<

AI-powered social recovery agent — confess what you did wrong, let Sir Alibi craft the perfect apology, draft a Gmail, schedule a Calendar follow-up, and send a real gift card (Amazon, Starbucks, Subway, and more) matched to the severity of what you did. Powered by OpenAI via Lava AI Gateway. Built at YHack 2026.

Quick Start · Architecture · API Reference · Tech Stack

You forgot. You flaked. You went MIA.

Sir Alibi is your autonomous relationship-repair agent. Tell it what you did wrong — it figures out the severity, constructs the alibi, writes the apology email, sends a real brand gift card (Amazon, Starbucks, Subway, etc.) scaled to how badly you messed up, and schedules the follow-up. You just show up.

Confess → Assess → Alibi → Apology → Gift → Follow-up

• Node.js 18+
• npm
• Google Cloud project (Gmail API enabled)
• OpenAI API key
• Lava account (AI gateway)
• Tremendous sandbox account

git clone https://github.com/your-org/the-alibi.git
cd the-alibi

# Root (agent test harness)
npm install

# Backend (Express API)
cd backend && npm install

# Frontend (React + Vite)
cd ../frontend && npm install

cp backend/.env.example backend/.env
# Edit backend/.env with your keys — see Environment Variables below

Open two terminals:

# Terminal 1 — Backend (port 3001)
cd backend
node server.js

# Terminal 2 — Frontend (port 5173)
cd frontend
npm run dev

Then open http://localhost:5173 and confess your crime.

┌─────────────────────────────────────────────────────────────────────────┐
│  Frontend (React + Vite, :5173)                                         │
│  • Form → POST /api/run-agent (SSE) — optional Auth0 Bearer token       │
│  • “Draft in Gmail” → POST /api/send-apology-email                      │
│  • “Schedule follow-up” → POST /api/schedule-followup (stub success)    │
└───────────────────────────────┬─────────────────────────────────────────┘
                                │
                                ▼
┌─────────────────────────────────────────────────────────────────────────┐
│  Backend (Express, :3001) — backend/server.js                           │
│  • Loads agent via dynamic import of agent/runAlibiAgent.ts (tsx/node)  │
│  • Google OAuth in-memory session → Gmail drafts (googleapis)           │
└───────────────────────────────┬─────────────────────────────────────────┘
                                │
                                ▼
┌─────────────────────────────────────────────────────────────────────────┐
│  Agent (TypeScript) — agent/runAlibiAgent.ts                            │
│  • Perception → Research → [pause gate] → Reason → Write                │
│  • LLM: Lava forward → OpenAI-compatible chat/completions (agent/llm.ts)│
└─────────────────────────────────────────────────────────────────────────┘

The workflow is deliberately staged. Each step has its own system prompt, user payload (including outputs from earlier steps), and Zod schema (agent/schemas.ts). The model is asked for one JSON object per step; callLLMJson (agent/llmJson.ts) parses and validates it, with one automatic retry if JSON or schema validation fails.

So “reasoning” is not one monolithic answer: it is factored so later steps cannot ignore structured earlier outputs unless the run is intentionally paused or a fallback activates (agent/repairFallbacks.ts).

If a caller attaches perceptionContext.calendar.nextFreeWindowISO, runAlibiAgent maps that timestamp into days-from-now for the bundled follow-up action (resolveFollowUpDays in agent/runAlibiAgent.ts). The HTTP path today passes text only from the backend (backend/services/runService.js); PerceptionContext is fully wired in agent/testHarness.ts for judges who want to experiment with rich metadata without changing the UI.

Set DEMO_MODE=true (see agent/demoMode.ts + agent/config.ts) for smaller prompts and token caps across steps.