PR https://git.cafe/pingdotgg/t3code/pulls/24763
Head 30ed8880d2336610a79e3987cc4e6760fd113f86
Baseline github/main 001d16e9699ad22863f065e5d9a3bd36305901fc
Linux VM; actual Chrome + repo Playwright; real npm-installed Codex; no mocks or fixtures.

SETUP
Added /home/user/provider-verification (real git repo) through Add project > Local folder.
Set provider Binary path through Settings > Providers to /home/user/pr24763-provider/bin/codex.
Installed Codex 0.164.0, refreshed provider status through the UI, then installed 0.165.0 externally without refreshing the app.
Clicked the existing stale Update available > Update now action.

BASELINE
Server 16733 / web 8693; /home/user/main/.t3; only baseline dev server running.
Real binary was already 0.165.0; client still displayed 0.164.0.
Process trace showed installer launch:
execve("/home/user/.nvm/versions/node/v24.22.1/bin/npm", ["npm", "install", "-g", "--prefix", "/home/user/pr24763-provider", "--allow-scripts=@openai/codex", "@openai/codex@latest"], 0xd2090d0 /* 73 vars */) = 0
Recording shows Running npm, Verifying the installed version, and v0.165.0.
Real gpt-6.1-sol turn returned: provider verification main ok.
Baseline stopped by its captured process-group PID before starting PR.

PR CURRENT
Server 13773 / web 5733; explicit isolated /home/user/t3code/.t3.
Same stale advisory / actual current binary scenario.
Installer execve count was 0 when checked immediately after this action (before the next outdated scenario).
Actual client WebSocket provider states:
[
  {
    "at": "2026-10-28T10:10:21.680Z",
    "version": "0.164.0",
    "advisory": "behind_latest",
    "updateState": {
      "status": "running",
      "startedAt": "2026-10-28T10:10:21.383Z",
      "finishedAt": null,
      "message": "Checking for the latest version",
      "output": null
    }
  },
  {
    "at": "2026-10-28T10:10:22.055Z",
    "version": "0.165.0",
    "updateState": {
      "status": "succeeded",
      "startedAt": "2026-10-28T10:10:21.383Z",
      "finishedAt": "2026-10-28T10:10:22.039Z",
      "message": "Already up to date.",
      "output": null
    }
  },
  {
    "at": "2026-10-28T10:10:22.318Z",
    "version": "0.165.0",
    "advisory": "current",
    "updateState": {
      "status": "succeeded",
      "startedAt": "2026-10-28T10:10:21.383Z",
      "finishedAt": "2026-10-28T10:10:22.039Z",
      "message": "Already up to date.",
      "output": null
    }
  }
]

PR OUTDATED NEIGHBOR
Reset real installation to 0.164.0; clicked Refresh provider status then Update now without externally upgrading.
Client received Running npm install -g --prefix /home/user/pr24763-provider --allow-scripts=@openai/codex @openai/codex@latest.
Same trace subsequently recorded a real successful npm exec for this action, and npm's log exited 0.
Final provider state after reloading the client:
[
  {
    "version": "0.165.0",
    "advisory": "current",
    "updateState": {
      "status": "succeeded",
      "startedAt": "2026-10-28T10:11:03.944Z",
      "finishedAt": "2026-10-28T10:11:12.033Z",
      "message": "Provider updated.",
      "output": "changed 2 packages in 7s"
    }
  }
]
Real gpt-6.1-sol turn returned: provider verification pr ok.

CHECKS
blacksmith-check run --id <task-owned-worker> "vp i && cd apps/server && ../../node_modules/.bin/vp test run src/provider/providerMaintenanceRunner.test.ts"
Test Files 1 passed (1); Tests 28 passed (28).

NOT VERIFIED
Other provider drivers; live unknown-version/probe-failure/advisory-failure/explicit-pin paths; Electron shell; native iOS/Android; macOS/Windows; relay/tunnel connections. Existing focused tests cover the fallback/pinned cases, not real-provider runtime checks.

SETUP INCIDENT
The first PR startup in the root checkout defaulted to ~/.t3. It was stopped when its log revealed that path, then restarted with the explicit isolated home above. No direct database edits were made.