Linux AppImage: mount dirs from earlier launches pile up on the server PATH; a dead one makes every git spawn fail with ENOTCONN #18790

What happened

Starting a new thread in a project failed with "workspace prep failed ENOTCONN".

Diagnosis

Every git spawn from the server failed with Error: spawn ENOTCONN (worktree prep, GitManager.branchPullRequest.remotes, ThreadPullRequestServiceV2.synchronize).

• The server process's PATH held 8 AppImage mount dirs (<cache>/t3code/tmp/.mount_T3-Cod* and …/usr/sbin), one per launch going back 5 days. All of them came before /usr/bin.
• 6 AppImage runtime processes from earlier launches (Oct 7–12) were still alive, each keeping its mount alive. The oldest mount was dead: ls returned Transport endpoint is not connected.
• PATH lookup during spawn stops at the first error other than ENOENT/EACCES/ENOTDIR. The dead mount returns ENOTCONN, so the search ended before /usr/bin/git. Reproduced with the server's exact PATH: env PATH="$SERVER_PATH" git --version → env: 'git': Transport endpoint is not connected.
• How it piles up: each new instance inherits the previous instance's environment (PATH already including the old mount), and the AppImage runtime adds its own APPDIR in front. Evidence: an Oct 11 runtime process had 4 mounts on PATH, and the Oct 12 one had 7 plus its own. Nightly auto-updates relaunch roughly daily. Inferred, not confirmed: the updater/relaunch path starts the new AppImage with process.env as is.
• Source (v0.0.46-nightly.20261012.3021, same on main): stripAppImageRuntimeEnv in apps/server/src/terminal/Manager.ts removes only the current APPDIR segments, and only for terminal PTYs. Nothing cleans .mount_* segments from earlier launches out of the desktop/server process env. DesktopShellEnvironment.installPosixEnvironment merges the login-shell PATH with the inherited process.env.PATH, which keeps the stale entries.

Suggested fix: at desktop startup, remove every PATH (and LD_LIBRARY_PATH/XDG_DATA_DIRS) segment that is under any .mount_* AppImage mount except the current APPDIR. Do this before the server is spawned, and also scrub the env passed to relaunch/update.

Steps to reproduce

1. Linux, run the AppImage build from a desktop launcher.
2. Restart it several times in a way that hands down the env (in-app relaunch or nightly auto-update).
3. Inspect the server's PATH: tr '\0' '\n' < /proc/<server-pid>/environ | grep ^PATH=. It shows one .mount_T3-* pair per earlier launch.
4. Kill the AppImage runtime of one older launch (or wait for it to die). Its mount turns into ENOTCONN.
5. Start a thread / worktree. Workspace prep fails with spawn ENOTCONN.

Version

0.0.46-nightly.20261012.3021 (desktop AppImage)

Environment

CachyOS Linux x64 (kernel 7.2.8), Wayland, Node v26.8.2, git 2.55.0

Evidence

runGitCommand  git.cwd=<home>/.t3/worktrees/<project>/t3-918b8cb8
  exit: Failure  Error: spawn ENOTCONN
    at ChildProcess.spawn (node:internal/child_process:458:11)
    at Module.spawn (node:child_process:813:9)
    at runGitCommand (…/apps/server/dist/binCli-*.mjs:120805:5)
    at GitManager.branchPullRequest.remotes (…)

ThreadPullRequestServiceV2.synchronize  event: "thread branch pull request lookup failed"  cause: Error: spawn ENOTCONN

Server PATH (excerpt):
  <home>/.local/bin : … : <cache>/t3code/tmp/.mount_T3-CodDlgiNJ : …/usr/sbin
  : .mount_T3-CodbBAAcc : .mount_T3-CodBcobcf : .mount_T3-CodJKbOcO : .mount_T3-CodkdaOdj
  : .mount_T3-CodFHbFnM : .mount_T3-CodloMHfl : .mount_T3-CodGLOGmC (dead) : /usr/local/bin : /usr/bin …

$ ls <cache>/t3code/tmp/.mount_T3-CodGLOGmC
ls: cannot open directory: Transport endpoint is not connected

Related issues

#14259 (AppImage mount disappears, children die with SIGBUS) is a related way an AppImage mount can die, but the bug there is the main process exiting. Here the main process stays alive and stale mounts pile up on PATH. Not a duplicate.

Fix applied or workaround

fusermount3 -uz on the dead mount (it becomes an empty dir, so PATH lookup skips it). Git spawns worked again right away with no restart. Lasting workaround: fully quit and relaunch from the desktop launcher so the app starts with a fresh environment.

Filed by

Claude Code (claude-opus-5-5) via t3 triage