test(e2e): Agent 行为全栈测试体系（serve 37 + hub 拓扑 4 场景）+ 入 CI 门禁

.github/workflows/ci.yml

@@ -56,6 +56,25 @@ jobs:
       - name: Rustfmt check
         run: bazel build //... --config=rustfmt
 
+  agent-e2e:
+    name: Agent E2E
+    runs-on: macos-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Bazel
+        uses: bazel-contrib/setup-bazel@0.14.0
+        with:
+          bazelisk-cache: true
+          repository-cache: true
+
+      - name: Agent behavior E2E (serve + hub topology, mock LLM wire)
+        run: >-
+          bazel test
+          //:cli_llm_e2e_test
+          //:hub_llm_e2e_test
+          --test_output=errors
+
   desktop-e2e:
     name: Desktop E2E
     runs-on: macos-latest

apps/desktop/e2e/real/provider/host-provider-adapter-failures.spec.ts

@@ -12,11 +12,11 @@ test('OpenAI retries a transport closed before response headers', async () => {
     const conversation = page.getByTestId('conversation')
     await ready(page)
     await send(page, 'Recover OpenAI before headers')
-    await expect(conversation).toContainText('Retrying in 2.0s', { timeout: 10_000 })
+    await expect(conversation).toContainText(/Retrying in \d+\.\ds/, { timeout: 10_000 })
     await expect(conversation).toContainText(
       'OpenAI recovered after its pre-header transport failure.', { timeout: 20_000 },
     )
-    await expect(conversation).not.toContainText('Retrying in 2.0s')
+    await expect(conversation).not.toContainText('Retrying in')
     expect((await desktop.llm!.requests()).map((request) => request.protocol))
       .toEqual(['openai_responses', 'openai_responses'])
     await expect.poll(() => desktop.llm!.state()).toMatchObject({

apps/desktop/e2e/real/provider/host-provider-recovery.spec.ts

@@ -10,11 +10,11 @@ test('renders retry, fatal, recovery, and retry-exhaustion HTTP states', async (
     await ready(page)
 
     await send(page, 'Recover one server error')
-    await expect(conversation).toContainText('Retrying in 2.0s', { timeout: 10_000 })
+    await expect(conversation).toContainText(/Retrying in \d+\.\ds/, { timeout: 10_000 })
     await expect(conversation).toContainText('Recovered after one HTTP 503.', {
       timeout: 20_000,
     })
-    await expect(conversation).not.toContainText('Retrying in 2.0s')
+    await expect(conversation).not.toContainText('Retrying in')
     await ready(page)
 
     await send(page, 'Exercise fatal authentication')

apps/desktop/e2e/real/provider/host-provider-runtime.spec.ts

@@ -12,12 +12,12 @@ test('retries a real Anthropic HTTP 429 and clears the retry state', async () =>
     await ready(page)
     await send(page, 'Recover from a scripted rate limit')
     await expect(page.getByTestId('conversation')).toContainText(
-      'Retrying in 2.0s', { timeout: 10_000 },
+      /Retrying in \d+\.\ds/, { timeout: 10_000 },
     )
     await expect(page.getByTestId('conversation')).toContainText(
       'Recovered through the real provider retry loop.', { timeout: 15_000 },
     )
-    await expect(page.getByTestId('conversation')).not.toContainText('Retrying in 2.0s')
+    await expect(page.getByTestId('conversation')).not.toContainText('Retrying in')
     expect(await desktop.llm!.requests()).toHaveLength(2)
     await expect.poll(() => desktop.llm!.state()).toMatchObject({
       served: 2, remaining: 0, verified: true,
@@ -84,7 +84,7 @@ test('retries a connection closed before HTTP headers', async () => {
     const conversation = desktop.page.getByTestId('conversation')
     await ready(desktop.page)
     await send(desktop.page, 'Recover from a pre-header disconnect')
-    await expect(conversation).toContainText('Retrying in 2.0s', { timeout: 10_000 })
+    await expect(conversation).toContainText(/Retrying in \d+\.\ds/, { timeout: 10_000 })
     await expect(conversation).toContainText(
       'Recovered after the transport disappeared before headers.', { timeout: 15_000 },
     )
@@ -103,13 +103,13 @@ test('interrupts a retry wait and accepts the next turn immediately', async () =
     const conversation = page.getByTestId('conversation')
     await ready(page)
     await send(page, 'Wait on a long provider retry')
-    await expect(conversation).toContainText('Retrying in 30.0s', { timeout: 10_000 })
+    await expect(conversation).toContainText(/Retrying in \d{2}\.\ds/, { timeout: 10_000 })
     await page.getByRole('button', { name: 'Settings' }).click()
     await selectSettingsSection(page, 'agent')
     await page.getByRole('group', { name: 'Agent controls' })
       .getByRole('button', { name: 'Interrupt' }).click()
     await page.getByRole('button', { name: 'Close settings' }).click()
-    await expect(conversation).not.toContainText('Retrying in 30.0s', { timeout: 10_000 })
+    await expect(conversation).not.toContainText('Retrying in', { timeout: 10_000 })
     await expect(page.getByLabel('Message Loopal')).toBeEnabled({ timeout: 10_000 })
     await send(page, 'Recover after interrupting retry wait')
     await expect(conversation).toContainText(

apps/desktop/src/workbench/browser/workbench.context.test.tsx

@@ -46,7 +46,10 @@ describe('Workbench session catalog', () => {
 
     fireEvent.click(list.getByText('Write the guide'))
     await waitFor(() => expect(openSession).toHaveBeenCalledWith(docsSession.id))
-    expect(screen.getByTestId('active-session-title')).toHaveTextContent('Write the guide')
+    // The header re-renders after the openSession promise resolves — waiting
+    // on the mock call alone races that commit.
+    await waitFor(() => expect(screen.getByTestId('active-session-title'))
+      .toHaveTextContent('Write the guide'))
     expect(screen.getByText('Conversation for Write the guide')).toBeInTheDocument()
   })
 

apps/desktop/src/workbench/contrib/agents/browser/workbench.child-agent.test.tsx

@@ -1,10 +1,22 @@
-import { act, fireEvent, render, screen, waitFor } from '@testing-library/react'
+import { act, fireEvent, render, screen, waitFor, within } from '@testing-library/react'
 import { richView } from '../../../../../test/fixtures/workbench/rich-session'
 import {
   createTestAPI, sessionDetail, sessionOne, updatedAt,
 } from '../../../../../test/support/workbench/api-stub'
 import { Workbench } from '../../../browser/workbench'
 
+// Clicking an already-selected panel tab COLLAPSES the panel (toggle
+// semantics), so a blind click races the zone's initial expansion state and
+// can hide the tree it meant to open. Converge instead: each retry expands
+// the panel if it is collapsed, then selects the agent node.
+async function selectAgent(root: HTMLElement, name: RegExp): Promise<void> {
+  await waitFor(() => {
+    const tab = within(root).getByRole('tab', { name: 'Agents' })
+    if (tab.getAttribute('aria-expanded') !== 'true') fireEvent.click(tab)
+    fireEvent.click(within(root).getByRole('treeitem', { name }))
+  }, { timeout: 15_000 })
+}
+
 describe('Workbench child agent selection', () => {
   it('routes messages and resources to a live child while retaining completed output', async () => {
     const detail = {
@@ -26,14 +38,15 @@ describe('Workbench child agent selection', () => {
     }
     const sendMessage = vi.fn(async () => undefined)
     const { api, events } = createTestAPI({ openSession: async () => detail, sendMessage })
-    render(<Workbench api={api} />)
+    const { container } = render(<Workbench api={api} />)
     await screen.findByText(`Conversation for ${sessionOne.title}`)
     fireEvent.change(screen.getByLabelText('Message Loopal'), {
       target: { value: 'Root-only draft' },
     })
-    fireEvent.click(screen.getByRole('tab', { name: 'Agents' }))
-    fireEvent.click(screen.getByRole('treeitem', { name: /Research child/ }))
-    expect(screen.getByText('Child context compacted.')).toBeInTheDocument()
+    await selectAgent(container, /Research child/)
+    expect(
+      await screen.findByText('Child context compacted.', {}, { timeout: 15_000 }),
+    ).toBeInTheDocument()
     expect(screen.getByText('Child-only goal')).toBeInTheDocument()
     const composer = screen.getByLabelText('Message Research child')
     expect(composer).toHaveValue('')
@@ -66,13 +79,12 @@ describe('Workbench child agent selection', () => {
       }],
     }
     const { api } = createTestAPI({ openSession: async () => detail })
-    render(<Workbench api={api} />)
+    const { container } = render(<Workbench api={api} />)
     await screen.findByText(`Conversation for ${sessionOne.title}`)
-    fireEvent.click(screen.getByRole('tab', { name: 'Agents' }))
-    fireEvent.click(screen.getByRole('treeitem', { name: /Starting child/ }))
-    expect(screen.getByTestId('conversation')).toHaveTextContent(
+    await selectAgent(container, /Starting child/)
+    await waitFor(() => expect(within(container).getByTestId('conversation')).toHaveTextContent(
       'Viewing Starting child · starting · waiting for registration',
-    )
+    ), { timeout: 15_000 })
   })
 })
 

apps/desktop/src/workbench/contrib/settings/browser/loopal-default-settings.test.tsx

@@ -63,7 +63,10 @@ describe('LoopalDefaultSettings', () => {
     expect(restartButton).toBeEnabled()
     fireEvent.click(restartButton)
     await waitFor(() => expect(restart).toHaveBeenCalledWith('session'))
-    expect(screen.getByRole('status')).toHaveTextContent('restarted with the saved')
+    // The status line updates only after the restart promise resolves and
+    // commits — waiting on the mock call alone races that commit.
+    await waitFor(() => expect(screen.getByRole('status'))
+      .toHaveTextContent('restarted with the saved'))
   })
 
   it('surfaces load, validation, save, and restart failures without leaking credentials', async () => {

apps/desktop/src/workbench/contrib/settings/browser/metahub-settings-section.test.tsx

@@ -82,6 +82,9 @@ describe('MetaHubSettingsSection', () => {
     })
     const view = render(<MetaHubSettingsSection api={api} target={target} />)
     const panel = await screen.findByTestId('metahub-settings')
+    // The form is disabled until the persisted settings replace the
+    // placeholder defaults; the loaded hub name marks that moment.
+    await within(panel).findByDisplayValue('desktop-ui')
     fireEvent.change(within(panel).getByLabelText('MetaHub address'), {
       target: { value: 'meta.internal:9000' },
     })

apps/desktop/src/workbench/contrib/settings/browser/metahub-settings-section.tsx

@@ -25,6 +25,11 @@ export function MetaHubSettingsSection(props: {
 }): React.JSX.Element | null {
   const { t } = useI18n()
   const [settings, setSettings] = useState(emptySettings)
+  // reason: until the persisted settings arrive, `settings` holds placeholder
+  // defaults — editing or saving in that window would overwrite the user's
+  // stored values (and join under the placeholder hub name), so the form and
+  // every settings-writing action stay disabled while !loaded.
+  const [loaded, setLoaded] = useState(false)
   const [token, setToken] = useState('')
   const [runtime, setRuntime] = useState(props.initialState)
   const [local, setLocal] = useState<LocalMetaHubStatus>({ state: 'stopped' })
@@ -34,7 +39,7 @@ export function MetaHubSettingsSection(props: {
     let active = true
     const acceptError = (value: unknown): void => { if (active) setError(message(value)) }
     void props.api.getMetaHubSettings().then((next) => {
-      if (active) setSettings(next)
+      if (active) { setSettings(next); setLoaded(true) }
     }).catch(acceptError)
     void props.api.getLocalMetaHubStatus().then((next) => {
       if (active) setLocal(next)
@@ -106,39 +111,42 @@ export function MetaHubSettingsSection(props: {
       <div className="metahub-fields">
         <Field label={t('settings.metahub.address')}>
           <input aria-label={t('settings.metahub.addressAria')} value={settings.address}
+            disabled={!loaded}
             onChange={(event) => setSettings({ ...settings, address: event.target.value })} />
         </Field>
         <Field label={t('settings.metahub.hubName')}>
           <input aria-label={t('settings.metahub.hubNameAria')} value={settings.hubName}
+            disabled={!loaded}
             onChange={(event) => setSettings({ ...settings, hubName: event.target.value })} />
         </Field>
         <Field label={t('settings.metahub.token')}>
           <input aria-label={t('settings.metahub.tokenAria')} type="password" value={token}
-            autoComplete="off" placeholder={t(settings.tokenConfigured
+            autoComplete="off" disabled={!loaded} placeholder={t(settings.tokenConfigured
               ? 'settings.metahub.tokenConfigured'
               : 'settings.metahub.tokenRequired')}
             onChange={(event) => setToken(event.target.value)} />
         </Field>
       </div>
       <label className="settings-check"><input aria-label={t('settings.metahub.joinOnStartAria')}
-        type="checkbox" checked={settings.joinOnStart}
+        type="checkbox" checked={settings.joinOnStart} disabled={!loaded}
         onChange={(event) => setSettings({ ...settings, joinOnStart: event.target.checked })} />
         <span>{t('settings.metahub.joinOnStart')}</span></label>
       <label className="settings-check"><input aria-label={t('settings.metahub.startLocalAria')}
-        type="checkbox" checked={settings.startLocalOnLaunch}
+        type="checkbox" checked={settings.startLocalOnLaunch} disabled={!loaded}
         onChange={(event) => setSettings({ ...settings, startLocalOnLaunch: event.target.checked })} />
         <span>{t('settings.metahub.startLocal')}</span></label>
       <div className="metahub-actions">
-        <button disabled={busy} onClick={() => void execute(async () => { await save() })}>
+        <button disabled={busy || !loaded} onClick={() => void execute(async () => { await save() })}>
           {t('common.save')}
         </button>
-        <button disabled={busy || !settings.tokenConfigured} onClick={() => void execute(async () => {
-          const next = UpdateMetaHubSettingsInputSchema.parse({
-            ...settings, clearToken: true,
-          })
-          setSettings(await props.api.updateMetaHubSettings(next)); setToken('')
-        })}>{t('settings.metahub.clearToken')}</button>
-        <button disabled={busy || !props.target} onClick={join}>
+        <button disabled={busy || !loaded || !settings.tokenConfigured}
+          onClick={() => void execute(async () => {
+            const next = UpdateMetaHubSettingsInputSchema.parse({
+              ...settings, clearToken: true,
+            })
+            setSettings(await props.api.updateMetaHubSettings(next)); setToken('')
+          })}>{t('settings.metahub.clearToken')}</button>
+        <button disabled={busy || !loaded || !props.target} onClick={join}>
           {t('settings.metahub.join')}
         </button>
         <button disabled={busy || !props.target || runtime?.state === 'disconnected'}

build_defs/rust/root_tests.bzl

@@ -1,6 +1,6 @@
 """Rust tests owned by the root Bazel package."""
 
-load("@rules_rust//rust:defs.bzl", "rust_test")
+load("@rules_rust//rust:defs.bzl", "rust_binary", "rust_test")
 load("//build_defs/rust:desktop_test.bzl", "desktop_serve_test")
 
 def _binary_e2e(name, src, deps):
@@ -63,6 +63,69 @@ def loopal_root_tests():
             "@crates//:tokio",
         ],
     )
+    rust_test(
+        name = "cli_llm_e2e_test",
+        srcs = native.glob(["tests/e2e/cli_llm/*.rs"]),
+        crate_root = "tests/e2e/cli_llm/suite.rs",
+        data = [":loopal"],
+        edition = "2024",
+        env = {"LOOPAL_BINARY": "$(rootpath :loopal)"},
+        local = True,
+        # reason: excluded from `//...` wildcards (three-OS CI matrix) and run
+        # by the dedicated Agent E2E gate job, mirroring the desktop e2e setup.
+        tags = [
+            "e2e",
+            "manual",
+        ],
+        deps = [
+            "//crates/loopal-ipc",
+            "//crates/loopal-mock-llm:loopal-mock-llm-lib",
+            "//crates/loopal-protocol",
+            "@crates//:chrono",
+            "@crates//:reqwest",
+            "@crates//:serde_json",
+            "@crates//:tempfile",
+            "@crates//:tokio",
+            "@crates//:uuid",
+        ],
+    )
+    rust_binary(
+        name = "mock_mcp_server",
+        srcs = ["tests/fixtures/mock_mcp_server/main.rs"],
+        edition = "2024",
+        deps = ["@crates//:serde_json"],
+    )
+    rust_test(
+        name = "hub_llm_e2e_test",
+        srcs = native.glob(["tests/e2e/hub_llm/*.rs"]),
+        crate_root = "tests/e2e/hub_llm/suite.rs",
+        data = [
+            ":loopal",
+            ":mock_mcp_server",
+        ],
+        edition = "2024",
+        env = {
+            "LOOPAL_BINARY": "$(rootpath :loopal)",
+            "LOOPAL_MOCK_MCP_BINARY": "$(rootpath :mock_mcp_server)",
+        },
+        local = True,
+        tags = [
+            "e2e",
+            "manual",
+        ],
+        deps = [
+            "//crates/loopal-ipc",
+            "//crates/loopal-mock-llm:loopal-mock-llm-lib",
+            "//crates/loopal-protocol",
+            "//crates/loopal-vault-age",
+            "//crates/loopal-vault-api",
+            "@crates//:reqwest",
+            "@crates//:secrecy",
+            "@crates//:serde_json",
+            "@crates//:tempfile",
+            "@crates//:tokio",
+        ],
+    )
     desktop_serve_test()
     rust_test(
         name = "loopal-unit-test",

crates/loopal-backend/src/lib.rs

@@ -27,6 +27,6 @@ pub use log_writer::{
 };
 pub use process_group::{KillOutcome, SpawnedChild, kill_process_group};
 pub use tmp_cleanup::{
-    cleanup_orphans, cleanup_orphans_in, cleanup_session_tmp, cleanup_session_tmp_in,
-    loopal_tmp_root, session_tmp_root, session_tmp_root_in,
+    ORPHAN_MIN_AGE, cleanup_orphans, cleanup_orphans_in, cleanup_session_tmp,
+    cleanup_session_tmp_in, loopal_tmp_root, session_tmp_root, session_tmp_root_in,
 };