I would like if it can update todo and check what is done or in progress - and make it have sub-todo ( so we can see the phase we creating until it done ) , and if can , make sure it can be edited before accept.
Has anyone encountered their model not switching when you click āBuildā button on a plan?
I use separate models for planning and agentic workflows. Gemini 2.5 Pro for plan mode and Claude Sonnet 4.5 for agent. However after I create a plan, when I click build - it switches to Agent mode but continues to use Gemini for implementing the plan - which is not what I want.
Therefore, I have to manually switch to agent mode before clicking on the Build button so that my Claude model gets used.
2 Likes
Last week I became addicted to plan mode. I honestly think itās a great feature, even though itās not very stable yet. Iām looking forward to more updates and improvements.
Iāve reported a similar bug here:
Could you attach evidence?
1 Like
Model selector on āBuildā button @condor ?
I found it great, but I came across the same problem as @RafeSacks mentioned - although the tasks were complete, the to-do list was not updated. Agent was telling me to do it manually, I couldnāt do it. When switching to Markdown, the to-do part is not there. The file ends before it. On top of it Agent could not find the file and update it, which was a bit odd.
Overall, I feel that it keeps updating the same .md file. Perhaps it would be good to have a log of previous plans, so we can keep track of the steps made along the way.
In my mind, the plan should not be just next generation advanced todo plan. It should be more like system plan, describing the system and what it does and how is it built. From that it should take and create actual todos to implement it.
I have not tested it long enough to do it, but this far all I can see others (and myself) doing with it is the advanced todo planning. I wonder if it can be bend to that purpose, where it would be most useful 
I would love to have standard templates for describing the system instead of just randomly starting from some corner and hoping it gets done.
Wow. Alibaba got in the ring too?
edit: Lol. What is this? Did Cursor copy from them or did they copy from Cursor?
1 Like
as far as I get, there is an issue changing provider (it has to reupload and burn tokens if provider changes (antropic/openai ex)
though I can see you may save a plan to workspace, then reference it in a chat with different model
Iād like to specify output dir when saving plan to workspace, and keeping authoritative plan.md and todos.md files in sync while plan is executed (i work from multiple locations and workstations). see Plan mode save to workspace workaround and Plan mode - ToDo list does not update. More fine-grained contol over todos would be a big step-up 
Keep up the good work!
1 Like
I change models by:
- Click stop
- Change model
- Send prompt: ācontinueā
ā¦seems to work, just annoying. Note I can change between some models when still in plan mode, like GPT-5 high ā GPT-5 codex.
1 Like
Has anyone had extreme slow-down issues with the plan preview?
- Giving it focus freezes Cursor for 3 seconds+, but not in markdown mode
- A few times I had to close it and click the chat link to reopen to see changes (though that might have just been the freezing and I hadnāt noticed it yet)
I donāt see resources spiking for my computer. A bit of CPU, but not enough to cause general performance issues.
Working on performance fixes for Plan!
1 Like
Is there any way to resolve the issue of not being written after requesting in PLAN mode?
One very short rule Iāve added to my user rules that has made plan mode a bit more enjoyable, specially using Sonnet 4.5:
# Plan Mode Rules
1. **Avoid follow-up questions**
WHEN: Ambiguities are detected in user request/task
FIRST: try to resolve them through codebase navigation / understanding and web research.
THEN: if doubts still remain -> ASK the user follow up questions.
2. **Plan write-up and delivery**
PRE: Define a objective evaluation of confidence, as weighted measurements of:
`codebaseUnderstading` + `taskDetailQuality` + `probabilityOfScopeFulfilment` + `externalLibraryUnderstanding` - `hallucinationProbability` - `probabilityOfBugsIntroduced`
WHEN: Iterative research and codebase understanding is complete
AND: Confidence score is high
THEN: Write plan, **always** include markdown formatted `[] tasks`.
1 Like
Pretty sure plan mode has the exact opposite prompt. Itās for wishi washi prompting.
Itās got a comment on āambiguity requiring clarificationā. Thatās why my prompt exploits the idea instead of contradicting it. My prompt does not ask the model to ignore ambiguity but to be proactive in resolving it first. Itās important to not contradict goals - complementing instructions are not divergent ones
