Delta Code docs

Run the AI migration review bot.

Start the control plane, worker, migration inbox, GPT-4o intelligence, and optional sandbox executor locally—or connect the hosted product to a GitHub App installation.

PrerequisitesPython 3.12+, Node.js 22.13+, Docker, Git, and Make.

Local quickstart

Install dependencies and start PostgreSQL:

make setup
make db-up
make db-schema

Then run the API, worker, and frontend in separate terminals:

make api
make worker
make frontend-dev LIVE_API_URL=http://localhost:8000

Architecture

Provider source→Impact analysis→GPT-4o→Sandbox→Draft PR

Local URLs

Dashboard	`http://localhost:3000`
Backend API	`http://localhost:8000`
Interactive API docs	`http://localhost:8000/docs`
Health endpoint	`http://localhost:8000/health`

Core API

GET/auth/meCurrent GitHub identity and repositoriesGET/migrationsWorkspace-scoped migration review inboxPOST/migrations/{id}/generateQueue a bounded AI migration attemptGET/providersConfigured provider sources and sync healthGET/migrations/{id}Migration evidence, patch, checks, and historyPOST/migrations/{id}/publishPublish an approved migration as a draft pull request

AI assistance

Set OPENAI_API_KEY on the Railway worker and WORKSPACE_INTELLIGENCE_ENABLED=true on the Railway web service. Vercel only needs the backend API URL; the key must never be exposed through a browser or NEXT_PUBLIC_ variable.

OPENAI_API_KEY=your-server-side-key
OPENAI_MODEL=gpt-4o
WORKSPACE_INTELLIGENCE_ENABLED=true
LLM_DAILY_BUDGET_USD=1.00
LLM_TOTAL_BUDGET_USD=9.00

Model requests use strict structured output, no model tools, bounded context and output, store: false, retry limits, and request-level cost limits. Repository and provider text is treated as untrusted data.