Skip to content

Desktop 2.3.0: background service restart loop (10-30s) causes 'socket connection was closed unexpectedly' / 'Transport: Failed to fetch' #64069

Open

Description

@zigih-1212

Environment

  • OpenCode Desktop 2.3.0 (win32 x64), bundled CLI 2.3.0
  • Windows (x64), Wi-Fi, RU ISP with DPI filtering (Happ TUN enabled)
  • The failing session's workspace is the filesystem root (C:\)

Summary

On Desktop 2.3.0 the managed background service (opencode-cli serve --service, 127.0.0.1:28894) is repeatedly detected as "missing" and restarted by the main process. Every restart drops in-flight requests, so the UI shows ClientError: Transport: Failed to fetch, ClientError: Transport: network error and ECONNRESET: The socket connection was closed unexpectedly.

The service itself is healthy: when started manually it stays up indefinitely and answers /api/info. Only the app-managed instance dies or gets restarted.

Steps to reproduce

  1. Start OpenCode Desktop 2.3.0, open a session whose workspace is C:\ (very large tree).
  2. Leave the app idle for ~30 s.
  3. Observe UI errors and repeated service restarts.

Evidence (from %APPDATA%\ai.opencode.desktop\logs\<session>\)

main.log (repeats every 10-30 s):

[info] (main) starting v2 background service
[info] (main) v2 CLI background service starting { reason: 'missing' }
[info] (main) v2 CLI background service ready { runningVersion: '2.3.0', probed: false, ... }

renderer.log:

[info]  [global-sdk] event stream disconnected            (exactly every 10.000 s)
[error] ClientError: Transport: network error
[error] ClientError: Transport: Failed to fetch
[warn]  [server-request-queue] server thrashing detected
[error] Failed to read server terminal support ClientError: UnexpectedStatus: 503
[error] Failed to list terminals Error: The server took too long to answer

opencode.log:

level=ERROR message="failed to fetch models.dev" ... operation timed out after '10s'
level=INFO  message="location services booted" durationMs=18481

Control tests performed

  • opencode-cli.exe serve --port 28894 (manual, standalone): runs >100 s without exit; /api/info responds.
  • opencode-cli.exe serve --service --port 28894 --log-level debug: runs >100 s without exit.
  • opencode-cli.exe service start + service status => service running and responding.
  • Disabling all global plugins in ~/.config/opencode/plugins does not stop the restart loop.
  • Launching the app with --no-proxy-server does not stop it.
  • Duplicate spawns produce Failed to start server. Is port 28894 in use? in the same window.

Root cause hypothesis

Two independent effects:

  1. The SSE stream /api/event is closed by the server every 10 s (Bun idleTimeout default), which the client treats as a transport failure.
  2. When a renderer request fails during service boot (503 while "location services" boot for a huge workspace), the main-process supervisor restarts the background service; each restart kills in-flight requests and produces the user-visible "connection closed" errors. With a large workspace boot takes 10-20 s, so the loop never converges.

Expected

  • No service restarts while the service is healthy; SSE reconnects should not be treated as fatal.
  • Either a configurable service-start timeout, or readiness should be probed until boot completes.

Related

#60399 covers model requests failing due to RU ISP DPI (ECONNRESET). It explains the error text but not this local supervisor loop: the loop reproduces with a working network path, and the service answers fine when run manually.

Activity

  1. opencodejr commented on Nov 15, 2026

    @opencodejr
    Contributor

    Thanks for the detailed logs and control tests. I couldn't reproduce this here: the restart loop comes from the Windows desktop app's supervisor, and my sandbox is Linux-only with no desktop build.

    What I could check, with the 2.3.0 CLI on Linux:

    • Event stream. /api/event, both from opencode serve and from the managed service started with opencode service start, sent server.connected and then a heartbeat every 15 s, and stayed open for the 40-45 s I watched. I did not see the server close it after 10 s. So the disconnect every 10 s in your renderer.log may come from something else, such as the app reconnecting or something between the app and 127.0.0.1.
    • Large location. With the filesystem root as the location, requests were answered quickly, /api/info kept returning 200 and the service kept the same PID. But the root here is tiny compared with C:\: it booted in about 1 s, not the ~18 s in your log, so I couldn't recreate the slow boot.

    These would help a maintainer with a Windows setup:

    • From DevTools in the desktop app, the error field of the [global-sdk] event stream disconnected console entry.
    • Whether the opencode-cli serve --service process keeps the same PID across the reason: 'missing' lines, or a new process starts each time.
    • Whether the loop stops if you open a session in a small project folder instead of C:\.
    • Whether the loop also happens with Happ TUN turned off.

    Related issues, most likely first:

    A maintainer with a Windows environment will need to look at this.

  2. zigih-1212 commented on Nov 15, 2026

    @zigih-1212
    Author

    Thanks. Here is the additional data you asked for, collected on Windows 11 (desktop 2.3.0) while the loop was running.

    1. PID stability across reason: 'missing' events — a new PID every cycle

    Sampling Get-NetTCPConnection -LocalPort 28894 and Get-Process opencode-cli every 2 s:

    17:50:31 listenPid=0     cliPids=9084
    17:50:35 listenPid=9084  cliPids=
    17:50:42 listenPid=0     cliPids=6588
    17:50:45 listenPid=6588  cliPids=
    17:50:53 listenPid=0     cliPids=9684
    17:50:59 listenPid=9684  cliPids=9684
    17:51:06 listenPid=9684  cliPids=
    17:51:13 listenPid=0     cliPids=10232
    17:51:17 listenPid=10232 cliPids=
    17:51:25 listenPid=0     cliPids=3800
    17:51:30 listenPid=3800  cliPids=
    17:51:40 listenPid=4296  cliPids=
    17:51:46 listenPid=0     cliPids=476
    

    Each opencode-cli lives only ~5–10 s, and the PID changes on every spawn. Correlated main.log:

    17:52:19 starting v2 background service
    17:52:24 v2 CLI background service starting { reason: 'missing' }
    17:52:27 v2 CLI background service ready
    17:52:30 starting v2 background service
    17:52:35 v2 CLI background service starting { reason: 'missing' }
    17:52:38 v2 CLI background service ready
    17:52:42 starting v2 background service
    ...
    

    Control: the same executable, same port, started outside the app, is stable and keeps the same PID:

    • opencode-cli.exe serve --port 28894 — >100 s, /api/info answers;
    • opencode-cli.exe serve --service --port 28894 --log-level debug — >100 s;
    • one manually started instance stayed alive 31 minutes (listenPid=6308 constant) while the desktop app was open, and disappeared only when the app was restarted.

    So the child process is healthy, but as soon as it is app-managed it is replaced every ~11 s.

    2. The error field of [global-sdk] event stream disconnected

    Not captured. renderer.log only records [object Object]; the DevTools console is not reachable from a script. If you tell me which console filter or how to dump the entry (or if the log call can JSON.stringify the argument), I will paste it.

    3. Small project folder instead of C:\

    Not tested through the UI yet (the app has to open a folder interactively). What I can confirm is the boot cost difference with the same CLI: location services booted durationMs is ~1000–1800 ms for a small folder vs 18481 ms for C:\. I can run the UI test on request.

    4. Loop with Happ TUN off

    Not tested: without the tunnel, model requests to opencode.ai / models.dev are RST by the ISP here (the #60399 condition), so the test would mix two failures. The loop itself is local (127.0.0.1) and reproduces while curl --noproxy "*" http://127.0.0.1:28894/api/info returns 401 immediately.

    Notes

  3. zigih-1212 commented on Nov 15, 2026

    @zigih-1212
    Author

    Additional data from the same Windows machine (desktop 2.3.0, CLI 2.3.0), trying to find a mode that survives:

    1. Managed service (default TUI)

    The TUI spawns serve --service roughly every 9–11 s:

    16:08:58 cli starting args=["serve","--service"]
    16:09:07 cli starting args=["serve","--service"]
    ...
    16:27:11 cli starting args=["serve","--service"]   (twice, same second)
    

    The listening socket on 28894 flaps in and out between spawns, and the TUI shows a permanent overlay:

    Restarting Service... Your session will resume automatically
    

    2. --standalone (private server)

    The TUI starts a private serve --stdio --port 0, which exits shortly after start. The client then loops on ECONNREFUSED once per second:

    16:15:33 event stream connecting attempt=352
    16:15:33 event stream disconnected attempt=353 error="Transport: ECONNREFUSED: Unable to connect..."
    16:15:34 event stream connecting attempt=353
    ...
    16:15:38 event stream connecting attempt=357
    

    UI shows connection lost / Reconnecting to the servers automatically. No level=ERROR from the server itself — the child simply disappears.

    3. Pre-started service + plain TUI

    opencode-cli service start alone is stable: the process keeps the same PID for 30+ minutes and answers service status.
    If the TUI is started while that service is already running, the TUI attaches, but within ~3 minutes the supervisor spawned serve --service twice and the original stable process was gone.

    4. --server <url>

    opencode-cli --server http://127.0.0.1:28894 fails immediately:

    cli process failed cause="... Server at http://127.0.0.1:28894 requires a password ..."
    

    There is no documented flag or environment variable to pass that password to --server (serve --help has no auth flag), so this path is unusable as a workaround.

    Conclusion

    The service binary itself is healthy — it survives indefinitely when nothing supervises it. The failure appears whenever a TUI client is attached: the client's lifecycle logic kills/replaces the server (managed and standalone alike) within ~10 s, which is also the observed cadence of the serve --service respawns. This points at the supervisor/readiness logic rather than at the server, config, plugins or the network.

    If a maintainer can reproduce on Windows, the interesting question is: why does the app-spawned serve --service/serve --stdio process exit after ~5–10 s while the same binary started directly from the shell lives for minutes?

  4. jerry155756294 commented on Nov 17, 2026

    @jerry155756294
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions