cli(win): 45s event-stream idle watchdog restarts the managed service, aborting all sessions/subagents #52049

Description

On Windows the v2 managed background service (opencode serve --service) is repeatedly killed and respawned during normal use, aborting every in-flight session and background subagent ("failed"). It is not a crash — the client is killing it.

Every restart is preceded by the client log line event stream disconnected ... error="Event stream stalled", then ~6s later a new cli starting ... args=["serve","--service"].

The direct trigger is the client's event-stream idle watchdog (values read from the shipped 2.0.19 binary):

• server sends : heartbeat on /api/event every 15s;
• client resets a 45s idle timer on any received byte (idleTimeout ?? 45000);
• after 45s with no bytes the client aborts with Event stream stalled, then treats the service as unresponsive and replaces it (Background service is unresponsive; recovery cannot preserve persistent terminals, Failed to shut down persistent terminals before service replacement) instead of reconnecting.

Measured correlation: server's last log line 08:47:12.749, stall logged 08:48:40.688 → 46.1s, i.e. the 45s watchdog, not the ~6s health probe.

This is the same class as #50594 (Windows: file-watcher re-registration storm on ~/.claude/skills → serve --service disappears with no error/dump) and #50591 (Linux: Event stream stalled + InterruptError on /api/event). The new information here is that the service is not dying on its own: the client kills it after a 45s heartbeat gap, so the "silent death" in #50594 is a client-side replacement, and the fix must also cover the watchdog margin / the replace-instead-of-reconnect policy.

Ruled out on this machine:

• Not a crash — no opencode.exe APPCRASH in the Windows Application log; no crash dumps under %LOCALAPPDATA%\CrashDumps or WER (no RADAR_PRE_LEAK_64 either).
• Not OOM — 23.7 GB RAM, 6.7 GB free, page file usage ~205 MB.
• Not graceful shutdown — 2 shutting down lines in 12 days; the old process logs nothing before disappearing.
• Not the subagents — they are victims of the replacement.

Frequency (opencode.log, one machine): Event stream stalled = 3× on 2026-09-28, 6× on 2026-09-29; serve --service starts = 7 and 6 on those days. It stops completely when idle (12+ min with zero restarts, liveness probe 0–2ms).

Suspected trigger for the ≥45s heartbeat gap (evidence below, hypothesis): concurrent cold-boot of many locations plus watcher re-registration churn, forming a feedback loop — replace service → all recently-active locations cold-boot → watcher/config/MCP churn blocks the loop → heartbeat gap ≥45s → watchdog replaces the service again.

Plugins

• local ~/.config/opencode/plugins/orca-opencode-status.js
• local ~/.config/opencode/plugins/win-notify/
• no plugin array in opencode.json

(~/.claude/skills and ~/.agents/skills are both 44-entry directories and are re-subscribed ~950× each — see Additional context.)

OpenCode version

2.0.19 (channel=latest, npm global @opencode/cli)

Steps to reproduce

Intermittent, load-dependent; not reproducible on demand.

1. On Windows with opencode 2.0.19 CLI/TUI, let the shared managed service run (~/.local/state/opencode/service.json → http://127.0.0.1:49372).

2. Have several locations active (home directory plus several projects on other drives) and run subagent-heavy turns.

3. Observe repeatedly in ~/.local/share/opencode/log/opencode.log:

   08:48:40  client  event stream disconnected ... error="Event stream stalled"
   08:48:46  cli     cli starting version=2.0.19 args=["serve","--service"]
   08:48:47  client  event stream connected
   

4. In-flight work is aborted: InterruptError: All fibers interrupted without error ... http.url=/api/event http.status=200.

5. Restarts recur every ~2.5–16 min while busy; zero restarts while idle.

Screenshot and/or share link

N/A — log excerpts, code snippets and counts are inline below.

Operating System

Windows 11 Home Single Language, build 26200

Terminal

Windows Terminal (WT_SESSION set, TERM=xterm-256color), shell C:\WINDOWS\system32\cmd.exe

Additional context

Relevant client code (from the shipped 2.0.19 binary):

// event stream idle watchdog
var yY = 2000, Pv = 1000, gY = 50, bY = 45000, EY = 20000;
function wv(t, n) {
  let u = n.idleTimeout ?? bY;              // 45000 ms
  ...
  let Ke = () => { x = Date.now(); ...; qe = setTimeout(() => te.abort(Error("Event stream stalled")), u) };
  // onActivity resets the timer on ANY received byte
}

// server SSE handler
let s = qK("15 seconds").pipe(wa(() => `: heartbeat\n\n`));

// SSE reader: any byte counts as activity, including the heartbeat comment
if (!y.done) n?.onActivity?.();

Watcher churn / location cold-boot evidence (hypothesis for the 45s gap):

stall 08:18:06 : watcher_subscribe=116  location_boot=8   spawns=40
stall 08:24:19 : watcher_subscribe=123  location_boot=9   spawns=12
stall 08:26:55 : watcher_subscribe=129  location_boot=10  spawns=6

location services booted: C:\Users\<user> 73×, plus ~20 other project locations. This matches #50594 (.claude/skills storm) and #51852 (intermittent main-thread spin correlated with watcher churn).

Next step being tested: OPENCODE_DISABLE_FILEWATCHER=true (per #51852). Will report whether restarts stop under the same load.

Related issues: #50594, #50591, #51852, #51216, #37795, #48588, #36285.