2,490 karma · joined June 6, 2011
AI tooling can also be a place where we start building our view of what maintainable software practices look like so we don't make decisions that have these same tail effort profiles. That can be things like building out tooling to handle maintenance updates
I think the real thing that comes out of AI tooling is probably that the tooling needs to be trained (or steered) towards activities that enhance human attention management.
> It isn't fair to blame TUIs.
> The real problem is that pretty much the whole stack has a terrible AX story.
> First, most GPU-rendered terminal emulators don't engage in system-provided accessibility APIs AT ALL. Because text is GPU-rendered, AX tooling can't "read" it, it just shows up as an image. This applies to Kitty, Alacritty, WezTerm. My own terminal Ghostty is AX-readable (on macOS), and so are others like iTerm2 and Terminal.app (which admittedly do it better than me, we have gaps to fill).
> Second, there are no terminal sequences or initiatives at all for TUIs to communicate AX information to the emulator, so the emulator itself can't do much more than display a blob of text to AX tooling. We need the equivalent of ARIA-style annotations but for terminal cells, runs, and regions. No such initiative exists. Even if TUIs do great things with the cursor, this is going to bite a lot of use cases.
> As an example of combining the above, I've been working on something with Ghostty where we integrate semantic prompt (OSC133) and AX APIs so that we can present each shell prompt, input, and command as structurally significant to AX tooling (rather than simply a text box where the cursor is somewhere else). This shows the importance of the relationship between terminal specs (OSC133), TUIs (which must emit OSC133), and terminal emulators (which must both understand OSC133 AND communicate it to AX APIs).
> The whole stack is rotten. And no one is earnestly trying to fix it (including me, I have limited time and I do my best but this is a WHOLE TOPIC that requires a huge amount of time and politicking the ecosystem and I don't have it, sorry).
Bonus: a simultaneously awesome and horrible reality is that AI is really helping to improve AX here. A lot of AI tooling uses/abuses AX APIs to make things happen. How is OpenAI reading your list of windows, typing into them, etc? Accessibility frameworks! So a lot more apps are taking AX integration a lot more seriously since its table stacks for AI using it... Sad it requires that but the glass half full is more software is doing that.
Result<Option<Result<Message, WsError>>, Elapsed>
That’s three independent “not the happy path” channels: timeout, stream closed,
and websocket error.The nicer version is not a cleverer match. It’s choosing a domain error shape and converting into it one layer at a time:
let timed = tokio::time::timeout(duration, receiver.next()).await;
let next = timed.map_err(|_| ReceiveError::Timeout)?;
let item = next.ok_or(ReceiveError::Closed)?;
let msg = item.map_err(ReceiveError::WebSocket)?;
The ugly line is what happens when you have not decided where to normalize the
shape yet.yes, and... the idea here is that it would be neat to extend the hermetic builds idea such that this can be run locally / anywhere where there's compute easily. The root problem that's being called out here is that idea of running something until the CI says it's green when there's a change, commit, network call, in the cycle is a pain in the ass. (The best way to avoid this churncycle is to just never write bugs! TFIC ;P)
> It quietly defines the ceiling of your ambition and the floor of your operational risk.
https://github.blog/open-source/git/get-up-to-speed-with-par...
Ha yup - I've felt this one before :D
Good docs are docs that make it easy to implement the next feature.
From an AI perspective, it's my observation that LLMs often write code with lower quantity / quality docs. At the same time, they are reasonably good at synthesizing / inferring meaning from code that lacks good docs. They often do so internally by forming a chain of thought / reasoning around how the code works. The docs that should be written as part of the code are probably the same things that an LLM would reasonably come to by spending tokens when modifying that code. I believe that this should be trained into model so that future LLM work starts with not having to build up context.
In the absence of that being built in, something I've been experimenting a little with is tuning what I want to see in docs that actually help source control / development. Currently that's at https://github.com/joshka/skills/tree/main/doc-steward - still needs a bunch of work, but it's generally better than nothing. YMMV
I have a PR up for jjk that does the full change as a review changes, and there's another user's PR that allows diffs over arbitrary ranges (i.e. when working out whether the commits that make up a PR are good as a whole rather than individually)
Go throw yourself at Stockfish or another chess engine for this. They tend to have good test harness that can give you an idea of whether the speedups are worth the effort.
This convention existed before clap came into being, but I don't recall when I first saw it. I have been using the command line for just shy of 40 years across various operating systems.
A good way to think of it is that jj new is an empty git staging area. There's still a `jj commit` command that allows you to desc then jj new.
> I also often end up with in a dirty repo state with multiple changes belonging to separate features or abstractions. I usually just pick the changes I want to group into a commit and clean up the state.
jj split allows you do to this pretty well.
> Since it's git compatible, it feels like it must work to add files and keep files uncommitted, but just by reading this tutorial I'm unsure.
In jj you always have a commit - it's just sometimes empty, sometimes full, has a stable changeid regardless. jj treats the commit as a calculated value based on the contents of your folder etc, rather than the unit of change.
I find it reasonably good with lots of tweaking over time. (With any agent - ask it to do a retrospective on the tool use and find ways to avoid pain points when you hit problems and add that to your skill/local agents.md).
I expect git has a lot more historical information about how to fix random problems with source control errors. JJ is better at the actual tasks, but the models don't have as much in their training data.
When reviewing, you can also mark individual files as reviewed (useful in larger reviews where you're incrementally reviewing files). If you do this, only files that are changed will be expanded when you come back to the review.