Desktop App: OAuth token refresh kills all active Claude Code sessions with exit code 143 (SIGTERM)

Bug

Every time the Desktop App refreshes its OAuth token, it sends SIGTERM to all running Claude Code subprocess sessions, causing them to exit with code 143. Active sessions are interrupted mid-task with "Session was interrupted / Claude Code process exited with code 143".

Steps to reproduce

1. Open Claude Desktop App
2. Start one or more local agent sessions
3. Wait for OAuth token refresh (roughly every ~60 minutes, or on reconnect)
4. All active sessions die simultaneously with exit code 143

Log evidence

From ~/Library/Logs/Claude/main.log — the token refresh immediately precedes the simultaneous kills:

2026-06-06 13:39:13 [info] [oauth] obtained new token for orgId=7eba2edf-04fd-4893-a7b5-9dda9973717d, caching
2026-06-06 13:39:15 [error] Session local_3cc101b5-cb3e-437c-92dd-aad2d5ff200c query error: Claude Code process exited with code 143
2026-06-06 13:39:15 [error] Session local_f8704b36-24ef-4f83-ad14-42f800c2f73d query error: Claude Code process exited with code 143
2026-06-06 13:40:16 [error] Session local_ac71aef1-9d9d-44ac-af33-10335e4395cb query error: Claude Code process exited with code 143
2026-06-06 13:41:18 [error] Session local_3cc101b5 ... exit 143  (all three again, simultaneously)

The simultaneous failure of all sessions on token refresh is consistent and reproducible. Between failures the app respawns sessions with fresh [CCD] Passing 3 plugin(s) to SDK log lines, then they die again on the next refresh cycle.

Expected behavior

Token refresh should hand off the new token to running subprocesses without killing them, or at minimum restart them gracefully without interrupting in-flight work.

Workaround

Running claude directly from the terminal is unaffected — the terminal process does not go through the Desktop App's token management layer.

Environment

• Claude Desktop App: 2.1.167
• macOS: Darwin 25.5.0
• Mode: local agent sessions (CCD)