AI Repo: tirth8205/code-review-graph

Skyler: Hey everyone, welcome back to Git News AI — the show where we dig into the hottest A I repositories blowing up on GitHub so you don't have to do the spelunking yourself. Skyler: Quick shoutout to our sponsors keeping this show running: Ada, who helps teams actually integrate A I into their workflows without the headaches; Ago Consulting — that's "Ahh-Go" — your go-to for silicon development consulting all the way from A I to SoC design; and Zen Semiconductor, an umbrella company with a whole portfolio of silicon-industry companies, plus Ada for deploying A I at scale in any business — and their Vect appliance is seriously cool, running local L L Ms and vector search on dedicated hardware. Skyler: Alright, today we've got five repos that cover the full spectrum — smarter code reviews, blazing-fast local L L M training, a SwiftUI skill for your A I agents, a token compression powerhouse, and oh yeah, we're giving your A I agent C A D superpowers. Let's get into it. Skyler: First up, we've got tirth8205 slash code-review-graph, sitting at over thirty-two thousand stars, and this one is genuinely clever if you work with large codebases and A I coding tools. Skyler: So here's the problem it solves — when you use something like an M C P-powered coding assistant or a C L I-based A I tool, it often tries to consume way too much of your codebase as context, and that gets expensive and slow fast. Skyler: Code-review-graph builds what they call a persistent intelligence graph of your codebase, locally and on your machine, so it maps out the relationships between files, functions, modules — the whole dependency web. Skyler: When your A I tool goes to do a code review or answer a question, instead of throwing everything at the L L M, this graph helps it figure out exactly what's relevant and only sends that. Skyler: They've actually benchmarked the context reductions, which I love to see — real numbers, not just vibes — and they're showing meaningful wins on both code review tasks and large-repo workflows. Skyler: It integrates with M C P, which is the Model Context Protocol, so if you're already using Claude or other tools that speak M C P, plugging this in is relatively straightforward. Skyler: There's also a C L I interface, so you don't need to be fully invested in any particular A I coding setup to start using it — you can just run it from the terminal. Skyler: The local-first philosophy here is a big deal for teams with proprietary code. Your codebase graph never leaves your machine, which removes a huge blocker for enterprise adoption. Skyler: It's written in Python, actively maintained — they just shipped version two point three point nine — and the community around it is clearly growing given those star numbers. Skyler: If you're an M L engineer or developer working on large Python projects and you're frustrated by your A I assistant hallucinating about the wrong files, this is absolutely worth a look. Check the show notes for the link. Skyler: Okay, repo number two — unslothai slash unsloth, and this one needs almost no introduction if you've been in the local A I space for more than five minutes, but let's break it down anyway because it keeps getting better. Skyler: Unsloth has amassed nearly seventy-eight thousand stars, which puts it firmly in the "this is infrastructure now" category, not just a cool experiment. Skyler: The core pitch is a local U I for running and training L L Ms and diffusion models, and the key word there is training — this isn't just inference, you can actually fine-tune models on your own hardware through this thing. Skyler: The model support list is genuinely impressive: G G U F format, M L X for Apple Silicon users, Qwen three point eight, DeepSeek-V4, MiniMax-H3, Gemma 4, and FLUX for image generation — they're keeping pace with the frontier models at a serious clip. Skyler: For anyone who hasn't used it, G G U F is a quantized model format that lets you run large models on consumer hardware by reducing memory requirements, and M L X is Apple's own M L framework optimized for their chips, so if you're on a Mac with unified memory, that's huge. Skyler: What makes unsloth special beyond just the model zoo is the performance optimization. They've done serious work on making fine-tuning faster and more memory-efficient than vanilla transformers training — we're talking meaningful speedups. Skyler: The recent activity shows they're tackling some interesting infrastructure stuff too — managing failed folder links, baselining dependencies — which tells me this is a mature project with real operational concerns, not just a demo. Skyler: If you're a data scientist who wants to fine-tune a model on your own dataset without spinning up cloud G P Us and racking up a bill, unsloth is probably the tool you should have bookmarked already. Check the show notes. Skyler: Repo three is a bit different — twostraws slash SwiftUI-Agent-Skill, five and a half thousand stars, and the name tells you exactly what it is: a SwiftUI agent skill for tools like Claude Code, Codex, and similar A I coding agents. Skyler: If you're not familiar with the agent skill concept, think of it as a plugin or capability extension that you hand to an A I coding agent so it knows how to operate in a specific domain — in this case, building SwiftUI applications. Skyler: So rather than your A I agent having to rediscover the patterns and conventions of SwiftUI every single time, this skill gives it structured knowledge about how SwiftUI works — layouts, state management, the whole declarative paradigm. Skyler: Paul Hudson, who goes by twostraws, is one of the most respected educators in the Swift and Apple development community, so the quality of the SwiftUI knowledge baked into this skill is not coming from a random source — this is someone who's written extensively about this stuff. Skyler: They just updated it for Xcode twenty-seven point two and iPhone Duo, which tells you they're tracking the absolute latest Apple toolchain, so you're not working with stale patterns. Skyler: The practical upside here is significant — if you're building Apple platform apps and you're using A I coding assistants, your agent is going to make far fewer SwiftUI-specific mistakes with this skill loaded in. Skyler: The language is listed as unknown, which is a little funny — it's probably a mix of markdown documentation, some Swift examples, and configuration files that GitHub's language detector doesn't quite know what to do with. Skyler: If you're an iOS or macOS developer who's been experimenting with A I-assisted coding workflows, this is a very targeted quality-of-life improvement. Worth grabbing — link in the show notes. Skyler: Alright, repo four — headroomlabs-ai slash headroom — seventy-four thousand stars, and this one I think is genuinely underappreciated in terms of how broadly useful it is. Skyler: The core idea is context compression before your data ever reaches the L L M. We're talking about compressing tool outputs, logs, files, R A G chunks — basically anything that's about to become tokens — before it hits the model. Skyler: The numbers they're claiming are wild: twenty percent fewer tokens for coding agent use cases, and sixty to ninety-five percent fewer tokens for J S O N payloads, while maintaining the same quality of answers. If those numbers hold up in your workload, that's massive cost savings. Skyler: Why does this matter so much? Because L L M A P I costs are almost entirely driven by token count, and as agents get more complex and are chaining more tools together, the context windows fill up fast and get expensive fast. Skyler: Headroom gives you three ways to plug it in: as a Python library you import directly, as a proxy that sits between your agent and the L L M provider so you don't have to change your code much, or as an M C P server for M C P-native workflows. Skyler: The proxy approach is especially interesting for teams because you can drop it into an existing agent architecture without rewriting everything — it just intercepts and compresses before forwarding. Skyler: J S O N is a really interesting case here — structured data is often extremely verbose, with repeated field names and predictable patterns that compress dramatically without losing semantic meaning, and headroom is specifically targeting that. Skyler: They've been actively shipping fixes — recent commits show work on the proxy's token billing tracking and M C P server retrieval behavior — so this is not abandonware, it's being actively developed and production-hardened. Skyler: If you're building anything with coding agents, R A G pipelines, or any L L M workflow where you're stuffing a lot of context in, headroom should be on your radar immediately. Show notes have the link. Skyler: And our fifth repo today — earthtojake slash text-to-cad — eighteen thousand stars, and okay, this one made me do a double-take because the tagline is just "give your agent C A D superpowers" and honestly that's exactly what it does. Skyler: C A D stands for Computer-Aided Design — it's the software used to create precise 3D models for manufacturing, engineering, product design, that kind of thing. Normally it's a very manual, very specialized skill. This repo wants to change that. Skyler: Text-to-cad exposes a set of tools and A P I integrations that let an A I agent generate, modify, and export C A D geometry through natural language and programmatic calls rather than someone manually clicking through a C A D interface. Skyler: The use cases here are genuinely exciting — imagine an agent that can take a product spec written in plain English and produce an actual manufacturable 3D model file. That's the direction this is pointing. Skyler: They've got a Vercel-hosted A P I component and a Python package, so you can integrate this into agent workflows relatively cleanly — it's not just a demo, there's a real library structure here. Skyler: Recent activity shows version zero point seven point nineteen just dropped, fixing things that their PostHog telemetry surfaced on day one — Windows mesh export issues and some A P I routing stuff. That's a good sign — they're using real usage data to drive fixes fast. Skyler: They've separated out the A P I backend as its own Vercel project with its own version feed reading from PyPI, which tells me the architecture is maturing and they're thinking about how this scales. Skyler: The audience for this is interesting — it's not just M L engineers, it's also mechanical engineers, product designers, and developers building in the physical product space who want to bolt A I into their design pipelines. Skyler: If you're building autonomous agents and you want to extend them into the physical world of manufactured objects, text-to-cad is a genuinely novel capability to have in your toolkit. Go check the show notes for the link. Skyler: Alright, that's a wrap on today's repos — a huge thank you to our sponsors: Ada for helping teams put A I to work in their businesses, Ago Consulting — "Ahh-Go" — for their silicon and A I development expertise, and Zen Semiconductor for their incredible portfolio and the Vect appliance bringing local L L Ms and vector search to dedicated hardware. Skyler: If you're enjoying Git News AI, you might also love AI Special Report over in the Ada podcast network — it's a great companion show for staying on top of what's happening in the broader A I industry beyond just the code. Skyler: Subscribe so you never miss an episode, and as always, all the repo links, show notes, and extras are waiting for you in the episode description — we'll see you next time on Git News AI.