Versioned source for the JorisJonkers-dev agent renderer, installer kit, and portable runner runtime templates.
agent-kit owns the checked-in agent templates, generated Claude/Codex surfaces,
council bundle, installer artifacts, and runner runtime manifests used by
JorisJonkers-dev automation.
uv sync --frozen
uv run python render-agent-kit.py --check
uv run python render-agent-kit.py --doctor
uv run ruff check .
uv run mypy
uv run pytest
uv run python scripts/validate_manifest.py
uv run python scripts/render_registry.py --checkUse uv run python render-agent-kit.py --write only when intentionally updating
checked-in generated surfaces from templates/ or council/. Use
scripts/build-runtime-package.sh v0.0.0-local dist to build the runtime-home
bundle locally.
./installer/setup-workstation.sh --check # report only
./installer/setup-workstation.sh # install/upgrade everythingClaude Code and Codex at their latest releases, every plugin, the language
servers those plugins drive, the MCP fleet, and the first-party skills. The
script is generated from registry/estate-tooling.yaml -- see
docs/SETUP.md.
- docs/SETUP.md: workstation setup, credentials, exit codes.
- docs/REGISTRY.md: the one file to edit to add a skill, plugin, CLI or MCP server to every surface at once.
- docs/HERMES.md: getting a registry entry onto the in-cluster gateway, and verifying it landed.
- docs/MEMORY.md: what replaces the retired knowledge base.
- docs/CLOUD.md: the same setup in a Claude Code cloud
environment, from the two files in
cloud/. - skills/critique/README.md: the critique pipeline for its user: what runs when, depth, budget, the gates, the kill switch and how to retune routing.
- docs/critique-pipeline/integration.md: what the critic CLIs were shown to do, where the critique pipeline lands in agent-kit, and where it departs from its brief.
The estate ships exactly three agent hooks, the critique pipeline's
commit gate, PR gate and session-start line
(ADR-0002).
setup-workstation.sh merges them by name into Claude Code's settings.json
and Codex's hooks.json, leaving your own hooks alone, and purges the retired
knowledge hooks from a machine that still has them, on every run.
registry/estate-tooling.yaml: the tooling registry (hand-edited).installer/setup-workstation.sh,registry/generated/hermes/: generated from it byscripts/render_registry.py. Never hand-edit these.skills/: first-party multi-file skills copied onto a surface.templates/repo/: source templates for.claude,.codex,.agents, and project Spec Kit seed payload.templates/runner-runtime/: source templates for runner runtime artifacts.installer/:setup-workstation.shandsetup-container.sh, generated as above, plus the hand-writtenport-forward-agent.shandcloud-session.shthey source.cloud/: the setup script and environment variables pasted into a Claude Code cloud environment.runner-manifests/: rendered runtime package artifacts and validation fixtures.council/: council driver, prompts, schemas, and default config.manifest.yaml: managed path, checksum, parity, and artifact ledger.
From a checkout:
./installer/setup-workstation.shOr fetched over HTTP; outside a checkout it fetches the kit bundle of the same release and runs that:
curl -fsSL https://github.com/JorisJonkers-dev/agent-kit/releases/latest/download/setup-workstation.sh | bash--uninstall removes what a machine set up by the retired install.sh /
install-agents.sh wrote (base skills, Spec Kit commands/skills, their MCP
entries), purges the retired knowledge hooks and removes the critique hooks;
a later run without it installs them again. See
PORTABILITY.md.
Copyright (c) Joris Jonkers. Source available for viewing only; use, copying, modification, redistribution, deployment, or reuse is not licensed. See LICENSE.