TODO tool

[zikajk]

We should have a TODO task so that the LLM can follow a plan and avoid getting lost.
I did some tests and it obviously helps.

The goal is something between the Claude Code Todo tool and the ScratchPad in ClojureMCP.
That is—to be able to have the TODOs stored (at the workspace hash level), but at the same time, not have them as deeply nested as they are in ClojureMCP.

I plan to start with creating a tool and improving the system prompt for both AGENT and PLAN behavior.
TODO will be saved in JSON transit for the workspace hash.

To upvote this issue, give it a thumbs up. See this list for the most upvoted issues.

[ericdallo]

Sounds like a good feature @zikajk ! I believe it will be a matter of prompt tune a lot, but LMK if you need help with the front / protocol later so we can pin the TODO current stage in the header or something, dunno

[ericdallo]

I believe we can indeed start in steps, like having the tool first + prompt tune, then discuss what's the UX we want to offer

[zikajk]

I didn't think about any protocol changes.
Right now it would be just a tool call.
But I agree that it might be nice to have a better UX for it (e.g. Antigravity has TODO in its own panel) later.

[ericdallo]

@zikajk yeah, for now it works great, but I think ECA's protocol already fit that possibility, via the custom details in the toolCall messages, which we can create a new one for the TODO, and when clients receive that, we display however we think it's the best :)