[Bug] Todo list does not update incrementally during long-running tasks

[anhyxdy]

Description

When running a long-running, multi-step task in DeepSeek Harness, the Todo list does not update incrementally during execution.

The agent may generate a complete Todo list at the beginning, but completed items remain in_progress or pending in the UI while the task is still running. After the task finishes, the Todo list suddenly updates and multiple items are marked as completed at once.

This makes the task progress shown in the UI inaccurate and difficult to follow.

Environment

• OS: Windows 11
• DeepSeek Harness: 0.1.2-alpha.5
• Mode: Web UI
• Task type: Long-running, multi-step tasks

Steps to Reproduce

1. Start a long-running task involving multiple sequential steps.
2. Let the agent generate a Todo list with several items.
3. Observe the Todo list while the agent is executing the task.
4. Wait for several individual steps to finish.
5. Observe the Todo list again.

Actual Behavior

• Todo items are not marked as completed immediately after the corresponding step finishes.
• The UI may remain unchanged for a long period while the agent continues working.
• After the task finishes, multiple Todo items may suddenly be marked as completed at once.
• The displayed Todo progress therefore does not reflect the actual execution progress.

Screenshot 1 — Todo list during execution

The task is still running, but the Todo list does not reflect the completed steps.

Screenshot 2 — Todo list after task completion

After the task finishes, multiple Todo items are updated at once immediately.

Expected Behavior

The Todo list should update incrementally during task execution:

• Each Todo item should be marked as completed shortly after the corresponding step is actually completed.
• The UI should reflect the agent's current progress in near real time.
• Todo completion should not be batched until the entire task finishes.

Impact

For long-running tasks, the current behavior makes it difficult to determine:

• Which steps have actually been completed.
• Which step the agent is currently working on.
• Whether the agent is making progress or is stuck.

Suggested Fix

Consider updating/pushing the Todo state whenever an individual step changes status, rather than only updating the UI after the task or Todo write operation finishes.

[anhyxdy]

Related Discussions

#5456 — Agent does not mark the final todo item as complete after multi-step tasks
Focuses on the last todo staying spinning after task completion. My issue differs: todos do not update incrementally; multiple finished subtasks are only marked complete when the whole task ends.

#2575 — Task list does not update while tasks are running — no mid-run push, subagent list is pull-only
Closely related. It also mentions missing real-time progress updates. My report shows the todo list lags and batch-updates once the task finishes.

This may share the same progress sync root cause, but specifically covers missing incremental todo state updates during execution.

Relationship between related issues

[anhyxdy]

描述

在 DeepSeek Harness 运行长耗时多步骤任务时，Todo 列表无法在任务执行过程中增量刷新。
Agent 初始会生成完整 Todo 清单，但已完成条目在界面上仍显示为进行中或待处理。仅当任务全部结束后，才会一次性批量更新多条 Todo 的完成状态。
这会导致界面进度失真，不便跟踪任务。

环境

• 系统：Windows 11
• DeepSeek Harness：0.1.2-alpha.5
• 界面：Web UI
• 任务类型：长耗时多步骤任务

复现步骤

1. 启动多步骤长任务
2. 让 Agent 生成 Todo 清单
3. 任务运行时观察 Todo 列表
4. 等待若干子步骤执行完毕
5. 再次查看 Todo 状态

实际表现

• 步骤完成后，对应 Todo 不会立刻标记为已完成
• Agent 运行期间界面长时间无变化
• 任务结束后多条 Todo 一次性更新为完成
• UI 进度无法反映真实执行情况

预期行为

任务运行期间 Todo 列表应增量更新：

• 每个步骤完成后，及时将对应 Todo 标记为已完成
• UI 近乎实时展示 Agent 当前进度
• 不要等到整个任务结束才批量更新 Todo 状态

影响

长任务场景下难以判断：

• 哪些步骤已经完成
• Agent 当前执行到哪一步
• 任务是正常推进还是卡住

建议修复方案

每当单个步骤状态变更时，推送 Todo 状态更新，而非等到任务或 Todo 写入操作结束后才刷新前端界面。

相关讨论

#5456 — Agent does not mark the final todo item as complete after multi-step tasks
该讨论描述任务结束后最后一条 Todo 持续转圈。本问题不同：任务运行时 Todo 不会增量刷新，多个子任务需等到整体任务结束才标记完成。
#2575 — Task list does not update while tasks are running — no mid-run push, subagent list is pull-only
高度相关，同样提到运行时缺少实时进度推送。本案例表现为 Todo 进度滞后，任务结束后批量刷新。
根源可能是同一个进度同步问题，但本问题重点是执行过程中缺少 Todo 的增量状态更新。

[PerryLink]

I checked the premise against the source, and it does not hold: the push-based
todo pipeline you ask for already exists — it existed in your own 0.1.2-alpha.5,
and it is unchanged on master (c291e79).

Verified on master:

1. Every todo_write call appends a whole-list todo/write snapshot to the
   session immediately (packages/todo/tool-todo/src/index.ts:210).
2. The todos projection folds every committed event eagerly: apply at
   tool-todo/src/index.ts:138-140 ("latest whole todo/write list"); the
   sessionProjections registry subscribes to session/event once and drives every
   registered unit per committed event (packages/extensions/tool-cordis/src/
   api-catalog.ts:1669-1671), and changed views are pushed to the client.
3. The TodoPanel reads that projection live via useProjection('todos')
   (packages/client/ui-conversation/src/client/skeleton/TodoPanel.tsx:127-130).
4. There is an e2e test proving mid-session live rendering: it appends a
   todo/write to a live session and expects the panel within 10s
   (apps/web/tests/queue-actions.e2e.ts:304-310).
5. 0.1.2-alpha.5 (tagged 2026-09-02) already contained both: TodoPanel.tsx and
   the tool-todo projection exist at tag dsh-v0.1.2-alpha.5, and no todo-related
   fix landed between that tag and master.

So the panel always renders the latest list the model wrote. If the model calls
todo_write once at the start and once at the end (marking several items
completed in that final write), the UI shows exactly that — several items
flipping to completed at once. There are no intermediate todo/write events to
render; the UI cannot display step completion the model never reported. This is
model cadence, not a push bug.

How to confirm, and what to change:

• Count todo/write events in the session log (session.v3.jsonl.zstd) for a
  long task. 1–2 writes over a long multi-step task = model cadence.
• Instruct the model: "after each step completes, call todo_write with updated
  statuses." Related [Bug] Agent does not mark the final todo item as complete after multi-step tasks #5456 (the final item stays spinning) is the same family —
  model-side todo hygiene, not the UI pipeline.

（中文摘要：master 与 0.1.2-alpha.5 的 Todo 面板本来就是事件推送式——
todo/write → 'todos' 投影 → useProjection 实时渲染，有 e2e 测试证明运行中即时
更新。你看到的"结束时批量完成"是模型只在最后一次性写 todo 列表造成的，UI
无法显示模型没有上报的中间完成。验证方法：数日志里 todo/write 的次数；对策：
明确要求模型每完成一步就调用 todo_write 更新状态。）