[BUG]

Preflight Checklist

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

What's Wrong?

Claude Desktop (Windows): main process + renderer freeze 5–30s while the machine is idle; misreported as "likely sleep"

Summary

The Electron main process and the renderer freeze together for 5–30 seconds, repeatedly,
while the rest of the machine stays completely idle and responsive. The bundled Claude Code
CLI child process is healthy throughout — only the desktop shell freezes.

Today alone: 94 stalls, 767 seconds frozen, worst single stall 26.2 s.
Over Aug 15–19: 3,602 stalls, median 5.2 s, p90 11.7 s, 13.2% at or above 10 s.

Environment

• Claude Desktop 1.32885.1 (MSIX / Microsoft Store build), CCD 2.1.234, Node 24.18.1
• Windows 11 Enterprise 10.0.26200, x64
• 13th Gen Intel Core i7-1360P, 31.7 GB RAM

Key evidence

1. The freeze is confined to Claude. The machine is idle.

I ran an independent PowerShell watchdog that sleeps 500 ms in a loop and reports how much
longer than 500 ms it was actually descheduled. If anything external were freezing the
machine, it would freeze the watchdog too.

11:32:13 CLAUDE-only drift=0ms freeRAM=16.4GB diskQ=0 diskIdle=99% cpuPerf=96%
claude reported: 16567ms

Claude froze for 16.5 s. The watchdog drifted 0 ms. 8 of 8 captured freezes were CLAUDE-only;
none showed system-wide starvation. Sequence 11:31:11 (15.4 s), 11:31:23 (6.5 s),
11:31:41 (13.3 s), 11:31:58 (16.6 s) — ~51 s frozen inside one minute, machine 99% idle.

2. No work is in flight when it freezes.

46 of today's 94 stalls report task none. When a task is named, its own duration is 0 ms:

blocked for 3742ms (task webrequest:before-send-headers=0ms, ...)
blocked for 4026ms (task LocalSessions.getGitInfo=0ms, ...)

The task label appears to be whatever was in flight when the watchdog fired, not the cause.

3. The renderer freezes at the same time.

mainView paint probe: page did not answer { outcome: 'no_js', probe_count: 7, elapsed_ms: 15683 }

4. Downstream timeouts are symptoms, not causes.

git status timed out is logged repeatedly, but git returns in 218 ms when measured directly
(repo is 31 tracked files, 1 MB). The app can't collect the result because its event loop is
blocked when the timer fires. Same for pty-host: no spawn ack within timeout,
device-registry: TimeoutError, PluginBridgeMcp apply exceeded 15000ms.

5. Possible telemetry blind spot.

50 of today's 94 stalls are tagged [likely sleep: duration_heuristic]. The machine did not
sleep — the watchdog process was running normally throughout, and Modern Standby transitions
show no correlation (4.2% of stalls within 60 s of one, vs 3.7% expected by chance). If this
heuristic suppresses or reclassifies these reports upstream, that may be why the issue isn't
visible in telemetry.

6. The CLI underneath is healthy.

The desktop spawns claude-code as a child process. That child is fine — the app's own health
check reports healthy cycle (466s, hadFirstResponse=true) during the freezing, and the CLI
binary starts in 172 ms consistently. The fault is in the Electron shell.

Ruled out by measurement

| Hypothesis | Result |
|---|---|
| Antivirus / process-creation interception | Ruled out — process spawn is 25 ms (normal) |
| Wi-Fi disruption | Ruled out — anti-correlated (2.6% within 60 s vs 4.1% by chance) |
| Which network (home vs corporate) | Ruled out — median 4.97 s vs 5.36 s, near-identical |
| Modern Standby | Ruled out — 4.2% vs 3.7% by chance |
| Network latency | Ruled out — 8–10 ms to api.anthropic.com, 5 hops, no VPN |
| CPU throttling | Ruled out — 76–97% processor performance state, 12–23% load |
| Memory / disk / repo size | Ruled out — 16 GB free, disk 99% idle, queue 0, 31-file repo |
| Backup/VSS snapshots | Minor contributor only — 1.34x enrichment |
| Main process memory leak | Ruled out — RSS stable ~250–300 MB, tree ~1.7 GB, flat over hours |

Timeline

Stall rate normalized per hour of active use (raw daily counts are misleading — the app runs
24/7 but is only used part of the day):

Jun 46 stalls/busy-hour
Jul 14
Aug 1-14 175
Aug 15-19 95

Severity, which is independent of usage volume, confirms the same step change:

Jun median 4.0 s p90 8.3 s 5.7% >= 10 s
Jul median 1.5 s p90 6.1 s 4.7% >= 10 s
Aug 1-14 median 5.2 s p90 13.9 s 19.8% >= 10 s
Aug 15-19 median 5.2 s p90 11.7 s 13.2% >= 10 s

The degradation began in early August, between Jul 31 and Aug 2. The stall detector
threshold was unchanged (~500 ms minimum) across all periods, so these are comparable.

Impact

Two app crashes today within 10 minutes. Before the first, that instance had accumulated
total 103, cumulative 374650ms — 103 freezes totalling 6 min 15 s. Typing a message locks
up mid-keystroke; opening the app takes ~2 minutes.

What Should Happen?

All information in what's wrong.

Error Messages/Logs

Steps to Reproduce

please view the earlier version

Claude Model

Not sure / Multiple models

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude Desktop 1.32885.1 (MSIX / Microsoft Store build), CCD 2.1.234, Node 24.18.1

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

_No response_