fix(web-shell): scope artifact actions to owning workspace

.qwen/e2e-tests/2026-08-04-secondary-artifact-workspace-ownership.md

@@ -0,0 +1,117 @@
+# Secondary artifact workspace ownership
+
+Issue: https://github.com/QwenLM/qwen-code/issues/8494
+
+## Baseline
+
+- Global CLI: `qwen 0.21.3`
+- `qwen serve --help` confirms repeated `--workspace` registration and Web
+  Shell serving are available.
+- Current component behavior is captured failure-first: an artifact tab with
+  no owner uses root workspace actions, and scheduled-task detail omits the
+  secondary `workspaceId`.
+
+## Local setup
+
+1. Create isolated primary and secondary temporary workspaces.
+2. Put the same relative sentinel filename in both workspaces with different
+   contents.
+3. Create distinct durable scheduled-task fixtures for both runtimes.
+4. Start the built `dist/cli.js serve` on loopback with both `--workspace`
+   flags and the production Web Shell bundle.
+
+## Scenarios
+
+### Secondary file action
+
+Open a secondary-session artifact and download/preview its sentinel.
+
+Expected:
+
+- The request targets `/workspaces/<secondary>/file...`.
+- The returned bytes are the secondary sentinel, never the primary sentinel.
+
+### Secondary scheduled task
+
+Open the secondary durable task, toggle it, edit it, and delete the test copy.
+
+Expected:
+
+- Every request targets `/workspaces/<secondary>/scheduled-tasks...`.
+- The primary task and primary task file are unchanged.
+
+### Ownership loss
+
+Keep an artifact tab open, then remove or mark the secondary workspace
+unavailable in the capability fixture before a delayed read resolves.
+
+Expected:
+
+- The panel displays the localized workspace-unavailable state.
+- The delayed response is ignored.
+- No request is retried against the primary route.
+
+## Evidence
+
+- Focused test output for resolver, turn outputs, artifact panel, nested
+  subagent, and app tab propagation.
+- Production build/typecheck/lint output.
+- Captured two-workspace HTTP route log and sentinel hashes.
+- Browser screenshot of the secondary artifact panel and the fail-closed
+  ownership-loss state when the local browser harness can represent it.
+
+## Results (2026-08-04)
+
+Result: PASS
+
+### Automated verification
+
+- Failure-first run: 3 expected failures and 29 existing passes. The primary
+  sentinel leaked into an ownerless tab, the secondary qualified client was
+  unused, and scheduled-task list omitted `workspaceId`.
+- Focused ownership suite: 413/413 tests passed across the six affected test
+  files.
+- Full Web Shell suite: 167 files and 2,777 tests passed.
+- Web Shell lint and TypeScript typecheck passed.
+- Package Web Shell build and repository root production build passed.
+- The changed files pass Prettier. The package-wide format check still reports
+  five unchanged baseline files:
+  `BranchPickerPopover.module.css`, `GitModePopover.module.css`,
+  `GitDialog.module.css`, `PlanExecutionView.module.css`, and `index.html`.
+
+### Production two-workspace verification
+
+The built CLI served the copied production bundle with two trusted workspaces.
+`GET /capabilities` advertised distinct primary and secondary runtime IDs, and
+the served JavaScript contained the new stale-owner guard.
+
+The same relative file was read through both routes:
+
+- `GET /file?path=artifact-owner.txt` returned
+  `PRIMARY_WORKSPACE_SENTINEL_8494`, SHA-256
+  `818d5f4f1fb9c7e3f9bdfd9a3ad39361c93a23ca3f3d0ba5e4a7c889b33b9127`.
+- `GET /workspaces/<secondary-id>/file?path=artifact-owner.txt` returned
+  `SECONDARY_WORKSPACE_SENTINEL_8494`, SHA-256
+  `0e9fac7909b16ff8b17014ea103ed5018c8b4e9ad21d2f95e128fef3a3544a12`.
+- The secondary bytes route returned only the secondary sentinel bytes.
+
+Durable task CRUD was exercised against isolated primary and secondary task
+fixtures:
+
+- Secondary list, update, and delete requests all used
+  `/workspaces/<secondary-id>/scheduled-tasks...`.
+- The secondary update changed only the secondary name/enabled state.
+- The primary task remained present, enabled, and unchanged after that update.
+- Deleting the secondary task emptied only the secondary list; the primary
+  task was still present. Both test tasks were then removed.
+
+The daemon request log independently recorded the qualified file, bytes,
+scheduled-task POST/GET/PATCH/DELETE routes and their successful statuses.
+
+### Visual evidence
+
+The in-app browser runtime reported no available browser instance, so a UI
+screenshot could not be captured in this environment. No synthetic screenshot
+was substituted. The production server returned the current Web Shell bundle
+with HTTP 200, and the DOM behavior is covered by the focused and full suites
+above.

docs/design/web-shell-artifact-workspace-ownership.md

@@ -0,0 +1,101 @@
+# Web Shell artifact workspace ownership
+
+Status: implemented and verified
+
+Issue: https://github.com/QwenLM/qwen-code/issues/8494
+
+## Problem
+
+The Web Shell has one app-level `DaemonWorkspaceProvider`. Its default
+workspace actions target the primary workspace. Session providers can attach to
+secondary workspaces, but artifact surfaces currently retain or fall back to
+the app-level actions. A file preview, download, review, nested subagent
+artifact, or scheduled-task mutation can therefore reach the primary runtime
+even though the producing session belongs to a secondary workspace.
+
+The unsafe fallback also survives ownership changes: action objects are stored
+in right-panel tabs, so removing or distrusting a workspace does not invalidate
+an already-open tab.
+
+## Ownership contract
+
+The producing session owns every turn output. Its `workspaceCwd`, resolved by
+the session connection, is the source of the owner claim. The current daemon
+capabilities are the authority that accepts or rejects that claim.
+
+A target is usable only when one of these cases holds:
+
+1. Exactly one advertised workspace has the same cwd and is trusted.
+2. For a legacy single-workspace daemon with no workspace list, the cwd exactly
+   matches `capabilities.workspaceCwd`.
+
+Unknown, duplicate, untrusted, removed, or identity-mismatched targets fail
+closed. They never fall back to the primary workspace.
+
+## Design
+
+### Resolve at use time
+
+Turn-output requests and right-panel tabs carry immutable owner identity
+(`workspaceCwd` and the advertised `workspaceId`) instead of long-lived action
+objects. `ArtifactPanel` resolves that identity against current capabilities on
+every render. This makes workspace removal, trust loss, and runtime replacement
+invalidate open tabs immediately.
+
+The resolver returns a small artifact action surface only:
+
+- `readWorkspaceFile`
+- `readFileBytes`
+- `stat`
+- scheduled-task list/update/delete operations
+
+Primary targets reuse the provider's primary actions. Trusted secondary
+targets use `client.workspaceByCwd(cwd)` for file operations. Scheduled-task
+operations still use the Web UI REST actions, but always receive the resolved
+workspace id explicitly.
+
+### Propagation
+
+- `TurnOutputs` derives the owner from its session `workspaceCwd`, uses the
+  scoped file actions for direct downloads, and stamps owner identity onto
+  review, artifact, and scheduled-task open requests.
+- `ChatPane` and `SubagentDetail` preserve the request identity and add only the
+  producing session id. They no longer stamp root workspace actions onto the
+  request or artifact snapshot.
+- `App` stores owner identity on each right-panel tab. Pane artifact snapshots
+  contain artifacts only; they no longer retain workspace clients.
+- `ArtifactPanel` renders the existing workspace-unavailable state when the
+  owner cannot be resolved or no longer matches. It never substitutes
+  `useWorkspaceActions()` for a missing owner.
+- Durable scheduled-task snapshots carry `workspaceId`, and every list,
+  update, toggle, and delete request passes that id.
+
+### Async invalidation
+
+Scoped action wrappers check that their captured owner is still current before
+starting an operation and again after each response. Effects and scheduled-task
+mutations also ignore results after unmount or owner replacement. A response
+started for one runtime cannot populate a tab after that runtime is removed or
+replaced.
+
+## Non-goals
+
+- No daemon route or workspace-registry semantics change.
+- No primary fallback for compatibility on multi-workspace daemons.
+- No redesign of artifact storage, session attachment, or scheduled-task data.
+- No attempt to make arbitrary app-level workspace actions session-scoped;
+  only the artifact surfaces in issue #8494 are changed.
+
+## Verification
+
+1. Failure-first component tests pin the current primary fallback and missing
+   scheduled-task workspace id.
+2. Resolver tests cover primary, trusted secondary, unknown, duplicate,
+   untrusted, removed, and runtime-replaced targets.
+3. Turn-output tests verify secondary downloads and open requests use the
+   secondary owner.
+4. Artifact-panel tests verify missing/stale ownership fails closed and all
+   durable scheduled-task operations include `workspaceId`.
+5. A production bundle is run against two registered local workspaces with
+   distinct sentinel files/tasks. Captured requests must use the secondary
+   workspace-qualified routes and leave the primary sentinels unchanged.

packages/web-shell/client/App.test.tsx

@@ -909,11 +909,7 @@ vi.doMock('./components/SplitView', async () => {
       onExit?: () => void;
       sessionIds?: string[];
       onPanesChange?: (ids: string[]) => void;
-      onPaneArtifactsChange?: (
-        sessionId: string,
-        artifacts: unknown[],
-        workspaceActions: unknown,
-      ) => void;
+      onPaneArtifactsChange?: (sessionId: string, artifacts: unknown[]) => void;
       onRightPanelOpen?: (request: unknown) => void;
       onOpenMonitor?: (
         task: DaemonSessionMonitorTaskStatus,
@@ -926,9 +922,6 @@ vi.doMock('./components/SplitView', async () => {
       }) => unknown;
       voiceWorkspaces?: readonly unknown[];
     }) => {
-      const paneActions = {
-        readWorkspaceFile: vi.fn().mockResolvedValue('<p>pane</p>'),
-      };
       const artifact = {
         id: 'pane-artifact',
         kind: 'report',
@@ -981,11 +974,7 @@ vi.doMock('./components/SplitView', async () => {
             'data-testid': 'split-report-artifact',
             type: 'button',
             onClick: () =>
-              props.onPaneArtifactsChange?.(
-                'pane-session',
-                [artifact],
-                paneActions,
-              ),
+              props.onPaneArtifactsChange?.('pane-session', [artifact]),
           },
           'artifact',
         ),
@@ -995,11 +984,7 @@ vi.doMock('./components/SplitView', async () => {
             'data-testid': 'split-report-updated-artifact',
             type: 'button',
             onClick: () =>
-              props.onPaneArtifactsChange?.(
-                'pane-session',
-                [updatedArtifact],
-                paneActions,
-              ),
+              props.onPaneArtifactsChange?.('pane-session', [updatedArtifact]),
           },
           'updated artifact',
         ),
@@ -1009,11 +994,7 @@ vi.doMock('./components/SplitView', async () => {
             'data-testid': 'split-report-changed-artifact',
             type: 'button',
             onClick: () =>
-              props.onPaneArtifactsChange?.(
-                'pane-session',
-                [changedArtifact],
-                paneActions,
-              ),
+              props.onPaneArtifactsChange?.('pane-session', [changedArtifact]),
           },
           'changed artifact',
         ),
@@ -1022,8 +1003,7 @@ vi.doMock('./components/SplitView', async () => {
           {
             'data-testid': 'split-clear-artifacts',
             type: 'button',
-            onClick: () =>
-              props.onPaneArtifactsChange?.('pane-session', [], paneActions),
+            onClick: () => props.onPaneArtifactsChange?.('pane-session', []),
           },
           'clear artifacts',
         ),
@@ -1040,7 +1020,9 @@ vi.doMock('./components/SplitView', async () => {
                 turnId: 'turn-1',
                 artifactId: artifact.id,
                 artifact,
-                workspaceActions: paneActions,
+                workspaceCwd: '/tmp/project',
+                workspaceId: 'primary',
+                sourceSessionId: 'pane-session',
                 previewContent: '<p>stale</p>',
               }),
           },
@@ -4554,6 +4536,17 @@ describe('App session callbacks', () => {
   });
 
   it('opens the latest reviewable turn from the empty right panel', () => {
+    mockWorkspace.capabilities = {
+      workspaceCwd: '/tmp/project',
+      workspaces: [
+        {
+          id: 'primary',
+          cwd: '/tmp/project',
+          primary: true,
+          trusted: true,
+        },
+      ],
+    } as typeof mockWorkspace.capabilities;
     testState.messages = [
       {
         id: 'user-1',
@@ -9898,6 +9891,17 @@ describe('App session callbacks', () => {
   });
 
   it('reconciles split pane artifact snapshots in the right panel', async () => {
+    mockWorkspace.capabilities = {
+      workspaceCwd: '/tmp/project',
+      workspaces: [
+        {
+          id: 'primary',
+          cwd: '/tmp/project',
+          primary: true,
+          trusted: true,
+        },
+      ],
+    } as typeof mockWorkspace.capabilities;
     const { container } = renderApp();
     await flush();