agent-workflow-kit

A Python CLI that packages, installs, and manages Claude Code subagent personas and workflow knowledge assets. It is the opinionated workflow layer that sits on top of awesome-claude-hooks -- the two projects complement each other but remain independent.

Why

Setting up a Claude Code workflow system on a new machine means manually copying 12+ markdown files into the right directories, editing ~/.claude/CLAUDE.md to add @ references, and hoping you remember which version is current. agent-workflow-kit automates that entire lifecycle: install, update, customize, and uninstall -- with checksums, backups, and conflict detection.

Installation

pipx install agent-workflow-kit

Or with uv:

uv tool install agent-workflow-kit

Requires Python 3.12+.

Quick Start

# Install all assets into ~/.claude/
agent-kit install

# Check installation health
agent-kit doctor

# See what's available
agent-kit list

Commands

Common Flags

agent-kit install --dry-run          # Preview without making changes
agent-kit install --force            # Overwrite existing files (creates backups)
agent-kit install --skip-claude-md   # Install files without touching CLAUDE.md
agent-kit install --claude-home /path/to/dir  # Override default ~/.claude/
agent-kit update --check-only        # Report status without modifying
agent-kit update --force             # Overwrite user-modified files
agent-kit uninstall --keep-files     # Only remove manifest and CLAUDE.md section

Bundled Assets

Agents (installed to ~/.claude/agents/)

Workflows (installed to ~/.claude/knowledge/workflows/)

Guidelines (installed to ~/.claude/knowledge/)

How It Works

Install Flow

1. Creates target directories under ~/.claude/ (agents, knowledge/workflows, etc.)
2. Copies each bundled markdown asset to its target location
3. Injects a sentinel-bounded managed section into ~/.claude/CLAUDE.md with @ references
4. Writes a manifest (~/.claude/.workflow-kit/manifest.json) tracking installed files with SHA-256 checksums

Update Flow

1. Reads the manifest to identify installed files
2. Compares on-disk checksums against manifest (detects user modifications)
3. Compares bundled assets against manifest (detects upstream changes)
4. For each file, reports one of: current, user modified, outdated, missing, conflict
5. Applies updates with backup, skipping user-modified files unless --force is used

CLAUDE.md Management

The kit uses HTML comment sentinels to manage its section:

<!-- agent-kit:begin -- managed by agent-workflow-kit, do not edit -->
## Agent Kit Agents
- Pair programming agent: @~/.claude/agents/Ada.md
...
## Agent Kit Workflows
...
## Agent Kit Guidelines
...
<!-- agent-kit:end -->

User content outside the sentinels is never touched. The section is cleanly removed on uninstall.

Development

This project uses just as a task runner and uv for package management.

Requirements

• Python 3.12+
• uv
• just

Setup

git clone https://github.com/mandelbro/agent-workflow-kit.git
cd agent-workflow-kit
uv sync --extra dev

Running locally

# Run via just
just run -- install --dry-run

# Or directly with uv
uv run agent-kit --help

Commands

# Run all checks (format, lint, typecheck, test)
just

# Individual commands
just test       # uv run pytest
just lint       # uv run ruff check --fix src/ tests/
just format     # uv run ruff format src/ tests/
just typecheck  # uv run pyright
just coverage   # pytest --cov --cov-fail-under=80

Releasing

Releases are automated via release-please. On every push to main, release-please opens (or updates) a release PR with a changelog derived from Conventional Commits. Merging that PR triggers a GitHub Release and publishes to PyPI via trusted OIDC publishing.

To publish manually: Actions > Release > Run workflow with a git tag (e.g. v0.2.0).

Project Structure

src/claude_workflow_kit/
├── cli/
│   ├── app.py          # Typer app entry point (7 commands)
│   ├── install.py      # install / uninstall
│   ├── update.py       # update / diff
│   └── status.py       # status / list / doctor
├── core/
│   ├── config.py       # Pydantic settings (CLAUDE_HOME, derived paths)
│   ├── manifest.py     # Manifest read/write, SHA-256 checksums, file status
│   ├── file_ops.py     # Atomic write, backup, copy, remove
│   ├── claude_md.py    # Sentinel-based CLAUDE.md section management
│   └── registry.py     # Asset registry mapping bundled files to targets
└── assets/             # 12 bundled markdown files
    ├── agents/         # 6 agent personas
    ├── workflows/      # 3 workflow definitions
    ├── guidelines/     # 2 guideline files
    └── coding-principles/  # 1 coding principle

Configuration

The default install target is ~/.claude/. Override it with:

# Environment variable
export CLAUDE_HOME=/path/to/custom/dir
agent-kit install

# CLI flag (per command)
agent-kit install --claude-home /path/to/custom/dir

License

Apache-2.0