Claude Desktop (macOS) randomly restarts mid-session, killing all in-flight Claude Code turns (renderer crash; Crashpad/Sentry captured dump at incident time)

Summary

Claude Desktop on macOS spontaneously "restarts" at random — the window reloads as if the app relaunched (the main process actually stays alive; the renderer layer dies and respawns). When it happens, every open Claude Code session's in-flight turn is killed mid-run: long agent tasks, running tools, everything the model was working on is dropped and unrecoverable. This is recurring and severely disrupts multi-session workflows.

The latest incident was 2026-07-08 ~14:45 EDT (18:45 UTC). This report was filed from the affected machine minutes afterward, so the forensics below are fresh.

Environment

• Claude Desktop: 1.19367.0 (com.anthropic.claudefordesktop)
• Embedded Claude Code runtime: 2.1.202 (~/Library/Application Support/Claude/claude-code/2.1.202/)
• Standalone Claude Code CLI on same machine: 2.1.204
• macOS: 26.5.1 (Build 25F80), Mac16,8, 12 cores, 48 GB RAM, arm64
• Heavy usage: multiple Claude Code sessions open concurrently in the desktop app

What happens (user-visible)

1. Without warning, the app UI resets as if freshly launched — no user action, no quit.
2. All sessions' in-progress model responses are dropped. Anything mid-turn (long agent runs, tool executions, multi-minute reasoning) is lost and has to be manually re-run.
3. It recurs at random intervals. It is not tied to any specific action.

Evidence from this incident (all times America/New_York, UTC-4)

1. The main process did NOT restart — the renderer layer did. At 14:48, the main app process was 14+ hours old while fresh renderer/plugin helpers had just spawned at the incident time:

ps -axo pid,lstart,etime,comm
 4671  Wed Jul  8 00:34:01 2026  14:14:36  .../Claude.app/Contents/MacOS/Claude          ← main, up since 00:34
48810  Wed Jul  8 14:45:22 2026  03:15     .../Claude Helper (Renderer)                  ← fresh renderer
48819  Wed Jul  8 14:45:23 2026  03:14     .../Claude Helper (Plugin)                    ← fresh plugin helper

The unified log additionally shows short-lived Claude Helper (Renderer) processes spawning at 14:43:59 (PID 48673), 14:45:22 (48802), 14:45:26 (48953), and 14:46:39 (49254) — renderer churn right through the incident window.

2. Crashpad captured a dump at exactly 14:45. All three dump directories were touched at the incident moment and are now empty (dump written → uploaded → swept):

~/Library/Application Support/Claude/Crashpad/
  completed/  new/  pending/   — all mtime Jul  8 14:45, all empty now

3. The Sentry crash queue was touched at the same moment:

~/Library/Application Support/Claude/sentry/queue/queue-v2.json — mtime Jul  8 14:45

Sentry session context from this device, for server-side lookup:

• did: 73bf29d3-5105-4b4b-8bc7-862c2be1cd82
• sid: 25570a55547c41449d35a584f9abe0a1 (session started 2026-07-08 00:34:06 EDT)

The crash event(s) for this incident should be findable in your Sentry around 2026-07-08 18:45 UTC for that device id. The dumps no longer exist locally.

4. macOS ReportCrash activated during the window — unified log shows runningboardd resolving com.apple.ReportCrash (14:44:05.061) and com.apple.ReportCrash.Root (14:44:05.724), and the app tearing down XPC/notification connections at 14:44:33–14:44:41.

5. Sustained resource pressure from the app this week. macOS has written resource-violation diagnostics for Claude Desktop on Jul 1, Jul 2, Jul 6 (Claude Helper), and Jul 8:

/Library/Logs/DiagnosticReports/Claude_2026-07-08-020516_*.diag
Event:         disk writes
Writes:        2148.22 MB of file-backed memory dirtied over 26789 seconds (80.19 KB/s avg),
               exceeding limit of 24.86 KB/s over 86400 seconds
PID:           4671   Version: 1.19367.0

This may or may not be related to the crash, but it shows the app is consistently blowing past macOS resource limits on this machine.

Impact

Multi-minute agent work across several concurrent sessions dies simultaneously and silently. There is no recovery path — sessions persist, but the in-flight turn does not, so all model effort in progress at crash time is simply gone. For someone running Claude Code as a primary development driver, each occurrence destroys real work and breaks flow.

Ask

1. Root-cause the crash — the Crashpad dumps were uploaded from this device (Sentry did above, ~18:45 UTC 2026-07-08); local copies were swept after upload.
2. Harden against it — an in-flight turn should survive a renderer reload, or auto-resume after one. Session history already persists across the restart; the running turn is the only thing lost, and it's the thing that matters most.

---
Filed via Claude Code from the affected machine immediately after the incident.