[nsd97]

What happened

On a Mac with Homebrew's coreutils first on PATH, a source install stopped before installing anything. That PATH resolved stat to /opt/homebrew/opt/coreutils/libexec/gnubin/stat (GNU coreutils 9.12):

==> Ending any stale session and checking the recovery journal

Install stopped: the backstop could not fully undo a previous session
(exit status 75).

The Mac had never run Insomnia. insomnia.log had one line:

backstop: recovery lock .../.recovery.lock still held after 10s (lockf exit 75); nothing changed, will retry

backstop.sh decides whether its fd 9 is the caller's handle on the lock with stat -f %i, and stat was the only tool it still took from PATH. GNU stat reads -f as --file-system, so both sides print file system text instead of an inode, and the two never match.

• The backstop that install.sh runs under its own recovery lock therefore reopens the lock and waits on its caller for LOCK_TIMEOUT_SECONDS, then exits 75.
• uninstall.sh hands its lock down the same way.
• The LaunchAgent is not affected: launchd runs it with the system PATH.
• The same bare stat reads the size of the app binary's --resume-frozen answer with -f %z.

What changes

• backstop.sh: STAT=/usr/bin/stat joins the fixed tool paths, and both calls use it. A comment says why.
• RecoveryScriptTests:
  • testBackstopSharesLockHandedDownOnFd9WithGNUStatFirstOnPath repeats the existing hand-down test with a GNU-style stat stub first on PATH. The stub logs and prints what GNU prints.
  • On the old script it fails the way the install did: exit 75, the stub ran, "still held" in the log.
  • With the change it passes, along with testBackstopSharesLockHandedDownOnFd9 and testBackstopWithUnrelatedFd9DoesNotShareAForeignLock.

Checks

• Tests: swift test on this branch: 1,224 tests, 0 failures, 0 skipped (tmux installed), the new test included. swift build -c release -Xswiftc -warnings-as-errors passes.
• Scripts: /bin/bash -n and the bash 4 grep on every script, and ShellCheck 0.11.0, pass.
• Real Mac: MacBook Pro (Mac14,6), M2 Max, macOS 27.0 (26A428), with GNU coreutils 9.12 first on PATH. With this change in a local build that also had macOS 27: allow the DisplayServices calls and read Low Power Mode from powermode #58, Menu bar: route status item clicks by the cursor on macOS 27 #76 and Add an explicit End session action to the menu #74, the same ./scripts/install.sh from the same Terminal finished, and logged "LaunchAgent com.insomnia.backstop loaded (launchctl print confirms)".
• Not run: uninstall.sh on that Mac, which goes through the same check.

Sources

• GNU stat's -f is --file-system: "Report information about the file systems where the given files are located instead of information about the files themselves." GNU Coreutils, stat invocation
• BSD stat's -f format, "Display information using the specified format", with i the inode number and z the size in bytes: FreeBSD stat(1). macOS's man 1 stat says the same.
• The fixed tool paths rule: .greptile/config.json, scripts-call-tools-through-fixed-path-variables
• Script and test requirements: CONTRIBUTING.md

🤖 Generated with Claude Code

Confidence Score: 5/5

This PR appears safe to merge; no actionable issue remains.

Summary

The backstop now calls /usr/bin/stat through STAT, so a GNU stat earlier on PATH cannot disrupt recovery.

• The backstop recognizes a handed-down lock when GNU stat is first on PATH.

_{Reviews (2) · Last reviewed commit: "Tests: write the GNU stat stub and its w..." · Reviewed by Greptile}

[krishhgg]

I pushed two commits to this branch:

• 7bf2569 merges main. It merged cleanly.
• 636a33f is a test-only fix from review. This branch started before Tests: run ShellCheck over every shell script the tests write #86, which added TestScripts.write, so the new test wrote the GNU stat stub and its wrapper straight to disk, and CI's ShellCheck step never saw them. Both now go through TestScripts.write, like the sibling hand-down test's wrapper. The stub passes ShellCheck, and the three hand-down tests pass locally.

The backstop change itself had no findings.

🤖 Generated with Claude Code

[krishhgg]

Thanks again for this fix. Before it ships in a release, I need test evidence from your Mac for the code as it is now on main. The evidence in this PR's description came from an earlier build. Since then the branch got review fixes and merges of main, so it doesn't cover the code that merged. Any change that can affect how Insomnia behaves needs this, so I'm asking every contributor from this round.

Checks for this change (on your macOS 27 MacBook Pro)

Test main alone this time. Your earlier build also had #76 and #74 applied, which aren't merged.

Run checks 1 to 3 in a Terminal with GNU coreutils first on PATH: export PATH="$(brew --prefix coreutils)/libexec/gnubin:$PATH". Post command -v stat and stat --version | head -1 from that shell.

1. Install over an existing install. Run ./scripts/install.sh. Expected: it finishes and prints "LaunchAgent com.insomnia.backstop loaded (launchctl print confirms)", and insomnia.log has no "still held" line. Post the full install output.
2. Uninstall. In the same shell, run ./scripts/uninstall.sh. Expected: it finishes and removes the app and the agent. Then reinstall the same way.
3. Resuming a frozen test process through the backstop. Install with INSOMNIA_LID_SIMULATION=1 ./scripts/install.sh from the same shell. Put a disposable app such as TextEdit on the freeze list, start a session, and run scripts/simulate-lid.sh closed. ps -o stat= -p <pid> should show T. Then kill -9 Insomnia and run ./scripts/uninstall.sh. Expected: the backstop resumes the test app through the installed Insomnia --resume-frozen (no T any more), and sleep is restored. This path now reads a size with /usr/bin/stat. Reinstall a normal build afterwards.
4. Normal PATH. In a new Terminal with the default PATH, run ./scripts/install.sh again. Expected: it finishes and prints the same line as check 1.
5. Recovery agent after a crash. Start a session, kill -9 Insomnia, and let the session's end time pass. Expected: the agent restores sleep (SleepDisabled reads 0), and the log shows its lines.

Build and what to post

• Build. A clean checkout of current main (db9085e or later), with no other PRs applied, installed with ./scripts/install.sh. Post git rev-parse HEAD and the output of git status --short, which should be empty.
• Mac. Post sw_vers, sysctl -n hw.model and sysctl -n machdep.cpu.brand_string.
• Samples. Before each check, start a sampler in a Terminal window and leave it running. It keeps logging while the lid is closed, because the Mac stays awake:

  while :; do date '+%T'; pmset -g | grep SleepDisabled; pmset -g custom | sed -n '/Battery Power/,/AC Power/p' | grep -E 'powermode|lowpowermode'; sleep 5; done | tee -a ~/insomnia-samples.txt

• For each check, post the steps you ran with times, pass or fail, and the raw output that shows it:
  • the lines of ~/Library/Logs/Insomnia/insomnia.log for that run;
  • the samples for that run;
  • pmset -g log | grep -E ' (Sleep|Wake|DarkWake) ' for that time window, which shows whether the Mac slept;
  • the extra readings named in the check.
• Not run. List anything you couldn't run, and why.

Post the results as a comment on this PR. Put long logs in a gist and link it. Run everything while you're with the Mac, on a stable, ventilated surface, never in a bag, and not while you have unsaved work open.

🤖 Generated with Claude Code