This past weekend, we built Full Self Driving (FSD) for a wheelchair.
Designed for people with disabilities or the elderly, a user can speak where they want to go (an address or even "take me to the bathroom") and the wheelchair safely avoids obstacles and navigates to the destination.
We built this on an Arduino UNO Q spoofing the joystick's sensor chip, and an open robotics model (OmniVLA) post-trained on drives from Google DeepMind's Gemini Robotics-ER 2
Today we turned off external PRs for Mastra
We noticed weโre almost always closing PRs and having our factory do them anyway
Instead weโre encouraging contributors to open issues with detailed repros
Open source is drowning in swarms of AI slop, and more and more projects are closing their doors for contributors.
For example, Cloudflare's Agents SDK doesn't take external PRs. When I hit bugs building on it, I filed 10 issues with reproductions instead.
So when @Cloudflare asked to rethink the Git platform for agents, I built Forkin:
> Issues can get automatically reproduced by tests running in a sandbox
> Any number of agents can find, triage, and fix bugs in parallel
> Only quality reaches the maintainer
Its live on real repos here: https://t.co/xiImRDNgYu
Open source is drowning in swarms of AI slop, and more and more projects are closing their doors for contributors.
For example, Cloudflare's Agents SDK doesn't take external PRs. When I hit bugs building on it, I filed 10 issues with reproductions instead.
So when @Cloudflare asked to rethink the Git platform for agents, I built Forkin:
> Issues can get automatically reproduced by tests running in a sandbox
> Any number of agents can find, triage, and fix bugs in parallel
> Only quality reaches the maintainer
Its live on real repos here: https://t.co/xiImRDNgYu
Every agent can spin up its own fork on Cloudflare Artifacts, so a thousand can work together without colliding. Over MCP or the CLI, they see every attempt, learn from rejected ones, and build on each other's forks.
And no agent gets to hallucinate about whether an issue is real and worth fixing. Forkin runs the code, and in seconds the issue is identified as reproduced, already fixed, or couldn't reproduce.
Built on Workers, Durable Objects, Sandboxes and Artifacts, with @EffectTS_
For the https://t.co/IIOggqfhlz
cc @dinasaur_404@zebassembly@dillon_mulroy
Iโm excited to announce the release of https://t.co/f04bmdz4Bv, the unified interface for document parsing!
It started as an internal tool I built for https://t.co/FUPuEaVLct, as a way to experiment with different providers and optimize cost, reliability, latency, and accuracy easily.
FileRouter sits between your app and document parsers. You get the primitives for a pipeline: parse, wait on jobs, and get normalized results through the SDK, CLI, and API. Our infrastructure handles the end-to-end lifecycle, or bring your own keys and runtime.
Available providers today: @llama_index , @firecrawl , @MistralAI , and @datalabto with more on the way.