RuneGärd

Autonomous Kubernetes runbook executor with continual learning.

What it does

1. Parses markdown runbooks into executable decision trees
2. Executes them step-by-step against a live K8s cluster
3. Asks for approval before any mutating action
4. Learns from failures via an CL loop that improves the skill over time

Quick Start

Install

uv sync
make setup  # installs git hooks

As a CLI

# Parse a runbook
uv run python -m runegard parse assets/runbooks/crashloop.md

# Execute (dry-run)
uv run python -m runegard run assets/runbooks/crashloop.md --dry-run

# Execute (live, against a K8s cluster)
uv run python -m runegard run assets/runbooks/crashloop.md

# Analyze failures and improve
uv run python -m runegard improve trace_log.json --runbook assets/runbooks/crashloop.md

As a Claude Code Skill

Add the skill.md to your Claude Code skills directory, then:

"Run the crashloop runbook against my cluster"

Development

make fmt        # format code
make lint       # run ruff linter
make typecheck  # run ty type checker
make test       # run all tests
make audit      # run all quality checks

Demo

# 1. Create kind cluster with seeded failures
./demo/setup_cluster.sh

# 2. Run the full demo
./demo/run_demo.sh

Environment

Requires:

• Python 3.12+
• uv for package management
• ANTHROPIC_API_KEY environment variable
• kubectl configured for your cluster (for live execution)