Repository navigation
[BUG] Desktop: conversation freezes mid-"thinking", no error, must abort and resend — missing watchdog/retry UX #41365
Description
Description
On the Windows desktop app, quite often — especially with models that do long/reasoning "thinking" (e.g. OpenAI/DeepSeek-style reasoning dumps) — the conversation freezes mid-thinking: the agent is stuck in "thinking" state, spinner/tool call never resolves, and no output arrives.
The only workaround is to abort/terminate the conversation and resend the message (losing the reasoning context), which is disruptive for long debugging tasks.
Details of the bug
- User sends a prompt (sometimes after long thinking, sometimes right at start).
- Client shows "thinking…" (or the thinking block dumps) indefinitely; the stream silently stops; no error toast.
- UI (desktop) is stuck; user must hit the abort/stop button and resent the exact same prompt; then it usually works — pointing to a transient stream/connection race, not the model being slow.
Associated issues (other symptoms, same "thinking stuck" family)
- Desktop stuck in "thinking" mode - sidecar stream never completes (v1.17.18, Windows) #36370 (Desktop stuck in "thinking" - sidecar stream never completes, Windows)
- Plan/Build mode sometimes hangs after long reasoning and never emits the final response #37609 (Plan/Build mode sometimes hangs after long reasoning, never emits final response)
- API 404 error silently hangs session in 'thinking' state with no error recovery #38951 (API 404 silently hangs in thinking state, no error recovery)
- bug: UI stuck on 'thinking' indefinitely after stream error, no error displayed or state recovery #32366 (UI stuck on thinking indefinitely after stream error, no error displayed)
- Silent failure: opencode provider returns 500 errors but client drops response without showing error #38644 (provider 500 drops response without showing error)
- Desktop app: "Auto-approve requests" setting toggle cannot be turned on and does not persist #41069 (previously closed as duplicate of Desktop v1.18.3 — "Auto-accept permissions" settings toggle permanently disabled (cursor: not-allowed) #37617 — free-tier model network blips treated as normal; user on this session reported similar)
Suggested improvements
- Watchdog / timeout on the thinking phase: if no chunks received for N seconds (e.g. 60s), surface an explicit error/info banner ("model hasn't responded; retry?") and offer "retry with cached thinking".
- Recover stuck turn automatically: detect stuck state and prompt user "Retry / discard and re-ask" instead of silently dangling.
- Show provider error in the UI (many of these are provider 5xx/empty chunk cases): single toast for "provider error, auto-retried".
- Non-lossy retry: when user aborts+resends, keep the previous reasoning/tool output as context so the retry isn't blind.
This is a frequent occurrence (dozens of similar reports) and the UX cost is high for long sessions.
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon Aug 9, 2026
This issue might be a duplicate of existing issues. Please check:
If your issue is specifically about the missing watchdog/timeout UX and retry mechanism (rather than just the stuck-thinking symptom), it may be worth keeping open as a feature/UX request. Otherwise, please consider commenting on one of the above.
liudongyan13701205717-source commentedon Aug 9, 2026
The auto-bot listed #36370 / #32366 / #38644 as possible duplicates — those cover the symptom (stuck "thinking", silent stream drop, provider errors not surfaced). This issue's distinct value is the UX/proposal side: a watchdog + timeout + explicit retry-with-cached-thinking UX, i.e. (1) detect a stalled turn, (2) offer "retry / discard and re-ask" instead of silent hanging, (3) keep reasoning context on retry. That's a feature request that complements the bug reports, not a duplicate. If maintainers prefer, I can reduce this to just the watchdog/retry spec and let the bug reports own the diagnosis.
mcj777 commentedon Aug 14, 2026
Freeze at ~14% context fill — same "thinking stuck" family, with log evidence
Environment: Windows 11, OpenCode v1.18.18 desktop (local sidecar server), DeepSeek V4 Pro via API, 1M context window,
reasoningEffort: max.Symptom: when a session reaches ~14% context fill (~140K input tokens), the chat and permission dialogs freeze (red server indicator). No error in UI, nothing in the log — the stream to
deepseek-v4-progoes completely silent (20 min in my last capture), followed byAbortedafter a manual Esc. No crash, no dumps. It self-recovers after 5–30 min, or after Esc → resend.What makes this case notable: it reproduces across different sessions at the same context fill percentage (~14%), suggesting the stall correlates with accumulated context size, not just long reasoning. Same family as #37609 / #36370 / #38951, but with a measurable trigger threshold.
Workarounds confirmed: Esc → resend succeeds; starting a new session before ~100K tokens prevents the freeze.
liudongyan13701205717-source commentedon Aug 14, 2026
@mcj777 thanks for the detailed repro — the ~14% context-fill threshold is a great new data point.
This is the same "stuck thinking / silent stream drop" family as #36370, #32366, #38644, #37609, #38951 — but your capture adds something the others lack: a reproducible trigger threshold (~14% / ~140K input tokens on a 1M window), which points at accumulated-context-size handling rather than just long-reasoning or a one-off provider error.
Where this issue stays distinct (and worth keeping open alongside those bug reports): the UX gap — there's no watchdog, no stall timeout, and no explicit "retry / discard-and-re-ask" affordance when a turn silently hangs. Today the only recovery is a manual
Esc → resendthat throws away reasoning context. A minimal, high-value fix would be:The diagnosis belongs to the bug issues; the watchdog/retry UX is the feature request here. Boosting visibility for both.
liudongyan13701205717-source commentedon Aug 27, 2026
I can reproduce the mid-"thinking" freeze on my end as well. Happy to provide local repro logs / context size at freeze time if it helps narrow down the ~14% context trigger mcj777 identified.
liudongyan13701205717-source commentedon Aug 30, 2026
跟进：目前已有两名真实用户复现——mcj777（~14% 上下文填充 ~140K token 阈值触发冻结，跨会话可复现）和我方（Windows Desktop）。症状归属 #36370/#32366/#37609/#38951 等 bug 报告，本 issue 的价值在于非诊断侧：watchdog + 停顿检测 + 显式"重试/丢弃重问"UX，且保留推理上下文重试。
诉求小结（也可作为最小可评审规格）：
维护者若有排期或需补充规格细节，请指示。愿意提供本地复现日志/冻结时的上下文大小。