An easy way to discard last generation, and re-edit last prompt to retry

[bew]

Hello!

Maybe I'm missing something 🤔

I'm trying to quickly test out things with the prompt and more often than not I don't like the way the model interpreted what I wanted to say and I want to stop it.
In this case (and others) I'd like to delete the last generation, edit my last prompt a bit and try again.

I've found [[ to jump to last header, but it's only in normal mode (does not work in operator mode or visual mode).
👉 Could you bind [[ / ]] in operator & visual mode so I can easily select / delete, e.g. with d[[

Alternatively, a bit like in most llm web UIs, we can edit an old prompt and it'll discard the response it generated and replace it with a new one.
👉 Could we be able to easily jump between prompts (not headers), edit it and re-run the generation of this prompt (only using previous context, no new ones that are below it)

And if this is too specific (but this would be really cool for the general case as well!)..
👉 Could you expose (and especially document) various functions to get semantic sections of the chat buffer (with line number, roles), information about the current chat/adapter/.. All so we can define our own custom extmark UI or keybindings for example 🤔
(If you're not avoidant of this, I might open a separate issue for this)

[olimorris]

Have you looked into gd? That may do what you need.

[bew]

Hmm but gd gives me a new window, I was mentioning access to details programatically: not for me the human, but for my keymaps / my config.

In the meantime I've found the CodeCompanion.buf_get_chat function to access the chat object, but that doesn't seem to help to find what line my last prompt starts/ends 🤔

[olimorris]

With gd you can "... delete the last generation, edit [your] last prompt" - It exposes the messages table that makes up the response to your LLM and allows you to fully edit the history.

Maybe a more fundamental question, are you trying to test out multiple LLM's responses to a specific prompt? Whereby you determine which LLM gives you a response you like and then perhaps you go into a more deeper conversation with that LLM?

It's easier to try and understand what you're trying to accomplish than work around your understanding of the plugin.

[0xdapper]

this is interesting but is it possible to directly just edit/delete sections in the chat buffer that will edit/remove the relevant messages sent to the llm too? that is the approach taken by zed and it is pretty helpful to control exactly what you are sending to the LLM and the UX is pretty nice too opposed to editing lua tables

[olimorris]

is it possible edit/delete sections in the chat buffer that will edit/remove the relevant messages sent to the llm too?

I'm sure it is possible and that's how the plugin originally functioned. I moved away from that because for large conversations, parsing the whole buffer led to some performance issues (Maybe Tree-sitter's new async features mitigate that somewhat). I also needed to embed more information into messages other than just role and content. Matching up the Tree-sitter node to the Lua table (which contains options and tool ids etc) proved to be non-trivial.

My view with a feature such as this is if a large proportion of the user base could benefit from it, then I'd consider its inclusion. However, given gd enables you to do this (albeit in a slightly cumbersome way), it's not worth it and I strongly suspect most users are not doing this often.

[bew]

I strongly suspect most users are not doing this often.

That sounds surprising...
I think there is a reason all web UI allows users to edit last prompt with UpArrow.
Here the process is rather unusual for what I think is a fairly common need (slightly edit last prompt and try again)

For the matching between text and TS nodes, a simpler alternative might be to use extmarks for all the headers/prompt/generated-text, and expose these extmarks' positions in the message objects (or another table idk). WDYT?