New in v11.11.0 - ✨buffers can now be watched

[olimorris]

All, please give a huge round of applause to @bassamsdata for #610, which adds the ability for buffers to be watched in the chat buffer itself.

Previously, when buffers were pinned to the chat buffer, ALL of their content was sent to the LLM along with a user's message. The impact of this was it consumed a significant amount of tokens, on every turn!

Thanks to @bassamsdata's PR, you can now choose to watch a buffer (with gw) and it will send only the added, edited or deleted lines to the LLM. What makes the PR so great is the sheer amount of testing which was done to make sure the feature works well with so many models.

[mystilleef]

@olimorris @bassamsdata kudos! This is huge! Very impressive work. Congratulations!

This should also significantly improve updating the buffer using the editor tools.

codecompanion is by far my favorite AI experience and I've tested the vast majority of them including paid ones.

Once again kudos to the developers and this community.

[olimorris]

Exactly that! I've noticed that you don't even need to save the buffer for the LLM to be notified of the changes too so it's very slick with the @editor tool.

codecompanion is by far my favorite AI experience and I've tested the vast majority of them including paid ones.

That means a lot, so thank you! And I still have so much more planned as well 😄.

[bassamsdata]

Thanks a bunch, @olimorris, for all the kind words and helpful advice while I was working on the PR.

If anyone has any issues with it, feel free to tag me and let me know.

[SDGLBL]

Nice work

[GitMurf]

This is amazing! Nice work!

A couple questions:

1. Does the "watching" / calculating of the diffs (changes) get calculated / watched realtime? Or is it being calculated when you submit a new message? I am ocd about autocmds and stuff running in the background as I work so am curious if processing/calculating is going on or if just when submitting messages?

2. Will the pin feature stay the same and available as a separate feature?

3. If we delete the file reference later will it also delete all the messages where there were provided updates / diffs provided for a file throughout the conversation?

Thanks!

[bassamsdata]

Thank you for the nice words.

1. TL;DR: Diffs and processing are calculated only when you submit a message.

When the watcher starts, it sets up a state for each watched buffer, including its initial content and changedtick. The diff calculation is triggered only at the time of submission to the LLM. At that point, it compares the buffer’s current content against the last sent content.

I am ocd about autocmds and stuff running in the background as I work

The only autocmd being used is BufDelete. No heavy processing or monitoring runs in real-time as you work. Your OCD about background tasks is respected :) (I’m the same when it comes to avoiding excessive TextChanged* autocommands.)

2. I believe Oli mentioned this might replace buffer pins at some point, though I’m unsure if that’s still his position. Personally, I don’t think it should (at least for now). As I noted in the PR, some enhancements would need to be implemented first.

3. The watcher rigorously respects the original references mechanism. From what I recall while working on it—and if I’ve understood your question correctly, the diffs/changes should remain in the chat message history table even if you delete the buffer (e.g., using :bdelete), in other words: they should still be in the LLM memory
   However, if you’re referring to deleting the lines where references were located in the chat buffer, the LLM should still retain them in the context window. That said, this is based on my memory and could be worth double-checking.

Thank you

[GitMurf]

Thanks for your responses. Wonderful!

1. Sounds great!
2. Makes sense
3. Sorry for not clarifying. I was referring to when deleting a reference from the chat by removing from the "sharing" blockquote.

[GitMurf]

After finally getting some time to test it, just wanted to say the watcher feature is working great! Nice work @bassamsdata and @olimorris 🎉