PR #27836 real dev verification, 2026-10-31 UTC
main: 5fe18d79ac (github/main)
PR: e15c50df4c3df1f9178fd8c487f6760531e23206
No PR source edits, commits or pushes.

Environment: disposable Linux Boat VM, Chrome /usr/bin/google-chrome,
Playwright 1.60.0 from repository dependencies, real Codex gpt-6.1-sol
through the configured local proxy. Projects, threads, script creation and
provider turns were driven through the real web app. No database fixtures.
Only one dev server ran at a time, each background launch PID retained.
main server/web: 16733/8693. PR server/web: 13773/5733.

Scenario: a real Node script appends a started entry, waits, then appends a
completed entry. A transient systemd user service owns it so the real file
write can survive a provider/server crash on Linux. A running native Codex
subagent provides interrupted background work. The server process is killed
by its captured port-owner PID, then its retained dev process group is stopped
and vp run dev is restarted against the same isolated state.

MAIN, surviving-service case:
21:27:32.051Z deploy.mjs started pid 24500 (180-second delay).
21:27:47 approximately: server crash/restart, native subagent interrupted.
21:28:02.911Z outgoing Codex turn/start, model gpt-6.1-sol:
  input[0].text = "Note: the T3 server restarted, and this background work was
  cancelled before it finished. It will not report back on its own, so resume
  or restart what you still need and wait for its result:
  - subagent: /root/restart_wait (id 01a193dc-6487-7db3-b992-cca89fc459f3)

  Continue where you left off."
21:28:11.296Z real provider command, exit 0:
  timeout 15 bash -c 'systemctl --user show run-r7d3818fffe964d28a4fa318a83fa1f50.service -p ActiveState -p SubState -p MainPID -p Result; cat deploys.jsonl; ...'
  MainPID=24499, Result=success, ActiveState=active, SubState=running
21:30:32.145Z deploy.mjs completed pid 24500.
Service pid 24500: one started, one completed. No duplicate deployment.
The archived main log also has two started-only exploratory command attempts,
pids 22293 and 24039. Those ordinary provider commands died on Linux crashes;
they are excluded from the surviving-service count.
21:30:37.313Z baseline provider returned:
  no active upstream websocket for response.interrupt (HTTP 400)
The baseline's resumed native-subagent final result was therefore not verified.
Missing safety guidance was reproduced. A duplicate side effect was not.

PR, same surviving-service mechanism:
21:33:04.403Z started pid 32144 (60-second delay to shorten this pass).
21:33:21 approximately: server crash/restart while service and subagent run.
  systemctl --user show run-r30f3405958fa4bcf82176731b112693a.service -p ActiveState -p MainPID
  MainPID=32143, ActiveState=active
21:33:36.282Z outgoing Codex turn/start, model gpt-6.1-sol:
  input[0].text contains the cancelled-background-work note followed by:
  "Continue where you left off. Your last turn was cut short, so a command or
  other action it started may have finished or may still be running. Before
  you repeat one that changes state, confirm it did not take effect and is not
  still running; if you cannot tell, say so instead of repeating it."
21:33:40.989Z real provider inspection, exit 0:
  timeout 10 bash -c 'cat deploy.mjs; wc -l main-deploys.jsonl deploys.jsonl; cat deploys.jsonl; systemctl --user status run-r30f3405958fa4bcf82176731b112693a.service --no-pager; ...'
  deploys.jsonl: one entry, pid 32144 started.
  service: active (running), Main PID 32143, child pid 32144.
21:34:04.456Z completed pid 32144.
21:34:05.755Z real provider wait/count command, exit 0:
  timeout 45 bash -c 'while systemctl --user is-active --quiet run-r30f3405958fa4bcf82176731b112693a.service; do sleep 2; done; wc -l main-deploys.jsonl deploys.jsonl; cat deploys.jsonl; ...'
  deploys.jsonl: two entries, one started and one completed, both pid 32144.
21:34:33.715Z resumed native subagent's read-only sleep completed, exit 0.
The subagent then reported git status; parent completed and the client showed Done.
Exactly one service launch in the provider trace. No repeated deployment.

Neighbor checks on PR:
21:34:55.420Z outgoing turn/start input contains only the new user's read-only
count request, with no restart note. Provider replied: 1 started, 1 completed.
21:36:37.531Z plain Codex restart (no interrupted background subagent):
  outgoing turn/start model=gpt-6.1-sol, input=[] (native continuation).
21:37:15.743Z read-only command completed, exit 0, returning the same two rows.
Client rendered: started: 1 completed: 1; command exited with code 0.

Blacksmith checks on exact PR head:
  vp test run apps/server/src/orchestration-v2/RestartContinuation.test.ts
  PASS 48 tests.
  vp run check:changed bf270e74e4c6144fc65648186e879d8bfdb411ed
  PASS format, lint, both knip passes, graphql, typecheck (3 packages),
  tests (49 passed across 2 files).
GitCafe CI snapshot: 30 successful, 0 failing, 0 pending.

Limits: no native Windows orphan-process reproduction, macOS, Electron shell,
iOS/Android, other providers, external deploy/push service, or statistical
model-reliability evaluation. Linux systemd used a real surviving local action.
This proves warning delivery and the observed behavior, not guaranteed model compliance.