Session Lifecycle
Every development session follows a structured lifecycle — boot, execute, close. State is read on entry, written on exit. No implicit knowledge.
Documentation-first protocol for solo developers working with AI. Work produces documents, documents feed the vector DB, the vector DB remembers everything.
Supervisor (human) Orchestrator (AI)
│ │
├── sets objectives ──────────────→ boot: read state, history, protocol
├── injects ideas ───────────────→ capture: log immediately, triage later
├── approves/steers ──────────────→ execute: dispatch agents, update OCD
└── reads status ←────────────── close: finalize docs, embed, snapshotDOSI sits between you and your AI coding partner. It gives the AI operational memory — where you are, what happened before, and how to work.
Without this, every session starts cold. With it, session 47 knows what session 1 decided.
| Component | What It Is | Analogy |
|---|---|---|
| DOSI | The system — protocol + embedding pipeline | Git workflow docs |
| OCD | Context unit — directory of everything from one session | A git commit |
| STATE.md | Live snapshot — where are we now | Working tree status |
| CLAUDE.md | Project config — what are we building | README for the AI |
| Knowledge Network | Vector memory — recall anything | Searchable brain |
Bepely — Created by Bepely. Battle-tested across 60+ sessions building production software with Claude.