jellydn/ai-flow：把 GitHub URL 变成结构化 AI 工作流

AI Flow

Turn GitHub URLs into structured AI workflows — without writing a prompt.

What is AI Flow?

AI Flow packages common software engineering tasks into reusable launchers. Pick a launcher, paste a GitHub URL, and watch as the queue worker gathers context, calls an AI provider, validates the structured response, and delivers a shareable report — all in real time.

🔍 Review PR	Identify risks, findings, and verification steps for any public pull request.
📋 Plan Issue	Turn a GitHub issue into an actionable implementation plan with tasks.
📖 Explain Repo	Summarize a public codebase — architecture, structure, and key patterns.
🩺 Laravel Doctor	Inspect a Laravel project and recommend improvements across conventions, security, and performance.
✨ Custom launchers	Signed-in users can create their own launchers with custom names, prompts, and JSON output schemas. Custom launchers appear alongside built-in ones on the home page.

Every run produces a structured, schema-validated report — not a wall of chat text. Each finding carries a severity label, file references, and concrete fix suggestions. Results are shareable by URL. Signed-in users can mark runs as private (default) or public via a visibility toggle.

✨ Why AI Flow?

	AI Flow	Chat-based AI tools
Output	Structured JSON → polished report	Free-form markdown
Prompting	Zero — pick a launcher	Must craft context manually
Context	Automatic GitHub REST fetches	Copy-paste code snippets
Validation	JSON Schema enforced	Hope the LLM follows format
Sharing	Shareable result URLs	Chat links, screenshots
Real-time	SSE progress streaming	Polling or manual refresh

🚀 Quick Start

git clone https://github.com/jellydn/ai-flow.git
cd ai-flow/backend

cp .env.example .env
composer install
npm install
php artisan key:generate
touch database/database.sqlite && php artisan migrate --seed

# Add your API keys
# Required: OPENROUTER_API_KEY (for guest runs)
# Optional: OPENAI_API_KEY, GITHUB_TOKEN

composer run dev

Open http://localhost:8000 and launch your first workflow.

Prerequisites: PHP 8.4+, Composer, Node.js 24+, SQLite. See backend/README.md for full setup, sign-in, credentials, and API details.

⚡ API in 30 seconds

# Discover available launchers
curl http://localhost:8000/api/launchers

# Start a run (returns 202 + UUID)
curl -X POST http://localhost:8000/api/runs \
  -H 'Content-Type: application/json' \
  -d '{"launcher": "explain-repository", "source_url": "https://github.com/laravel/framework"}'

# Check status or stream progress
curl http://localhost:8000/api/runs/RUN_UUID
curl -N -H 'Accept: text/event-stream' http://localhost:8000/api/runs/RUN_UUID/stream

🧱 Architecture

Browser                                 Queue Worker
  │                                         │
  ├─ POST /api/runs ──▶ API ──▶ DB Queue ──▶ ExecuteLauncherJob
  │                         │                    │
  ├─ SSE /api/runs/{id}/stream               GitHubService
  │                         │                    │
  └─ GET /api/runs/{id}    │              AIProviderInterface
                                        (OpenAI · OpenRouter
                                         Anthropic · Gemini)
                                               │
                                          JsonSchemaValidator
                                               │
                                          runs.result ✅

Layer	Stack
UI	React 19, TypeScript, Vite
API	Laravel 13, PHP 8.4
Async	Laravel database queue + SSE
AI	OpenAI, OpenRouter, Anthropic, Gemini
Storage	SQLite (dev), PostgreSQL/MySQL (prod)
Deploy	Dokku · Laravel Cloud

🔐 Authentication & Providers

Access	Behavior
Guest	Uses OpenRouter's free model router. No setup needed.
Signed in	Choose any provider (OpenAI, OpenRouter, Anthropic, Gemini), pick a model, and use a one-time key or encrypted saved credential.

Provider keys are never stored on runs, logged, or returned by the API. Saved credentials are encrypted at rest and decrypted only by the worker executing the run. Only public https://github.com/... URLs are accepted.

🧪 Quality

# Backend
php artisan test                    # 216 tests, 635 assertions
./vendor/bin/pint --test            # PSR-12

# Frontend
npm run typecheck                   # tsc --noEmit (strict)
npm run lint                        # oxlint + oxfmt
npm run konsistent                  # Structural TS conventions
npm run build                       # Production build
npm test                            # Vitest + React Testing Library
npm run test:e2e                    # Playwright (real backend)

CI runs the backend suite on PHP 8.4 and the frontend suite on Node 24. Pre-commit hooks via prek.

📚 Documentation

Document	What it covers
`backend/README.md`	Full app setup, API reference, auth, credentials, Docker, cloud deploy
`doc/adr/`	Architecture decision records (24 ADRs)
`DESIGN.md`	Visual identity — colors, typography, components, spacing
`AGENTS.md`	AI coding assistant guide — conventions, gotchas, commands
`.planning/codebase/`	Codebase map — stack, architecture, conventions, testing, concerns

🌍 Demo

Try it live at ai-flow-staging.itman.fyi.

🤝 Contributing

Issues and pull requests are welcome! Here's how:

Fork the repo and create a feature branch
Make your changes — keep them focused and include tests
Run the checks — php artisan test, ./vendor/bin/pint --test, npm run typecheck, npm run lint
Open a PR — describe what, why, and how

All contributions are under the MIT License. See AGENTS.md for AI-assisted development conventions.

👤 Author

Dung Huynh (@jellydn)

🔗 GitHub · Website · X

📄 License

MIT © jellydn

_{Built with ❤️ using Laravel, React, and a lot of ☕}