PR #35394 real environment verification, 2026-11-18 UTC main: b648395fa5 (github/main), /home/user/main PR: cf84ba22cff05dad06b0b698cf5d7689961f3ad4, /home/user/t3code Installed both with vp i. One dev server at a time, isolated empty .t3 homes. Main ports: web 8693, server 16733. PR ports: web 5733, server 13773. Browser: repo Playwright 1.64.1, /usr/bin/google-chrome, 1440x1000. Project, thread, messages, model selection and watches created through client. Real Codex model: gpt-6.1-sol. Initial reply on both: codex provider path verified. $ gh run rerun 43677422364 --repo pingdotgg/t3code $ gh api -X POST repos/pingdotgg/t3code/actions/runs/43677422364/cancel Actual Agent context budget jobs, all at the unchanged PR head: A 134694310726 cancelled 09:08:15 B 134695243123 cancelled 09:10:03 C 134700140692 cancelled 09:19:21 Each URL: https://github.com/pingdotgg/t3code/actions/runs/43677422364/job/ MAIN Initial cancellation A delivered a checks-failed notice and a real Codex response. Rewatching after the new job C forced a fresh host read. Client Details opened the actual C URL (134700140692). At 09:24:48, watch active, failedChecks=[Agent context budget]. The thread still had only the earlier A notice and reply; C did not wake Codex. Read-only state inspection: {"observedAt":"2026-11-18T09:24:48.425Z","active":true,"failedChecks":["Agent context budget"],"reportedFailures":null} PR HEAD Initial baseline consumed B without a notice, persisting B's immutable key. Restarted only the task-owned PR server after C existed, forcing a fresh read. 09:25: checks-failed C delivered through the running server and rendered in web. Codex resumed and replied checks failed. No fixes or code edits by the provider. Consumed history: ["[\"cf84ba22cff05dad06b0b698cf5d7689961f3ad4\",\"https://github.com/pingdotgg/t3code/actions/runs/43677422364/job/134695243123\",\"cancelled\"]","[\"cf84ba22cff05dad06b0b698cf5d7689961f3ad4\",\"https://github.com/pingdotgg/t3code/actions/runs/43677422364/job/134700140692\",\"cancelled\"]"] Rewatched at 09:26:52 through the client. Fresh read completed by 09:27:22. No second C notice or provider response; both immutable keys retained. Read-only state inspection: {"observedAt":"2026-11-18T09:27:22.012Z","active":true,"reportedFailures":["[\"cf84ba22cff05dad06b0b698cf5d7689961f3ad4\",\"https://github.com/pingdotgg/t3code/actions/runs/43677422364/job/134695243123\",\"cancelled\"]","[\"cf84ba22cff05dad06b0b698cf5d7689961f3ad4\",\"https://github.com/pingdotgg/t3code/actions/runs/43677422364/job/134700140692\",\"cancelled\"]"],"checksPending":false} Stopped watch through the UI afterward. $ gh run rerun 43677422364 --repo pingdotgg/t3code Restoration job 134704537992: success. Final GitHub checks: 23 SUCCESS, 9 SKIPPED, no failures or pending checks. Blacksmith remote checks: vp test run apps/server/src/orchestration-v2/pullRequestWatch.test.ts 48 passed. vp test run apps/server/src/orchestration-v2/runtimeLayer.test.ts -t 'starts, records, and stops a pull request watch|watching a dismissed stack layer again' 2 passed, 174 not selected. Limits: exact failed-A/pending/same-failed-A host-read fluctuation and URL-decoration replay were not induced live; focused policy tests cover them. Real same-name new cancelled attempts, baseline suppression, persisted restart history, stop/rewatch and provider delivery were exercised. Native clients, remote/tunnel connections and new-head transitions were not exercised. No source edits, commits, pushes, labels, merge, close, rebase, or force push.