evolve-loop или comprendo

A self-learning development pipeline that gets smarter with every cycle. It calibrates itself against your codebase, dynamically selects the highest-impact tasks, implements them in isolated git worktrees, and blocks bad code through eval gates — then learns from each cycle to make better decisions next time. 4 specialized agents (Scout, Builder, Auditor, Operator) work across 5 phases: DISCOVER → BUILD → AUDIT → SHIP → LEARN. The Scout uses multi-armed bandit scoring and 8-dimension quality benchmarks to target your codebase's weakest areas. The Builder implements in isolation. The Auditor blocks MEDIUM+ severity issues. After each cycle, the pipeline extracts reusable "instincts" — patterns about what worked and what failed — that compound over time. Every 5 cycles, a meta-cycle evaluates the pipeline itself and evolves its own agent prompts based on measured performance. It ran 151 cycles on its own codebase, shipping 67+ improvements and raising its quality score from 83.8 to 94.4 — all autonomously. No e

Comprendo gates AI implementation behind a visual developer approval step. When you type /comprendo <task>, Claude reads your codebase, builds a structured plan with before/after architecture diagrams and code previews, then opens a browser UI where you review everything before a single file is touched. No more coming back to a multi-file diff you have to reverse-engineer — you see exactly what's about to change, why, and how, then you decide.