📦 faiface/ai-pair

AI Pair: Watch Your AI Type at Human Speed

AI Pair is the VS Code extension that turns AI coding assistance into a live, watchable collaboration. Instead of just handing off a task and reviewing a giant block of finished code later, this tool places a second, visible cursor in your editor. The AI agent types character by character at a comfortable, human speed, while explaining its reasoning out loud in a side panel. You can pause it, interrupt it, or even take over the keyboard to type yourself while the agent watches. This is built for people who want to deeply understand the code they are working on, rather than just accepting it blindly.

🆔 @hackernewsgithubprojects

📦 rlaope/deepseek-bot

deepseek-bot: The AI Teammate That Actually Remembers

A self-hosted AI teammate that lives inside Slack, Discord, and Telegram, and actually remembers what you discussed last week without you repeating yourself. DeepSeek bot plugs directly into the chats you already use, so you never have to open a new tab to ask for help. The coolest part is that it keeps a persistent memory of your conversations, meaning it can recall specific decisions from past threads and apply them to new questions instantly. It even drafts real documents like spreadsheets and slides right inside the chat.

🆔 @hackernewsgithubprojects

📦 datologyai/zephon

Zephon: The Data Loader That Doesn't Lie

Zephon keeps your training runs honest when your hardware changes mid-job. Standard data loaders often shuffle the deck or crash when you resume a training run on a different number of GPUs, which ruins your ability to compare experiments fairly. Zephon solves this by guaranteeing that every batch of data is identical no matter how many machines you throw at it. It handles tokenization and data mixing on the fly, so you skip the expensive pre-processing step entirely. If your cluster gets interrupted and you restart on fewer cards, it picks up exactly where it left off with the same data.

🆔 @hackernewsgithubprojects

📦 steventsvik/ghosttyextreme

Ghostty Extreme

Ghostty Extreme is the macOS terminal that lets you watch AI coding agents work live in real time. Built as a specialized version of the fast, native Ghostty terminal, this app connects directly to tools like Claude Code and Codex to show you exactly what your agents are doing across multiple projects. As an agent works, a dedicated sidebar shows its thinking process, while a built-in code editor physically types out the agent's edits as they happen. It even maps out which files are being read, tracks your backend services, and lets you approve or deny permissions with a single click.

🆔 @hackernewsgithubprojects

📦 tencentcloud/octop-harness

Octop Harness: The Missing Layer for Real AI Agents

Tencent Cloud released Octop Harness, a production-grade runtime that finally turns experimental AI agents into reliable, working software. It wraps the deepagents foundation in a layer of real-world engineering, handling memory, security, and multi-provider model routing so your agents actually run without breaking. You get a full terminal interface, pluggable storage backends, and a registry to manage multiple isolated agents in one go. It solves the classic problem: demo agents work, but production agents fall apart. This is the glue that makes them stick. If you are building agents, this is the safety net you have been waiting for. Stop building sandboxes and start building products. This is the one.

🆔 @hackernewsgithubprojects

📦 startluxlabs/decisionpivot

DecisionPivot: The Sneaky Test for AI Logic

DecisionPivot exposes when artificial intelligence fakes its reasoning. Instead of asking open-ended questions, this diagnostic benchmark feeds models pairs of scenarios that differ by just one tiny detail. If a model gets both versions right, it truly understood the situation. If it fails, it likely guessed based on surface patterns. This catches shortcuts that standard tests miss. It proves whether a system can handle subtle changes in real-world data. It’s the difference between a smart assistant and a parrot. Use this to check if your AI actually thinks or just vibes. Stop relying on basic benchmarks.

🆔 @hackernewsgithubprojects

📦 zeamoxwang/pdmd

Generate Videos in Two Steps with PDMD

Generate a complete video with sound in just two network evaluations. That is the core promise of PDMD, a new training method for video diffusion models. Standard fast generation often leads to blurry, oversaturated images because the learning process drifts over time. PDMD solves this by filtering out the specific error component that causes that degradation. It is essentially a one-line fix to the training loop that makes the student model learn much more stably. The result is a video generator that is both incredibly fast and surprisingly high quality. You can even run it on a single, modest graphics card.

🆔 @hackernewsgithubprojects