What problem does it solve? AI-assisted coding degrades across long sessions: requirements drift, earlier promises get silently broken, and trust rests on plausible-looking diffs. ADD keeps all project state on disk in a .add/ bundle so every change is driven by a frozen specification, red tests, and a recorded pass/fail receipt rather than conversation memory. ## Core Features & Use Cases - 3-beat task loop: Direction (rules, contract, red tests, one human freeze approval) → Build (code to green inside frozen scope) → Verify (evidence receipt, residue checks, recorded gate verdict of PASS, RISK-ACCEPTED, or HARD-STOP). - Intake lanes and session resume: routes requests to Quick, Task, Explore, or Milestone lanes based on size and sensitivity, and resumes any session from the bundle alone via add status. - Living specs and personas: lessons are filed as tagged deltas (DDD/SDD/UDD/TDD/ADD) that humans fold into specs, and project-fit persona lenses advise beats without ever lowering a gate. - Use Case: A team building a multi-milestone feature asks the agent to "start a task" for each slice; ADD freezes an approved contract per task, runs the test suite red then green, and records a JUnit-based receipt so milestone six never silently breaks milestone one. ## Quick Start Ask the agent to run add status to orient from the .add bundle, then describe the change you want so it can size the request and create a task.