Question: How are we supposed to update a plan before the implementation phase?

[d3ming]

I had conductor start a new track with a plan document, but upon review I wanted to give more feedback. Assuming it's a non-trivial feature, I can imagine there could be a round of brainstorming/ interactive back and forth before we have a good plan in place. However, when I gave it the feedback and asked the Gemini CLI to update the plan, it went ahead and started implementing. So I don't have a chance to even try the conductor:implement feature.

I'm just curious on how you guys see the workflow work for this case, and in general how should one manage work between conductor and non-conductor prompts?

Thanks!

[sherzat3]

Thanks for the feedback! To make sure I understand correctly: are you looking for an interactive questionnaire session for plan.md—similar to how product.md and tech-stack.md work—to allow for brainstorming before implementation begins?

[d3ming]

@sherzat3 I think that's part of it and I think that would be a good feature to have.

When I was writing this issue, I was mostly just confused at the "right" way to use conductor since it started to implement the plan before I felt I fully signed off on it. Having a questionnaire mode or some other way for me to collaborate more in the planning phase would be great.

I might have been on an earlier version of conductor too, it seemed to have auto-updated now and is doing a much better job deciding when to open a new track vs. edit the plan for the existing track.

[sherzat3]

@d3ming That makes a lot of sense, and it’s a great idea. We could introduce a pause to present a summary of the drafted plan.md and ask if the user would like to make any changes using natural language.

We're still evaluating whether a structured questionnaire is necessary for plan.md. Based on the flow I described above, would you strongly prefer a questionnaire, or would a natural language feedback loop be sufficient for your brainstorming process?

[d3ming]

Hey @sherzat3, I'm not sure I have strong preferences. Questionnaires could work as long as one of the choices is to enter freeform text, I notice that's a pattern you guys already use so that seems safe. Could be interesting as a config option later, my guess is that for most users questionnaire is simpler though as you won't need to type as much - so long as the AI can give good suggestions that have high hit rate.