You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
One concrete cause of "stuck on 思考中…" with the activity list showing everything complete, filed separately as #345 with a provider-free repro.
The last expectation in this issue is "失败或超时时明确提示,而不是一直停在「思考中…」". That case is reachable without any provider being at fault: the model gateway applies a hard-coded 120s read timeout (model_gateway.py:127). If the provider's first byte takes longer, the gateway returns 502, the bundled OpenCode runtime re-sends the same body, and nothing terminal ever reaches the panel.
Trace of one turn against a stub endpoint that accepts the request and never answers (<project>/agent-runs/<turn>.model.jsonl):
user status=complete text='你好'
assistant status=interrupted text=''
So the persisted status is already interrupted and AssistantPanel.tsx:490 renders a label for it, but the panel only picks that up on reload. During the live turn the user sees the earlier tool calls marked complete and an indefinite "thinking" line, exactly as described in the report. Repro script and charts: #345.
How this relates to #324: that PR collapses the repeated opendirector job wait rows and fails the activity when the job fails, which covers the long-wait half of this issue. The timeout path above is a different half — the turn itself never ends — and is not addressed by it. The remaining items listed here (step labels, streamed plan, elapsed time) are separate again.
I measured first-byte latency for three providers before filing, to check whether 120s is a realistic ceiling rather than a theoretical one: DeepSeek deepseek-v4-pro median 0.2s, Zhipu glm-5 median 1.0s, Moonshot kimi-k3 median 13.3s with 4 of 12 calls failing to connect at all; in an earlier window the same kimi-k3 body returned its first byte after 73-180s, 2 of 6 above the timeout. Numbers and method are in #345. I hold keys for those three only, so self-hosted vLLM, proxied endpoints and long-context reasoning models remain unmeasured.
Happy to take the timeout path as a PR if maintainers agree with the direction proposed in #345; the repro needs no paid credentials, so it can carry a regression test that fails on the current revision.
Activity
Birfy commentedon Nov 6, 2026
后续证据(2026-11-06 约 16:30–16:58 Asia/Shanghai)
同一项目「未命名项目 2」、用户发「继续」后的助手 run
d8713e62-3ff8-453e-921f-114d30b982b9:opendirector job wait就有约 586 条工具事件(约 293 次 complete),针对同一个视频 job56b8d594-…status: running/phase: poll,界面几乎看不到「还在等供应商 / 已等多久 / 失败原因」providerCode: 1027),工具状态仍记为complete(无 error 字段),更不利于用户理解相关新开:#136(1026/1027 文案)、#137(长 wait 后会话过期)。
[bugfix] agent: localize job-wait activity errors
[bugfix] agent: show one job wait as one activity and fail it when th…
[bugfix] agent: localize job-wait activity errors
Zhangbo-cn commentedon Nov 16, 2026
One concrete cause of "stuck on 思考中…" with the activity list showing everything complete, filed separately as #345 with a provider-free repro.
The last expectation in this issue is "失败或超时时明确提示,而不是一直停在「思考中…」". That case is reachable without any provider being at fault: the model gateway applies a hard-coded 120s read timeout (
model_gateway.py:127). If the provider's first byte takes longer, the gateway returns 502, the bundled OpenCode runtime re-sends the same body, and nothing terminal ever reaches the panel.Trace of one turn against a stub endpoint that accepts the request and never answers (
<project>/agent-runs/<turn>.model.jsonl):Stored conversation afterwards:
So the persisted status is already
interruptedandAssistantPanel.tsx:490renders a label for it, but the panel only picks that up on reload. During the live turn the user sees the earlier tool calls marked complete and an indefinite "thinking" line, exactly as described in the report. Repro script and charts: #345.How this relates to #324: that PR collapses the repeated
opendirector job waitrows and fails the activity when the job fails, which covers the long-wait half of this issue. The timeout path above is a different half — the turn itself never ends — and is not addressed by it. The remaining items listed here (step labels, streamed plan, elapsed time) are separate again.I measured first-byte latency for three providers before filing, to check whether 120s is a realistic ceiling rather than a theoretical one: DeepSeek
deepseek-v4-promedian 0.2s, Zhipuglm-5median 1.0s, Moonshotkimi-k3median 13.3s with 4 of 12 calls failing to connect at all; in an earlier window the samekimi-k3body returned its first byte after 73-180s, 2 of 6 above the timeout. Numbers and method are in #345. I hold keys for those three only, so self-hosted vLLM, proxied endpoints and long-context reasoning models remain unmeasured.Happy to take the timeout path as a PR if maintainers agree with the direction proposed in #345; the repro needs no paid credentials, so it can carry a regression test that fails on the current revision.