Skip to content

The Windows logon trigger, fired

The last thing install.ps1 did that nobody had watched. The Scheduled Task it registers has always been started by hand; whether Windows would start it at logon was a prediction. It is now an observation, and the run showed more than it was set up to show.

What was run

A real install at the default location: %APPDATA%\laneward\.env, %LOCALAPPDATA%\laneward\app, LANEWARD_AGENT=claude. A logon-check lane was registered and left pending, the hub was stopped, and neither task was started by hand. Then the operator logged off and back on.

The result

11:36:45  laneward-conductor-hub   Running   LastTaskResult 0x41301
11:36:45  laneward-conductor       Running   LastTaskResult 0x41301
          three bun processes alive, nothing started by a human

0x41301 is "the task is currently running", not an error code.

The Podman machine does not survive a logout, so the database was down when both tasks started. Neither of them died of it. The hub answered HTTP 500 rather than refusing the connection, which is the distinction that matters: it was up and serving, it simply had nothing to read from. The conductor logged its failed passes and kept looping, which is exactly what Stage 5a's error tolerance was written for.

Starting the Podman machine and the database container was the only thing done afterwards, and nothing else was touched:

+10s   logon-check: running
+20s   logon-check: completed

The hub reconnected on its own, the conductor picked the lane up on its next pass, and a real agent drove it to completion. notes.txt carried the requested line, git status showed the one file the lane owned, the evidence row was recorded, and the Git guard log held seven refusals, all of them reads and none a mutation.

So one cycle answered three questions rather than one: the trigger fires, both services tolerate a database that is not there yet, and they recover without help when it arrives.

The defect it found first

The cycle was set up twice. The first attempt proved nothing, because no task was registered at the time: the S3 run's task had been removed and the real install had not yet been done. That is a process mistake rather than a defect.

The defect came out of preparing the second attempt. install.ps1 registered only the conductor at logon, never the hub. On Linux install.sh installs two units and the conductor's declares Requires=laneward.service, so a boot brings the API up with it. Registering one task here looked equivalent and was not: after a logon the loop would have run against nothing, logged a connection failure every interval, and dispatched no lane. What the operator would have seen is a task reporting Running and a dashboard that never loads.

install.ps1 now registers laneward-conductor-hub alongside it. Two tasks rather than one wrapper, because a Scheduled Task has no ordering guarantee either; the conductor already tolerates an absent hub by design (D-038), which is what makes starting them independently safe, and this run is the evidence for that tolerance rather than the assumption of it.

What this does not establish

  • A reboot. A logoff and logon is not a cold boot. The Podman machine does not survive either, so a reboot has the same shape, but it has not been run.
  • Linux. loginctl enable-linger is still untested, and it is a persistent host change that needs its own approval.
  • The database surviving with it. Nothing here starts the Podman machine at logon. That is left to the operator on purpose: it is a machine-wide service, not Laneward's to register.