Yes, as long as we inject the updated context there is no issue. But if an older context fetched from messages above and this input changes and you edit a message down below, this will generate an issue in this destructive behaviour. I prefer immutability more.
Another issue I observe is that the model could be changed in between and this starts a new context from this point on, but the past messages and the respective context stay visible.
Like Kynlo said, if it will become some kind of notebook style, maybe a toggle where the user can choose their preference conversation UI behavior could be an option.
Folks, I have a PR that fixes some of the feedback raised here: toolbar buttons are tabbable, the “send without context” is more discoverable, etc. Check out the video in that PR and comment on the PR with any feedback on it specifically.
We are also working on better behavior around focusing the follow-up input after you send a message (and many other things, keep the feedback coming!).
OK, None of the things in the previous PR actually addressed the primary issues with the new UI, as far as I can see.
After “working” (being used loosely here today) with the pre-release, which, I do understand is a “WIP”, the flow is basically non-existent.
The input box takes forever to register you actually pressed send/return so you’re always unsure if the message actually went through (intuitively, the input box would at least unfocus after pressing send to indicate anything is happening).
Once the chat reaches the bottom of the window, the rest is yeeted off into the void.
Adding follow-up messages when the chat window is “full” causes Cody to have a little bounce before spitting out the new reply (which, if you’re lucky, you can see the first line of before having to scoll to find the rest).
I did a short video here as a little bit more of a visual explanation to the above:
I didn’t actually notice until towards the end of recording the video - where have half the @ commands gone? Why can I no longer @ in a domain for context or anything like that?
Yes, the PR only addresses some of the issues. Thank you for sharing this. Someone else just showed me the scroll bounce issue as well. We’re still working through these and will have stuff polished by the upcoming release at the end of May.
There is a bug where the Web URL context source does not show up if you have the OpenCtx (formerly known as OpenCodeGraph) VS Code extension installed. We’ll fix that, but please uninstall it in the meantime if you have it installed. Then you will see Web URLs in the @-mention menu. And many more context sources coming soon there!!
I have neither of those installed and still dont see the web-options.
Also something else I just noticed: If you press enter twice on a submitted input (due to the lag/delay in after pressing it the first time, or even by accident), it will either give you a completely blank output, or a broken one
@sqs I just listened back to my recording and realised I may have sounded more blunt than I intended. Please don’t see what I’m saying as any sort of disrespect, just my care towards Cody and my love for the product, and certainly no reflection on the work you are doing (as I mentioned previously, what you have made here in the pre-release could absolutely be a product in its own right, and honestly, is actually quite good if it’s looked at as a different product separate but related to Cody!)
No worries, I appreciate and we all at Sourcegraph appreciate you and the direct feedback! We want to make Cody the absolute best, and direct feedback is gold.
Chat features I’m missing the most with the latest build (v1.19.1716397784)
Sticky auto-scroll if I manually scroll to the bottom of the viewport
Ability to stop generation, ideally with the keyboard (and we should display this keyboard while the reply is streaming)
Option-/ shortcut should close/hide the chat tab when the input is focused. I’d like to be able to use this shortcut to toggle the visibility of the sidebar. Currently, it inserts the character ÷ for me.
Why do you want this? I agree it seems like it’s needed (and I thought I wanted it), but doesn’t it just interfere with your ability to read the message?
Esc is an undocumented way to do this for now. Will show it in the UI and document it.
I like the ChatGPT UX where there’s a down-arrow button to enable auto-scroll. Auto-scroll is sometimes convenient when I know the answer I’m looking for is coming later so I want to see the latest streaming text without having to manually scroll all the time. This is a minor issue though, not a high priority.
I recall discovering this at one point but then forgetting about it. I like how Cursor shows the shortcut while the reply is streaming. It would be a good UX improvement to document this shortcut and it would also be good give stronger visual hints in the UI after you have cancelled the generation (currently, it just stops streaming).
I think the new design is brilliant. This means you can always go back up if you take a wrong turn.
As already mentioned, autoscrolling would be good if you could switch it on. And deselecting to send the context is also a bit difficult. I didn’t know how to deselect it, only when I read it here. I didn’t realize that “opt” was the button. Maybe selecting/deselecting using a button would be better.
Just adding to the points from earlier, my biggest is that the top chat box continues to be the focus after submitting my first question. So if I want to follow up, it feels like I’m restarting a new chat? In my video, I simply start typing my follow up question, but because the top chat box is the active one, the text ends up there. And it also looks like Cody doesn’t retain the history as I’m editing the first prompt (you see this in the video where it actually doesn’t do the rewrite in Python)
The first top box is and will be the first message in the history. If you want to write a followup message you need to do this in the second chat box after Cody’s answer below on the bottom.
autoscrolling would be good if you could switch it on
Why do you (and others) want autoscrolling? I know it is common in AI chat apps, but as we’ve looked at it with a critical eye, it seems to get in the way of users reading the response more than it helps them.
You are right, but after pressing Enter, the input field should be deselected so the user can more clearly recognize that the submission was successful.
