# MoAI - Master of AIs

## What This Is

A multi-AI collaborative brainstorming Telegram bot where multiple AI models (Claude, GPT, Gemini) discuss topics together, see each other's responses, and work toward consensus. Users can run parallel queries, conduct sequential discussions, @mention specific models, and export conversations with AI-generated consensus summaries.

## Core Value

Get richer, more diverse AI insights through structured multi-model discussions—ask a team of AIs instead of just one.

## Requirements

### Validated

- ✓ Project scaffolding (pyproject.toml, ruff, pre-commit, src layout) — v1.0
- ✓ M1: Bot responds to /help, /status — v1.0
- ✓ M2: Project CRUD (/projects, /project new, select, delete, models, info) — v1.0
- ✓ M3: Single model Q&A working (/ask command) — v1.0
- ✓ M4: Open mode (parallel) with multiple models — v1.0
- ✓ M5: Discuss mode (sequential rounds) — v1.0
- ✓ M6: Consensus generation (/consensus) — v1.0
- ✓ M7: Export to markdown (/export) — v1.0
- ✓ M8: @mention direct messages — v1.0

### Active

(None — define requirements for next milestone)

### Out of Scope

- Web UI — Phase 2, after Telegram POC is validated
- Multi-user collaboration — Phase 3 future
- Personas (optimist/critic/pragmatist modes) — future enhancement
- Voting/tallying — future enhancement
- Cross-project memory — future enhancement
- Automated triggers/webhooks — future enhancement
- Voice memo transcription — future enhancement

## Context

**Current state:** v1.0 shipped. Telegram bot fully functional with 2,732 LOC Python.

**Tech stack:** Python 3.11+, python-telegram-bot, SQLAlchemy + SQLite, OpenAI SDK (Requesty/OpenRouter compatible), pytest, ruff.

**Architecture:** Telegram command → Handler → Service → Orchestrator → AI Client → Requesty/OpenRouter → AI APIs

**Known tech debt (from v1.0 audit):**
- Missing test coverage for database error paths
- Error handling enhancement deferred (comprehensive retry/timeout)
- Allowed users middleware not enforced (defined but unchecked)
- Minor code patterns: orphaned export, inconsistent get_selected_project, missing re-exports

## Constraints

- **Python version**: 3.11+ — required for modern async patterns
- **Bot framework**: python-telegram-bot (async) — spec requirement
- **Database**: SQLAlchemy + SQLite — upgrades to PostgreSQL in Phase 2
- **AI routing**: Modular abstraction layer — Requesty first, support OpenRouter and others
- **Linting**: ruff (line length 100) — enforced via pre-commit
- **Testing**: pytest, 80%+ coverage on core logic
- **Type hints**: Required on all public functions
- **Docstrings**: Required on modules and classes
- **Logging**: logging module only, no print()
- **Dependencies**: Unpinned unless security requires it

## Key Decisions

| Decision | Rationale | Outcome |
|----------|-----------|---------|
| AI client as abstraction layer | Support Requesty, OpenRouter, direct APIs without changing core code | ✓ Good — OpenAI SDK works with both routers |
| Full project scaffolding first | Consistent tooling from day one; prevents tech debt | ✓ Good — ruff/pre-commit caught issues early |
| User allowlist auth (Phase 1) | Simple for single-user POC | ⚠️ Revisit — middleware defined but not enforced |
| hatchling as build backend | Explicit src layout config | ✓ Good |
| String(36) for UUID storage | SQLite compatibility | ✓ Good |
| Module-level singletons | Simple pattern for AI client and database | ✓ Good — consistent access everywhere |
| asyncio.gather for parallel | Concurrent model queries in /open | ✓ Good — graceful per-model error handling |
| user_data for state | Store discussion context across commands | ✓ Good — clean multi-command flows |
| JSON format for consensus | Structured AI output parsing | ✓ Good — reliable extraction |
| BytesIO for export | In-memory file generation | ✓ Good — no temp files needed |

---
*Last updated: 2026-01-17 after v1.0 milestone*