The operator reported that automatic resume was not working during the supervised TheWell pilot implementation. Gitea candidate4167 succeeded at exact dev merge 1fa1a135b538c7c6a115d4d75591edf0f491407a, but watch63 stayed resume_waiting. Watch62 had already started turn 01a0f39e-67fb-7103-9b0f-60b868e9ac85, yet remained continuation_queued, blocking the next event.
The timer's service used PrivateTmp=true. Codex's profile socket is a symlink into /tmp/codex-daemon-0/; its target was absent inside the service namespace. Local dry-run succeeded while the same invocation with the exact unit isolation raised FileNotFoundError at socket.connect. This was a service mount issue, not a failed CI run or competing writer.
ProfileBackups PR25 integrates reviewed head 9139081f75205320ddfcc198085eef6fc08d9fe7 into configured main as f5f610f1c0e6156c4f59871d05191cdefbd09146. Only the service unit and application README changed. The service now bind-mounts the resolved root and ax_julian profile control sockets read-only using optional paths, retaining private temporary storage, NoNewPrivileges, ProtectSystem, daemon process/peer checks and existing watch/Goal rules. No daemon, CLI, writer lock, permission, model, spending or automation stop was changed. No historical event was replayed.
Validation: systemd-analyze verify and git diff --check passed. The exact isolated dry-run succeeds with both narrow mounts. Reviewed unit bytes were atomically installed at /etc/systemd/system/codex-pipeline-monitor.service, preserving original owner/group/mode, and systemd reloaded. At 19:16:09 UTC the actual service reconciled watch62 to delivered through exact client-ID/persisted-text history evidence, retained queue and turn IDs, and released watch63. Its resume service durably queued one matching event with client pipeline-watch-63, queue 01a0f3be-e5da-7522-8472-0ba602d78f5f, no error. Supported thread/queue/list independently confirms exactly one matching receipt. The current active CLI must finish its turn before consuming that event; automatic consumption and subsequent delivered reconciliation remain pending.
Next acceptance: finish the owning turn, then verify watch63 starts exactly one automatic continuation and the timer records its matching consumed turn. Continue the original supervised TheWell task from this terminal evidence: verify candidate promotion/deployment, publish/verify the matching full worker image, arm the scoped pilot and exercise one normally selected case. No first TheWell case has yet been selected; general automation stays paused and TheWell production protected.
Rollback: restore host/pipeline-monitor/codex-pipeline-monitor.service from pre-repair main 09ae702e118a5f90456e37106b8abe1a5962aa00 with existing ownership/mode and reload systemd. Preserve all watch receipts/history and uncertain states; do not replay events. Prior cumulative task runtime is incompletely recorded and is not treated as zero; this host repair took about15 active minutes, excluding earlier scheduled waiting.
The original CLI automatically consumed watch63 and started turn 01a0f3c0-c9bf-7f61-852e-8d77d6d85df8. The existing scheduled timer reconciled it to delivered, retaining queue 01a0f3be-e5da-7522-8472-0ba602d78f5f with no error. Supported history independently returns exactly one matching client-ID/persisted-text user message at that turn, and the shared queue has no matching receipt. The original task resumed in the same conversation without human login, writer replacement or manual event injection. Resume repair acceptance is complete; the original TheWell pilot delivery remains separate and unfinished.