ireadmevs

Deterministic assurances on top of nondeterministic tools

> The entire source code, without tests, is about 15,000 lines, which comes out to about 150,000 tokens with GPT and about 250,000 with Claude.

Woah, that big of a difference when it comes to token counting?

If it’s just filesystem sandboxing that you are after, then bubblewrap can help you stay in nix by just wrapping the agent’s derivation

Try checking what M-x magit-toggle-verbose-refresh has to say, you may be able to identify what's slow and optimize around it

> the changes that end up in nixpkgs are all reviewed by maintainers

They review the “recipes”, sure. But I don’t know if I’d trust that every diff of the source of every package update is reviewed by them.

Really fun! Feels great when your hunch pays off, but most of the time it didn’t. I’ll be certainly playing it more often! Joining the share here:

Anthropeum.com · Jul 28 2026

YYBGYYRGGB

61,616 · top 66% of players today!

I guess there are 2 sides of ACP that are valid then. What was asked is for Juggler to act as a ACP client. I, on the other hand, would appreciate it to be a ACP server, so that I can interact with it from my editor when I’m on my machine and from the web when on the go or for more advanced functionality. That’s my current approach with opencode.

For me is the ability to maintain the same keyboard workflow, no matter if I’m dealing with files or terminal (switching buffers, navigating/searching/selecting/copying/pasting text, etc.), I get to keep the same set of shortcuts and mental model. And i get to do some small automation on top of it with elisp, which I already do for the rest of my editor anyway.

Yes, a link requires user interaction. But what if the attacker decides to render an image instead and put the secret data in the query params? Loading an image is a way to trigger a request without user interaction

I would even go one step further: why would an AI tasked with summarizing external comments require any tool access, especially being able to reach into non public data?

The solution to prompt injection is not more AI on top of it, it starts with data access controls

Thank you for sharing!

For public transportation I use an app called Moovit

> SANA-WM uses only ~213K public video clips with metric-scale pose supervision, completes training in 15 days on 64 H100s, and generates each 60-second clip on a single GPU; its distilled variant runs on a single RTX 5090 with NVFP4 quantization to denoise a 60s 720p clip in 34s.