diff --git a/mobile/app/h/[hostId]/session/[worktreeId].tsx b/mobile/app/h/[hostId]/session/[worktreeId].tsx
index 76b0055d1..dee14cdd3 100644
--- a/mobile/app/h/[hostId]/session/[worktreeId].tsx
+++ b/mobile/app/h/[hostId]/session/[worktreeId].tsx
@@ -46,6 +46,7 @@ import {
 } from 'lucide-react-native'
 import type { RpcClient } from '../../../../src/transport/rpc-client'
 import { loadHosts } from '../../../../src/transport/host-store'
+import { startRuntimeCapabilityProbe } from '../../../../src/transport/runtime-capability-probe'
 import {
   loadTerminalAutocompleteEnabled,
   loadTerminalLinkOpenMode,
@@ -250,7 +251,6 @@ import type {
   MobileSessionTabType,
   RenderableDiffLine,
   RuntimeRepoSummary,
-  RuntimeStatusResult,
   SessionTabsResult,
   Terminal,
   TerminalCreateResult,
@@ -2317,40 +2317,23 @@ export default function SessionScreen() {
     }
     // Why: a client swap can keep the route connected while moving to an older
     // host; clear the prior capability before exposing host-specific actions.
+    setBrowserScreencastSupported(null)
+    setAgentSessionHistorySupported(null)
     setQuickCommandsSupported(null)
     setShowQuickCommands(false)
-    let stale = false
-    void client
-      .sendRequest('status.get')
-      .then((response) => {
-        if (stale || !response.ok) {
-          return
-        }
-        const status = (response as RpcSuccess).result as RuntimeStatusResult
-        setBrowserScreencastSupported(
-          status.capabilities?.includes('browser.screencast.v1') === true
-        )
-        setAgentSessionHistorySupported(
-          status.capabilities?.includes(MOBILE_AI_VAULT_CAPABILITY) === true
-        )
-        setQuickCommandsSupported(supportsMobileQuickCommands(status.capabilities))
-        // Why: hosts without this capability strip inputKind from terminal.send,
-        // so a forwarded xterm reply would become floor-stealing shell input.
-        hostQueryReplyInputSupportedRef.current =
-          status.capabilities?.includes(TERMINAL_QUERY_REPLY_INPUT_RUNTIME_CAPABILITY) === true
-      })
-      .catch(() => {
-        if (!stale) {
-          setBrowserScreencastSupported(false)
-          setAgentSessionHistorySupported(false)
-          setQuickCommandsSupported(false)
-          setShowQuickCommands(false)
-          hostQueryReplyInputSupportedRef.current = false
-        }
-      })
-    return () => {
-      stale = true
-    }
+    hostQueryReplyInputSupportedRef.current = false
+    // Why: the probe retries — a relay→direct cutover or request timeout rejects
+    // status.get without changing connState, which used to latch these hidden.
+    return startRuntimeCapabilityProbe(client, (capabilities) => {
+      setBrowserScreencastSupported(capabilities.includes('browser.screencast.v1'))
+      setAgentSessionHistorySupported(capabilities.includes(MOBILE_AI_VAULT_CAPABILITY))
+      setQuickCommandsSupported(supportsMobileQuickCommands(capabilities))
+      // Why: hosts without this capability strip inputKind from terminal.send,
+      // so a forwarded xterm reply would become floor-stealing shell input.
+      hostQueryReplyInputSupportedRef.current = capabilities.includes(
+        TERMINAL_QUERY_REPLY_INPUT_RUNTIME_CAPABILITY
+      )
+    })
   }, [client, connState])
 
   // Why: read deviceToken from host record so code can pass client.id on subscribe/send for driver-state-machine identity.
diff --git a/mobile/src/session/mobile-session-startup-source.test.ts b/mobile/src/session/mobile-session-startup-source.test.ts
index 58454c1d9..bad6bd330 100644
--- a/mobile/src/session/mobile-session-startup-source.test.ts
+++ b/mobile/src/session/mobile-session-startup-source.test.ts
@@ -48,6 +48,27 @@ describe('mobile session startup', () => {
     expect(startupEffect).toContain("showToast('Open Orca on the host to wake sleeping agents.'")
   })
 
+  it('fails runtime capability gates closed before probing a replacement client', () => {
+    const capabilityEffect = sliceBetween(
+      'const hostQueryReplyInputSupportedRef = useRef(false)',
+      '// Why: read deviceToken from host record'
+    )
+    const probeStart = capabilityEffect.indexOf('startRuntimeCapabilityProbe(client,')
+
+    expect(probeStart).toBeGreaterThanOrEqual(0)
+    for (const reset of [
+      'setBrowserScreencastSupported(null)',
+      'setAgentSessionHistorySupported(null)',
+      'setQuickCommandsSupported(null)',
+      'setShowQuickCommands(false)',
+      'hostQueryReplyInputSupportedRef.current = false'
+    ]) {
+      const resetIndex = capabilityEffect.lastIndexOf(reset)
+      expect(resetIndex).toBeGreaterThanOrEqual(0)
+      expect(resetIndex).toBeLessThan(probeStart)
+    }
+  })
+
   it('activates an already-selected pending terminal tab after hydration', () => {
     expect(source).toContain(
       'const pendingTerminalActivationAttemptRef = useRef<string | null>(null)'
diff --git a/mobile/src/tasks/worktree-create-capability.ts b/mobile/src/tasks/worktree-create-capability.ts
index 20500c10a..d9b9f75c0 100644
--- a/mobile/src/tasks/worktree-create-capability.ts
+++ b/mobile/src/tasks/worktree-create-capability.ts
@@ -1,6 +1,6 @@
 import { useCallback, useEffect, useRef, useState } from 'react'
 import type { RpcClient } from '../transport/rpc-client'
-import { LogicalClientCutoverError } from '../transport/stable-logical-rpc-client'
+import { isLogicalClientCutoverError } from '../transport/stable-logical-rpc-client'
 import type { RpcSuccess } from '../transport/types'
 import { MOBILE_TASKS_CAPABILITY } from './mobile-tasks-capability'
 
@@ -95,10 +95,3 @@ export function useNewWorktreeRuntimeCapabilities(
   )
   return { tasksSupported, getWorktreeCreateCutoverSupport }
 }
-
-function isLogicalClientCutoverError(error: unknown): boolean {
-  return (
-    error instanceof LogicalClientCutoverError ||
-    (error instanceof Error && error.message === 'RPC interrupted by connection migration')
-  )
-}
diff --git a/mobile/src/transport/runtime-capability-probe.test.ts b/mobile/src/transport/runtime-capability-probe.test.ts
new file mode 100644
index 000000000..cffc97f30
--- /dev/null
+++ b/mobile/src/transport/runtime-capability-probe.test.ts
@@ -0,0 +1,137 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
+import { startRuntimeCapabilityProbe } from './runtime-capability-probe'
+import { LogicalClientCutoverError } from './stable-logical-rpc-client'
+import type { RpcClient } from './rpc-client'
+import type { RpcResponse } from './types'
+
+type ProbeOutcome = RpcResponse | Error
+
+function makeClient(outcomes: ProbeOutcome[]): { client: RpcClient; calls: () => number } {
+  let calls = 0
+  const client = {
+    sendRequest: () => {
+      const outcome = outcomes[Math.min(calls, outcomes.length - 1)]
+      calls += 1
+      return outcome instanceof Error ? Promise.reject(outcome) : Promise.resolve(outcome)
+    }
+  } as unknown as RpcClient
+  return { client, calls: () => calls }
+}
+
+const ok = (capabilities: string[]): RpcResponse => ({
+  ok: true,
+  id: '1',
+  result: { capabilities },
+  _meta: { runtimeId: 'r1' }
+})
+
+async function flushMicrotasks(): Promise<void> {
+  await Promise.resolve()
+  await Promise.resolve()
+}
+
+describe('startRuntimeCapabilityProbe', () => {
+  beforeEach(() => {
+    vi.useFakeTimers()
+  })
+  afterEach(() => {
+    vi.useRealTimers()
+  })
+
+  it('delivers capabilities on first success', async () => {
+    const { client, calls } = makeClient([ok(['a.v1'])])
+    const seen: (readonly string[])[] = []
+    const cancel = startRuntimeCapabilityProbe(client, (capabilities) => seen.push(capabilities))
+    await flushMicrotasks()
+    expect(seen).toEqual([['a.v1']])
+    expect(calls()).toBe(1)
+    cancel()
+  })
+
+  it('retries promptly after a logical-client cutover rejection', async () => {
+    const { client, calls } = makeClient([new LogicalClientCutoverError(), ok(['a.v1'])])
+    const seen: (readonly string[])[] = []
+    const cancel = startRuntimeCapabilityProbe(client, (capabilities) => seen.push(capabilities))
+    await flushMicrotasks()
+    expect(seen).toEqual([])
+    await vi.advanceTimersByTimeAsync(250)
+    expect(seen).toEqual([['a.v1']])
+    expect(calls()).toBe(2)
+    cancel()
+  })
+
+  it('backs off on other failures and eventually recovers', async () => {
+    const { client, calls } = makeClient([
+      new Error('Request timed out: status.get'),
+      new Error('Request timed out: status.get'),
+      ok(['a.v1'])
+    ])
+    const seen: (readonly string[])[] = []
+    const cancel = startRuntimeCapabilityProbe(client, (capabilities) => seen.push(capabilities))
+    await flushMicrotasks()
+    await vi.advanceTimersByTimeAsync(1_000)
+    expect(seen).toEqual([])
+    await vi.advanceTimersByTimeAsync(2_000)
+    expect(seen).toEqual([['a.v1']])
+    expect(calls()).toBe(3)
+    cancel()
+  })
+
+  it('retries an ok:false response instead of settling', async () => {
+    const failure: RpcResponse = {
+      ok: false,
+      id: '1',
+      error: { code: 'internal', message: 'nope' },
+      _meta: { runtimeId: 'r1' }
+    }
+    const { client } = makeClient([failure, ok(['a.v1'])])
+    const seen: (readonly string[])[] = []
+    const cancel = startRuntimeCapabilityProbe(client, (capabilities) => seen.push(capabilities))
+    await flushMicrotasks()
+    expect(seen).toEqual([])
+    await vi.advanceTimersByTimeAsync(1_000)
+    expect(seen).toEqual([['a.v1']])
+    cancel()
+  })
+
+  it('caps the failure backoff', async () => {
+    const outcomes: ProbeOutcome[] = Array.from({ length: 10 }, () => new Error('timeout'))
+    outcomes.push(ok(['a.v1']))
+    const { client, calls } = makeClient(outcomes)
+    const seen: (readonly string[])[] = []
+    const cancel = startRuntimeCapabilityProbe(client, (capabilities) => seen.push(capabilities))
+    await flushMicrotasks()
+    // Why: 1s+2s+4s+8s then 15s cap; ten failures fit well inside 8 capped waits.
+    await vi.advanceTimersByTimeAsync(15_000 * 10)
+    expect(seen).toEqual([['a.v1']])
+    expect(calls()).toBe(11)
+    cancel()
+  })
+
+  it('stops retrying and dropping results once cancelled', async () => {
+    const { client, calls } = makeClient([new Error('timeout'), ok(['a.v1'])])
+    const seen: (readonly string[])[] = []
+    const cancel = startRuntimeCapabilityProbe(client, (capabilities) => seen.push(capabilities))
+    await flushMicrotasks()
+    cancel()
+    await vi.advanceTimersByTimeAsync(60_000)
+    expect(seen).toEqual([])
+    expect(calls()).toBe(1)
+  })
+
+  it('ignores a success that resolves after cancellation', async () => {
+    let resolveRequest: ((response: RpcResponse) => void) | null = null
+    const client = {
+      sendRequest: () =>
+        new Promise<RpcResponse>((resolve) => {
+          resolveRequest = resolve
+        })
+    } as unknown as RpcClient
+    const seen: (readonly string[])[] = []
+    const cancel = startRuntimeCapabilityProbe(client, (capabilities) => seen.push(capabilities))
+    cancel()
+    resolveRequest?.(ok(['a.v1']))
+    await flushMicrotasks()
+    expect(seen).toEqual([])
+  })
+})
diff --git a/mobile/src/transport/runtime-capability-probe.ts b/mobile/src/transport/runtime-capability-probe.ts
new file mode 100644
index 000000000..b0f4ad115
--- /dev/null
+++ b/mobile/src/transport/runtime-capability-probe.ts
@@ -0,0 +1,58 @@
+import type { RpcClient } from './rpc-client'
+import type { RpcSuccess } from './types'
+import { isLogicalClientCutoverError } from './stable-logical-rpc-client'
+
+// Why: a relay→direct cutover or request timeout can reject an in-flight
+// status.get without ever changing connState, so a one-shot probe would latch
+// capability-gated UI hidden until the screen remounts; retry until one lands.
+const CUTOVER_RETRY_DELAY_MS = 250
+const FAILURE_RETRY_BASE_DELAY_MS = 1_000
+const FAILURE_RETRY_MAX_DELAY_MS = 15_000
+
+export function startRuntimeCapabilityProbe(
+  client: RpcClient,
+  onCapabilities: (capabilities: readonly string[]) => void
+): () => void {
+  let cancelled = false
+  let retryTimer: ReturnType<typeof setTimeout> | null = null
+  let failureRetries = 0
+
+  function attempt(): void {
+    void client.sendRequest('status.get').then(
+      (response) => {
+        if (cancelled) {
+          return
+        }
+        if (!response.ok) {
+          scheduleRetry(false)
+          return
+        }
+        const status = (response as RpcSuccess).result as { capabilities?: string[] }
+        onCapabilities(status.capabilities ?? [])
+      },
+      (error: unknown) => {
+        if (cancelled) {
+          return
+        }
+        scheduleRetry(isLogicalClientCutoverError(error))
+      }
+    )
+  }
+
+  function scheduleRetry(cutover: boolean): void {
+    // Why: cutover means the replacement transport is already authenticated —
+    // re-ask promptly; other failures back off so a wedged host isn't hammered.
+    const delay = cutover
+      ? CUTOVER_RETRY_DELAY_MS
+      : Math.min(FAILURE_RETRY_BASE_DELAY_MS * 2 ** failureRetries++, FAILURE_RETRY_MAX_DELAY_MS)
+    retryTimer = setTimeout(attempt, delay)
+  }
+
+  attempt()
+  return () => {
+    cancelled = true
+    if (retryTimer) {
+      clearTimeout(retryTimer)
+    }
+  }
+}
diff --git a/mobile/src/transport/stable-logical-rpc-client.ts b/mobile/src/transport/stable-logical-rpc-client.ts
index 0931a6636..901263768 100644
--- a/mobile/src/transport/stable-logical-rpc-client.ts
+++ b/mobile/src/transport/stable-logical-rpc-client.ts
@@ -9,6 +9,14 @@ export class LogicalClientCutoverError extends Error {
   }
 }
 
+// Why: instanceof can miss across bundle copies, so also match by message.
+export function isLogicalClientCutoverError(error: unknown): boolean {
+  return (
+    error instanceof LogicalClientCutoverError ||
+    (error instanceof Error && error.message === 'RPC interrupted by connection migration')
+  )
+}
+
 type SubscriptionRecord = {
   method: string
   params: unknown