Replies: 4 comments
|
try switching models it will help |
|
在 master(c291e7961)核对:恢复机制存在,但你的"0 次 compaction/start"是四个静默出口之一造成的,且失败都发生在事务开始之前。 机制: 四个静默出口:
建议:先确认 preset 挂载;再看会话日志里是否出现过 |
|
感谢 @PerryLink 的逐行核对与缓解建议。补三组现场数据: 1) 预设挂载确认(你标为"第一优先确认项"的那项):已挂载。我们实测中 host 平面与 preset 隔离域两个 2) 在"静默出口"里,还有一个比阈值逻辑更早的位置(两条自动腿+手动腿都会中): 3) 修复后现场(供验证参考):把阈值调至 0.65(我们改的是 preset 行配置;你提的 建议(供参考):核心侧按自身文档做降级守卫( 需要 CDP 捕获原文/逐步复现,随时可提供。 |
|
Complementary case on the same harness version, from the opposite direction — recovery ran, repeatedly, and could not succeed.
39 of the 59 are exact equality, and 48 attempts happened inside a single turn. So where your session shows zero attempts, this one shows attempts that are structurally incapable of succeeding: the selected span is the checkpoint the previous compaction wrote (19 of 20 successful compactions selected a one-node span equal to the prior checkpoint). Details: #7212 There is also a second reason recovery can be skipped in this session, which may be relevant to your diagnosis of "no if (contextWindow && message.stopReason === "length" && message.usage.output === 0) { … }One token and the check misses, so Worth recording in the session log either way, per your question 2: both failure modes are silent. In this session the only trace of the 59 rejections is Evidence: https://gist.github.com/gorban/4725c87390cb391047a2f14b56b98d28 |
Uh oh!
There was an error while loading. Please reload this page.
现象
会话在收到 provider 确认的
CONTEXT_WINDOW_EXCEEDED后,未记录任何压缩尝试;后续每次唤醒/重试仍以同类报错结束(同一会话累计 8 次CONTEXT_WINDOW_EXCEEDED),直至turn/end.reason.kind=error终止,会话进入"每次唤醒都撞同一面墙"的状态。环境 / 证据
CONTEXT_WINDOW_EXCEEDED×8;maximum context×8compaction/start= 0;compaction/summary= 0;compacted-summary= 0requested 1048596 = 792596 messages + 256000 completion)→ 至 turn 44 仍为同类报错后 terminated(另一样本会话序列一致)。期望 vs 实测
compaction-basicREADME 描述:agent/request-error监听器对 provider 确认的CONTEXT_WINDOW_EXCEEDED会绕过常规阈值与保留策略、做一次最大平衡头部缩减,并在 surface replacement generation 前进后授权重试。实测:未观察到任何压缩尝试落地(无
compaction/start),会话直接死亡。疑问
影响
长会话一旦错过压缩窗口即进入死亡循环(每次唤醒都撞同一面墙),只能靠外部重建会话恢复;对长驻 agent 工作流影响大。
All reactions