PR 22291 real runtime verification, 2026-10-31, Linux
Main: 15fe283272 (github/main)
PR: 823993e69e41b198c37df694f5d6d5b68cdb5a28 (exact contributor head)
Client: dev web app, repository Playwright, /usr/bin/google-chrome
Provider: Codex gpt-6.1-sol through the configured local proxy
Project: /home/user/question-verification, initialized real git repository, added through UI
No mocks, replay provider, fixtures, direct database writes, or source changes.

Main worktree: git worktree add ../main github/main; timeout 600 vp i
Background server PID 23334, ports 16733/8693, stopped before PR started.
PR checkout: maria/takeover-pr-22291 at exact head; timeout 600 vp i
Background server PID 39662, ports 13773/5733, isolated .t3 home.

UI reproduction on both revisions:
Ask Codex to call request_user_input_async with question:
Which color should the sample use? Options Blue and Green.
Leave it open, run 510 separate pwd calls, then timeout 125 sleep 120.
Reload the active thread after the command rows accumulate.
Main: question panel absent. PR: question panel visible, Blue and Green intact.

Main read-only SQL at the missing-panel capture:
select status,count(*) from orchestration_v2_projection_runtime_requests
where thread_id='eb970881-ea1d-4482-a47d-0e5d97814c48' group by status;
pending|1
select type,count(*) from orchestration_v2_projection_turn_items
where thread_id='eb970881-ea1d-4482-a47d-0e5d97814c48' group by type;
assistant_message|1
command_execution|512
reasoning|1
user_input_request|1
user_message|1

PR read-only SQL at the retained-panel capture:
assistant_message|1
command_execution|512
user_input_request|1
user_message|1
UI exact question count: 1 after reload; 1 after Load earlier turns.

Repeated PR flow in a fresh browser context after 510 later pwd calls:
question visible; selecting Green auto-submitted; provider replied "you chose green."
Reload preserved that reply and no pending question panel.

Current PR runtime records (read-only queries follow):
849e6620-0209-428e-92cc-23a6a9f8d975|codex|gpt-6.1-sol|completed
849e6620-0209-428e-92cc-23a6a9f8d975|codex|gpt-6.1-sol|completed
bd6d2ceb-a4b3-4afd-a0f3-2976347f78c0|codex|gpt-6.1-sol|completed
bd6d2ceb-a4b3-4afd-a0f3-2976347f78c0|codex|gpt-6.1-sol|completed
849e6620-0209-428e-92cc-23a6a9f8d975|resolved|{"0":"Green"}
bd6d2ceb-a4b3-4afd-a0f3-2976347f78c0|resolved|{"0":"Green"}

Blacksmith checks:
vp test run apps/server/src/orchestration-v2/ProjectionStore.test.ts apps/server/src/orchestration-v2/threadHistoryPaging.test.ts
2 files, 62 tests passed.
vp run check:changed 056b4457cc687b053e15f1fc89afb900078a127e
Passed format, lint, both knip passes, server GraphQL generation check,
server and scripts typechecks, and the 62 focused tests.
The first aggregate attempt lacked base history in the remote shallow checkout;
fetching the PR history fixed setup, then the aggregate command passed.

Limits: native desktop/mobile, other providers, remote/relay/tunnel connections,
and approval-request aging were not exercised live. Existing focused tests cover
paging and resolved-request removal. Video omits startup and a reload loading screen.