diff --git a/.ai/todo.md b/.ai/todo.md
index 960883acf..b5835fc66 100644
--- a/.ai/todo.md
+++ b/.ai/todo.md
@@ -1,7 +1,16 @@
-# PR #16 main conflict resolution
+# Codex plan approval backend fix
 
-- [x] Merge main 2e23c71d, retain managed Claude installer and upstream settings behavior.
-- [x] Resolve icon migration and removed smoke-test import.
-- [x] Full check:all: 2,876 frontend tests, 1,339 Rust tests passed (one ignored); jean:dev rebuilt.
-- [x] Agent-browser smoke check: Settings → Claude shows Install latest with PATH selected.
-- [x] Pushed merge 342c6083 to open PR #16; GitHub reports MERGEABLE.
+- [x] Add focused regression test for active-chat YOLO approval backend.
+- [x] Pass active backend through persisted and streaming Build/YOLO approval sends.
+- [x] Run targeted tests and frontend gates; `check:all` reaches Rust then stops because local Cargo is unavailable (covered by Docker build).
+- [x] Build `jean:dev` and smoke-test the built UI with agent-browser at 1440×1000.
+- [x] Review diff, commit, push, and create/update PR.
+
+## Review
+
+- Regression test failed before the fix because the YOLO payload omitted `backend`, then passed with `backend: codex`.
+- Active-chat persisted and streaming Build/YOLO paths now pass the backend captured with the approved session.
+- Fresh review caught and fixed an async ref race; the regression now switches the active backend while persistence is pending.
+- Typecheck, lint, targeted tests, and all 2,877 frontend tests pass (full suite used a 10s timeout after two unrelated tests exceeded 5s under load; both also passed in isolation).
+- `make docker-build` produced `jean:dev`; agent-browser confirmed the built UI renders cleanly at 1440×1000.
+- `bun run check:all` cannot run local Rust commands because Cargo is unavailable; the Docker build compiled the Rust release binary successfully.
diff --git a/src/components/chat/hooks/useMessageHandlers.test.tsx b/src/components/chat/hooks/useMessageHandlers.test.tsx
new file mode 100644
index 000000000..57671240b
--- /dev/null
+++ b/src/components/chat/hooks/useMessageHandlers.test.tsx
@@ -0,0 +1,111 @@
+import type { PropsWithChildren, RefObject } from 'react'
+import { act, renderHook, waitFor } from '@testing-library/react'
+import { QueryClient, QueryClientProvider } from '@tanstack/react-query'
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+import type { EffortLevel, ExecutionMode, ThinkingLevel } from '@/types/chat'
+import type { CliBackend } from '@/types/preferences'
+import { useChatStore } from '@/store/chat-store'
+import { useMessageHandlers } from './useMessageHandlers'
+
+const mocks = vi.hoisted(() => ({
+  invoke: vi.fn(() => Promise.resolve()),
+  listen: vi.fn(),
+  markPlanApproved: vi.fn(() => Promise.resolve()),
+}))
+
+vi.mock('@/lib/transport', () => ({
+  invoke: mocks.invoke,
+  listen: mocks.listen,
+}))
+
+vi.mock('@/services/chat', async () => {
+  const actual = await vi.importActual('@/services/chat')
+  return {
+    ...(actual as object),
+    markPlanApproved: mocks.markPlanApproved,
+  }
+})
+
+const ref = <T,>(current: T): RefObject<T> => ({ current })
+
+function renderMessageHandlers() {
+  const queryClient = new QueryClient({
+    defaultOptions: { queries: { retry: false }, mutations: { retry: false } },
+  })
+  const sendMessage = { mutate: vi.fn() }
+  const selectedBackendRef = ref('codex' as CliBackend)
+  const wrapper = ({ children }: PropsWithChildren) => (
+    <QueryClientProvider client={queryClient}>{children}</QueryClientProvider>
+  )
+  const hook = renderHook(
+    () =>
+      useMessageHandlers({
+        activeSessionIdRef: ref('session-1'),
+        activeWorktreeIdRef: ref('worktree-1'),
+        activeWorktreePathRef: ref('/repo/worktree'),
+        selectedModelRef: ref('gpt-6.1-sol'),
+        buildModelRef: ref(null),
+        buildBackendRef: ref(null),
+        buildThinkingLevelRef: ref(null),
+        buildEffortLevelRef: ref(null),
+        yoloModelRef: ref(null),
+        yoloBackendRef: ref(null),
+        yoloThinkingLevelRef: ref(null),
+        yoloEffortLevelRef: ref(null),
+        selectedBackendRef,
+        getCustomProfileName: () => undefined,
+        executionModeRef: ref('plan' as ExecutionMode),
+        selectedThinkingLevelRef: ref('off' as ThinkingLevel),
+        selectedEffortLevelRef: ref('medium' as EffortLevel),
+        useAdaptiveThinkingRef: ref(true),
+        getMcpConfig: () => undefined,
+        sendMessage,
+        createSession: { mutateAsync: vi.fn() },
+        queryClient,
+        scrollToBottom: vi.fn(),
+        markAtBottom: vi.fn(),
+        inputRef: ref(null),
+        pendingPlanMessage: null,
+        projectIdRef: ref(null),
+      }),
+    { wrapper }
+  )
+  return { hook, selectedBackendRef, sendMessage }
+}
+
+describe('useMessageHandlers plan approval', () => {
+  beforeEach(() => {
+    vi.clearAllMocks()
+    useChatStore.setState({
+      sendingSessionIds: {},
+      executionModes: {},
+      selectedModels: {},
+      lastSentMessages: {},
+      errors: {},
+    })
+  })
+
+  it('preserves the Codex backend when approving a plan in yolo mode', async () => {
+    let finishPersistence!: () => void
+    mocks.markPlanApproved.mockReturnValueOnce(
+      new Promise<void>(resolve => {
+        finishPersistence = resolve
+      })
+    )
+    const { hook, selectedBackendRef, sendMessage } = renderMessageHandlers()
+
+    act(() => hook.result.current.handlePlanApprovalYolo('plan-message-1'))
+    selectedBackendRef.current = 'claude'
+    finishPersistence()
+
+    await waitFor(() => expect(sendMessage.mutate).toHaveBeenCalled())
+    expect(sendMessage.mutate).toHaveBeenCalledWith(
+      expect.objectContaining({
+        model: 'gpt-6.1-sol',
+        backend: 'codex',
+        executionMode: 'yolo',
+      }),
+      expect.any(Object)
+    )
+  })
+})
diff --git a/src/components/chat/hooks/useMessageHandlers.ts b/src/components/chat/hooks/useMessageHandlers.ts
index 520e336f1..0d1576c74 100644
--- a/src/components/chat/hooks/useMessageHandlers.ts
+++ b/src/components/chat/hooks/useMessageHandlers.ts
@@ -745,6 +745,7 @@ export function useMessageHandlers({
               worktreePath,
               message,
               model: buildModel,
+              backend: sessionBackend,
               executionMode: 'build',
               thinkingLevel: buildThinking,
               effortLevel: useAdaptiveThinkingRef.current
@@ -917,6 +918,7 @@ export function useMessageHandlers({
               worktreePath,
               message,
               model: yoloModel,
+              backend: sessionBackendYolo,
               executionMode: 'yolo',
               thinkingLevel: yoloThinking,
               effortLevel: useAdaptiveThinkingRef.current
@@ -1031,6 +1033,7 @@ export function useMessageHandlers({
         worktreePath,
         message: buildApprovalMsg,
         model: streamBuildModel,
+        backend: streamBuildSessionBackend,
         executionMode: 'build',
         thinkingLevel: streamBuildThinking,
         effortLevel: useAdaptiveThinkingRef.current
@@ -1131,6 +1134,7 @@ export function useMessageHandlers({
         worktreePath,
         message: yoloApprovalMsg,
         model: streamYoloModel,
+        backend: streamYoloSessionBackend,
         executionMode: 'yolo',
         thinkingLevel: streamYoloThinking,
         effortLevel: useAdaptiveThinkingRef.current