Introduction

Last Light is an open-source GitHub repository maintenance agent. It sits in front of one or more repositories, listens for events, and runs an AI agent against them to triage issues, review pull requests, respond to comments, and — when you ask — build features end-to-end through an Architect → Executor → Reviewer cycle.

It is built on the Claude Agent SDK with a lightweight TypeScript harness that handles webhook ingestion, Slack messaging, cron scheduling, sandbox isolation, and an admin dashboard. Every piece of non-trivial behavior is expressed as a YAML workflow — the harness itself knows nothing about "build" or "triage" specifically; it just executes the phases a workflow declares.

What it does out of the box

• Triages new issues — labels, deduplicates, chases missing info.
• Reviews pull requests — posts severity-ranked feedback with file:line references.
• Builds features on request — Architect plans, Executor implements, Reviewer verifies, fix-feedback loops, PR opened.
• Writes weekly health reports — open/stale issues, PR velocity, action items.
• Talks to you on Slack — per-thread conversation, DB-backed session history.
• Gives you a dashboard — live activity, pipeline DAGs, full session drill-down.

Who it's for

Last Light is aimed at maintainers of repos that have moved into maintenance mode — libraries, side projects, and internal tools that still matter but aren't getting daily attention. It keeps the lights on after you've moved on.

It runs as a single container you deploy somewhere you control. There is no SaaS — you own the webhook endpoint, the Docker volume, and the audit trail.

How this documentation is organized

• Getting Started walks you through prerequisites, creating a GitHub App, running locally, deploying to production, and wiring up Slack.
• Reference documents every environment variable the harness reads.