claude-code - 💡(How to fix) Fix # [BUG] [Windows] Auto-updater stuck in reinstall loop: daemon self-restarts on mtime change and reinstalls the same version repeatedly, flashing cmd windows every few minutes

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

Summary

On Windows, the Claude Code auto-updater has entered a self-perpetuating loop that repeatedly runs npm install -g @anthropic-ai/claude-code@<latest> for a version that is already installed. Each reinstall rewrites claude.exe, changing its mtime; the daemon's binary-watcher interprets the mtime change as a new upgrade and self-restarts, which re-arms the update check, which reinstalls again. The visible symptom is a black cmd.exe/node.exe console window flashing on screen every 1-3 minutes, continuously, for over 24 hours.

The version never advances because it is already the latest - the loop is driven purely by mtime churn, not by an actual version difference.

Environment

• OS: Windows 11 Home (10.0.26200)
• Claude Code: 2.1.158 (installed via npm global)
• Latest published: 2.1.158 (identical - already up to date)
• Node: v22.19.0
• npm: 10.9.3
• npm global prefix: C:\Users<user>\AppData\Roaming\npm (confirmed writable)
• Concurrent claude.exe sessions: ~33 (long-standing setup, not new)

Symptom

A black console window briefly appears and disappears every couple of minutes. Process tracing shows the spawned command is always one of:

cmd /c "npm view @anthropic-ai/claude-code@latest version --prefer-online"
cmd /c "npm install -g @anthropic-ai/[email protected]"

Every such process is parented to a claude.exe session.

Impact

• Constant flashing console windows (UX/annoyance, steals focus on some configs).
• Continuous redundant npm installs and daemon restarts (wasted CPU/network/disk churn).
• Not a failed update and not a permissions problem - the install succeeds every time; there is simply no terminal/"already current" state.

Suggested fixes

1. Skip npm install when the installed version already equals the resolved latest.
2. Use a content hash (or version string) instead of mtime to decide whether the binary actually changed.
3. Add an "already current" short-circuit so a no-op reinstall cannot re-arm the watcher.

Workaround

Setting DISABLE_AUTOUPDATER=1 stops the loop (already on latest, so nothing lost), but this disables automatic updates entirely, which is not desirable as a permanent fix.

What Should Happen?

Expected behavior

• The updater should treat "already on latest" as a no-op and not reinstall.
• The daemon's "binary changed" detection should key off version/content hash rather than mtime, so an identical reinstall does not trigger a self-restart-for-upgrade.

Error Messages/Logs

Steps to Reproduce

Root cause (evidence from ~/.claude/daemon.log)

The daemon log shows a tight repeating cycle. Representative excerpt:

[2026-05-30T00:49:31Z] [supervisor] binary at ...\claude.exe changed (mtime changed) - self-restarting for upgrade
[2026-05-30T00:49:31Z] [supervisor] shutting down
[2026-05-30T00:49:33Z] [supervisor] daemon start version=2.1.158
[2026-05-30T00:52:34Z] [supervisor] binary at ...\claude.exe changed (mtime changed) - self-restarting for upgrade
[2026-05-30T00:52:34Z] [supervisor] shutting down
[2026-05-30T00:52:37Z] [supervisor] daemon start version=2.1.158
... repeats indefinitely, always landing on the same version 2.1.158

The loop:

1. A session's auto-updater runs npm install -g @anthropic-ai/claude-code@latest.
2. The reinstall rewrites claude.exe - same version, new mtime.
3. The daemon's binary-watcher sees the mtime change and self-restarts "for upgrade".
4. On restart the update check re-arms and reinstalls again -> back to step 1.

It never converges because the reinstall changes the file mtime even though the contents/version are unchanged.

Contributing factor: multi-daemon lock races

With multiple concurrent sessions, the log also shows daemons racing over the singleton lock, e.g.:

[supervisor] another daemon won the lock race (pid=...) - exiting
[supervisor] another daemon is already running (pid=..., version=2.1.158, origin=transient ...). Stop it with `taskkill /PID ...`, then retry.

When one session reinstalls and bumps the binary mtime, it appears to trip the "binary changed" self-restart in the other sessions' daemons too, so they each re-check/reinstall. This amplifies the loop across the session fleet.

Timeline / trigger

• daemon.log begins 2026-05-29T23:53 running version 2.1.157.
• First "self-restarting for upgrade" fires 3 minutes later at 2026-05-29T23:56.
• 45 restart events logged between 2026-05-29T23:56 and 2026-05-31T01:17 (and ongoing).
• This coincides with a burst of releases (2.1.153 on 27 May -> 2.1.158 on 30 May). The rapid upgrades appear to have kicked off the mtime-churn loop, which then sustains itself even after reaching the latest version.

The session count is long-standing and unchanged; the new variable was the release burst, not the number of sessions.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

2.1.158 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

No response