little-coder

About

Small local models fail at agentic coding mostly because harnesses assume frontier-model context discipline, so little-coder wraps the pi agent with extensions that enforce read-before-edit, gate permissions, inject skills per turn, watch for compaction, and monitor output quality — all as lifecycle hooks rather than core patches. Plan mode dispatches isolated read-only sub-coders for research and hands a written plan to a fresh session on the action model, with separate /plan-model and /action-model commands for big-plan/small-implement economics. Everything ships as extensions and skills around pi, keeping cold-start context near 7k tokens. Hobbyists running Qwen-class models on consumer laptops are the target audience, with published Terminal-Bench and GAIA results.