Somasundaram

Your repository shouldn't adapt to an AI reviewer. The reviewer should adapt to your repository. Today I'm releasing Pipr: code-owned AI review, powered by @pidotdev. Keep the reviewer in .pipr/config.ts. Run it in your environment. pipr.run 1/8

This is so cool!

Built a monster-battling game you play over SSH. Pokémon Red nostalgia got me started. LLMs made me willing to attempt it. I'm glad I did. Original monsters and multiplayer battles. ssh termon.sh termon.sh

found out today that codex can turn the conversation rails into an audio visualizer lol Settings → Toys → Audio visualizer, if you want your codex to vibe with you

Astra is definitely a step up from sol but it is still prone to over engineering things if not steered right, same as it's predecessor

I care less about the happy path demo. What does the agent do on the second failure. Retry the same broken step, ask you, or rewrite the plan. That branch is the actual product.

Sneak peek at what I've been hacking on this weekend. I wanted to recapture the nostalgia of playing Pokémon Red as a kid, plus a little PvP over SSH. Pretty happy with how it's coming along. I'll be open-sourcing it soon. Coming to a terminal near you soon. #Termon #GameDev #SSH #buildinpublic

I started by trying to find the best LLM for code review. The results kept pointing elsewhere. Once a model crossed the capability floor, context, tools, validation, and evaluation mattered more. I wrote up the numbers and what they changed for Pipr.

Pipr 0.7 is out! You can now run the same AI code review workflow across GitHub Enterprise, GitLab Self-Managed, Gitea/Codeberg, Azure DevOps Server, and Bitbucket Data Center. Run diagnostics also make failed reviews much easier to debug.

I spent the last few weeks dogfooding Pipr on real repositories, and I kept holding off on the release until the boring parts felt solid: building the diff, running a review, validating what the model returned, and deciding what should reach a PR. I also ran it against @withmartian's Code Review Bench. It did reasonably well in my testing, which was reassuring, but I was never trying to build for a benchmark. I don't think a single reviewer can work for every codebase. Each repo has its own conventions, risks, and idea of what a useful review looks like. That is why Pipr has a small core you can shape around the project. Next, I want to make every review run observable. You should be able to see what Pipr caught, what it missed, and why, then turn those failures into evals. Over time, your Pipr setup should improve alongside the repository.