Refactor path utilities and update model path usage across services

- Replaced direct usage of MODELS_PATH with getModelsPath() in ModelService and VectorModelDownloadService.
- Updated LanceDbCore and SqliteCore to use getLanceDbPath() and getSqliteDbPath() respectively.
- Refactored WorkspaceKnowledgeCache to utilize getUserDataPath().
- Introduced new utility functions in paths.ts for better path management.
- Updated various tests to reflect changes in path handling.
- Enhanced file change tracking in terminal tests to ensure isolation between runs.
- Added unit tests for file change run scope functionality.

doc/test-plan/ai-test.md

@@ -102,6 +102,7 @@ AI unit / 辅助测试目录：
 ### 3.4 覆盖归属与去重规则
 
 - `ai/tool` 负责生产工具 runtime 的完整输入输出、读写链、错误边界、原子性、修改记录与回退契约。
+- `ChatRunService` 合同测试验证 run 结束不会扫描并归属未受控的工作区变化；只有受控文件工具与 terminal 命令窗口能创建 run 级修改记录。
 - `ai/human/mock-llm` 负责真实 human graph 下的工具学习、工具选择、串联执行、暂停恢复和结果交付；只允许固定 LLM 输出，不允许替换工具执行结果。
 - `ai/human/live` 负责真实模型在自然语言输入下能否自主选择并正确使用能力；用户输入不得包含内部工具名或固定调用顺序。
 - Workbench E2E 负责用户可见的附件、预览、Diff、跨 Sheet 交互与回退闭环，不重复验证工具内部算法。
@@ -790,6 +791,15 @@ AI unit / 辅助测试目录：
 - Oracle：命令真实生效并经过审批；terminal 工具结果不包含 `artifacts`；命令完成后产生带当前 Run ID 的修改记录和 `file_artifact`。
 - Not：不根据命令文本推断文件路径；差异只来自命令前后工作区快照。
 
+##### `terminal-tracking-start-failure-cleanup`
+
+- 测试类型：`backend`
+- What：终端命令已建立追踪基线但 PTY 启动失败时，追踪上下文会被立即丢弃。
+- Why：启动失败不应留下无法完成的快照、修改记录或文件 artifact。
+- How：真实构建 terminal 工具，替换终端会话启动为明确失败的实现，并记录开始、完成和丢弃追踪调用。
+- Oracle：恰好开始并丢弃同一个命令追踪 ID；不调用完成追踪，不产生修改记录或 `file_artifact`。
+- Not：不验证 PTY 本身的启动错误归因；该能力由终端会话管理测试覆盖。
+
 ##### `terminal-inline-interpreter-execution`
 
 - 测试类型：`backend`
@@ -802,10 +812,10 @@ AI unit / 辅助测试目录：
 ##### `terminal-persistent-session-contract`
 
 - 测试类型：`backend`
-- What：真实 PTY 持久终端保留当前路径、长命令让出控制权、增量读取和停止。
+- What：真实 PTY 持久终端保留当前路径、长命令让出控制权、增量读取、停止和命令级文件追踪。
 - Why：后台服务和长任务不能继续依赖单次 shell 超时或系统孤儿进程。
-- How：在同一 terminalId 中真实执行 cd、pwd 和持续输出命令，同时新开第二终端；短命令若先返回 running，则通过 terminal_read 等到 idle 后再复用；长命令使用 terminal_read 读取输出并用 terminal_stop 停止。
-- Oracle：同一终端保留路径，新终端从 workspace 根目录启动；长命令返回 command_still_running 且进程仍在；读取能取得真实增量输出；正常停止后终端回到 idle，强制停止后终端进入 closed，且均无子进程残留。
+- How：在同一 terminalId 中真实执行 cd、pwd 和持续输出命令，同时新开第二终端；从子目录分别执行“长命令正常完成”“写入后停止”和“写入后非零退出”，通过 terminal_read 等到 idle 或调用 terminal_stop。
+- Oracle：同一终端保留路径，新终端从 workspace 根目录启动；长命令返回 command_still_running 且进程仍在；读取能取得真实增量输出；正常停止后终端回到 idle，强制停止后终端进入 closed，且均无子进程残留。三条终端写命令均以工作区根目录为扫描范围，生成当前 run 的 `terminal_tool` 修改记录和 `file_artifact`。
 - Not：不验证 Workbench 浏览器联动，该能力由真实 e2e 覆盖。
 
 #### web-browsing 类别

doc/test-plan/workbench-test.md

@@ -400,6 +400,15 @@ tests/case/workbench/
 - Oracle：AI 写入后普通预览显示最新内容且不误报外部修改；普通预览与审核模式共用一个文件 Tab；Run 终态后才展示修改入口；审核模式展示差异数量、修改状态和回退，预览模式全部隐藏；已应用记录打开外部应用后显示重新打开提醒；文件夹入口打开的文件列表从工具栏底部贴右悬浮，不改变预览宽度，可通过外部点击、Escape 或选择文件关闭；当前已经是最新记录且处于预览模式时，再次点击文件总分类仍会重置到最新审核模式；历史批次、结构化 Diff、不可变展开和回退刷新整体对齐 `WB-PANEL-003`、`WB-APPLY-001`、`WB-APPLY-002`、`WB-APPLY-003`；确认回退后 bytes 与原始输入一致且有效改动标记消失。
 - Not：不验证 unified diff 字符串；UI 只消费结构化 diff artifact。
 
+#### `file-change-run-scope`
+
+- 测试类型：`e2e / live`
+- What：两个会话各完成一次真实终端写入后，每条 assistant 修改入口与右侧“本次运行改动”只展示该 `{ sessionId, runId }` 的文件。
+- Why：文件改动归属已经从会话级改为 run 级，后台或历史会话完成时不能把其他 run 的记录混入当前审核列表。
+- How：在首个会话让真实 AI 只通过 `terminal_run` 创建第一个文件；新建并绑定同一工作区的第二个会话，再让真实 AI 创建另一个文件；分别打开两条修改入口，再回到首个会话复查右侧变更列表。
+- Oracle：两条 run 和两个会话 ID 均不同；每条 assistant 卡片只含自身文件；右侧“本次运行改动 1”只含当前 run 文件，切回历史会话后仍不出现另一 run 的文件，对齐 `WB-PANEL-003`、`WB-APPLY-003`。
+- Not：不要求两个终端命令实际时间重叠；重叠命令的允许重复归属由后端并发单测覆盖。
+
 #### `ppt-create-diff-and-revert`
 
 - 测试类型：`e2e / live`

src/main/index.ts

@@ -17,7 +17,7 @@ import { CdpEndpointService } from './services/browser/CdpEndpointService'
 import { createWebCrawlerService } from './services/webCrawler/createWebCrawlerService'
 import { AppUpdateService } from './services/update/AppUpdateService'
 import { APP_ID } from '../shared/constants'
-import { MODELS_PATH } from './utils/paths'
+import { getModelsPath } from './utils/paths'
 import { terminalSessionManager } from './services/ai/human/tools/terminal/TerminalSessionManager'
 import { AppShutdownCoordinator } from './services/AppShutdownCoordinator'
 import { SystemNotificationService } from './services/SystemNotificationService'
@@ -201,7 +201,7 @@ if (!singleInstanceLock) {
     const ingestionService = IngestionService.getInstance()
     const embeddingRuntimeClient = new EmbeddingRuntimeClient({
       entryPath: fileURLToPath(new URL('./embeddingRuntimeProcess.js', import.meta.url)),
-      modelsPath: MODELS_PATH,
+      modelsPath: getModelsPath(),
       modelName: DEFAULT_EMBEDDING_MODEL_NAME
     })
     const embeddingService = new EmbeddingService({

src/main/services/model/ModelService.ts

@@ -3,7 +3,7 @@ import path from 'path'
 import { Readable, Transform } from 'node:stream'
 import { pipeline } from 'node:stream/promises'
 import type { ReadableStream as NodeReadableStream } from 'node:stream/web'
-import { listModelPathCandidates, MODELS_PATH } from '../../utils/paths'
+import { getModelsPath, listModelPathCandidates } from '../../utils/paths'
 
 const HF_ENDPOINT = 'https://hf-mirror.com'
 const DOWNLOAD_LOCK_POLL_MS = 1000
@@ -111,11 +111,11 @@ export class ModelService {
   }
 
   private resolveModelPath(repoId: string): string {
-    return path.join(MODELS_PATH, repoId)
+    return path.join(getModelsPath(), repoId)
   }
 
   private resolveDownloadLockPath(repoId: string): string {
-    return path.join(MODELS_PATH, `${repoId.replaceAll('/', '__')}.download.lock`)
+    return path.join(getModelsPath(), `${repoId.replaceAll('/', '__')}.download.lock`)
   }
 
   private resolveDownloadStatePath(repoId: string): string {

src/main/services/model/VectorModelDownloadService.ts

@@ -1,7 +1,7 @@
 import fs from 'fs'
 import path from 'path'
 import type { EmbeddingService } from '../vector/EmbeddingService'
-import { MODELS_PATH } from '../../utils/paths'
+import { getModelsPath } from '../../utils/paths'
 import type { ModelDownloadProgress } from './ModelService'
 import { ModelService } from './ModelService'
 
@@ -211,7 +211,7 @@ export class VectorModelDownloadService {
   }
 
   private resolveSnapshotPath(): string {
-    return path.join(MODELS_PATH, ACTIVE_DOWNLOAD_SNAPSHOT_FILE)
+    return path.join(getModelsPath(), ACTIVE_DOWNLOAD_SNAPSHOT_FILE)
   }
 
   private readSnapshot(): ActiveVectorDownloadSnapshot | null {
@@ -238,7 +238,7 @@ export class VectorModelDownloadService {
 
   private setSnapshot(snapshot: ActiveVectorDownloadSnapshot): void {
     this.activeSnapshot = snapshot
-    fs.mkdirSync(MODELS_PATH, { recursive: true })
+    fs.mkdirSync(getModelsPath(), { recursive: true })
     fs.writeFileSync(this.resolveSnapshotPath(), JSON.stringify(snapshot, null, 2), 'utf8')
   }
 

src/main/services/storage/core/LanceDbCore.ts

@@ -1,7 +1,7 @@
 import * as lancedb from '@lancedb/lancedb'
 import fs from 'fs'
 import * as arrow from 'apache-arrow'
-import { LANCE_DB_PATH } from '../../../utils/paths'
+import { getLanceDbPath } from '../../../utils/paths'
 import type { TableConfig } from '../../../types/store'
 
 export enum LanceDbCoreStatus {
@@ -19,12 +19,15 @@ export type LanceTableName = (typeof LANCE_TABLES)[keyof typeof LANCE_TABLES]
 
 export class LanceDbCore {
   private db: lancedb.Connection | null = null
+  private readonly dbPath: string
   private status: LanceDbCoreStatus = LanceDbCoreStatus.UNINITIALIZED
   private tableNamesCache: Set<string> | null = null
   private tableCache = new Map<string, object>()
   private tableWriteQueues = new Map<string, Promise<unknown>>()
 
-  public constructor(private readonly dbPath: string = LANCE_DB_PATH) {}
+  public constructor(dbPath?: string) {
+    this.dbPath = dbPath || getLanceDbPath()
+  }
 
   public getStatus(): LanceDbCoreStatus {
     return this.status

src/main/services/storage/sqlite/SqliteCore.ts

@@ -1,7 +1,7 @@
 import Database from 'better-sqlite3'
 import fs from 'fs'
 import path from 'path'
-import { SQLITE_DB_PATH } from '../../../utils/paths'
+import { getSqliteDbPath } from '../../../utils/paths'
 import { initialSchemaSql } from './migrations/001_initial_schema'
 
 const notificationSettingsMigration = `
@@ -122,8 +122,11 @@ const resolveBetterSqliteNativeBinding = (): string => {
 
 export class SqliteCore {
   private db: SqliteDatabase | null = null
+  private readonly dbPath: string
 
-  public constructor(private readonly dbPath: string = SQLITE_DB_PATH) {}
+  public constructor(dbPath?: string) {
+    this.dbPath = dbPath || getSqliteDbPath()
+  }
 
   public initialize(): void {
     if (this.db) return

src/main/services/workspace/WorkspaceKnowledgeCache.ts

@@ -2,7 +2,7 @@ import { createHash, randomUUID } from 'node:crypto'
 import { mkdir, readFile, rename, rm, writeFile } from 'node:fs/promises'
 import path from 'node:path'
 import { SPLITTER_BIG_CHUNK_SETTING, SPLITTER_MINI_CHUNK_SIZE } from '../../utils/constant'
-import { USER_DATA_PATH } from '../../utils/paths'
+import { getUserDataPath } from '../../utils/paths'
 import type {
   WorkspaceKnowledgeCache,
   WorkspaceKnowledgeSplitSnapshot,
@@ -46,7 +46,7 @@ const isWorkspaceKnowledgeCacheEnabled = (): boolean =>
 const resolveWorkspaceKnowledgeCacheRoot = (): string => {
   const explicitPath = String(process.env.NCURATOR_WORKSPACE_KNOWLEDGE_CACHE_PATH || '').trim()
   if (explicitPath) return path.resolve(explicitPath)
-  return path.join(USER_DATA_PATH, 'workspace-knowledge-cache')
+  return path.join(getUserDataPath(), 'workspace-knowledge-cache')
 }
 
 const resolveWorkspaceKnowledgeCacheStatsPath = (cacheRoot: string): string => {

src/main/utils/paths.ts

@@ -2,8 +2,6 @@ import { app } from 'electron'
 import os from 'os'
 import path from 'path'
 
-const E2E_USER_DATA_PATH = String(process.env.NCURATOR_USER_DATA_PATH || '').trim()
-const E2E_MODELS_PATH = String(process.env.NCURATOR_MODELS_PATH || '').trim()
 const FALLBACK_USER_DATA_PATH = path.join(os.tmpdir(), 'ncurator-user-data')
 const KNOWN_USER_DATA_DIRS = ['ncurator', 'ncurator-desktop'] as const
 
@@ -54,14 +52,17 @@ const resolveElectronUserDataPath = (): string => {
   return FALLBACK_USER_DATA_PATH
 }
 
-export const USER_DATA_PATH = E2E_USER_DATA_PATH
-  ? path.resolve(E2E_USER_DATA_PATH)
-  : resolveElectronUserDataPath()
-export const SQLITE_DB_PATH = path.join(USER_DATA_PATH, 'app.sqlite')
-export const LANCE_DB_PATH = path.join(USER_DATA_PATH, 'lancedb')
+export const getUserDataPath = (): string => {
+  const overridePath = String(process.env.NCURATOR_USER_DATA_PATH || '').trim()
+  return overridePath ? path.resolve(overridePath) : resolveElectronUserDataPath()
+}
+
+export const getSqliteDbPath = (): string => path.join(getUserDataPath(), 'app.sqlite')
+
+export const getLanceDbPath = (): string => path.join(getUserDataPath(), 'lancedb')
 
 export const listModelPathCandidates = (
-  userDataPath: string = USER_DATA_PATH,
+  userDataPath: string = getUserDataPath(),
   platform: NodeJS.Platform = process.platform,
   homeDir: string = os.homedir(),
   tmpDir: string = os.tmpdir()
@@ -75,11 +76,13 @@ export const listModelPathCandidates = (
     : [activeModelsPath, ...persistentModelPaths]
 
   const candidates = [
-    E2E_MODELS_PATH ? path.resolve(E2E_MODELS_PATH) : '',
+    String(process.env.NCURATOR_MODELS_PATH || '').trim()
+      ? path.resolve(String(process.env.NCURATOR_MODELS_PATH).trim())
+      : '',
     ...orderedModelPaths
   ].filter(Boolean)
 
   return dedupePaths(candidates)
 }
 
-export const MODELS_PATH = listModelPathCandidates()[0]
+export const getModelsPath = (): string => listModelPathCandidates()[0]

src/renderer/src/components/workbench/WorkspaceFileExplorer.tsx

@@ -326,7 +326,7 @@ const WorkspaceFileExplorer: React.FC<WorkspaceFileExplorerProps> = ({
             )
           })
         ) : (
-          <div className="px-3 py-8 text-center text-xs text-[#91918E]">本会话暂无文件改动</div>
+          <div className="px-3 py-8 text-center text-xs text-[#91918E]">本次运行暂无文件改动</div>
         )}
       </div>
     </div>

src/renderer/src/components/workbench/applyTargetUtils.ts

@@ -1,4 +1,5 @@
 import type { AiFileChangeRecord } from '../../../../shared/types'
+import { groupFileChangeRecordsByRun } from '../../../../shared/aiFileChangeScope'
 import type { AssistantTurnApplyTarget } from './types'
 
 export const buildApplyTarget = (record: AiFileChangeRecord): AssistantTurnApplyTarget => ({
@@ -23,12 +24,7 @@ export const mergeApplyTargets = (
 export const groupApplyTargetsByRunId = (
   records: AiFileChangeRecord[]
 ): Map<string, AssistantTurnApplyTarget[]> => {
-  const grouped = new Map<string, AiFileChangeRecord[]>()
-  records.forEach((record) => {
-    const existing = grouped.get(record.runId) || []
-    existing.push(record)
-    grouped.set(record.runId, existing)
-  })
+  const grouped = groupFileChangeRecordsByRun(records)
 
   return new Map(
     [...grouped.entries()].map(([runId, runRecords]) => [

src/renderer/src/pages/ChatPage.tsx

@@ -96,6 +96,13 @@ import {
   groupApplyTargetsByRunId,
   mergeApplyTargets
 } from '../components/workbench/applyTargetUtils'
+import {
+  createFileChangeListRequest,
+  createFileChangeRunScope,
+  filterFileChangeRecordsForScope,
+  isCurrentFileChangeRunScope,
+  type FileChangeRunScope
+} from '../../../shared/aiFileChangeScope'
 import { chatSelectors, workbenchSelectors } from '../testing/selectors'
 import SidebarPanelIcon from '../components/SidebarPanelIcon'
 import brandIcon from '../../../../logo/png/ncurator-logo-24.png'
@@ -667,7 +674,7 @@ const ChatPage: React.FC = () => {
 
   const [activeApplyRecord, setActiveApplyRecord] = useState<AiFileChangeRecord | null>(null)
   const [fileChangeRecords, setFileChangeRecords] = useState<AiFileChangeRecord[]>([])
-  const fileChangeRunScopeRef = useRef<{ sessionId: string; runId: string } | null>(null)
+  const fileChangeRunScopeRef = useRef<FileChangeRunScope | null>(null)
   const [applyLoading, setApplyLoading] = useState(false)
   const [applyReverting, setApplyReverting] = useState(false)
   const [applyRevertConfirm, setApplyRevertConfirm] = useState<ApplyRevertConfirmState | null>(null)
@@ -1512,19 +1519,13 @@ const ChatPage: React.FC = () => {
       }
 
       setActiveApplyRecord(nextRecord)
-      if (
-        fileChangeRunScopeRef.current?.sessionId === nextRecord.sessionId &&
-        fileChangeRunScopeRef.current.runId === nextRecord.runId
-      ) {
+      if (isCurrentFileChangeRunScope(fileChangeRunScopeRef.current, nextRecord)) {
         setFileChangeRecords((current) =>
           current.map((record) => (record.id === nextRecord.id ? nextRecord : record))
         )
       }
 
-      if (
-        fileChangeRunScopeRef.current?.sessionId !== nextRecord.sessionId ||
-        fileChangeRunScopeRef.current.runId !== nextRecord.runId
-      ) {
+      if (!isCurrentFileChangeRunScope(fileChangeRunScopeRef.current, nextRecord)) {
         void loadFileChangesForSessionRef.current(nextRecord.sessionId, nextRecord.runId, {
           preferredApplyId: nextRecord.id,
           skipApplyRecordReload: true
@@ -1556,7 +1557,8 @@ const ChatPage: React.FC = () => {
   ): Promise<void> => {
     const requestId = fileChangeLoadRequestRef.current + 1
     fileChangeLoadRequestRef.current = requestId
-    fileChangeRunScopeRef.current = { sessionId, runId }
+    const scope = createFileChangeRunScope({ sessionId, runId })
+    fileChangeRunScopeRef.current = scope
     try {
       const shouldWaitForRecords =
         Boolean(options?.forceOpenWorkspace) || Boolean(options?.preferredApplyId)
@@ -1565,8 +1567,8 @@ const ChatPage: React.FC = () => {
       let activeRecords: AiFileChangeRecord[] = []
 
       for (let attempt = 0; attempt < maxAttempts; attempt += 1) {
-        const response = await window.api.aiFileChangeList({ sessionId, runId })
-        const listedItems = sortFileChangeRecords(response.items)
+        const response = await window.api.aiFileChangeList(createFileChangeListRequest(scope))
+        const listedItems = sortFileChangeRecords(filterFileChangeRecordsForScope(response.items, scope))
         const noEffectiveChangeRecords = listedItems.filter(
           (item) => item.status === 'failed' && item.error === NO_EFFECTIVE_FILE_CHANGE_MESSAGE
         )
@@ -2402,8 +2404,10 @@ const ChatPage: React.FC = () => {
         if (
           eventSessionId &&
           eventSessionId === currentSessionIdRef.current &&
-          fileChangeRunScopeRef.current?.sessionId === eventSessionId &&
-          fileChangeRunScopeRef.current.runId === event.runId
+          isCurrentFileChangeRunScope(fileChangeRunScopeRef.current, {
+            sessionId: eventSessionId,
+            runId: event.runId
+          })
         ) {
           void loadFileChangesForSession(eventSessionId, event.runId)
         }
@@ -2439,8 +2443,10 @@ const ChatPage: React.FC = () => {
         if (
           eventSessionId &&
           eventSessionId === currentSessionIdRef.current &&
-          fileChangeRunScopeRef.current?.sessionId === eventSessionId &&
-          fileChangeRunScopeRef.current.runId === event.runId
+          isCurrentFileChangeRunScope(fileChangeRunScopeRef.current, {
+            sessionId: eventSessionId,
+            runId: event.runId
+          })
         ) {
           void loadFileChangesForSession(eventSessionId, event.runId)
         }
@@ -2459,8 +2465,10 @@ const ChatPage: React.FC = () => {
         if (
           eventSessionId &&
           eventSessionId === currentSessionIdRef.current &&
-          fileChangeRunScopeRef.current?.sessionId === eventSessionId &&
-          fileChangeRunScopeRef.current.runId === event.runId
+          isCurrentFileChangeRunScope(fileChangeRunScopeRef.current, {
+            sessionId: eventSessionId,
+            runId: event.runId
+          })
         ) {
           void loadFileChangesForSession(eventSessionId, event.runId, {
             notifyNoEffectiveChange: true

src/shared/aiFileChangeScope.ts

@@ -0,0 +1,41 @@
+import type { AiFileChangeRecord, ListAiFileChangeRecordsRequest } from './types'
+
+export type FileChangeRunScope = {
+  sessionId: string
+  runId: string
+}
+
+export const createFileChangeRunScope = (input: FileChangeRunScope): FileChangeRunScope => ({
+  sessionId: input.sessionId,
+  runId: input.runId
+})
+
+export const isCurrentFileChangeRunScope = (
+  scope: FileChangeRunScope | null,
+  input: FileChangeRunScope
+): boolean => scope?.sessionId === input.sessionId && scope.runId === input.runId
+
+export const createFileChangeListRequest = (
+  scope: FileChangeRunScope
+): ListAiFileChangeRecordsRequest => ({
+  sessionId: scope.sessionId,
+  runId: scope.runId
+})
+
+export const filterFileChangeRecordsForScope = (
+  records: readonly AiFileChangeRecord[],
+  scope: FileChangeRunScope
+): AiFileChangeRecord[] =>
+  records.filter((record) => record.sessionId === scope.sessionId && record.runId === scope.runId)
+
+export const groupFileChangeRecordsByRun = (
+  records: readonly AiFileChangeRecord[]
+): Map<string, AiFileChangeRecord[]> => {
+  const grouped = new Map<string, AiFileChangeRecord[]>()
+  records.forEach((record) => {
+    const current = grouped.get(record.runId) || []
+    current.push(record)
+    grouped.set(record.runId, current)
+  })
+  return grouped
+}

tests/case/ai/tool/terminal/terminal-inline-edit-artifact-contract.test.ts

@@ -1,6 +1,11 @@
 import fs from 'node:fs/promises'
 import path from 'node:path'
 import { describe, it } from 'vitest'
+import type { AiRunEvent } from '../../../../../src/shared/types'
+import type { TerminalToolContext } from '../../../../../src/main/services/ai/types'
+import { createHumanRuntimeCache } from '../../../../../src/main/services/ai/human/node/utils/runtimeCache'
+import { buildTerminalCapabilityTools } from '../../../../../src/main/services/ai/human/tools/terminal/terminalTools'
+import type { TerminalSessionManager } from '../../../../../src/main/services/ai/human/tools/terminal/TerminalSessionManager'
 import { createTerminalToolHarness } from '../../../../harness/backend/aiToolHarness'
 import { runBackendCase } from '../../../../harness/backend/run-backend-case'
 import { filterEventsByType } from '../testHelpers'
@@ -82,4 +87,88 @@ describe('ai tool terminal backend', () => {
       }
     })
   })
+
+  it('terminal-tracking-start-failure-cleanup', async () => {
+    await runBackendCase({
+      caseId: 'terminal-tracking-start-failure-cleanup',
+      harnessMode: 'tool-real',
+      domain: 'ai',
+      suite: 'tool',
+      groups: ['terminal'],
+      description: '验证终端启动失败时丢弃已创建的命令级文件追踪。',
+      fixtures: [],
+      runMode: 'single',
+      execute: async (runtime) => {
+        const baseCtx = runtime.createAiRunContext()
+        const events: AiRunEvent[] = []
+        const begunTrackingIds: string[] = []
+        const discardedTrackingIds: string[] = []
+        let finishCalls = 0
+        const ctx = {
+          runId: runtime.runId,
+          platform: baseCtx.platform,
+          checkCancelled: baseCtx.checkCancelled,
+          humanRuntimeOverrides: baseCtx.humanRuntimeOverrides,
+          sendEvent: (event: AiRunEvent) => events.push(event),
+          requestApprovalBatch: baseCtx.requestApprovalBatch,
+          saveFileChangeRecord: baseCtx.saveFileChangeRecord,
+          beginTerminalCommandTracking: async ({ commandId }) => {
+            begunTrackingIds.push(commandId)
+          },
+          finishTerminalCommandTracking: async () => {
+            finishCalls += 1
+            return []
+          },
+          discardTerminalCommandTracking: (commandId) => {
+            discardedTrackingIds.push(commandId)
+          }
+        } satisfies TerminalToolContext
+        const failedSessionManager = {
+          list: () => [],
+          run: async () => {
+            throw new Error('terminal startup failed')
+          }
+        } as unknown as TerminalSessionManager
+        const terminalRun = buildTerminalCapabilityTools({
+          ctx,
+          cache: createHumanRuntimeCache(),
+          runId: runtime.runId,
+          sessionId: 'test-session',
+          taskId: runtime.taskId,
+          workspaceRootPath: runtime.workspaceRoot,
+          workspaceId: 'test-workspace',
+          sessionManager: failedSessionManager,
+          aiResultTokenLimitRatio: 0.25,
+          emitActivity: () => undefined
+        }).tools.find((tool) => tool.name === 'terminal_run')
+        if (!terminalRun) throw new Error('terminal_run tool is missing')
+        const result = await terminalRun.invoke({
+          command: isWindows ? 'Get-Location' : 'pwd',
+          commandExplanation: '验证终端启动失败时的追踪清理。'
+        })
+
+        return {
+          result,
+          begunTrackingIds,
+          discardedTrackingIds,
+          finishCalls,
+          artifactEvents: events.filter((event) => event.type === 'file_artifact')
+        }
+      },
+      assert: async (result, runtime) => {
+        runtime.assertThat(
+          result.begunTrackingIds.length === 1 &&
+            result.discardedTrackingIds.length === 1 &&
+            result.discardedTrackingIds[0] === result.begunTrackingIds[0] &&
+            result.finishCalls === 0 &&
+            result.artifactEvents.length === 0,
+          {
+            name: 'terminal-startup-failure-discards-file-tracking',
+            message: '终端启动失败后必须丢弃命令快照，不得保存修改记录或发送文件 artifact。',
+            data: result
+          }
+        )
+      }
+    })
+  })
 })

tests/case/ai/tool/terminal/terminal-persistent-session-contract.test.ts

@@ -1,6 +1,7 @@
 import fs from 'node:fs/promises'
 import path from 'node:path'
 import { describe, it } from 'vitest'
+import type { AiFileChangeRecord, AiRunEvent } from '../../../../../src/shared/types'
 import { createTerminalToolHarness } from '../../../../harness/backend/aiToolHarness'
 import { runBackendCase } from '../../../../harness/backend/run-backend-case'
 
@@ -40,7 +41,7 @@ describe('ai tool terminal backend', () => {
       domain: 'ai',
       suite: 'tool',
       groups: ['terminal'],
-      description: '验证真实持久终端保留当前路径、长命令让出控制权、增量读取和停止。',
+      description: '验证真实持久终端保留当前路径、命令级文件追踪、长命令让出控制权和停止。',
       fixtures: [],
       runMode: 'single',
       execute: async (runtime) => {
@@ -100,6 +101,18 @@ describe('ai tool terminal backend', () => {
           }
           throw new Error(`terminal output did not include ${expected}`)
         }
+        const waitForTrackedRecords = async (
+          fileNames: readonly string[]
+        ): Promise<AiFileChangeRecord[]> => {
+          for (let attempt = 0; attempt < 60; attempt += 1) {
+            const records = await harness.ctx.listFileChangeRecords({ runId: runtime.runId })
+            if (fileNames.every((fileName) => records.some((record) => record.fileName === fileName))) {
+              return records
+            }
+            await new Promise<void>((resolve) => setTimeout(resolve, 50))
+          }
+          throw new Error(`terminal change records were not saved: ${fileNames.join(', ')}`)
+        }
         const changed = await waitForIdle(
           await harness.invoke<TerminalRunResult>('terminal_run', {
             command: isWindows ? 'Set-Location nested' : 'cd nested'
@@ -116,17 +129,43 @@ describe('ai tool terminal backend', () => {
             command: isWindows ? 'Get-Location' : 'pwd'
           })
         )
+        const completedLongCommand = await waitForIdle(
+          await harness.invoke<TerminalRunResult>('terminal_run', {
+            terminalId: changed.terminalId,
+            command: isWindows
+              ? 'Set-Content -LiteralPath ..\\completed-from-nested.txt -Value completed; Start-Sleep -Milliseconds 500'
+              : 'printf "completed\\n" > ../completed-from-nested.txt; sleep 0.5'
+          })
+        )
         const snapshotsBeforeStop = harness.sessionManager.list({ chatSessionId: 'test-session' })
         const running = await harness.invoke<TerminalRunResult>('terminal_run', {
           terminalId: changed.terminalId,
           command: isWindows
-            ? 'while ($true) { Write-Output "terminal-tick"; Start-Sleep -Milliseconds 50 }'
-            : 'while true; do echo terminal-tick; sleep 0.05; done'
+            ? 'Set-Content -LiteralPath ..\\stopped-from-nested.txt -Value stopped; while ($true) { Write-Output "terminal-tick"; Start-Sleep -Milliseconds 50 }'
+            : 'printf "stopped\\n" > ../stopped-from-nested.txt; while true; do echo terminal-tick; sleep 0.05; done'
         })
         const read = await readUntilOutput(running, 'terminal-tick')
         const stopped = await harness.invoke<TerminalStopResult>('terminal_stop', {
           terminalId: changed.terminalId
         })
+        const nonZero = await waitForIdle(
+          await harness.invoke<TerminalRunResult>('terminal_run', {
+            terminalId: changed.terminalId,
+            command: isWindows
+              ? 'Set-Content -LiteralPath ..\\nonzero-from-nested.txt -Value nonzero; cmd /c exit 1'
+              : 'printf "nonzero\\n" > ../nonzero-from-nested.txt; false'
+          })
+        )
+        const records = await waitForTrackedRecords([
+          'completed-from-nested.txt',
+          'stopped-from-nested.txt',
+          'nonzero-from-nested.txt'
+        ])
+        const artifactEvents = runtime.eventCollector
+          .snapshot()
+          .filter((event): event is Extract<AiRunEvent, { type: 'file_artifact' }> =>
+            event.type === 'file_artifact'
+          )
         const detachedRejected = await harness.invoke<TerminalFailureResult>('terminal_run', {
           command: isWindows ? 'Start-Process powershell' : 'sleep 1 &'
         })
@@ -136,9 +175,13 @@ describe('ai tool terminal backend', () => {
           changed,
           continued,
           fresh,
+          completedLongCommand,
           running,
           read,
           stopped,
+          nonZero,
+          records,
+          artifactEvents,
           detachedRejected,
           snapshotsBeforeStop,
           snapshotsAfterStop
@@ -187,6 +230,31 @@ describe('ai tool terminal backend', () => {
             data: result
           }
         )
+        const trackedFileNames = new Set(result.records.map((record) => record.fileName))
+        const artifactFileNames = new Set(result.artifactEvents.map((event) => event.fileName))
+        runtime.assertThat(
+          result.completedLongCommand.executionStatus === 'completed' &&
+            result.nonZero.executionStatus === 'completed' &&
+            result.nonZero.exitCode === 1 &&
+            ['completed-from-nested.txt', 'stopped-from-nested.txt', 'nonzero-from-nested.txt'].every(
+              (fileName) => trackedFileNames.has(fileName) && artifactFileNames.has(fileName)
+            ) &&
+            result.records
+              .filter((record) => trackedFileNames.has(record.fileName))
+              .every((record) => record.source === 'terminal_tool' && record.runId === runtime.runId),
+          {
+            name: 'persistent-terminal-finalizes-command-scoped-file-changes',
+            message:
+              '长命令完成、停止和非零退出都必须按命令窗口保存文件变化并发出当前 run 的 artifact。',
+            data: {
+              completedLongCommand: result.completedLongCommand,
+              stopped: result.stopped,
+              nonZero: result.nonZero,
+              records: result.records,
+              artifactEvents: result.artifactEvents
+            }
+          }
+        )
       }
     })
   })

tests/case/unit/ai/file-change-run-scope.test.ts

@@ -0,0 +1,60 @@
+import { describe, expect, it } from 'vitest'
+import type { AiFileChangeRecord } from '../../../../src/shared/types'
+import {
+  createFileChangeListRequest,
+  filterFileChangeRecordsForScope,
+  groupFileChangeRecordsByRun,
+  isCurrentFileChangeRunScope
+} from '../../../../src/shared/aiFileChangeScope'
+
+const record = (input: Partial<AiFileChangeRecord>): AiFileChangeRecord => ({
+  id: input.id || 'change-1',
+  sessionId: input.sessionId || 'session-a',
+  runId: input.runId || 'run-a',
+  status: 'applied',
+  capability: 'text_file',
+  path: input.path || '/workspace/file.txt',
+  fileName: input.fileName || 'file.txt',
+  operation: 'updated',
+  changeCount: 1,
+  insertCount: 1,
+  deleteCount: 0,
+  createdAt: 1,
+  updatedAt: input.updatedAt || 1
+})
+
+describe('ai unit file change run scope', () => {
+  it('keeps the right panel request and records within one session and run', () => {
+    const scope = { sessionId: 'session-a', runId: 'run-a' }
+    const records = [
+      record({ id: 'a', sessionId: 'session-a', runId: 'run-a', path: '/workspace/a.txt' }),
+      record({ id: 'b', sessionId: 'session-a', runId: 'run-b', path: '/workspace/b.txt' }),
+      record({ id: 'c', sessionId: 'session-b', runId: 'run-a', path: '/workspace/c.txt' })
+    ]
+
+    expect(createFileChangeListRequest(scope)).toEqual({ sessionId: 'session-a', runId: 'run-a' })
+    expect(filterFileChangeRecordsForScope(records, scope).map((item) => item.id)).toEqual(['a'])
+  })
+
+  it('ignores completion events for another session or run', () => {
+    const scope = { sessionId: 'session-a', runId: 'run-a' }
+
+    expect(isCurrentFileChangeRunScope(scope, scope)).toBe(true)
+    expect(isCurrentFileChangeRunScope(scope, { sessionId: 'session-a', runId: 'run-b' })).toBe(
+      false
+    )
+    expect(isCurrentFileChangeRunScope(scope, { sessionId: 'session-b', runId: 'run-a' })).toBe(
+      false
+    )
+  })
+
+  it('keeps assistant history entries grouped by their own run', () => {
+    const groups = groupFileChangeRecordsByRun([
+      record({ id: 'a', runId: 'run-a', path: '/workspace/a.txt', updatedAt: 2 }),
+      record({ id: 'b', runId: 'run-b', path: '/workspace/b.txt', updatedAt: 3 })
+    ])
+
+    expect(groups.get('run-a')?.map((item) => item.id)).toEqual(['a'])
+    expect(groups.get('run-b')?.map((item) => item.id)).toEqual(['b'])
+  })
+})

tests/case/unit/ai/model-path-candidates.test.ts

@@ -2,11 +2,36 @@ import os from 'os'
 import path from 'path'
 import { describe, expect, it } from 'vitest'
 import {
+  getLanceDbPath,
+  getSqliteDbPath,
+  getUserDataPath,
   listKnownPersistentUserDataPaths,
   listModelPathCandidates
 } from '../../../../src/main/utils/paths'
 
 describe('ai unit model path candidates', () => {
+  it('resolves user data paths when called instead of at module load time', () => {
+    const original = process.env.NCURATOR_USER_DATA_PATH
+    const firstPath = path.join(os.tmpdir(), 'ncurator-paths-first')
+    const secondPath = path.join(os.tmpdir(), 'ncurator-paths-second')
+
+    try {
+      process.env.NCURATOR_USER_DATA_PATH = firstPath
+      expect(getUserDataPath()).toBe(firstPath)
+      expect(getSqliteDbPath()).toBe(path.join(firstPath, 'app.sqlite'))
+      expect(getLanceDbPath()).toBe(path.join(firstPath, 'lancedb'))
+
+      process.env.NCURATOR_USER_DATA_PATH = secondPath
+      expect(getUserDataPath()).toBe(secondPath)
+    } finally {
+      if (original === undefined) {
+        delete process.env.NCURATOR_USER_DATA_PATH
+      } else {
+        process.env.NCURATOR_USER_DATA_PATH = original
+      }
+    }
+  })
+
   it('includes known persistent macOS user data directories', () => {
     const homeDir = path.join('/Users', 'tester')
     const candidates = listKnownPersistentUserDataPaths('darwin', homeDir)

tests/case/unit/ai/persistence/chat-run-service-contract.test.ts

@@ -1,9 +1,11 @@
-import { mkdir, mkdtemp, rm } from 'node:fs/promises'
+import { access, mkdir, mkdtemp, rm, writeFile } from 'node:fs/promises'
 import os from 'node:os'
 import path from 'node:path'
 import { afterEach, describe, expect, it, vi } from 'vitest'
 import { ChatRunService } from '../../../../../src/main/services/ai/ChatRunService'
+import { AiRunService } from '../../../../../src/main/services/ai/AiRunService'
 import { AiRunResourceManager } from '../../../../../src/main/services/ai/run/AiRunResourceManager'
+import { createCompletedAiRunState } from '../../../../../src/main/services/ai/run/state'
 import {
   CONVERSATION_MEMORY_RECENT_TURNS,
   CONVERSATION_MEMORY_SUMMARY_TRIGGER_TURNS
@@ -689,6 +691,35 @@ describe('unit ai persistence backend', () => {
     }
   })
 
+  it('does-not-attribute-uncontrolled-workspace-changes-when-a-run-finishes', async () => {
+    const setup = await createTestService({ autoStartRuns: true })
+    const runner = vi.spyOn(AiRunService, 'getInstance').mockReturnValue({
+      run: async (payload) => {
+        await writeFile(path.join(setup.workspaceRoot, 'uncontrolled-change.txt'), 'external change')
+        return createCompletedAiRunState({
+          runId: payload.runId,
+          sessionId: payload.sessionId,
+          input: payload.input,
+          finalDeliveryMarkdown: '完成，但没有调用文件工具或终端。'
+        })
+      }
+    } as unknown as AiRunService)
+    try {
+      const run = await submitAiRun(setup)
+      await waitForCondition(() => setup.events.some((event) => event.type === 'run_completed'))
+
+      const records = await setup.storage.aiFileChange.listRecords({
+        sessionId: run.sessionId,
+        runId: run.id
+      })
+      expect(await access(path.join(setup.workspaceRoot, 'uncontrolled-change.txt')).then(() => true)).toBe(true)
+      expect(records.items).toEqual([])
+      expect(setup.events.filter((event) => event.type === 'file_artifact')).toEqual([])
+    } finally {
+      runner.mockRestore()
+    }
+  })
+
   it('approval-required-is-restored-through-session-snapshot', async () => {
     const setup = await createTestService()
     const run = await submitAiRun(setup)

tests/case/unit/vector/vector-download-single-task-contract.test.ts

@@ -7,7 +7,7 @@ import type {
   ModelDownloadProgressReporter
 } from '../../../../src/main/services/model/ModelService'
 import { runBackendCase } from '../../../harness/backend/run-backend-case'
-import { MODELS_PATH } from '../../../../src/main/utils/paths'
+import { getModelsPath } from '../../../../src/main/utils/paths'
 import { TransformersEmbeddingExecutor } from '../../../../src/main/services/vector/runtime/TransformersEmbeddingExecutor'
 
 const REPO_A = 'jinaai/jina-embeddings-v2-base-zh'
@@ -36,7 +36,7 @@ describe('unit vector backend', () => {
         const modelService = ModelService.getInstance()
         const embeddingService = new EmbeddingService({
           executor: new TransformersEmbeddingExecutor({
-            modelsPath: MODELS_PATH,
+            modelsPath: getModelsPath(),
             modelName: REPO_A
           })
         })

tests/case/unit/vector/vector-download-startup-resume-contract.test.ts

@@ -1,7 +1,7 @@
 import fs from 'fs'
 import path from 'path'
 import { afterEach, describe, it, vi } from 'vitest'
-import { MODELS_PATH } from '../../../../src/main/utils/paths'
+import { getModelsPath } from '../../../../src/main/utils/paths'
 import { EmbeddingService } from '../../../../src/main/services/vector/EmbeddingService'
 import { ModelService } from '../../../../src/main/services/model/ModelService'
 import { VectorModelDownloadService } from '../../../../src/main/services/model/VectorModelDownloadService'
@@ -19,7 +19,7 @@ type VectorModelDownloadServiceStatics = typeof VectorModelDownloadService & {
   instance: VectorModelDownloadService | null
 }
 
-const resolveSnapshotPath = (): string => path.join(MODELS_PATH, ACTIVE_DOWNLOAD_SNAPSHOT_FILE)
+const resolveSnapshotPath = (): string => path.join(getModelsPath(), ACTIVE_DOWNLOAD_SNAPSHOT_FILE)
 
 const waitForCondition = async (predicate: () => boolean, timeoutMs = 5000): Promise<void> => {
   const deadline = Date.now() + timeoutMs
@@ -55,7 +55,7 @@ describe('unit vector backend', () => {
         const modelService = ModelService.getInstance()
         const embeddingService = new EmbeddingService({
           executor: new TransformersEmbeddingExecutor({
-            modelsPath: MODELS_PATH,
+            modelsPath: getModelsPath(),
             modelName: SNAPSHOT_REPO_ID
           })
         })
@@ -69,7 +69,7 @@ describe('unit vector backend', () => {
           releaseDownload = resolve
         })
 
-        await fs.promises.mkdir(MODELS_PATH, { recursive: true })
+        await fs.promises.mkdir(getModelsPath(), { recursive: true })
         await fs.promises.writeFile(
           resolveSnapshotPath(),
           JSON.stringify(

tests/case/workbench/result/apply/file-change-run-scope.e2e.test.ts

@@ -0,0 +1,172 @@
+import { describe, it } from 'vitest'
+import { runE2ECase } from '../../../../harness/e2e/run-e2e-case'
+import {
+  WORKBENCH_LIVE_AI_E2E_TIMEOUT_MS,
+  assertNoWorkbenchFakeSignals,
+  createWorkbenchNewSession,
+  getWorkbenchLocator,
+  recordWorkbenchAiEvidence,
+  selectWorkbenchWorkspace,
+  sendWorkbenchMessage,
+  setupWorkbenchWithTempWorkspace,
+  waitForWorkbenchAiRunCompletion,
+  workbenchSelectors
+} from '../../../../harness/e2e/workbench'
+
+const CASE_ID = 'file-change-run-scope'
+const FIRST_FILE = 'run-scope-first.txt'
+const SECOND_FILE = 'run-scope-second.txt'
+
+const getRunStarted = (events: import('../../../../../src/shared/types').AiRunEvent[]) => {
+  const event = events.find(
+    (item): item is Extract<import('../../../../../src/shared/types').AiRunEvent, { type: 'run_started' }> =>
+      item.type === 'run_started'
+  )
+  if (!event) throw new Error('run_started event is missing')
+  return event
+}
+
+const getApplyEntryForFile = async (page: import('playwright').Page, fileName: string) => {
+  const entry = page
+    .locator(`[data-testid^="${workbenchSelectors.assistantTurnApplyPrefix}-"]`)
+    .filter({ hasText: fileName })
+    .first()
+  await entry.waitFor({ state: 'visible', timeout: 30_000 })
+  return entry
+}
+
+const readChangedFileList = async (page: import('playwright').Page) => {
+  await getWorkbenchLocator(page, workbenchSelectors.workspaceExplorerToggle).click()
+  const explorer = getWorkbenchLocator(page, workbenchSelectors.workspaceExplorerOverlay)
+  await explorer.waitFor({ state: 'visible', timeout: 30_000 })
+  const text = ((await explorer.textContent()) || '').trim()
+  await page.keyboard.press('Escape')
+  return text
+}
+
+describe('workbench result e2e', () => {
+  it(
+    CASE_ID,
+    async () => {
+      await runE2ECase({
+        caseId: CASE_ID,
+        domain: 'workbench',
+        suite: 'result',
+        groups: ['apply'],
+        description:
+          '真实 UI 在不同会话的两个 run 分别写入文件后，右侧本次运行改动不得串入另一个 run 的记录。',
+        fixtures: ['tests/fixtures/workspace'],
+        serviceMode: 'real',
+        execution: 'live',
+        windowMode: 'visible',
+        runMode: 'single',
+        execute: async (runtime) => {
+          const { page } = runtime
+          const setup = await setupWorkbenchWithTempWorkspace(page, { caseId: CASE_ID })
+
+          const firstStartedAtMs = Date.now()
+          await sendWorkbenchMessage(
+            page,
+            `请只使用 terminal_run，在工作区根目录执行 printf 'first-run' > ${FIRST_FILE}。不要修改其他文件。`
+          )
+          const firstEvidence = await waitForWorkbenchAiRunCompletion(
+            runtime,
+            firstStartedAtMs,
+            setup.env.modelName,
+            WORKBENCH_LIVE_AI_E2E_TIMEOUT_MS,
+            { page, blockerPolicy: 'auto_approve' }
+          )
+          recordWorkbenchAiEvidence(runtime, firstEvidence, 0.8)
+          const firstRun = getRunStarted(firstEvidence.events)
+          const firstEntry = await getApplyEntryForFile(page, FIRST_FILE)
+          const firstCardText = ((await firstEntry.textContent()) || '').trim()
+          await firstEntry.click()
+          const firstChangedFileList = await readChangedFileList(page)
+
+          await createWorkbenchNewSession(page)
+          await selectWorkbenchWorkspace(page, setup.workspaceName)
+          const secondStartedAtMs = Date.now()
+          await sendWorkbenchMessage(
+            page,
+            `请只使用 terminal_run，在工作区根目录执行 printf 'second-run' > ${SECOND_FILE}。不要修改其他文件。`
+          )
+          const secondEvidence = await waitForWorkbenchAiRunCompletion(
+            runtime,
+            secondStartedAtMs,
+            setup.env.modelName,
+            WORKBENCH_LIVE_AI_E2E_TIMEOUT_MS,
+            { page, blockerPolicy: 'auto_approve' }
+          )
+          recordWorkbenchAiEvidence(runtime, secondEvidence, 0.8)
+          const secondRun = getRunStarted(secondEvidence.events)
+          const secondEntry = await getApplyEntryForFile(page, SECOND_FILE)
+          const secondCardText = ((await secondEntry.textContent()) || '').trim()
+          await secondEntry.click()
+          const secondChangedFileList = await readChangedFileList(page)
+
+          await getWorkbenchLocator(
+            page,
+            `${workbenchSelectors.sessionItemPrefix}-${firstRun.sessionId}`
+          ).click()
+          const restoredFirstEntry = await getApplyEntryForFile(page, FIRST_FILE)
+          await restoredFirstEntry.click()
+          const restoredFirstChangedFileList = await readChangedFileList(page)
+
+          return {
+            firstRunId: firstRun.runId,
+            secondRunId: secondRun.runId,
+            firstSessionId: firstRun.sessionId,
+            secondSessionId: secondRun.sessionId,
+            firstCardText,
+            secondCardText,
+            firstChangedFileList,
+            secondChangedFileList,
+            restoredFirstChangedFileList,
+            firstTerminalCalls: firstEvidence.events.filter(
+              (event) => event.type === 'tool_call_started' && event.toolName === 'terminal_run'
+            ).length,
+            secondTerminalCalls: secondEvidence.events.filter(
+              (event) => event.type === 'tool_call_started' && event.toolName === 'terminal_run'
+            ).length
+          }
+        },
+        assert: async (result, runtime) => {
+          const firstListIsIsolated =
+            result.firstChangedFileList.includes('本次运行改动 1') &&
+            result.firstChangedFileList.includes(FIRST_FILE) &&
+            !result.firstChangedFileList.includes(SECOND_FILE)
+          const secondListIsIsolated =
+            result.secondChangedFileList.includes('本次运行改动 1') &&
+            result.secondChangedFileList.includes(SECOND_FILE) &&
+            !result.secondChangedFileList.includes(FIRST_FILE)
+          const restoredFirstListIsIsolated =
+            result.restoredFirstChangedFileList.includes('本次运行改动 1') &&
+            result.restoredFirstChangedFileList.includes(FIRST_FILE) &&
+            !result.restoredFirstChangedFileList.includes(SECOND_FILE)
+
+          await runtime.assertThat(
+            result.firstRunId !== result.secondRunId &&
+              result.firstSessionId !== result.secondSessionId &&
+              result.firstCardText.includes(FIRST_FILE) &&
+              !result.firstCardText.includes(SECOND_FILE) &&
+              result.secondCardText.includes(SECOND_FILE) &&
+              !result.secondCardText.includes(FIRST_FILE) &&
+              firstListIsIsolated &&
+              secondListIsIsolated &&
+              restoredFirstListIsIsolated &&
+              result.firstTerminalCalls >= 1 &&
+              result.secondTerminalCalls >= 1,
+            {
+              name: 'workbench-file-change-run-scope',
+              message:
+                '每条 assistant 修改入口和右侧本次运行改动必须只显示自身 sessionId 与 runId 的文件记录。',
+              data: result
+            }
+          )
+          await assertNoWorkbenchFakeSignals(runtime)
+        }
+      })
+    },
+    WORKBENCH_LIVE_AI_E2E_TIMEOUT_MS * 2
+  )
+})

tests/harness/backend/realAiDepsFactory.ts

@@ -17,7 +17,7 @@ import { TransformersEmbeddingExecutor } from '../../../src/main/services/vector
 import { ensureEmbeddingReady } from '../../../src/main/services/vector/ensureEmbeddingReady'
 import { WorkspaceKnowledgePrepareService } from '../../../src/main/services/workspace/WorkspaceKnowledgePrepareService'
 import { WorkspaceSourceService } from '../../../src/main/services/workspace/WorkspaceSourceService'
-import { MODELS_PATH } from '../../../src/main/utils/paths'
+import { getModelsPath } from '../../../src/main/utils/paths'
 import type { AiRuntimeSettings } from '../../../src/shared/types'
 import type { BackendHarnessMode, CreateAiRunContextInput, LiveLlmConfig } from './test-types'
 import type { EventCollector } from '../core/eventCollector'
@@ -113,7 +113,7 @@ export const createRealAiDepsFactory = async (input: {
 
   const embeddingService = new EmbeddingService({
     executor: new TransformersEmbeddingExecutor({
-      modelsPath: MODELS_PATH,
+      modelsPath: getModelsPath(),
       modelName: 'jinaai/jina-embeddings-v2-base-zh'
     })
   })