Add ability to iterate on a TUI (like v0.dev)

[shobrook]

No description provided.

[apockill]

I'm working on something to address #3, do you have any preferences in how to do it?

I was thinking instead of saving the script to .termite/####_script_name.py, there's a project directory for each script.

.
└── .termite
    ├── another_cool_tool
    │   ├── metadata.json
    │   └── tui.py
    └── cool_tool
        ├── metadata.json
        └── tui.py

When termite is called, it could create a dto a la:

@dataclass
class Project:
	_metadata_path: Path
    _tui_path: Path

    @property
    def script(self) -> str:
         ...
   
    @property 
    def metadata(self) -> Metadata:
        ...

    @staticmethod
    def from_dir() -> "Project":
        ...

    @staticmethod
    def create(script, design) -> "Project":
         ...

Then metadata can contain converstations and context from past runs.

If the user calls termite <prompt> --name "cool_tool" it can use that name. In the future, instead of ####_script_name/ an LLM can generate a useful name for the tool.

To run a tool, perhaps termite --run cool_tool would suffice.

---

Let me know what you think, I'm happy to implement this. Cheers!

[shobrook]

This is a good idea. Only thing I'm wondering is how the actual UI should look for iterating on a TUI. In v0.dev, the chat is side-by-side with the output. That might not be so doable in the terminal. Any ideas?

[apockill]

It's a good point. In this application, I suppose you don't really care to see what the LLM has to say, only what the TUI looks like. So maybe the workflow is:

What do you want to make? (Ctrl-C to exit)
> Make a GPU monitor using nvidia-smi

<The usual loading bars...>
Press `q` to exit your TUI and return to the conversation
<TUI pops up>

User presses q

What do you want to make? (Ctrl-C to exit)
> Make a GPU monitor using nvidia-smi
What would you like to change?
> It doesn't seem to be refreshing live, there must be some kind of issue with your callbacks ... blah blah ...

Then it jumps back to loading bars / showing the TUI. Thoughts? On the next repitition it might show the same prompt:

What do you want to make? (Ctrl-C to exit)
> Make a GPU monitor using nvidia-smi
What would you like to change?
> It doesn't seem to be refreshing live, there must be some kind of issue with your callbacks ... blah blah ...
What would you like to change?
> Add more colors, I want it to be NVIDIA-green themed

etc etc

[shobrook]

That makes sense. How do we guarantee that pressing q will always quit the TUI and return the user to the chat session?

[apockill]

That I do not know 😁

It might be easier to just have Ctrl+C exit the TUI, we catch the KeyboardException in the main chat loop, and then go back to prompting. If they then Ctrl+C while in the chat loop, it exits the program. That's nice from a UX standpoint, it keeps the "exit" mechanics the same.

If you like the general proposal, I can play around with implementations and see if I can experiment with possible ways to exit the TUI reliably.

[shobrook]

That seems clean and simple. We should make sure that Ctrl+C and q (if implemented) behave the same way, though.

[shobrook]

Also, how do you think issue #6 should be supported in this new workflow? Maybe just a simple Y/n input prompt before opening the TUI?

[silopolis]

Hello there :)

First, congrats and thanks for sharing this pretty exciting tool 🙏

After giving it a few tries, this is indeed the first feature I thought I'd like because it is not that simple to drop the perfect prompt the first time and of course, each new "answer" seeds new ideas... and bugs 😅

The second, which is related to things discussed here is that I wished Termite would:

• use the current directory (or one specified on the CLI) as the root of the Python project (with all usual conventions and good practices) it is assisting in building,
• use ./.termite/ to store its special sauce (config, metadata, prompts/conversations histories, design document, etc),
• generate code and project files (deps, venv) in CWD instead of burying it deep into $HOME

TY
J

[shobrook]

I'd merge any PR that allows iteration on the TUI.

And as for your other ideas, those also sound good to me, although much lower priority.

[chrisdlees]

I dunno about you guys, but I'm thinking that the newest "thinking" models might replace an agentic approach.... I may just have to try a stab at PR myself

[shobrook]

I wouldn't call Termite an agent, but yes I think using a reasoning model would improve results. @chrisdlees If you give this a go, I'd recommend using LiteLLM.