[itrycodinganditsbad]

The low findings left from the re-review of #6.

What changed

• Old tray on upgrade (L3): craft-cloud --close tells the running window and tray to quit, waits for them to release their locks, and prints which ones were running. It checks both the new folder and the old loose file names.

  • install.sh runs it before copying and starts the new tray at the end.
  • If the install fails, the EXIT trap starts the installed tray again.

• Restart waiter (L5): where the user has systemd, the waiter runs under systemd-run --user with a fixed unit name. It survives logind's KillUserProcesses, and a second install that gives up finds the first waiter and doesn't add another. Without systemd it still runs under nohup.

• cmd remover fallback (N1–N3):

  • tasklist writes to a temp file first, so a tasklist that can't run means "wait", not "the setup has ended".
  • find and ping are called by full path.
  • The removal is retried every second for 30 seconds.

• Uninstall failing after the stop (N5): if the agent's program is still there, its Run values are put back and, if it was running, it is started again. The error message says so.

• Fix round from this PR's review:

  • Restart waiter, no agent unit (M1): the waiter only waits. It starts the agent outside itself (in a unit of its own, else detached) and ends, so a later install that gives up adds a waiter again.
  • Agent's folder variables (M2): CRAFT_CLOUD_AGENT_DIR and the XDG folder variables that are set reach the units through --setenv.
  • Reused pid (L6): the wait checks the process name as well as the pid, so a pid reused by another program can't hold the waiter.
  • Uninstall, held file (M3): the program folder's files are moved aside (Craft Cloud.removing-<pid>) before anything is deleted. The whole folder moves in one step, or file by file with roll-back when the setup runs from inside it. If a file held open blocks the move, nothing is touched, the Run values are put back and the agent is started again. On Windows a running exe can be moved; what can't be deleted yet goes once the setup has ended.
  • Put-back (L2): it needs all three exes in place, and says when a Run value couldn't be restored.
  • cmd remover (L1, L4): only find decides, so a failing ping no longer counts as "ended". The test now checks that each round waited.
  • Tray after upgrade (L5): the tray is restarted only where the session bus is reachable; elsewhere the script says how to get it back. Ctrl-C, TERM and HUP run the EXIT trap.

• Second review round:

  • Roll-back: a roll-back that can't move every file back names the ones left aside.
  • Leftovers: the next setup run finishes any Craft Cloud.removing-<pid> folder left behind (its setup no longer running). An install moves back what the folder lacks and then installs over it; an uninstall deletes it.
  • Deleting the moved-aside files: this can no longer fail the uninstall; whatever is left over is cleared by the next run.
  • cmd removers: each has its own temp file.
  • Waiter inside its own unit: if it can't create the agent's unit, it reports that to the journal instead of starting a nohup process the unit would kill.

• After the real-machine install:

  • ssh … ./install.sh never returned. Under bash (Arch's sh), a background a && b & list keeps the script's streams open while b runs. Every background start now goes through detach: stdin, stdout and stderr on /dev/null, and setsid where it's available. A test runs install.sh under both sh and bash with its output as a pipe.
  • Start at login pointed at another copy. Every install now points both Run values at the installed copy. That covers a tray value left by a dev build that ran once. A value that isn't there stays away.

Testing

• Linux: cargo xtask ci passes.

• New install.sh tests:

  • closing the old tray and starting the new one;
  • a second install that gives up adds no second waiter;
  • the nohup fallback without systemd-run.

  Stand-ins for systemd-run and systemctl keep the tests from touching the machine's real systemd.

• Windows: clippy --all-targets -D warnings and the tests for the desktop, setup and xtask crates pass. The cmd remover now also has a test with a tasklist that can't run: the folder stays.

• New unit tests cover --close (a tray in the new folder and an old-version window at the old names) and the uninstall that fails after the stop.

Left

• The cmd remover's temp file stays in TEMP if it gives up after an hour. If TEMP isn't writable, the remover waits the hour and removes nothing.

• close_running_tells_the_window_and_the_tray_new_and_old takes about 6 s: its fake listeners never release their locks.

• N4 (needs a decision): clipboard-win ships no licence file, so LICENSES.txt gives it the BSL-1.0 text from error-code, which has the same author and licence. That is a deliberate exception to "no file means plain MIT/Apache or fail". Please accept it or rule otherwise.

• Without systemd, the nohup waiter still dies with the session under KillUserProcesses. Systems without systemd don't have that setting.

• A final uninstall message left open for over an hour still leaves the program folder behind (L7).

🤖 Generated with Claude Code

https://claude.ai/code/session_01RC1oqfy9kZznj94RjMojyz

[itrycodinganditsbad]

Tried on the two test machines:

• Windows, at d61f2e7: the setup took 2.3 s from a WSL pipe, over a running install and a dev-build window. The window was closed, and the agent started again and was in sync. The tray value still pointed at the dev build then, which led to 49e1220.
• Arch/Hyprland, at 49e1220: ssh … ./install.sh returns in 2.8 s. The old window and tray were closed, the new tray started, and the agent restarted and is in sync. At d61f2e7 the same ssh never returned (fixed in e362101).

Not tried: uninstalling from Apps & features.