Ran into this a bunch while vibecoding with longer agent runs.
The agent confidently starts building something a teammate already tried months ago and closed the PR on. It has no way to know, because git only keeps what got merged. Every rejected attempt, and the reason it was rejected, just isn't there. So the agent happily walks into the same wall.
The reasoning behind a codebase (why the timeout is 30, why an approach got dropped) lives in closed PRs and reviews nobody reads again. Your agent can't see any of it.
How are you all handling this? Do you paste past decisions into context by hand, keep a decisions doc, or just catch it in review? Genuinely curious what's working, because review-only has not been enough for me.
The moment that started this: I watched a coding agent confidently begin rewriting something a teammate had closed a PR on months earlier. It had no way to know. Git keeps what got merged and throws away what didn't, so every rejected attempt, and the reason behind it, just disappears.
I spent the last two months building a fix for exactly that. It reads a repo's pull requests and reviews and answers why the code is the way it is. And when there's no real answer, it says "no one wrote this down" instead of making one up.
Launched it today. No manifesto, just a tool I kept wishing existed. Happy to talk to anyone building in this space.
After 2 months of building, Icarus is live on PH today.
It reads your repo's PRs, reviews and issues and answers why the code is the way it is, with the exact evidence on screen. When nobody wrote the reason down, it says so instead of guessing.
Works as a Mac app, a Chrome extension, and an MCP connector for coding agents.
Would genuinely love feedback, especially where it gets something wrong: https://try-icarus.vercel.app
