RusticStack

Building simple, reliable tools I actually want to use.
Agent harnesses • Pi providers • Rust/TS tooling
Shipping in public at

[RusticStack @RusticStack]

@finkd 1,000+ tool calls over 24 hours means context is getting recycled or summarized somewhere. Curious how Muse handles that.

[RusticStack @RusticStack]

@theo Does snoozing a thread pause any background work the agent was doing, or just hide it from the sidebar?

[RusticStack @RusticStack]

@biwanczuk @deno_land I've noticed the same thing building. A few ms on an empty script feels nice in benchmarks but once you import a real dependency tree the startup cost is dominated by module resolution and initialization, not the runtime bootstrap.

[RusticStack @RusticStack]

@thdxr The IaC angle is what makes this actually scalable for agent-generated tools. I tried doing basic auth via env vars first and it got messy fast once you have more than one tool running. Cloudflare Access keeps the auth layer completely outside the app code.

[RusticStack @RusticStack]

@mattpocockuk Have you tried just chaining separate /goal calls with a shared state file? Curious whether the "dumb zone" is mostly context length or if it's the agent losing the thread of what it already accomplished.

[RusticStack @RusticStack]

Rust 1.97.1 just dropped a fix for an LLVM miscompilation that’s been quietly sitting there since 1.87. Point release, but if you’ve ever hit weird codegen bugs it’s worth grabbing immediately.

Anyone else run into that specific issue?

[RusticStack @RusticStack]

@mattpocockuk I've been using Wezterm with Pi Harness. It works well enough for me managing with tab system but it's not ideal for most people. I tend to T3 code when I'm managing a lot more different agents at once.

[RusticStack @RusticStack]

One of the things I use ChatGPT Work @ChatGPTapp for most is just getting my own head straight.

I throw in call notes, random links, half-finished ideas, and the things I said I’d come back to. Then I use it to turn that mess into a plan and figure out what actually needs doing first.

Every now and then I also use it to get the first version of a doc or spreadsheet down. I still go through it properly afterwards, but editing something rough is much easier than staring at a blank page.

I don’t really use it for some huge, complicated workflow. I mostly use it to get organised and start moving.

[RusticStack @RusticStack]

@serg_vecher It looks good. Do you know how many tokens it used for that amount of usage?

[RusticStack @RusticStack]

I've been seeing the new Kimi K3 from @Kimi_Moonshot all over my timeline, so I wanted to try it somewhere I actually work, not in another model playground.

@ZenMuxAI has it "free"(your account balance needs to be higher than 0$) through their API for a limited time . There was some ZenMux provider for @pidotdev, but I wasn't feeling confident on their implementations.

I could have hard-coded K3, but it felt more useful to do the provider properly. It discovers ZenMux's live model catalog, maps pricing and model capabilities, handles /login, and includes a /zenmux connection check, tests, and an offline fallback.

It's on npm now:

pi install npm:pi-zenmux-provider

Then run /login, choose ZenMux AI, and select MoonshotAI: Kimi K3 (Free) in /model.

https://t.co/gnSHYdIleh

[RusticStack @RusticStack]

Novita has Tencent's new Hy3 model free through their API right now. I wanted to try it in Pi, not just in a playground.

There wasn't a Novita provider, so I ended up making one.

It started as the quickest way to get Hy3 running, then I kept going until it had proper /login support, live model discovery, useful errors, tests, CI and checks for the actual npm package.

You can try it with:

pi install npm:pi-novita-ai

Then choose Novita AI in /login and tencent/hy3 [novita] in /model.

https://t.co/iPuIC7ZhaA

@novita_labs @pidotdev

[RusticStack @RusticStack]

Quick note: the website and current GitHub releases are still for the old Wails 3 version.

The new Rust/GPUI app has not been packaged yet, but you can already compile it from source if you want to try it or contribute.

Built with GPT-5.6 Sol on High Thinking and @zeddotdev's GPUI. Tagging @OpenAIDevs here since I missed both in the original post.

[RusticStack @RusticStack]

I spent an afternoon yesterday with GPT-5.6 Sol and came out the other side with a native rewrite of my ESO addon manager. I wanted to share what building with it actually felt like.

I pointed it at a Go/Wails/Svelte app and said: rewrite this in Rust using Zed's GPUI, keep the behavior, and make it fast.

It held the whole architecture in context at once: settings, the MMOUI catalog, dependency resolution, staged installs with rollback, and the rkyv archive codec. The code was mostly right the first time, including lifetimes and unsafe boundaries. I spent more time reviewing and steering than writing.

What surprised me most was its performance instinct. It pushed the catalog hot path onto validated rkyv archives, and the numbers got silly.

Warm load went from 840k allocations to 693, and from 12.4 ms to 3.1 ms.

Scanner parse went from 122 ms and 392 MB to 24 ms and 18 MB.

The catalog is ready around 6.6 ms before the first frame.

It understood "don't sort 7,000 addons in the render path" without me having to say it twice.

The UI/UX is still WIP. Dialogs, context menus, accessibility, and a lot of polish are still left to do. But the core logic and performance are essentially done, and that happened in a single afternoon.

https://t.co/fUcVxuo2SO

[RusticStack @RusticStack]

@debug_w_vikas @heyblake appreciate you letting me know. proxy went down and took the whole service with it. just fixed the issue so it should be flying now for everyone. my bad on that, would love for you to give it another try

[RusticStack @RusticStack]

@theo I love it. Way better than looking at the CLI. During the questions and answers when on a smaller window it can clip horizontally but that's just nitpicking. Some simple quick(sub 0.5s) animations would make it feel smoother while keeping it snappy, kind of like T3 Chat.