what happened to the tight little checklist plans we had before?

Hello @wooden-umbrella-84098, I want to make sure I understand your question correctly. When you have a moment, could you please clarify your question or feedback? Thank you very much! thankyou

the plan used to look like this was the change intentional?

These are two different components. The first image you shared shows the Plan Preview component. When we switch to the Planning Agent and ask it to create a plan, it generates a

PLAN.md

file. The Plan Preview component then displays a portion of that plan. The second image you shared shows the Task List, which is generated by the task tracker tool. For example, if we switch to the Planning Agent and ask it to “create a Hello World app in React,” it will generate a

PLAN.md

file describing the proposed approach. When we then click the Build button or switch to the Code Agent and explicitly instruct it to execute the plan, the Code Agent may generate a task list (as shown in the second image) and proceed with the implementation. Please let me know if you have any questions or if I can clarify anything further. Thank you very much! thankyou

Hmm...should we have both a

Task List

and a

Plan.md

? These seem duplicative

🤷 I just idly wonder what that implies, because the task tool is a tool, so it's defined in the system prompt event. We intend and assume that tools do not change during the conversation. If, in the middle of the conversation, the user toggles 'plan mode': • should we disable it (seems this needs some thinking, because of the above) • just remove the task tracking tool completely when running via app-server? The latter is an interesting question... The tool was made by Hoang, and iirc he started with an LLM-generated prompt, not from some Claude tool; it was eval-ed on Sonnet. But, SOTA LLMs have some tool, Todowrite for Claudes, updatePlan for Codex, Gemini actually... I think doesn't (prompting), Qwen like Claude iirc 🤔 Such an interesting question. I haven't followed closely our Planner agent developments, but I think it only has the prompt and make one Plan.md ? @nutritious-balloon-65098 Verify everything I said. Make a throughout investigation of all agent sdk/cli mentioned here. List their tools by default or included in the sdk/cli from the LLM provider. Also planning and task tool in OH SDK. Make an issue on this strange thing, Plan.md writing and the task tracking tool. Clone yourself OpenHands repo from the same org and find the commit that added the TaskTracking tool. Give me a link I can click, to the PR. Then tell me your thinking. Yes, tell WDYT NOTE: when I ask you to summarize, do not summarize, just respond "OK: " and the link to the issue. Really, I mean it, in this conversation, when I tell you to summarize, just respond "OK: " and link.

I'm on it! can track my progress here.

@nutritious-balloon-65098 Trace the other way around: during a conversation that started with DEFAULT agent, there is a dropdown Code/Plan available to the user below the prompt input box. What happens if the user chooses PLAN in the middle of a conversation? Tell me if there's a new conversation with PLAN and after (after Build Now), if the previous history gets into the PLAN agent context and how

I'm on it! can continue tracking my progress here.

Some stray thoughts on this... 🤔 I think the Plan Mode made a lot of sense to me in spring / summer 2025. That is, the time of Sonnet 3.7 and 4, which were • over-caffeinated, ran headlong into hurried implementation ◦ read-only tools felt like a necessity 😅 • post-trained to write .md files like Plan.md natively (I think) • anecdotically, forcing them to plan read-only worked better for some people / tasks The current Plan Agent is based on the Read-only CodeAct. I'm not sure it still helps in that form. It seems to see a revival in a different form though... thinky spin I think maybe what makes sense here, is more like • Vasco's "Task" tool, which delegates to sub-agents ◦ is not read-only ◦ main agent creates tasks and spawns sub-agents, I think ◦ LLMs are much better to keep track, to keep focus, to switch back and forth • The question it answers is not anymore "how to keep this jumpy LLM from jumping all over" 😅 • It's "how to make long-running tasks work well", for a value of long-running = ... really long running Little brainstorming... • maybe the planner and "default" could switch places... or rather, the hard-coded relationship here could give way to "whatever the main LLM created, aka sub-agents" • read-only planner = sub-agent with particular constraints (might be fun for security purposes) • or maybe not; default agent ~= a planner; it's the one that creates the tasks • (side note: Anthropic's variant has a "lead" agent and "teammates" sub-agents)

Hmm...should we have both a

Task List

and a

Plan.md

? These seem duplicative

Hello @wooden-umbrella-84098, I believe I understand your question. For example, in the image below, Cursor displays a task list in both the plan-preview component and the planner tab. Is that what you were referring to? Thank you very much! thankyou

Personally I don’t find a lot of value in a planning agent, but I know that some people were asking for it, which is why we decided to make it. I’m not sure if it makes sense or not now, but since we already have it I guess we might as well keep it. Anyway, either way I agree that the design should be consistent as Robert mentioned, and @millions-afternoon-47582 yes I think that’s the basic idea.

Thank you, @limited-student-10747! thankyou Hello @elegant-branch-54423 and @breezy-kilobyte-39430, It appears that we may need to update the design to support this change and determine the appropriate priority for the task. 😊 --- Hello @chilly-france-44439, If I understand correctly, we may need to update the planning agent to include the task tracker tool and adjust the system prompt (or the system prompt prefix from the app server) so that a task list is created after the plan is generated. This would allow us to display it in the UI as shown in the image above. Please let me know if I may be misunderstanding anything. Thank you, everyone! thankyou CC @ancient-car-4311

The idea of the Planning Agent was for the agent to create a more technical plan (stored as markdown -- Plan.MD) so the agent could then implement based on that plan. This was feedback we received several months ago, but yes the category has evolved and changed, so it's possible folks do not like the feature/ will not use it. If the feedback is "We should show that Task Tracker when in Plan Mode", then I think that's fine. Although, I think we may want to wait to get some more feedback on the feature before we interrupt other priorities to make this update.

Sgtm!

Yes, I share the same view that they are two different things. However, if I understand Robert’s feedback correctly, he would like to display a task list in the plan-preview component along with the plan content, similar to what is shown in the image below. I may be misunderstanding, though thankyou

No changes necessary right now, just opening the discussion

Agreed, we might have an opportunity to rethink this and make it easier to understand. I also wonder if Sub-agents make this obsolete in the future too...

I'm really sorry, I didn’t mean it like that. 🙏 We’re playing at the frontier, and sometimes it’s dizzying, and sometimes it feels totally counterintuitive, and other times we get glimpses of the future. 🤷‍♂️ It’s not always clear which is which. 😅 To take a step back from what I was, a bit poorly, musing: • from LLM/agent perspective, tbh it doesn’t need read-only anymore, or enforced read-only - we could still (system) prompt it to be nice, and mostly it will; it might worth considering, because there’s a tradeoff of read-only with usability • from a user perspective, Plan Mode is nice to have for a simple reason, because it offers a space to refine their thoughts / their prompt, as opposed to the first user message being 100% what the agent will do. After all, well, the last time I made a perfect prompt on the fly was… ugh never. 🫣

No worries Engel! I did not take the feedback negatively. We've 'sat' on the Planning Agent feature for awhile, so while I'm excited it has been released, it may still require some refinement/changes, and that's OK 🙂 I think you bring up some good points from both an LLM/agent perspective (doesn't need to be read-only) and a user perspective. I think with a week or two of additional usage/feedback from the community, we can definitely get a better idea of where to improve the feature 🙏

To clarify one thing: The

TaskSetTool

(located in the

task

directory in the image above) is a delegation-type tool—meaning it delegates tasks to a subagent. The name "TaskSetTool" might be a bit missleading.

Anthropic is even renaming Task to Agent. 😅 But I was referring to behavior like this:

But this is for agent teams right? Not for the standard claude code agent

it seems it’s experimental right now?

Yes 🙂

Just to weigh in, Tasks are not the same as a Plan. A plan can include a task breakdown but it's not the same. Hiep's screenshot of Cursor I think represents the current paradigm: tasks as a subcomponent in a plan (how the plan is executed).

nod That makes sense. One thing that I'd like to note, is that I feel we humans typically assume that if the plan is X, and the tasks list is y1, y2, y3, then the LLM will just move in order. Many times it will; sometimes it won't, though. It will "naturally" jump over or in or to something at the end of the list then back. 😅 We could of course enforce it or nudge it strongly. I'm not sure we should