
      expect.objectContaining({
        role: 'assistant',
        content: null,
        reasoning_content: '先读取文件',
        tool_calls: [
          expect.objectContaining({
            id: 'call-streamed',
            function: {
              name: 'workspace_read_text',
              arguments: '{"path":"README.md"}'
            }
          })
        ]
      })
    )
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('synthesizes and pairs a missing OpenAI Chat tool call id', async () => {
    const responses = [
      {
        id: 'chatcmpl-missing-call-id-1',
        model: 'qwen3',
        choices: [
          {
            message: {
              role: 'assistant',
              content: null,
              tool_calls: [
                {
                  type: 'function',
                  function: {
                    name: 'workspace_read_text',
                    arguments: '{"path":"README.md"}'
                  }
                }
              ]
            }
          }
        ]
      },
      {
        id: 'chatcmpl-missing-call-id-2',
        model: 'qwen3',
        choices: [
          {
            message: {
              role: 'assistant',
              content: '读取完成。'
            }
          }
        ]
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider: createToolProvider()
    })

    for await (const _event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed143',
        conversationId: 'conversation-chat-fallback-id',
        prompt: '读取 README',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      void _event
    }

    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as { messages: Array<Record<string, unknown>> }
    const assistant = secondBody.messages.at(-2) as {
      tool_calls: Array<Record<string, unknown>>
    }
    const result = secondBody.messages.at(-1) as {
      tool_call_id: string
    }
    const toolCallId = assistant.tool_calls[0]?.id
    expect(toolCallId).toEqual(
      expect.stringMatching(/^goodbuddy_call_[0-9a-f]{32}$/u)
    )
    expect(result).toMatchObject({
      role: 'tool',
      tool_call_id: toolCallId
    })
  })

  it('uses refreshed tool definitions in subsequent model rounds', async () => {
    const loadTool: ModelToolDefinition = {
      name: 'mcp_load_tools',
      displayName: 'CRM / load tools',
      description: 'Load CRM tools',
      inputSchema: { type: 'object' },
      source: 'mcp',
      serverName: 'CRM'
    }
    const dynamicTool: ModelToolDefinition = {
      name: 'mcp_list_opportunities',
      displayName: 'CRM / list opportunities',
      description: 'List opportunities',
      inputSchema: { type: 'object' },
      source: 'mcp',
      serverName: 'CRM'
    }
    const listTools = vi
      .fn<ModelToolProviderLike['listTools']>()
      .mockResolvedValueOnce([loadTool])
      .mockResolvedValueOnce([loadTool, dynamicTool])
      .mockResolvedValueOnce([loadTool, dynamicTool])
    const toolProvider = createToolProvider({ listTools })
    const responses = [
      {
        choices: [{
          message: {
            role: 'assistant',
            content: null,
            tool_calls: [{
              id: 'call-load',
              type: 'function',
              function: {
                name: loadTool.name,
                arguments: '{}'
              }
            }]
          }
        }]
      },
      {
        choices: [{
          message: {
            role: 'assistant',
            content: null,
            tool_calls: [{
              id: 'call-list',
              type: 'function',
              function: {
                name: dynamicTool.name,
                arguments: '{}'
              }
            }]
          }
        }]
      },
      {
        choices: [{
          message: {
            role: 'assistant',
            content: '已读取商机。'
          }
        }]
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })

    for await (const _event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed139',
        conversationId: 'conversation-dynamic-tools',
        prompt: '列出商机',
        workMode: 'execute'
      },
      new AbortController().signal,
      vi.fn(async () => 'once' as const)
    )) {
      void _event
    }

    expect(listTools).toHaveBeenCalledTimes(3)
    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as {
      tools: Array<{ function: { name: string } }>
    }
    expect(secondBody.tools.map((tool) => tool.function.name)).toContain(
      dynamicTool.name
    )
    expect(toolProvider.callTool).toHaveBeenCalledTimes(2)
  })

  it('continues without tools when the refreshed inventory is empty', async () => {
    const progressTool: ModelToolDefinition = {
      name: 'record_progress',
      displayName: 'Record progress',
      description: 'Record the final required step',
      inputSchema: { type: 'object' },
      source: 'builtin'
    }
    const listTools = vi
      .fn<ModelToolProviderLike['listTools']>()
      .mockResolvedValueOnce([progressTool])
      .mockResolvedValueOnce([])
    const toolProvider = createToolProvider({ listTools })
    const responses = [
      {
        choices: [{
          message: {
            role: 'assistant',
            content: null,
            tool_calls: [{
              id: 'call-progress',
              type: 'function',
              function: {
                name: progressTool.name,
                arguments: '{}'
              }
            }]
          }
        }]
      },
      {
        choices: [{
          message: {
            role: 'assistant',
            content: '进度记录完成。'
          }
        }]
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })

    const events = []
    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed151',
        conversationId: 'conversation-empty-refreshed-tools',
        prompt: '记录进度后完成',
        workMode: 'execute'
      },
      new AbortController().signal,
      vi.fn(async () => 'once' as const)
    )) {
      events.push(event)
    }

    expect(listTools).toHaveBeenCalledTimes(2)
    expect(toolProvider.callTool).toHaveBeenCalledOnce()
    expect(
      JSON.parse(fetcher.mock.calls[1]?.[1]?.body as string)
    ).not.toHaveProperty('tools')
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'text',
        delta: '进度记录完成。'
      })
    )
  })

  it('runs only scoped knowledge in Ask without requesting approval', async () => {
    const responses = [
      {
        choices: [
          {
            message: {
              role: 'assistant',
              content: null,
              tool_calls: [
                {
                  id: 'knowledge-list-call',
                  type: 'function',
                  function: {
                    name: 'knowledge_list',
                    arguments: '{}'
                  }
                },
                {
                  id: 'knowledge-call',
                  type: 'function',
                  function: {
                    name: 'knowledge_search',
                    arguments: '{"query":"release notes","limit":3}'
                  }
                }
              ]
            }
          }
        ]
      },
      {
        choices: [
          {
            message: {
              role: 'assistant',
              content: '基于知识库证据回答。'
            }
          }
        ]
      }
    ]
    const knowledgeListTool: ModelToolDefinition = {
      name: 'knowledge_list',
      displayName: '知识库列表',
      description: 'Scoped library metadata',
      inputSchema: {
        type: 'object',
        properties: {},
        additionalProperties: false
      },
      source: 'builtin'
    }
    const knowledgeTool: ModelToolDefinition = {
      name: 'knowledge_search',
      displayName: '知识库搜索',
      description: 'Scoped evidence',
      inputSchema: {
        type: 'object',
        properties: { query: { type: 'string' } },
        required: ['query'],
        additionalProperties: false
      },
      source: 'builtin'
    }
    const toolProvider = createToolProvider({
      listTools: vi.fn(async () => [
        knowledgeListTool,
        knowledgeTool
      ])
    })
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })
    const authorize = vi.fn(async () => 'deny' as const)
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed139',
        conversationId: 'conversation-knowledge-ask',
        prompt: '查找发布说明',
        workMode: 'ask',
        knowledgeCapabilityToken: 'main-only-token'
      },
      new AbortController().signal,
      authorize
    )) {
      events.push(event)
    }

    expect(toolProvider.listTools).toHaveBeenCalledWith(
      expect.objectContaining({
        conversationId: 'conversation-knowledge-ask',
        workMode: 'ask',
        knowledgeCapabilityToken: 'main-only-token'
      }),
      expect.any(AbortSignal)
    )
    expect(toolProvider.callTool).toHaveBeenCalledWith(
      'knowledge_list',
      {},
      expect.any(AbortSignal),
      expect.objectContaining({
        workMode: 'ask',
        knowledgeCapabilityToken: 'main-only-token'
      })
    )
    expect(toolProvider.callTool).toHaveBeenCalledWith(
      'knowledge_search',
      { query: 'release notes', limit: 3 },
      expect.any(AbortSignal),
      expect.objectContaining({
        workMode: 'ask',
        knowledgeCapabilityToken: 'main-only-token'
      })
    )
    expect(authorize).not.toHaveBeenCalled()
    expect(toolProvider.getApproval).not.toHaveBeenCalled()
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('runs enabled web search in Ask without per-call approval', async () => {
    const responses = [
      {
        choices: [
          {
            message: {
              role: 'assistant',
              content: null,
              tool_calls: [
                {
                  id: 'web-search-call',
                  type: 'function',
                  function: {
                    name: 'web_search',
                    arguments: '{"query":"current release","numResults":2}'
                  }
                }
              ]
            }
          }
        ]
      },
      {
        choices: [
          {
            message: {
              role: 'assistant',
              content: '基于联网搜索结果回答。'
            }
          }
        ]
      }
    ]
    const webSearchTool: ModelToolDefinition = {
      name: 'web_search',
      displayName: '联网搜索',
      description: 'Search public web',
      inputSchema: {
        type: 'object',
        properties: { query: { type: 'string' } },
        required: ['query'],
        additionalProperties: false
      },
      source: 'builtin'
    }
    const toolProvider = createToolProvider({
      listTools: vi.fn(async () => [webSearchTool])
    })
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher: vi.fn<typeof fetch>(async () =>
        Response.json(responses.shift())
      ),
      toolProvider,
      webSearchEnabled: true
    })
    const authorize = vi.fn(async () => 'deny' as const)

    const events = []
    for await (const event of runtime.run(
      {
        requestId: 'f0370284-5933-4743-892c-98263b8a44ae',
        conversationId: 'conversation-web-search-ask',
        prompt: '查找当前版本',
        workMode: 'ask'
      },
      new AbortController().signal,
      authorize
    )) {
      events.push(event)
    }

    expect(toolProvider.callTool).toHaveBeenCalledWith(
      'web_search',
      { query: 'current release', numResults: 2 },
      expect.any(AbortSignal),
      expect.objectContaining({ workMode: 'ask' })
    )
    expect(authorize).not.toHaveBeenCalled()
    expect(toolProvider.getApproval).not.toHaveBeenCalled()
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('returns recoverable tool failures to the model instead of aborting the run', async () => {
    const responses = [
      {
        choices: [
          {
            message: {
              role: 'assistant',
              content: null,
              tool_calls: [
                {
                  id: 'call-stale-ref',
                  type: 'function',
                  function: {
                    name: 'workspace_read_text',
                    arguments: '{"path":"README.md"}'
                  }
                }
              ]
            }
          }
        ]
      },
      {
        choices: [
          {
            message: {
              role: 'assistant',
              content: '已获取新快照并继续。'
            }
          }
        ]
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const toolProvider = createToolProvider({
      callTool: vi.fn(async () => {
        throw new RecoverableModelToolError(
          '浏览器元素引用已失效，请重新获取快照',
          '调用 browser_snapshot 后重试'
        )
      })
    })
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed130',
        conversationId: 'conversation-recoverable-tool-error',
        prompt: '继续浏览器操作',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledTimes(2)
    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as { messages: Array<Record<string, unknown>> }
    const toolMessage = secondBody.messages.find(
      (message) => message.role === 'tool'
    )
    expect(JSON.parse(toolMessage?.content as string)).toEqual({
      ok: false,
      recoverable: true,
      error: '浏览器元素引用已失效，请重新获取快照',
      nextAction: '调用 browser_snapshot 后重试'
    })
    expect(
      events
        .filter((event) => event.type === 'tool')
        .map((event) => event.state)
    ).toEqual(['pending', 'running', 'recoverable'])
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'text',
        delta: '已获取新快照并继续。'
      })
    )
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('streams OpenAI Responses text and reasoning through tool rounds', async () => {
    const streams = [
      createSseEventStream([
        {
          type: 'response.reasoning_summary_text.delta',
          delta: '先分析。'
        },
        {
          type: 'response.output_text.delta',
          delta: '准备读取。'
        },
        {
          type: 'response.completed',
          response: {
            id: 'resp-stream-tool-1',
            model: 'gpt-5',
            status: 'completed',
            output: [
              {
                id: 'reasoning-stream-1',
                type: 'reasoning',
                summary: [
                  { type: 'summary_text', text: '先分析。' }
                ]
              },
              {
                id: 'message-stream-1',
                type: 'message',
                role: 'assistant',
                status: 'completed',
                content: [
                  { type: 'output_text', text: '准备读取。' }
                ]
              },
              {
                id: 'function-stream-1',
                type: 'function_call',
                call_id: 'call-stream-1',
                name: 'workspace_read_text',
                arguments: '{"path":"README.md"}'
              }
            ],
            usage: { input_tokens: 12, output_tokens: 4 }
          }
        }
      ]),
      createSseEventStream([
        {
          type: 'response.output_text.delta',
          delta: '读取完成。'
        },
        {
          type: 'response.completed',
          response: {
            id: 'resp-stream-tool-2',
            model: 'gpt-5',
            status: 'completed',
            output: [
              {
                id: 'message-stream-2',
                type: 'message',
                role: 'assistant',
                status: 'completed',
                content: [
                  { type: 'output_text', text: '读取完成。' }
                ]
              }
            ],
            usage: { input_tokens: 20, output_tokens: 5 }
          }
        }
      ])
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(streams.shift(), {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.openai.com/v1',
      model: 'gpt-5',
      protocol: 'openai-responses',
      authentication: 'api-key',
      fetcher,
      toolProvider
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed151',
        conversationId: 'conversation-responses-streaming-tools',
        prompt: '读取 README',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    for (const [, init] of fetcher.mock.calls) {
      expect(JSON.parse(init?.body as string)).toMatchObject({
        stream: true
      })
    }
    expect(
      events
        .filter((event) => event.type === 'reasoning')
        .map((event) => event.delta)
    ).toEqual(['先分析。'])
    expect(
      events
        .filter((event) => event.type === 'text')
        .map((event) => event.delta)
    ).toEqual(['准备读取。', '读取完成。'])
    expect(
      events.findIndex((event) => event.type === 'text')
    ).toBeLessThan(
      events.findIndex(
        (event) =>
          event.type === 'tool' && event.state === 'running'
      )
    )
    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as { input: Array<Record<string, unknown>> }
    expect(secondBody.input).toEqual([
      {
        role: 'user',
        content: '读取 README'
      },
      {
        id: 'reasoning-stream-1',
        type: 'reasoning',
        summary: [{ type: 'summary_text', text: '先分析。' }]
      },
      {
        id: 'message-stream-1',
        type: 'message',
        role: 'assistant',
        status: 'completed',
        content: [
          { type: 'output_text', text: '准备读取。' }
        ]
      },
      {
        id: 'function-stream-1',
        type: 'function_call',
        call_id: 'call-stream-1',
        name: 'workspace_read_text',
        arguments: '{"path":"README.md"}'
      },
      {
        type: 'function_call_output',
        call_id: 'call-stream-1',
        output: [
          {
            type: 'input_text',
            text: 'tool result'
          }
        ]
      }
    ])
    expect(toolProvider.callTool).toHaveBeenCalledOnce()
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('rejects an incomplete OpenAI Responses tool stream', async () => {
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(
        [
          'event: response.output_text.delta',
          `data: ${JSON.stringify({
            type: 'response.output_text.delta',
            delta: 'partial'
          })}`,
          '',
          'data: [DONE]',
          '',
          ''
        ].join('\n'),
        {
          status: 200,
          headers: { 'content-type': 'text/event-stream' }
        }
      )
    )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.openai.com/v1',
      model: 'gpt-5',
      protocol: 'openai-responses',
      authentication: 'api-key',
      fetcher,
      toolProvider
    })
    const consume = async (): Promise<void> => {
      for await (const _event of runtime.run(
        {
          requestId: 'a431666e-5ec8-45e6-beb4-654132eed153',
          conversationId: 'conversation-responses-incomplete-tools',
          prompt: '读取 README',
          workMode: 'execute'
        },
        new AbortController().signal,
        async () => 'once'
      )) {
        void _event
      }
    }

    await expect(consume()).rejects.toThrow('流式响应意外中断')
    expect(toolProvider.callTool).not.toHaveBeenCalled()
    expect(fetcher).toHaveBeenCalledOnce()
  })

  it('continues OpenAI Responses with function_call_output', async () => {
    const responses = [
      {
        id: 'resp-tool-1',
        model: 'gpt-5',
        output: [
          {
            id: 'msg-responses-1',
            type: 'message',
            role: 'assistant',
            status: 'completed',
            content: [
              {
                type: 'output_text',
                text: '先读取 README。'
              }
            ]
          },
          {
            id: 'fc-responses-1',
            type: 'function_call',
            call_id: 'call-responses-1',
            name: 'workspace_read_text',
            arguments: '{"path":"README.md"}'
          }
        ],
        usage: { input_tokens: 14, output_tokens: 3 }
      },
      {
        id: 'resp-tool-2',
        model: 'gpt-5',
        output: [
          {
            id: 'fc-responses-2',
            type: 'function_call',
            call_id: 'call-responses-2',
            name: 'workspace_read_text',
            arguments: '{"path":"DESIGN.md"}'
          }
        ],
        usage: { input_tokens: 21, output_tokens: 4 }
      },
      {
        id: 'resp-tool-3',
        model: 'gpt-5',
        output: [
          {
            type: 'message',
            role: 'assistant',
            content: [
              {
                type: 'output_text',
                text: 'Responses 工具调用完成。'
              }
            ]
          }
        ],
        usage: { input_tokens: 30, output_tokens: 6 }
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.openai.com/v1',
      model: 'gpt-5',
      protocol: 'openai-responses',
      authentication: 'api-key',
      fetcher,
      toolProvider: createToolProvider({
        callTool: vi.fn(async () => createMultimodalToolResult())
      })
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed134',
        conversationId: 'conversation-responses-tools',
        prompt: '读取 README',
        history: [
          { role: 'user', content: '此前问题' },
          { role: 'assistant', content: '此前回答' }
        ],
        historyMessageIds: [
          '00000000-0000-4000-8000-000000000341',
          '00000000-0000-4000-8000-000000000342'
        ],
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    const firstBody = JSON.parse(
      fetcher.mock.calls[0]?.[1]?.body as string
    ) as Record<string, unknown>
    expect(firstBody).toMatchObject({
      model: 'gpt-5',
      stream: true,
      input: [
        {
          role: 'user',
          content: '此前问题'
        },
        {
          role: 'assistant',
          content: '此前回答'
        },
        {
          role: 'user',
          content: '读取 README'
        }
      ],
      tools: [
        {
          type: 'function',
          name: 'workspace_read_text',
          strict: false
        }
      ]
    })
    expect(firstBody).not.toHaveProperty('previous_response_id')
    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as Record<string, unknown>
    expect(secondBody).toMatchObject({
      input: [
        {
          role: 'user',
          content: '此前问题'
        },
        {
          role: 'assistant',
          content: '此前回答'
        },
        {
          role: 'user',
          content: '读取 README'
        },
        {
          id: 'msg-responses-1',
          type: 'message',
          role: 'assistant',
          status: 'completed',
          content: [
            {
              type: 'output_text',
              text: '先读取 README。'
            }
          ]
        },
        {
          id: 'fc-responses-1',
          type: 'function_call',
          call_id: 'call-responses-1',
          name: 'workspace_read_text',
          arguments: '{"path":"README.md"}'
        },
        {
          type: 'function_call_output',
          call_id: 'call-responses-1',
          output: [
            {
              type: 'input_text',
              text: 'tool result'
            },
            {
              type: 'input_image',
              image_url: `data:image/png;base64,${toolPng}`
            }
          ]
        }
      ]
    })
    const thirdBody = JSON.parse(
      fetcher.mock.calls[2]?.[1]?.body as string
    ) as {
      input: Array<Record<string, unknown>>
    }
    expect(thirdBody.input).toEqual([
      ...(secondBody.input as Array<Record<string, unknown>>),
      {
        id: 'fc-responses-2',
        type: 'function_call',
        call_id: 'call-responses-2',
        name: 'workspace_read_text',
        arguments: '{"path":"DESIGN.md"}'
      },
      {
        type: 'function_call_output',
        call_id: 'call-responses-2',
        output: [
          {
            type: 'input_text',
            text: 'tool result'
          },
          {
            type: 'input_image',
            image_url: `data:image/png;base64,${toolPng}`
          }
        ]
      }
    ])
    for (const [, init] of fetcher.mock.calls) {
      const body = JSON.parse(init?.body as string)
      expect(body).not.toHaveProperty(
        'previous_response_id'
      )
      expect(JSON.stringify(body)).not.toContain(
        '00000000-0000-4000-8000-000000000341'
      )
      expect(JSON.stringify(body)).not.toContain(
        '00000000-0000-4000-8000-000000000342'
      )
    }
    expect(
      events
        .filter((event) => event.type === 'tool')
        .map((event) => event.state)
    ).toEqual([
      'pending',
      'running',
      'completed',
      'pending',
      'running',
      'completed'
    ])
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'text',
        delta: 'Responses 工具调用完成。'
      })
    )
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('pairs a missing Responses call_id with the function-call item id', async () => {
    const responses = [
      {
        id: 'resp-tool-fallback-1',
        model: 'gpt-5',
        output: [
          {
            id: 'fc-responses-fallback-1',
            type: 'function_call',
            name: 'workspace_read_text',
            arguments: '{"path":"README.md"}'
          }
        ]
      },
      {
        id: 'resp-tool-fallback-2',
        model: 'gpt-5',
        output: [
          {
            type: 'message',
            role: 'assistant',
            content: [
              {
                type: 'output_text',
                text: '读取完成。'
              }
            ]
          }
        ]
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.openai.com/v1',
      model: 'gpt-5',
      protocol: 'openai-responses',
      authentication: 'api-key',
      fetcher,
      toolProvider: createToolProvider()
    })

    for await (const _event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed141',
        conversationId: 'conversation-responses-fallback-id',
        prompt: '读取 README',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      void _event
    }

    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as { input: Array<Record<string, unknown>> }
    expect(secondBody.input).toContainEqual(
      expect.objectContaining({
        id: 'fc-responses-fallback-1',
        type: 'function_call',
        call_id: 'fc-responses-fallback-1'
      })
    )
    expect(secondBody.input).toContainEqual(
      expect.objectContaining({
        type: 'function_call_output',
        call_id: 'fc-responses-fallback-1'
      })
    )
  })

  it('fails closed when a direct-model tool is denied', async () => {
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json({
        choices: [
          {
            message: {
              role: 'assistant',
              content: null,
              tool_calls: [
                {
                  id: 'call-denied',
                  type: 'function',
                  function: {
                    name: 'workspace_read_text',
                    arguments: '{"path":"secret.txt"}'
                  }
                }
              ]
            }
          }
        ]
      })
    )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })
    const events: Array<{ type: string; state?: string }> = []
    const consume = async (): Promise<void> => {
      for await (const event of runtime.run(
        {
          requestId: 'a431666e-5ec8-45e6-beb4-654132eed131',
          conversationId: 'conversation-denied',
          prompt: '读取 secret',
          workMode: 'execute'
        },
        new AbortController().signal,
        async () => 'deny'
      )) {
        events.push(event)
      }
    }

    await expect(consume()).rejects.toThrow('用户拒绝')
    expect(
      events
        .filter((event) => event.type === 'tool')
        .map((event) => event.state)
    ).toEqual(['pending', 'failed'])
    expect(toolProvider.callTool).not.toHaveBeenCalled()
  })

  it('uses Anthropic tool_use and tool_result messages in Execute mode', async () => {
    const responses = [
      {
        id: 'message-tool-1',
        model: 'claude',
        content: [
          {
            type: 'tool_use',
            id: 'toolu-1',
            name: 'workspace_read_text',
            input: { path: 'notes.md' }
          }
        ],
        usage: { input_tokens: 12, output_tokens: 3 }
      },
      {
        id: 'message-tool-2',
        model: 'claude',
        content: [{ type: 'text', text: '读取完成。' }],
        usage: { input_tokens: 20, output_tokens: 5 }
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'claude',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      toolProvider: createToolProvider({
        callTool: vi.fn(async () => createMultimodalToolResult())
      })
    })

    for await (const _event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed132',
        conversationId: 'conversation-anthropic-tools',
        prompt: '读取 notes',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      void _event
    }

    const firstBody = JSON.parse(
      fetcher.mock.calls[0]?.[1]?.body as string
    ) as Record<string, unknown>
    expect(firstBody).toMatchObject({
      stream: true,
      tools: [
        {
          name: 'workspace_read_text',
          input_schema: expect.objectContaining({ type: 'object' })
        }
      ]
    })
    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as { messages: Array<Record<string, unknown>> }
    expect(secondBody.messages.at(-1)).toEqual({
      role: 'user',
      content: [
        {
          type: 'tool_result',
          tool_use_id: 'toolu-1',
          content: [
            {
              type: 'text',
              text: 'tool result'
            },
            {
              type: 'image',
              source: {
                type: 'base64',
                media_type: 'image/png',
                data: toolPng
              }
            }
          ]
        }
      ]
    })
  })

  it('streams Anthropic text and thinking through tool rounds', async () => {
    const streams = [
      createSseEventStream([
        {
          type: 'message_start',
          message: {
            id: 'message-stream-tool-1',
            model: 'claude',
            usage: { input_tokens: 10 }
          }
        },
        {
          type: 'content_block_start',
          index: 0,
          content_block: { type: 'thinking', thinking: '' }
        },
        {
          type: 'content_block_delta',
          index: 0,
          delta: { type: 'thinking_delta', thinking: '先分析。' }
        },
        {
          type: 'content_block_delta',
          index: 0,
          delta: { type: 'signature_delta', signature: 'signed' }
        },
        { type: 'content_block_stop', index: 0 },
        {
          type: 'content_block_start',
          index: 1,
          content_block: { type: 'text', text: '' }
        },
        {
          type: 'content_block_delta',
          index: 1,
          delta: { type: 'text_delta', text: '准备读取。' }
        },
        { type: 'content_block_stop', index: 1 },
        {
          type: 'content_block_start',
          index: 2,
          content_block: {
            type: 'tool_use',
            id: 'toolu-stream-1',
            name: 'workspace_read_text',
            input: {}
          }
        },
        {
          type: 'content_block_delta',
          index: 2,
          delta: {
            type: 'input_json_delta',
            partial_json: '{"path":'
          }
        },
        {
          type: 'content_block_delta',
          index: 2,
          delta: {
            type: 'input_json_delta',
            partial_json: '"notes.md"}'
          }
        },
        { type: 'content_block_stop', index: 2 },
        {
          type: 'message_delta',
          delta: { stop_reason: 'tool_use' },
          usage: { output_tokens: 4 }
        },
        { type: 'message_stop' }
      ]),
      createSseEventStream([
        {
          type: 'message_start',
          message: {
            id: 'message-stream-tool-2',
            model: 'claude',
            usage: { input_tokens: 18 }
          }
        },
        {
          type: 'content_block_start',
          index: 0,
          content_block: { type: 'text', text: '' }
        },
        {
          type: 'content_block_delta',
          index: 0,
          delta: { type: 'text_delta', text: '读取完成。' }
        },
        { type: 'content_block_stop', index: 0 },
        {
          type: 'message_delta',
          delta: { stop_reason: 'end_turn' },
          usage: { output_tokens: 5 }
        },
        { type: 'message_stop' }
      ])
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(streams.shift(), {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.anthropic.com',
      model: 'claude',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      toolProvider
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed152',
        conversationId: 'conversation-anthropic-streaming-tools',
        prompt: '读取 notes',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    for (const [, init] of fetcher.mock.calls) {
      expect(JSON.parse(init?.body as string)).toMatchObject({
        stream: true
      })
    }
    expect(
      events
        .filter((event) => event.type === 'reasoning')
        .map((event) => event.delta)
    ).toEqual(['先分析。'])
    expect(
      events
        .filter((event) => event.type === 'text')
        .map((event) => event.delta)
    ).toEqual(['准备读取。', '读取完成。'])
    expect(
      events.findIndex((event) => event.type === 'text')
    ).toBeLessThan(
      events.findIndex(
        (event) =>
          event.type === 'tool' && event.state === 'running'
      )
    )
    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as { messages: Array<Record<string, unknown>> }
    expect(secondBody.messages.at(-2)).toEqual({
      role: 'assistant',
      content: [
        {
          type: 'thinking',
          thinking: '先分析。',
          signature: 'signed'
        },
        {
          type: 'text',
          text: '准备读取。'
        },
        {
          type: 'tool_use',
          id: 'toolu-stream-1',
          name: 'workspace_read_text',
          input: { path: 'notes.md' }
        }
      ]
    })
    expect(secondBody.messages.at(-1)).toEqual({
      role: 'user',
      content: [
        {
          type: 'tool_result',
          tool_use_id: 'toolu-stream-1',
          content: [{ type: 'text', text: 'tool result' }]
        }
      ]
    })
    expect(toolProvider.callTool).toHaveBeenCalledOnce()
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('rejects malformed streamed Anthropic tool arguments', async () => {
    const stream = createSseEventStream([
      {
        type: 'message_start',
        message: {
          id: 'message-invalid-tool-1',
          model: 'claude',
          usage: { input_tokens: 8 }
        }
      },
      {
        type: 'content_block_start',
        index: 0,
        content_block: {
          type: 'tool_use',
          id: 'toolu-invalid-1',
          name: 'workspace_read_text',
          input: {}
        }
      },
      {
        type: 'content_block_delta',
        index: 0,
        delta: {
          type: 'input_json_delta',
          partial_json: '{"path":'
        }
      },
      { type: 'content_block_stop', index: 0 },
      { type: 'message_stop' }
    ])
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(stream, {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.anthropic.com',
      model: 'claude',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      toolProvider
    })
    const consume = async (): Promise<void> => {
      for await (const _event of runtime.run(
        {
          requestId: 'a431666e-5ec8-45e6-beb4-654132eed154',
          conversationId: 'conversation-anthropic-invalid-tools',
          prompt: '读取 notes',
          workMode: 'execute'
        },
        new AbortController().signal,
        async () => 'once'
      )) {
        void _event
      }
    }

    await expect(consume()).rejects.toThrow(
      '模型返回了无效的工具参数 JSON'
    )
    expect(toolProvider.callTool).not.toHaveBeenCalled()
    expect(fetcher).toHaveBeenCalledOnce()
  })

  it('rejects a truncated streamed Anthropic tool round', async () => {
    const stream = createSseEventStream([
      {
        type: 'message_start',
        message: {
          id: 'message-truncated-stream-1',
          model: 'claude',
          usage: { input_tokens: 8 }
        }
      },
      {
        type: 'content_block_start',
        index: 0,
        content_block: { type: 'text', text: '' }
      },
      {
        type: 'content_block_delta',
        index: 0,
        delta: { type: 'text_delta', text: 'partial' }
      },
      { type: 'content_block_stop', index: 0 },
      {
        type: 'message_delta',
        delta: { stop_reason: 'max_tokens' },
        usage: { output_tokens: 4 }
      },
      { type: 'message_stop' }
    ])
    const fetcher = vi.fn<typeof fetch>(async () =>
      new Response(stream, {
        status: 200,
        headers: { 'content-type': 'text/event-stream' }
      })
    )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.anthropic.com',
      model: 'claude',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      toolProvider
    })
    const consume = async (): Promise<void> => {
      for await (const _event of runtime.run(
        {
          requestId: 'a431666e-5ec8-45e6-beb4-654132eed155',
          conversationId: 'conversation-anthropic-truncated-stream',
          prompt: '读取 notes',
          workMode: 'execute'
        },
        new AbortController().signal,
        async () => 'once'
      )) {
        void _event
      }
    }

    await expect(consume()).rejects.toThrow(
      'Anthropic 返回未完成结果：max_tokens'
    )
    expect(toolProvider.callTool).not.toHaveBeenCalled()
  })

  it('rejects a truncated Anthropic JSON fallback', async () => {
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json({
        id: 'message-truncated-json-1',
        model: 'claude',
        content: [{ type: 'text', text: 'partial' }],
        stop_reason: 'model_context_window_exceeded',
        usage: { input_tokens: 8, output_tokens: 4 }
      })
    )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://api.anthropic.com',
      model: 'claude',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      toolProvider
    })
    const consume = async (): Promise<void> => {
      for await (const _event of runtime.run(
        {
          requestId: 'a431666e-5ec8-45e6-beb4-654132eed156',
          conversationId: 'conversation-anthropic-truncated-json',
          prompt: '读取 notes',
          workMode: 'execute'
        },
        new AbortController().signal,
        async () => 'once'
      )) {
        void _event
      }
    }

    await expect(consume()).rejects.toThrow(
      'Anthropic 返回未完成结果：model_context_window_exceeded'
    )
    expect(toolProvider.callTool).not.toHaveBeenCalled()
  })

  it('synthesizes and pairs a missing Anthropic tool_use id', async () => {
    const responses = [
      {
        id: 'message-tool-missing-id-1',
        model: 'claude',
        content: [
          {
            type: 'tool_use',
            name: 'workspace_read_text',
            input: { path: 'notes.md' }
          }
        ]
      },
      {
        id: 'message-tool-missing-id-2',
        model: 'claude',
        content: [{ type: 'text', text: '读取完成。' }]
      }
    ]
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json(responses.shift())
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai',
      model: 'claude',
      protocol: 'anthropic-messages',
      authentication: 'api-key',
      fetcher,
      toolProvider: createToolProvider()
    })

    for await (const _event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed142',
        conversationId: 'conversation-anthropic-fallback-id',
        prompt: '读取 notes',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      void _event
    }

    const secondBody = JSON.parse(
      fetcher.mock.calls[1]?.[1]?.body as string
    ) as { messages: Array<Record<string, unknown>> }
    const assistant = secondBody.messages.at(-2) as {
      content: Array<Record<string, unknown>>
    }
    const result = secondBody.messages.at(-1) as {
      content: Array<Record<string, unknown>>
    }
    const toolUseId = assistant.content[0]?.id
    expect(toolUseId).toEqual(
      expect.stringMatching(/^goodbuddy_call_[0-9a-f]{32}$/u)
    )
    expect(result.content[0]).toMatchObject({
      type: 'tool_result',
      tool_use_id: toolUseId
    })
  })

  it('does not issue a follow-up model request after tool cancellation', async () => {
    const response = {
      choices: [
        {
          message: {
            role: 'assistant',
            content: null,
            tool_calls: [
              {
                id: 'call-aborted',
                type: 'function',
                function: {
                  name: 'workspace_read_text',
                  arguments: '{}'
                }
              }
            ]
          }
        }
      ]
    }
    const fetcher = vi.fn<typeof fetch>(async () => Response.json(response))
    const controller = new AbortController()
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider: createToolProvider({
        callTool: vi.fn(async () => {
          controller.abort()
          return createTextToolResult('late result')
        })
      })
    })
    const consume = async (): Promise<void> => {
      for await (const _event of runtime.run(
        {
          requestId: crypto.randomUUID(),
          conversationId: crypto.randomUUID(),
          prompt: 'run',
          workMode: 'execute'
        },
        controller.signal,
        async () => 'once'
      )) {
        void _event
      }
    }

    await expect(consume()).rejects.toThrow()
    expect(fetcher).toHaveBeenCalledOnce()
  })

  it('allows repeated tool rounds until the run is cancelled', async () => {
    let callId = 0
    const controller = new AbortController()
    const fetcher = vi.fn<typeof fetch>(async () => {
      callId += 1
      return Response.json({
        choices: [
          {
            message: {
              role: 'assistant',
              content: null,
              tool_calls: [
                {
                  id: `call-repeat-${callId}`,
                  type: 'function',
                  function: {
                    name: 'workspace_read_text',
                    arguments: '{"path":"README.md"}'
                  }
                }
              ]
            }
          }
        ]
      })
    })
    let completedCalls = 0
    const toolProvider = createToolProvider({
      callTool: vi.fn(async () => {
        completedCalls += 1
        if (completedCalls === 5) {
          controller.abort()
        }
        return createTextToolResult('same result')
      })
    })
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })
    const consume = async (): Promise<void> => {
      for await (const _event of runtime.run(
        {
          requestId: crypto.randomUUID(),
          conversationId: 'conversation-repeat',
          prompt: 'repeat',
          workMode: 'execute'
        },
        controller.signal,
        async () => 'once'
      )) {
        void _event
      }
    }

    await expect(consume()).rejects.toThrow()
    expect(fetcher).toHaveBeenCalledTimes(5)
    expect(toolProvider.callTool).toHaveBeenCalledTimes(5)
  })

  it('allows more than the legacy 40 tool calls in one model round', async () => {
    const toolCalls = Array.from({ length: 45 }, (_, index) => ({
      id: `call-bulk-${index}`,
      type: 'function',
      function: {
        name: 'workspace_read_text',
        arguments: JSON.stringify({ path: `file-${index}.txt` })
      }
    }))
    const fetcher = vi
      .fn<typeof fetch>()
      .mockResolvedValueOnce(
        Response.json({
          choices: [
            {
              message: {
                role: 'assistant',
                content: null,
                tool_calls: toolCalls
              }
            }
          ]
        })
      )
      .mockResolvedValueOnce(
        Response.json({
          choices: [
            {
              message: {
                role: 'assistant',
                content: 'all tools completed'
              }
            }
          ]
        })
      )
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: 'conversation-bulk-tools',
        prompt: 'run all tools',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    expect(toolProvider.callTool).toHaveBeenCalledTimes(45)
    expect(fetcher).toHaveBeenCalledTimes(2)
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('continues beyond the legacy 24 model tool rounds', async () => {
    let modelRound = 0
    const fetcher = vi.fn<typeof fetch>(async () => {
      modelRound += 1
      if (modelRound <= 25) {
        return Response.json({
          choices: [
            {
              message: {
                role: 'assistant',
                content: null,
                tool_calls: [
                  {
                    id: `call-round-${modelRound}`,
                    type: 'function',
                    function: {
                      name: 'workspace_read_text',
                      arguments: JSON.stringify({
                        path: `round-${modelRound}.txt`
                      })
                    }
                  }
                ]
              }
            }
          ]
        })
      }
      return Response.json({
        choices: [
          {
            message: {
              role: 'assistant',
              content: 'completed after 25 rounds'
            }
          }
        ]
      })
    })
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: 'conversation-many-rounds',
        prompt: 'continue until done',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    expect(toolProvider.callTool).toHaveBeenCalledTimes(25)
    expect(fetcher).toHaveBeenCalledTimes(26)
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('releases provider state for only the requested conversation', async () => {
    const toolProvider = createToolProvider()
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      toolProvider
    })

    await runtime.releaseConversation('conversation-release')

    expect(toolProvider.releaseConversation).toHaveBeenCalledOnce()
    expect(toolProvider.releaseConversation).toHaveBeenCalledWith(
      'conversation-release'
    )
  })

  it('releases known conversations before provider disposal and permits replacement reuse', async () => {
    const lifecycle: string[] = []
    const released = new Set<string>()
    const createProvider = (): ModelToolProviderLike =>
      createToolProvider({
        releaseConversation: vi.fn(async (conversationId) => {
          lifecycle.push(`release:${conversationId}`)
          released.add(conversationId)
        }),
        dispose: vi.fn(async () => {
          lifecycle.push('dispose')
        })
      })
    const createRuntime = (toolProvider: ModelToolProviderLike) =>
      new ModelAgentRuntime({
        baseUrl: 'http://127.0.0.1:11434/v1',
        model: 'qwen3',
        protocol: 'openai-chat-completions',
        authentication: 'none',
        fetcher: vi.fn<typeof fetch>(async () =>
          new Response(
            'data: {"choices":[{"delta":{"content":"ok"}}]}\n\ndata: [DONE]\n\n',
            {
            status: 200,
            headers: { 'content-type': 'text/event-stream' }
            }
          )
        ),
        toolProvider
      })
    const request = {
      requestId: crypto.randomUUID(),
      conversationId: 'conversation-replacement',
      prompt: 'hello',
      workMode: 'ask' as const
    }
    const firstProvider = createProvider()
    const firstRuntime = createRuntime(firstProvider)
    for await (const _event of firstRuntime.run(
      request,
      new AbortController().signal
    )) {
      void _event
    }

    await firstRuntime.dispose()
    expect(lifecycle).toEqual([
      'release:conversation-replacement',
      'dispose'
    ])
    expect(released).toContain('conversation-replacement')

    const replacement = createRuntime(createProvider())
    const replacementEvents = []
    for await (const event of replacement.run(
      { ...request, requestId: crypto.randomUUID() },
      new AbortController().signal
    )) {
      replacementEvents.push(event)
    }
    expect(replacementEvents.at(-1)).toMatchObject({ type: 'done' })
    await replacement.dispose()
  })

  it('keeps an oversized image tool result in the model round', async () => {
    const imageData = Buffer.alloc(1024 * 1024 + 1).toString('base64')
    const fetcher = vi
      .fn<typeof fetch>()
      .mockResolvedValueOnce(
        Response.json({
          choices: [
            {
              message: {
                role: 'assistant',
                content: null,
                tool_calls: [
                  {
                    id: 'call-large-image',
                    type: 'function',
                    function: {
                      name: 'workspace_read_text',
                      arguments: '{}'
                    }
                  }
                ]
              }
            }
          ]
        })
      )
      .mockResolvedValueOnce(
        Response.json({
          choices: [
            {
              message: {
                role: 'assistant',
                content: '图片已读取。'
              }
            }
          ]
        })
      )
    const runtime = new ModelAgentRuntime({
      baseUrl: 'http://127.0.0.1:11434/v1',
      model: 'qwen3',
      protocol: 'openai-chat-completions',
      authentication: 'none',
      fetcher,
      toolProvider: createToolProvider({
        callTool: vi.fn(async () => ({
          parts: [
            {
              type: 'image' as const,
              mimeType: 'image/png' as const,
              data: imageData
            }
          ],
          contextBytes: Buffer.byteLength(imageData)
        }))
      })
    })
    const events = []
    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: 'run',
        workMode: 'execute'
      },
      new AbortController().signal,
      async () => 'once'
    )) {
      events.push(event)
    }

    expect(fetcher).toHaveBeenCalledTimes(2)
    expect(events.at(-1)).toMatchObject({ type: 'done' })
    expect(fetcher.mock.calls[1]?.[1]?.body as string).toContain(
      imageData
    )
  })

  it('performs and validates a real image generation when testing an image connection', async () => {
    const png = Buffer.from([
      0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a,
      0x00
    ]).toString('base64')
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json({ data: [{ b64_json: png }] })
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai/v1',
      model: 'gpt-image-2',
      protocol: 'openai-images-generations',
      authentication: 'api-key',
      imageGenerationQuality: 'medium',
      requestHeaders: { 'x-image-route': 'internal' },
      requestBody: {
        background: 'transparent',
        quality: 'high'
      },
      fetcher
    })

    await expect(runtime.testConnection()).resolves.toMatchObject({
      available: true,
      capability: 'image-generation',
      detail: expect.stringContaining(
        '已完成真实图像生成测试'
      )
    })
    expect(fetcher).toHaveBeenCalledOnce()
    expect(
      JSON.parse(fetcher.mock.calls[0]?.[1]?.body as string)
    ).toMatchObject({
      model: 'gpt-image-2',
      n: 1,
      quality: 'medium',
      background: 'transparent',
      response_format: 'b64_json'
    })
    expect(
      new Headers(fetcher.mock.calls[0]?.[1]?.headers).get(
        'x-image-route'
      )
    ).toBe('internal')
  })

  it('generates a bounded image through the BigToken-compatible endpoint', async () => {
    const png = Buffer.from([
      0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a,
      0x00
    ]).toString('base64')
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json({
        id: 'image-provider-1',
        model: 'gpt-image-provider',
        usage: {
          input_tokens: 17,
          output_tokens: 29,
          total_tokens: 46
        },
        data: [{ b64_json: png }]
      })
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai/v1',
      model: 'gpt-image-2',
      protocol: 'openai-images-generations',
      authentication: 'api-key',
      imageGenerationQuality: 'high',
      fetcher
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed128',
        conversationId: 'conversation-image',
        prompt: '一只在窗边睡觉的猫'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    const [input, init] = fetcher.mock.calls[0] ?? []
    expect(input?.toString()).toBe(
      'https://bigtoken.ai/v1/images/generations'
    )
    expect(init?.headers).toEqual({
      authorization: 'Bearer test-key',
      'content-type': 'application/json'
    })
    expect(JSON.parse(init?.body as string)).toEqual({
      model: 'gpt-image-2',
      prompt: '一只在窗边睡觉的猫',
      n: 1,
      quality: 'high',
      response_format: 'b64_json'
    })
    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'generated-image',
        mimeType: 'image/png',
        data: png
      })
    )
    expect(events.filter((event) => event.type === 'model-usage')).toEqual([
      {
        requestId: 'a431666e-5ec8-45e6-beb4-654132eed128',
        type: 'model-usage',
        callId: expect.stringMatching(
          /^model-call:[0-9a-f]{32}:image-provider-1$/u
        ),
        runtime: 'model',
        provider: 'openai',
        model: 'gpt-image-provider',
        inputTokens: 17,
        outputTokens: 29,
        cacheReadTokens: 0,
        cacheWriteTokens: 0,
        reportedTotalTokens: 46
      }
    ])
    expect(events.findIndex((event) => event.type === 'model-usage')).toBeLessThan(
      events.findIndex((event) => event.type === 'generated-image')
    )
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('edits reference images through the multipart image edits endpoint', async () => {
    const source = Buffer.from([
      0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a,
      0x00
    ])
    const fetcher = vi.fn<typeof fetch>(async () =>
      Response.json({ data: [{ b64_json: source.toString('base64') }] })
    )
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai/v1',
      model: 'gpt-image-2',
      protocol: 'openai-images-generations',
      authentication: 'api-key',
      imageGenerationQuality: 'medium',
      requestHeaders: { 'x-image-route': 'internal' },
      requestBody: { background: 'transparent' },
      fetcher
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: '把背景改成白色',
        images: [
          {
            name: 'source.png',
            mediaType: 'image/png',
            data: source.toString('base64')
          }
        ]
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    const [input, init] = fetcher.mock.calls[0] ?? []
    expect(input?.toString()).toBe(
      'https://bigtoken.ai/v1/images/edits'
    )
    expect(new Headers(init?.headers).get('authorization')).toBe(
      'Bearer test-key'
    )
    expect(new Headers(init?.headers).get('content-type')).toBeNull()
    expect(new Headers(init?.headers).get('x-image-route')).toBe(
      'internal'
    )
    const form = init?.body as FormData
    expect(form.get('model')).toBe('gpt-image-2')
    expect(form.get('prompt')).toBe('把背景改成白色')
    expect(form.get('n')).toBe('1')
    expect(form.get('quality')).toBe('medium')
    expect(form.get('response_format')).toBeNull()
    expect(form.get('background')).toBe('transparent')
    const uploaded = form.get('image')
    expect(uploaded).toBeInstanceOf(Blob)
    expect((uploaded as Blob).type).toBe('image/png')
    expect(Buffer.from(await (uploaded as Blob).arrayBuffer())).toEqual(
      source
    )
    expect(events).toContainEqual(
      expect.objectContaining({ type: 'generated-image' })
    )
    expect(events.at(-1)).toMatchObject({ type: 'done' })
  })

  it('rejects remote image URLs instead of fetching provider output', async () => {
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai/v1',
      model: 'gpt-image-2',
      protocol: 'openai-images-generations',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async () =>
        Response.json({
          data: [{ url: 'https://untrusted.example/image.png' }]
        })
      )
    })

    const consume = async (): Promise<void> => {
      for await (const _event of runtime.run(
        {
          requestId: 'a431666e-5ec8-45e6-beb4-654132eed129',
          conversationId: 'conversation-image-url',
          prompt: '测试图片'
        },
        new AbortController().signal
      )) {
        void _event
      }
    }

    await expect(consume()).rejects.toThrow('未返回 base64 图片')
  })

  it('accepts a bounded inline image data URL from compatible gateways', async () => {
    const png = Buffer.from([
      0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a,
      0x00
    ]).toString('base64')
    const runtime = new ModelAgentRuntime({
      apiKey: 'test-key',
      baseUrl: 'https://bigtoken.ai/v1',
      model: 'gpt-image-2',
      protocol: 'openai-images-generations',
      authentication: 'api-key',
      fetcher: vi.fn<typeof fetch>(async () =>
        Response.json({
          data: [{ url: `data:image/png;base64,${png}` }]
        })
      )
    })
    const events = []

    for await (const event of runtime.run(
      {
        requestId: crypto.randomUUID(),
        conversationId: crypto.randomUUID(),
        prompt: '测试内联图片'
      },
      new AbortController().signal
    )) {
      events.push(event)
    }

    expect(events).toContainEqual(
      expect.objectContaining({
        type: 'generated-image',
        mimeType: 'image/png',
        data: png
      })
    )
  })

  it.each([
    {
      body: JSON.stringify({
        error: {
          message:
            'upstream unavailable Authorization: Bearer secret-token'
        }
      }),
      headers: {
        'content-type': 'application/json',
        'x-request-id': 'image-request-502'
      },
      expected:
        'upstream unavailable Authorization: Bearer secret-token（HTTP 502，请求 ID image-request-502）'
    },
    {
      body: '<html>Bad Gateway</html>',
      headers: { 'content-type': 'text/html' },
      expected: '图像生成请求失败（HTTP 502）'
    },
    {
      body: JSON.stringify({
        error: '模型接口请求失败（HTTP 502）'
      }),
      headers: { 'content-type': 'application/json' },
      expected:
        '上游图像服务暂时不可用，请稍后重试或联系服务商（HTTP 502）'
    }
  ])(
    'retains HTTP status for image gateway failures',
    async ({ body, headers, expected }) => {
      const runtime = new ModelAgentRuntime({
        apiKey: 'test-key',
        baseUrl: 'https://bigtoken.ai/v1',
        model: 'gpt-image-2',
        protocol: 'openai-images-generations',
        authentication: 'api-key',
        fetcher: vi.fn<typeof fetch>(async () =>
          new Response(body, { status: 502, headers })
        )
      })
      const consume = async (): Promise<void> => {
        for await (const _event of runtime.run(
          {
            requestId: crypto.randomUUID(),
            conversationId: crypto.randomUUID(),
            prompt: '测试网关错误'
          },
          new AbortController().signal
        )) {
          void _event
        }
      }

      await expect(consume()).rejects.toThrow(expected)
    }
  )

  it.runIf(
    process.env.GOODBUDDY_BIGTOKEN_IMAGE_INTEGRATION === '1'
  )(
    'generates a real synthetic image with BigToken gpt-image-2',
    async () => {
      const apiKey = process.env.GOODBUDDY_BIGTOKEN_API_KEY
      if (!apiKey) {
        throw new Error('GOODBUDDY_BIGTOKEN_API_KEY is required')
      }
      const runtime = new ModelAgentRuntime({
        apiKey,
        baseUrl: 'https://bigtoken.ai/v1',
        model: 'gpt-image-2',
        protocol: 'openai-images-generations',
        authentication: 'api-key'
      })
      const events = []
      for await (const event of runtime.run(
        {
          requestId: crypto.randomUUID(),
          conversationId: crypto.randomUUID(),
          prompt:
            'A simple solid blue circle centered on a plain white background.'
        },
        new AbortController().signal
      )) {
        events.push(event)
      }
      expect(events).toContainEqual(
        expect.objectContaining({
          type: 'generated-image',
          mimeType: expect.stringMatching(/^image\//u)
        })
      )
      await runtime.dispose()
    },
    180_000
  )
})