[BUG]任务已全部输出完成，任务列表最后一步仍显示进行中

[tzy168]

摘要

任务已全部输出完成，任务列表最后一步仍显示进行中

预期结果

任务列表显示全部已完成，无进行中状态

详情 / 复现步骤

1. 输入对话信息
2. LLM输出计划列表，页面展示出来，LLM输出完成后，任务列表仍显示进行中

环境信息

• dsh-v0.1.0-rc.7 master
• Chorme
• win10

截图

2.

解决方案

应该是todo_write最后一步计划在LLM输出完后，没有再继续调用todo_write，导致状态未更新
应当虑优化此tooL的调用提示词

[denial123789]

I traced this through rc.7 (99f6f02). The Web UI is normally rendering the last durable state correctly rather than failing to stop an animation:

• every successful todo_write appends a complete todo/write snapshot;
• the todos projection is last-write-wins;
• turn/end intentionally retains that snapshot;
• only the next turn/start clears the standing plan.

So if the model sends final prose without a final todo_write, the last in_progress value remains visible by design. Also, turn/end: completed means the last assistant message requested no more tools; it does not prove every external task succeeded. Automatically converting all active items on turn end would misreport aborted, blocked, errored, max-token, interrupted, and prematurely claimed work.

The existing tool description already asks the model to mark work completed immediately, so prompt reinforcement is useful but probabilistic. A stronger design is a bounded agent/turn-stopping guard: inspect the latest snapshot, and when active items remain, issue one logged agent.steer() reminder so the model verifies outcomes and explicitly calls todo_write. Cap it at one continuation per turn and never mutate statuses silently.

I documented the event-order diagnosis, both Web surfaces, safe recovery, proposed guard, and regression matrix here: https://sandbaseai.github.io/deepseek-harness-handbook/todo-stuck-in-progress.html