[BUG] Desktop randomly marks running turn as interrupted (AbortError: Aborted, no retry, no UI error)

[romg974]

Description

During normal Desktop usage on Windows, a running turn is randomly marked as "interrupted" even though I never pressed ESC or submitted anything. No error is shown in the UI — the turn just silently stops. Sending a new message right after works fine, and the session context is intact (no compaction/reset).

This matches the "Session Interrupted Without Retry" pattern described in #26180 (and #29589, which was deleted along with its author's account): a transient stream drop aborts the fiber, no retry happens, and the user sees a silently stopped turn.

Log evidence from ~/.local/share/opencode/log/opencode.log — two occurrences today on a live session, no user interrupt:

timestamp=2026-09-13T02:30:20.729Z level=ERROR run=312b5db8 message=process session.id=ses_f67e8f861ffet3LexvzrhZJB0V messageID=msg_09899f0f7001cZ4a7IFlBx4Fr0 error=Aborted stack="AbortError: Aborted
    at new DOMException (node:internal/per_context/domexception:76:18)
    at Array.<anonymous> (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:489828:27)
    at ~effect/Effect/successCont (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7896:29)
    at ~effect/Effect/evaluate (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7909:25)
    at FiberImpl.runLoop (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7716:112)
    at FiberImpl.evaluate (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7686:26)
    at file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7854:19
    at file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:6602:55
    at process.processTicksAndRejections (node:internal/process/task_queues:104:5)"

timestamp=2026-09-13T15:23:00.426Z level=ERROR run=6096202d message=process session.id=ses_f67e8f861ffet3LexvzrhZJB0V messageID=msg_09b5d2f96001cqCcVOTuegoCGR error=Aborted stack="AbortError: Aborted
    at new DOMException (node:internal/per_context/domexception:76:18)
    at Array.<anonymous> (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:489828:27)
    at Generator.next (<anonymous>)
    at ~effect/Effect/successCont (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7896:29)
    at ~effect/Effect/evaluate (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7909:25)
    at FiberImpl.runLoop (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7716:112)
    at FiberImpl.evaluate (file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7686:26)
    at file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:7854:19
    at file:///C:/Users/<redacted>/AppData/Local/Programs/@opencode-aidesktop/resources/app.asar/out/main/chunks/node-BIz0dqH1.js:6602:55
    at process.processTicksAndRejections (node:internal/process/task_queues:104:5)"

Expected: transient provider/network stream failures should be retried automatically, or at least surfaced as a visible recoverable error instead of a silent "interrupted" state.

Plugins

none

OpenCode version

1.18.30 (Desktop, Windows)

Steps to reproduce

1. Start a session in OpenCode Desktop and give the agent a task
2. Occasionally (randomly), the turn stops and shows as interrupted with no error in the UI
3. No ESC pressed, no message submitted at that moment
4. Sending a new message resumes work normally

Screenshot and/or share link

Operating System

Windows 11

Terminal

N/A - OpenCode Desktop

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Aborted provider stream recorded as clean stop (finish=unknown, zero usage, no text) — subagent returns empty with no error #37852: Similar symptom — provider stream aborts mid-turn with no retry and no visible error surfaced to the user. That issue covers the case where the stream ends without a finish reason; this issue covers an AbortError that silently sets interrupted state. The two issues share the same gap: transient stream failures are not retried or surfaced as recoverable errors.

[xylophoneengine]

I don't think this really is a duplicate issue. On my end the gateway itself seems to be down or very degraded:

curl -sS -o /dev/null -w "HTTP %{http_code} in oftime_total}s (connect time_connect}s) \n" -connect-timeout 10 --max-time 30 https://opencode.ai/zen/go/v1/model

curl: (28) Connection timed out after 30002 milliseconds
HTTP 000 in oftime_total}s (connect time_connect}s)

the model selection request even times out completely. Using my Zen API code in Hermes also leads to timeouts and when it works 2 tokens per second when trying to say "Hello". Using opencode the tool it seems to work but also way slower than usual with interrupts here and there

[romg974]

Thanks for the triage. Three points:

1. Not a strict duplicate of Aborted provider stream recorded as clean stop (finish=unknown, zero usage, no text) — subagent returns empty with no error #37852 That report covers a stream ending mid-turn being recorded as a clean stop (finish=unknown, zero usage, no text, nothing logged). This report is the AbortError path: the fiber is interrupted, the turn is marked "interrupted" in the UI, and the error only shows up in the sidecar log. Different failure mode, but same underlying gap — transient stream failures are neither retried nor surfaced as recoverable errors. Happy to have this consolidated under Aborted provider stream recorded as clean stop (finish=unknown, zero usage, no text) — subagent returns empty with no error #37852 (or a shared fix) if the maintainers prefer.

2. On @xylophoneengine's degraded-gateway hypothesis It would explain the trigger on my side too — my logs confirm I'm on the same opencode/opencode-go gateway (providerID=opencode-go, modelID=glm-5.3-flash for the main agent; the gpt-5.4-nano entries are only opencode's internal small model for auto-titling sessions). But the actionable bug on opencode's side stands regardless of gateway health: no automatic retry, and no visible error — just a silent "interrupted" turn.

3. Update on the retryability track (Sessions die on transient socket closes because @ai-sdk/gateway is pinned 87 versions behind the retryability fix #48761 / PR chore(deps): bump @ai-sdk/gateway to 3.0.191 #48710) Sessions die on transient socket closes because @ai-sdk/gateway is pinned 87 versions behind the retryability fix #48761 was just fixed by PR chore(deps): bump @ai-sdk/gateway to 3.0.191 #48710 (merged today, 2026-09-13), which bumps @ai-sdk/gateway to 3.0.191 so response-body network errors are now classified as retryable. My Desktop build (1.18.30) predates that bump, so this AbortError path may be covered by it. I'll retest on the next Desktop release — if random "interrupted" turns still reproduce after the bump, this report stands (the PR covers response-body network errors, not the Effect.onInterrupt fiber path specifically); if it's gone, feel free to close this as fixed by chore(deps): bump @ai-sdk/gateway to 3.0.191 #48710.