Skip to content

D-014, demonstrated

The claim this project is built around, approved lanes continue after Claude exits, had never been observed on any platform. It has now. This is A2 from the remaining-work plan, run on 2026-08-15 against 65b8b9d plus the three fixes it forced.

What was run

Fedora Linux 44 under WSL, systemd as PID 1, systemctl --user. The database was the existing postgres:16-alpine on the Windows Podman machine, reachable from WSL at <podman-machine-ip>:5433, with a throwaway laneward_a2 database.

No packages were installed. The plan assumed A2 needed podman and notify-send in WSL. It needed neither: the database already existed and LANEWARD_NOTIFY= switches the notifier off deliberately. What actually blocked it was install.sh demanding podman unconditionally; see below.

CODEX_BIN pointed at tests/fixtures/fake-codex.ts. The worker was a fixture, and that is a real limit on this result. What D-014 claims is that the conductor keeps dispatching and recording after the operator walks away; the worker's own behaviour is a separate question.

To be exact about what that leaves open: a real Codex worker has driven a lane to a recorded verdict on this project, on 2026-08-08 (the neura-system pilot). What has never happened is a real Codex worker under a supervised service, which is a genuinely different environment: constructed PATH, no terminal, no login shell, and a deliberately stripped environment from buildWorkerEnv. Codex's own credential lookup has never been exercised there.

The result

19:59:20  lane a2-demo2 registered through the API
19:59:20  a2-demo2: started          (conductor journal)
19:59:21  a2-demo2: completed

Between those lines nothing was touched. The conductor had been running as laneward-conductor.service since 18:33 (86 minutes idle), picked the lane up on its own, dispatched the worker, ran the evidence check, and recorded the lane completed with its lane-check evidence attached (overall: not_configured, since the throwaway repository declares no checks).

systemctl --user status reported both units active, and the API answered GET /pending and GET /lanes over HTTP throughout.

The first lane failed, correctly

a2-demo was registered owning only work.txt. The fixture also writes .prompt, so check-evidence exited 1 with violation: .prompt and the lane failed. Nothing was wrong: an unattended worker touched a path it did not own and the ownership gate stopped it without a human present. The second lane declared .prompt and completed.

That is a better first result than a clean pass would have been.

Clean shutdown, under a real service manager

A third lane was left running and the unit stopped mid-lane:

20:00:07  SIGTERM received - stopping 1 running lane(s)
20:00:07  a2-slow: handed back to HUB
          unit state: inactive     unit Result: success
          lane a2-slow: pending

Result: success rather than failed is the SIGTERM exit-0 change from Stage 5a doing what it was written for. The lane returned to pending with no manual cleanup, which is what makes a reboot safe. This also closes the half of B2 that asked for a strand-and-recover round trip: on Linux the lane is handed back rather than stranded, so reset-stranded was not needed at all.

Three defects it found, none of which tests could

Every one of these passes a full test suite on both platforms.

1. install.sh demanded podman unconditionally. The quadlet exists to run Postgres locally; a DATABASE_URL pointing elsewhere means the operator already has one. The check refused an install that would have worked. It is now decided from the config: a local URL requires podman and installs the quadlet, a remote one installs neither and says so. laneward.service's Requires=laneward-db.service is a substituted placeholder for the same reason: a unit that requires a container nobody installed never starts.

2. Both units exited 127. A systemd user service does not inherit the login shell's PATH, and bun lives under $HOME. Addressing bun absolutely in ExecStart was not enough: run start executes a package script that invokes bare bun. The fix is Environment=PATH= including bun's directory rather than a rewritten ExecStart, because the lane checks spawn whatever a project declares (bun test among them) and need it on PATH too.

3. PORT and HUB_URL could disagree silently. Setting PORT=8799 moved the API and left the conductor defaulting to 8787. The only symptom was drain pass failed: Unable to connect every interval, forever. defaultConfig now derives the default hub address from PORT, with HUB_URL still the override.

Defect 3 also produced the first field evidence for Stage 5a's error tolerance: the conductor logged the failure every two seconds and kept looping instead of dying, which is exactly why a failed pass was made non-fatal.

Cleaned up

Both units stopped and removed, daemon-reload run, ~/.config/laneward, ~/.local/share/laneward and ~/.local/state/laneward deleted, the three lane worktrees deleted, and laneward_a2 dropped. systemctl --user list-unit-files lists no laneward units. Nothing was left running.

What this does and does not establish

Establishes: Laneward installs from install.sh, runs as two systemd user services, dispatches and completes a lane with no human attached, records its evidence, enforces path ownership unattended, and stops cleanly with its lanes handed back.

Does not establish: that a real Codex worker behaves under a service, as distinct from under a hand-run conductor, which it has; that any of this survives a reboot or a logout without loginctl enable-linger; that the quadlet database path works, since an external database was used; or anything at all about Windows, where the Scheduled Task has been registered but never run.