Inspiration
AI coding agents are brilliant and forgetful. You fix a bug on Monday, and by Friday the agent writes the very same bug again. A line in a rules file doesn't stop it: nothing enforces it, and nothing tells you whether it worked.
What it does
TapWire turns every mistake an AI agent makes into a permanent, enforced rule, and measures how often the agent repeats a mistake you already fixed (the Amnesia Rate).
tapwire capturereads the commit that fixed a bug and saves a "scar": what went wrong, the bad pattern, the correct pattern and the severity. It writes the scar into AGENTS.md so the agent reads past mistakes before it writes code.tapwire agent-checkruns right after an agent edits a file. If the edit repeats a past mistake, the agent gets a message with what it wrote, why it's wrong and what to do instead, and can fix it in the same turn.- A pre-commit hook refuses a commit that repeats a scar (naming file and line), and
tapwire check --basecan fail a pull request in CI. - A Next.js dashboard charts the Amnesia Rate, lists every scar and shows how each repeat was caught. You can also paste code into the Try it page and see the exact line that repeats a scar.
How I built it
A TypeScript monorepo: a core library, the tapwire CLI, and the Next.js 15 dashboard. I planned it with the Devpost Learn Skill Pack (scope.md, prd.md, spec.md and a build checklist are in the repo), then built it step by step. There are 53 unit tests plus an install test that packs the CLI and runs it in an empty project. The demo video was made with Remotion from real terminal output and real dashboard captures, with an AI voice-over.
Challenges
Making check precise enough to catch real repeats without flagging corrected code, ignoring TapWire's own managed files, and making the hook fail safe so it never blocks a commit by accident.
What I learned
Telling an agent "don't do this" isn't enough; you need to enforce it and measure it. Honest limits: TapWire can't stop a model from writing a mistake, it makes sure a known mistake is noticed, named and fixed. The demo runs in mock mode (no API key); live mode uses the Anthropic API.
What's next
Live-mode testing with more languages, more agent integrations, and team-shared scars.
Code: https://github.com/Kaushik2210/tapwire | Live demo: https://tapwire-xi.vercel.app
Built With
- next.js
- node.js
- react
- remotion
- typescript