GitHub radar
Top 5 GitHub Repos This Week
This week's picks are mostly about agents that do not just answer in a chat window but actually carry a task through: one fixes a GitHub repository on its own, one works apps on a phone, and most of the code behind a Switch game ported to PC was written by agents too. And there is a fix for anyone who vibecoded an interface and winced at how it looked.
I opened the README and the authors warn right away: this is a research prototype, not an official DeepSeek product, and most of the code was written by coding agents, so bugs and rough spots are expected. They call the project itself a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness. It listens to GitHub, and optionally to the Feishu chat tool: when someone allowed asks it for work, it opens a work item, handles it in its own session, and answers back in the same thread, for example with a pull request. A separate web panel shows only what actually needs a person to look at. The stated goal is blunt: take as little of a team's attention as possible and give back as much value as possible, so it stays quiet until it is actually needed. I do not have DeepSeek Harness set up myself, so I judged this from the README and the numbered docs in the repo, not from running it.
Why a vibe-coder should care
For a team running a single repository, this removes the step of opening a chat and asking an agent by hand every time: the agent already watches GitHub and responds when someone asks, without a person sitting at a terminal.
I opened the README, and the authors say plainly what this is: an AI character reference and composition workbench, a 3-in-1 Chinese version. They explain that the original is three separate Japanese sites by one author, poses, costumes and portraits, scattered across different domains with metadata only in Japanese; here it is all merged into one, translated into Chinese (so it is also readable through any translator), plus a studio where you pick a pose, a costume and a portrait, combine them, and send the combination straight into Gemini, GPT Image, Qwen, Doubao or your own local ComfyUI. The library itself is not small: 11,568 local images, 1,439 poses, each with both a plain figure and a depth map for precise pose control, 1,407 costume sets, 1,499 portraits. It also works as a ready ComfyUI plugin: drop the folder into custom_nodes, and nodes appear with a visual picker instead of a file name. One thing worth saying plainly: the code is MIT, but the images themselves stay the original author's, so the gallery can be used and referenced but not resold.
Why a vibe-coder should care
Instead of hunting for pose references across Japanese sites by hand and translating the labels yourself, everything here is already in Chinese, in one place, and a finished combination of references goes straight into an image generator without extra copying.
The author calls this a skill for Claude Code and Codex that turns vibecoded, AI-looking UI into a designed interface with its own character. In the README he lists exactly what it hunts for first: purple gradients, glows, emoji used as icons in labels, and shouty headlines, the things that give away an agent built this at a glance. Instead of fixing right away, it first proposes a direction, a character, a couple of moments that carry it, one signature detail, and only after that works through the screen in passes: spacing first, then color, icons, copy, typography. It works with whatever the screen is built in, Next.js, React, Vue, Svelte, plain HTML, and even on someone else's live site from just a link, with no access to the code. It can also put together a short before-and-after video of the screen changing, clearly meant for posting. The author started this as a series of posts about unslopping on X, and it grew into this skill.
Why a vibe-coder should care
This is exactly what happens to anyone who builds an interface with an agent: it works, but it looks instantly AI-made. Here you just type /unslop, and the agent finds and fixes those giveaway signs itself, instead of you having to learn design.
The authors describe it plainly: an AI assistant on your phone, say what you need, and it works your apps and the web until it is done. It runs on a separate background screen, so you can keep using your phone as normal, with a small live view if you want to watch. In the README demos it reads through a WeChat group chat and summarizes the day, picks a charger under $30 rated 4.5 stars or more and adds it to the cart but stops before paying, reads a creator's recent posts and sums them up. It is not tied to specific apps: it looks at the screen and finds buttons the way a person would, with no per-app API integration, so in theory it works with any app. Mika has no server of its own: whatever the agent sees goes only to the model provider you connect yourself, by API key or by signing in with a ChatGPT, xAI or OpenRouter account. I do not have a compatible Android phone on hand, so I did not test the setup or the app control through wireless debugging myself, I judged this from the README, the demo clips and the docs.
Why a vibe-coder should care
Normally, automating a phone needs a separate integration for every app, or an API that most apps simply do not offer. Here the agent looks at the screen the way a person does, so it works anywhere, and you do not pay for a separate subscription, it uses whichever AI model you already have.
The authors call this a static recompilation: instead of emulating the Switch game on the fly, it turns the game code into an ordinary native application for Windows and Linux once, and after that it runs like any native program, with no emulation layer on top. The README is upfront that no game files are included, you need your own legally dumped copy of Tears of the Kingdom from your own console, and instructions for dumping it are not given. What that buys you: 60 fps by default and up to 144 on high refresh monitors, 21:9 and 32:9 ultrawide support, HDR10, plus cheats and mods carried over from the emulator scene. It is early development, so visual bugs and crashes are still expected. One more thing worth saying: most of the code and documentation is written by AI coding agents, Claude Code specifically, under the maintainer's direction, and the maintainer plays every build before deciding what gets merged. I do not have a dumped cartridge image, so I did not run the game myself, I read the README, the feature list and the status notes.
Why a vibe-coder should care
For anyone who already owns a legal copy of the game but wants to play it on a PC with a proper frame rate and a wide screen instead of on the Switch, there is no emulator in the way here, the game runs as a native app. And the fact that most of the code is written by agents under one person's direction is interesting on its own, as a sign of how far agents can carry a project this size.
▌ More finds