[BUG] Desktop app (Windows Store 1.28929.0): every background session turn dies silently (zero output, no error, exclamation badge); interactive session unaffected

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?

Since ~06:30 ET on 2026-08-12, EVERY turn in a BACKGROUND session (one the human is not actively typing in) dies silently in the Windows desktop app: the turn starts, shows "running" for 20-90 minutes, then ends with zero transcript events, zero file writes, and no error message anywhere. The app shows an exclamation badge next to the session row. The fleet of ~14 parallel sessions (1 orchestrator + 13 workers) worked normally before that morning.

Details:

• Messages delivered to affected sessions via the session-management MCP land in the transcript as user turns but never produce a processed response.
• The session the human actively interacts with works perfectly all day, including its subagents.
• Reproduced on: existing long-lived sessions, freshly created sessions, Opus and Fable models, medium and high effort - 30+ turns across 14 sessions, 100% of background turns affected.

What we tried (all failed to cure it):

1. Full Windows restart + app relaunch - first background turn after restart died the same way (22 min, zero output).
2. claude update npm-global 2.1.205 -> 2.1.228 (changelog suggested background-session fixes in 2.1.216/2.1.217) - post-update turns on fresh app processes died identically at ~34 min. This suggests the Store app's bundled runtime carries the bug, not the npm-global CLI.
3. Fresh sessions on both model families, staggered kicks, single kicks - no difference.

claude doctor: clean, npm-global 2.1.228, no installation issues found.

Environment: Claude Code desktop app for Windows, Microsoft Store MSIX package Claude 1.28929.0.0 (C:\Program Files\WindowsApps\Claude_1.28929.0.0_x64__pzs8sxrjxfjjc\app\Claude.exe), Windows 11 Home 10.0.26200. Timeline suggests a Store auto-update landed overnight right before the breakage.

Possibly related: #68625, #51052, #48045

Question: if the Store app bundles its own CLI runtime, does build 1.28929.0 predate the 2.1.216/2.1.217 background-session fixes, and when does the Store build pick them up? Happy to provide debug logs on the next occurrence.

What Should Happen?

A turn started in a background session (via cross-session message or scheduled task) should run to completion and produce its assistant response and tool side effects, exactly as it does when the human types into that session directly - or at minimum surface an error when it fails, instead of dying silently after tens of minutes.

Error Messages/Logs

No error output exists anywhere - that is the core of the report. The transcript shows the delivered user message with nothing after it; the session row gets an exclamation badge; `claude doctor` reports "No installation issues found" (npm-global 2.1.228). One session did render "You've hit your session limit - resets 11:30pm (America/New_York)" during a genuine limit event the previous night, but the owner's usage dashboard (weekly 65%, 5-hour 7%) ruled out limits for today's failures, and fresh sessions with zero usage failed identically.

Steps to Reproduce

1. Windows desktop app (Store build 1.28929.0), 2+ sessions open in one workspace.
2. From session A (the one you are typing in), send a work order to session B via the session-management MCP send_message tool (or a scheduled kick) - anything that starts a turn in a session the human is not focused on.
3. Session B shows a running turn (list_events reports "(running)"), and the session row eventually gets an exclamation badge.
4. Wait 20-90 minutes: the turn ends with no assistant message in the transcript, no tool side effects on disk, no error text. list_events shows the delivered user message with nothing after it.
5. Repeat with a freshly created session, different model, different effort - same result every time (30+ turns reproduced today).

Interactive control: type the same order directly into session B's own window - in that case the turn completes normally. Only background-initiated turns are affected.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Unknown - fleet worked normally 2026-08-11 on the prior Store desktop build; app auto-updated overnight

Claude Code Version

2.1.228 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

Terminal/Shell "Other" = this is the Claude Code DESKTOP APP for Windows (Microsoft Store MSIX build 1.28929.0), not a terminal session. "Claude Code Version" above is the npm-global CLI on the same machine; the desktop app bundles its own runtime whose version we cannot read, which is the suspected buggy component since updating the npm-global CLI changed nothing.