PR: https://git.cafe/pingdotgg/t3code/pulls/23748
GitHub: https://github.com/pingdotgg/t3code/pull/23748
Main: 0deb3837090af3a6fc8b2247af57a3322573138d
PR head: 95055375a01ae30f665be27cfae411f43e87386a

Runtime setup
- Installed dependencies with vp i in /home/user/main and /home/user/t3code.
- Ran one background vp run dev server at a time, with isolated .t3 state.
- Main: web 8693, server 16733. Stopped its captured process group before starting PR.
- PR: web 5733, server 13773.
- Created a real git project /home/user/pr23748-scenario through Add project in the app.
- Used repo Playwright 1.60.0 and /usr/bin/google-chrome.
- Configured SSH device host through Settings > Devices: user@127.0.0.1:2222.
  Port 2222 uses a real separate sshd listener, so the app takes SshDeviceHost rather than its localhost shortcut.
- Installed real Android emulator and Android 35 Google APIs image; created PR23748_Pixel using avdmanager.
- Observed remote execve calls by strace on this task's sshd, without replacing subprocesses or outputs.
- Waited for each device.list WebSocket Exit Success before counting commands.

Main reproduction: click Device > Refresh devices, three times
refresh 1: Success; platform probes=1; emulator -list-avds=1; SSH AVD present
refresh 2: Success; platform probes=1; emulator -list-avds=1; SSH AVD present
refresh 3: Success; platform probes=1; emulator -list-avds=1; SSH AVD present

Exact PR head: same app interaction, three times
refresh 1: Success; platform probes=0; emulator -list-avds=1; SSH AVD present
refresh 2: Success; platform probes=0; emulator -list-avds=1; SSH AVD present
refresh 3: Success; platform probes=0; emulator -list-avds=1; SSH AVD present
Local AVD discovery also returned PR23748_Pixel on PR.
These are command counts through the app, not a latency benchmark.

Nearby behavior through authenticated running-server WebSocket RPC
POST /api/auth/websocket-ticket/rpc; /ws?orchestrationProtocol=2
Request tag=device.list, payload={inspectOnly:true}:
  Success; 1 remote platform probe; Android available=true.
Temporarily removed real adb executables from SDK and system lookup paths, leaving SSH connected:
  inspectOnly: Success; 1 remote probe; Android available=false.
  listing 1: Success; 1 remote probe; 0 AVD commands; Android available=false.
  listing 2: Success; 1 remote probe; 0 AVD commands; Android available=false.
Restored the actual adb executables, without reconnecting SSH:
  listing 1: Success; 1 remote probe; 1 AVD command; Android available=true; SSH AVD found.
  listing 2: Success; 0 remote probes; 1 AVD command; SSH AVD found.
Counted npm --version as the bootstrap marker while adb was absent.
All executable paths restored after the check.

Real provider/client path
- Main: Codex gpt-6.1-sol returned MAIN_PROVIDER_OK in the web client.
- PR: Codex gpt-6.1-sol returned PR_PROVIDER_OK in the web client.
- Enabled Agent device access, then created a fresh app thread.
- In that fresh gpt-6.1-sol thread, requested device_list twice for the SSH host.
- Both actual MCP calls succeeded and returned PR23748_Pixel, booted=false, host ready.
- Final rendered reply: DEVICE_LIST_OK.
- The provider recovered from retries; the successful turn took 3m43s, below the 10-minute stall limit.
- An older session created before enabling access correctly retained its disabled device capability.

Checks, run remotely through blacksmith-check
cd apps/server && vp test run src/device/SshDeviceHost.test.ts
  PASS: 2 tests.
vp run check:changed 220eb73524ca7882d53b6a12c824ccbcd97ad279
  PASS: format, lint, both knip passes, graphql, server/scripts typecheck, focused tests.
  The first attempt lacked merge-base history in the shallow worker; fetching PR history resolved it.
GitCafe status on the unchanged head: 28 successful, 0 pending, 0 failing checks.

Not verified
- Native Electron desktop, native iOS/Android clients, macOS or Windows hosts.
- A physically separate remote machine, network RTT effects, relay/tunnel, or reconnect recovery.
- Device booting, streaming, screenshots, and interactive device control.
- SDK removal failure while a positive report remains cached, before explicit inspection.
- No measured latency improvement is claimed.
No repository source changes, commits, or pushes were needed.