service: background shell silently killed when idle location is evicted; completion notification never delivered #58049

Summary

Background shells (shell tool with background=true) are tied to the location's lifecycle. When the service evicts an idle location, running background shells are killed, their output files are deleted, and the promised "you will be notified when it completes" event is never delivered — the agent waits forever with no signal.

Environment

• opencode version: 2.0.32
• OS: Windows_NT 10.0.19045 (win32 x64)
• Terminal: TUI (TERM=xterm-256color)
• Shell: PowerShell (pwsh) spawned by the service
• Install/channel: latest (autoupdate disabled)
• Active plugins: none configured; one MCP server (playwright, via npx)

Reproduction

1. In a session, start a background shell that outlives the location idle window (e.g. a watcher loop polling a remote host every 3 min; observed death after ~60 min idle).
2. Let the session go idle between turns so the service evicts the location.
3. Service log shows message="location services evicted" and later message="location services booted" on next access.
4. The background shell process is gone; its output file under ~/.local/share/opencode/shell/<hash>/ is deleted; no completion/termination event reaches the session. GET /api/shell/<id>/output returns 4xx.

Expected Behavior

The documented contract "you will be notified automatically when the command finishes" should not silently break. Any of: (a) don't evict a location with running background shells; (b) deliver a termination event (with reason) to the owning session on eviction; (c) own background shells at service level instead of location level.

Actual Behavior

Shell killed without notice, output file removed, session never woken. A second identical watcher that completed before the eviction delivered its notification normally, isolating the failure to eviction.

Additional Context

• The service process itself never restarted (same PID before/after) — not a crash, purely location lifecycle.
• Timeline: watcher started 15:21 -> location services evicted 16:22:24 (INFO) -> user's GET /api/shell/<id>/output 404 at 16:22:24 -> location services booted.
• After the re-boot, MCP tools re-registered and the next user message showed "Instructions updated: core/codemode" — consistent with full location re-initialization.
• Frequency: observed once, timing-consistent; presumably deterministic whenever eviction happens while a background shell runs.
• Workaround: re-arming the watcher only postpones the same outcome.