PR #24475 live runtime verification, 2026-10-28 UTC

Baseline: github/main 0deb3837090af3a6fc8b2247af57a3322573138d
PR head: 2fff50ced4982b840c7037d7758a3e38934a25f0

Both: vp i; setsid nohup vp run dev (bounded by timeout).
Main worktree: /home/user/main; web 8693; server 16733.
PR checkout: /home/user/t3code; web 5733; server 13773;
explicit task home /home/user/t3code/.t3. Only one dev server ran at once.

Playwright from repo dependencies + /usr/bin/google-chrome.
Created projects and parent threads in the web UI, selected Codex gpt-6.1-sol,
and submitted prompts requesting real delegated Codex gpt-6.1-sol children.
No mocks, replay providers, seeded fixtures, or database writes.
Children executed timeout-bounded shell commands and returned child-runtime-ok.

MAIN, captured from actual provider tool items + read-only run projection:
06:53:14.024 delegate_task response: status=running, workState=working
06:53:14.126 actual child run startedAt (102ms AFTER the response)
06:54:44.159 actual child run completedAt
06:54:49.863 task_status: completed, result_available, summary=child-runtime-ok
FAIL: a child whose provider had not started was already reported running.

PR, captured from actual provider tool items + read-only run projection:
07:15:35.352 delegate_task response: status=starting, workState=working
07:15:35.420 actual child run startedAt
07:17:07.037 actual child run completedAt
07:18:15.319 task_status: completed, result_available, summary=child-runtime-ok
PASS: live parent delegation distinguishes provider startup; completion wakes parent.

Additional authenticated HTTP MCP probe against the same real PR server and parent:
07:18:28.135 delegate_task: starting, working
  actualRun: starting, startedAt=null
07:18:28.154 task_status: starting, working
  actualRun: starting, startedAt=null
07:18:28.174 actual child run startedAt
07:18:29.665 task_status: running, working
  actualRun: running, startedAt=07:18:28.174
Final task_status: completed, result_available, summary=child-runtime-ok
PASS: task_status transitions starting -> running -> completed.
The diagnostic provider wrapper only tees stdin to the unchanged real Codex binary;
its captured MCP credential was used privately for these HTTP requests.
SQLite was opened read-only for observations. No run states were injected.

Checks, through blacksmith-check on this exact PR head:
vp run check:changed 72dde22b6f092491fac4ddc480bc609bb8092bfd
PASS format, lint, both knip passes, generated GraphQL contract checks,
typechecks for 30 packages, and 39 tests in 3 existing test files.
PR CI: 27 successful, 0 failing, 0 pending.

Not verified: a provider permanently blocked in thread/start; queued/preparing,
cancellation/error branches; native desktop/mobile clients; remote/relay/tunnel modes.
One real child recovered from provider retries and completed within two minutes.
No source changes, commits, pushes, labels, merges, or rebases were made.