Puddles

Puddles makes coding agents better at working your way, so you don’t have to repeat the same corrections. Built as a fork of T3 Code, an open-source workspace for coding agents, it supports Claude Code, Codex, Muse Code, Gemini CLI and OpenCode. Its learning loop turns your corrections into candidate rules. It tests each proposed change against the current version on held-out tasks and previously accepted work. Changes must demonstrate improvement and pass regression and cost checks before being retained. Every decision is recorded, with uncertain cases sent for your review. Subquadratic’s Snap identifies corrections and supports decisions. Meta’s Muse Spark proposes rule variants, Muse Code runs evaluations, and Google Gemini judges checks that scripts cannot score. In my hackathon evaluation on Muse Code, the proportion of failed lesson checks fell from 35% to 11% across 10 held-out tasks, run twice per version. Three rules were retained after testing and human approval. Puddles also uses Snap to choose better moments for Claude Code to compact its context. Its interface shows what it learned, which changes passed, and why others were rejected.