Caretaker

Caretaker is an autonomous repository-maintenance system. A Python control plane on AKS receives GitHub App webhooks, decides what should happen next, and dispatches code changes to one of four coding backends — GitHub Copilot, an in-process LLM tool loop (Foundry), the opencode_local / claude-code-action hand-off, or a per-task Kubernetes Job.

It watches pull requests, issues, CI failures, upgrades, stale work, and security findings, then routes the right next action instead of waiting for a human to play inbox pinball.

What it does

• triages pull requests and CI failures
• classifies and dispatches issues
• opens upgrade issues when newer caretaker releases exist
• raises dependency and security findings for remediation
• produces documentation and escalation digests
• tracks state across runs using GitHub itself

Start here

• Getting started
• Architecture — the runtime topology, webhook event pipeline, coding-job lifecycle, and 18-agent inventory
• Configuration reference
• Agent overview
• Development workflow

Repository pointers

• setup guide: setup-templates/SETUP_AGENT.md
• default config template: setup-templates/templates/config-default.yml
• schema: schema/config.v1.schema.json
• source package: src/caretaker

Dogfooding

This repository is also the central caretaker repository, which means it uses its own automation to maintain itself. Very on-brand. Slightly terrifying. Mostly useful.