Sprocket

Nooooo, I forgot to devlog on time again. I made soooooo many improvements to the app. I properly released the app on npm by adding full release workflows for dev, canary, and latest. I worked on reducing the Convex db i/o as it was increasing very quickly which would have costed me a lot of money. There were also a whole bunch of problems with agent runs randomly ending or erroring out, I fixed all of them. I made it possible for the agent runs to continue even after you close the Sprocket browser tab. I added support for Claude Opus 5 and AWS bedrock mantle. I added support for agent skills, which you can store at ~/.agents/skills, ~/.sprocket/skills, and ./.sprocket/skills, alongside having skills built-in to Sprocket. There are no built-in skills yet, but I will be adding some soon. I also added context compaction to help reduce costs and make longer-running tasks possible. I also added light mode support and made the UI look better.

I started using Sprocket to create Sprocket and discovered a few random crashes that needed fixing. After fixing them, I started working on getting the app ready for production. I added convex-test to the CI which allows me to add tests for convex functions that actually talk with the backend database. Alongside this, I also made convex deployments automatic for preview deployments on PRs and production deployments whenever something is merged into main. I also worked on improving the system prompt of the agent to make it work better, this mostly involved dividing everything into better sections and explicitly telling it to prefer the scrape_url tool over web_search as web_search is more expensive. I also worked on adding proper usage limits and subscription tiers. The subscription tiers only exist in the backend right now since I don’t have a way to accept payments fully set up. I also created a new page in the settings which shows the amount of usage left, when it resets, etc.

Ok, so apparently, I worked for 33h and 43m and completely forgot to devlog. I was sooo deep into making the project that I completely forgot about Stardance. I created around 30 PRs/commits with wayyyyy too much work done in them. I started this project before Stardance (during Flavortown) but never shipped it as it was barely complete. Before my current work, the project was just a vibe-coded experiment to see if it would even be possible for me to create my own coding/robotics AI agent that could actually beat what the frontier labs are making. The real work started with these 30 commits, I basically rewrote the entire app. The current features include a full AI coding agent which can run shell commands, edit files, surf the web, and fetch web pages as markdown. The app allows one to use the SOTA AI models from OpenAI, Anthropic, and xAI, switch between different reasoning and service tier modes, send images to the agent, run multiple agents/threads in parallel across multiple projects, and archive threads. The work featured solving various edge cases and bugs, and doing a ton of testing. Even after all this, for me personally, this project is only 1% done. There is so much more work that I want to do on this and improve it with. Here’s to more consistent devlogs! lol.