MoonshotAI / MoonshotAI/kimi-code
中断恢复后模型接错话题(回合无收尾 + turnId 重置 + TodoList 主子共享残留)
Nobody has claimed this yet.
- Dominant language
- TypeScript
- Stars
- 7.5k
- Forks
- 1.2k
- Avg merge
- 11h 53m
- Merged PRs (30d)
- 350
Description
中断恢复后模型接错话题(上下文抓错)+ turnId 重置 + TodoList 残留
环境:Kimi Code CLI 0.36.0,Windows 11,模型 kimi-code/k3
现象
长会话中一次子代理调用被 403(5 小时额度上限)打断,该轮没有产出任何面向用户的总结就结束了。用户随后问了一句"做完了还是?"(指当前任务),模型却把它接到了压缩保留下来的两天前旧话题上,答非所问。
wire.jsonl 证据(session_bc28eb7c-d1a4-4f50-ba8f-68c9966d300a):
- 行 14142:turn 210 step 31,Agent 工具调用(子代理,随后 403)
- 行 14145:用户新输入"做完了还是?"
- 行 14149-14153:回复被标记为 turnId 25(前一事件是 turnId 210——turnId 在压缩/恢复后重置,交错出现)
- 行 14152 模型的 think 自认在"猜"用户问的是旧话题
- 行 14137 vs 14148:两条 TodoList reminder 内容来自不同时代(旧窗口残留 todo 与当前 todo 并存)——子代理的 TodoList 与主 agent 共享 store,子代理中断后留下脏状态
三个可拆的问题
- 任务异常中断(如额度 403)后回合无收尾——没有提示用户"中断了",下一条输入进来时模型要自己猜断点在哪
- turnId 在压缩/恢复后重置,wire 里出现 210 → 25 的交错,给排查制造混乱
- TodoList store 在主子代理间共享,子代理中断后其 todo 残留进主上下文的 reminder
建议
- 子代理/工具异常终止时给用户一个明确的"回合中断"收尾消息
- turnId 保持全局单调(或压缩后带前缀区分时代)
- TodoList 按 agent 隔离,或子代理退出时恢复主 agent 的 todo 快照
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the supplied wire.jsonl evidence for session_bc28eb7c-d1a4-4f50-ba8f-68c9966d300a, especially lines 14142-14153 and the TodoList reminders at lines 14137 and 14148. Trace interruption recovery, turnId assignment after compression, and TodoList ownership between the main agent and sub-agent. Done means interrupted turns are explicitly closed, turn IDs remain distinguishable, and sub-agent todo state does not leak into the main context.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- typescript
- Domain
- cli, tooling
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100