OSSFactory-Scaler

Autonomous AI agent team that scales, maintains, and improves open-source repositories. Runs 25 VDays/day on a $5/day OpenRouter budget with 5 specialized agents.

Agents

Agent	Role	Model Tier
Scout	Scans repos, audits quality, prioritizes work queue	micro
Builder	Clones repo, generates upgrades, repair cascade, pushes releases	fast -> standard -> engineering
Demo	Creates/updates SEO-optimized demo pages, deploys to VPS	fast
Maintainer	Triages issues, labels, auto-responds, health scoring	micro
Critic	Reviews Builder output, runs quality gates, daily observation	micro

Install

git clone https://github.com/AdametherzLab/OSSFactory-Scaler.git
cd OSSFactory-Scaler
bash scripts/setup.sh

Configuration

Copy .env.example to .env and fill in your keys:

cp .env.example .env

Required:

• OPENROUTER_API_KEY — Your OpenRouter API key
• GITHUB_ORG — GitHub org/user to scan (default: AdametherzLab)

Optional:

• OSS_BOT_TOKEN / TELEGRAM_USER_ID — Telegram notifications
• VPS_HOST / VPS_USER / SSH_KEY_PATH — Demo page deployment
• DAILY_BUDGET_USD — Daily spending limit (default: $5)
• VDAYS_PER_DAY — Virtual days per real day (default: 25)

Usage

# Run directly
bun run src/index.ts

# Run with PM2
npx pm2 start ecosystem.config.cjs

# Run tests
bun test

How It Works

Each VDay (~58 minutes), the agents run sequentially:

1. Scout scans repos via gh repo list, audits quality, queues work items
2. Builder picks top work item, clones repo, generates upgrade via AI cascade, runs quality gates
3. Critic reviews recent work, writes 333-char observation
4. Demo generates SEO demo pages for recently shipped repos, deploys via SCP
5. Maintainer triages open issues, labels them, computes health scores

Every 5th VDay, a team meeting fires with a Slicing Pie leaderboard.

Slicing Pie Points

Action	Points
Ship a release	+10
Create demo page	+8
Fix/triage an issue	+5
Update demo page	+5
Quality improvement	+3
Successful review	+2
Failed ship	-3
Regression introduced	-5
Budget overrun	-2

Quality Gates

1. Compile — bun build passes
2. Tests — bun test with >= 50% pass rate
3. README — >= 800 chars with install + usage sections
4. Security — No eval(), hardcoded secrets, or IP addresses
5. Ship-ready — Composite >= 70/100

Model Tiers

Tier	Model	Cost (in/out per 1M)
micro	Gemini 2.5 Flash Lite	$0.075 / $0.30
fast	Gemini 2.5 Flash	$0.15 / $0.60
standard	Kimi K2.5	$0.45 / $2.20
engineering	DeepSeek R1	$0.55 / $2.19

Data Files

All persisted in data/ (gitignored):

• scaler-state.json — Work queue, audits, completed work
• token-usage.json — Per-model spend log
• slicing-pie.json — Agent reward history
• builds/ — Temporary clone directory

Requirements

• Bun runtime
• GitHub CLI (gh) authenticated
• OpenRouter API key

License

MIT