HCF (Halt and Catch Fire)

Autonomous development plugin for Claude Code. Define requirements with a PM, then let parallel workers implement everything using TDD.

Table of Contents

• Overview
• Getting Started
  • Installation
  • Quick Start
• How It Works
  • Four Phases
  • Pipeline
  • Dependency Graph
  • Parallel Execution
  • TDD Methodology
• Reference
  • Files Created
  • Task Format
  • Skills
  • Outputs
• Architecture
• Design Principles
• Development
  • Local Testing
  • Testing Workflow
  • Debugging
• Changelog
• Contributing
• License

Overview

HCF separates planning (human-in-the-loop) from execution (fully autonomous):

Planning → Human + AI collaborate on requirements
Execution → Parallel TDD workers implement autonomously

Getting Started

Installation

Add the marketplace, install, then reload:

/plugin marketplace add markshust/hcf
/plugin install hcf@hcf
/reload-plugins

Quick Start

1. Setup Project (one-time)

/project-setup

Interactively configures your project:

• Auto-detects tech stack (Laravel, React, etc.)
• Asks about testing, linting, architecture
• Creates .claude/ config directory

2. Plan a Feature

Just describe what you want:

"Help me implement user authentication with JWT"

The plan-create skill activates automatically to:

• Discover & brainstorm — explore the codebase and enumerate scope/assumption permutations a senior engineer would catch
• Create a feature/{plan-name} branch for the work
• Ask grounded clarifying questions categorized as must-answer vs will-default-if-silent
• Break down into tasks with dependencies
• Write requirements as test descriptions
• Create .claude/plans/user-auth/ with task files
• Run post-plan pipeline agents (devil's advocate by default) to find gaps before execution

After planning completes, you'll be asked:

Ready to begin autonomous implementation?

1. Yes, start now
2. No, I'll run it later

3. Execute Autonomously

If you chose "later" or want to re-run, say:

"Run the user-auth plan" or "Execute the plan"

The plan-orchestrate skill auto-triggers. It verifies you're on the correct feature/{plan-name} branch before starting, and automatically uses ralph-wiggum for session persistence if installed. If not installed, you'll see a warning but execution continues.

After completion, you'll be prompted to push the branch and create a PR (never done without your permission).

ralph-wiggum is prompted during /project-setup, or install manually:

/plugin marketplace add anthropics/claude-code
/plugin install ralph-wiggum@claude-code-plugins

How It Works

Four Phases

Phase	Type	What Happens
Setup	One-time	Configure project for autonomous dev
Planning	Interactive	Discover codebase, brainstorm scope, then define tasks with human guidance
Post-Plan Hooks	Automated	Agents enrolled at the post-plan hook review the plan
Execution	Autonomous	Parallel TDD implementation, with agents enrolled at the implementation hooks

Pipeline

The pipeline controls which agents run at fixed points in the plan/implementation flow. HCF uses convention over configuration: there is no central registry — each agent enrolls itself by declaring a phase in its own YAML frontmatter. If an agent declares a phase, it runs at that hook; if it has no phase, it never runs via a hook.

Hook points:

There are exactly 8 hook points where enrolled agents can run:

Hook	Fires
pre-plan	Before planning Discovery begins
post-plan	After the plan is built and validated, before user review
pre-implementation	Before the first implementation batch
pre-batch	Before each batch of TDD workers is spawned
post-batch	After each batch of TDD workers completes
post-implementation	After all tasks complete
pre-commit	After the full test suite passes, before the commit
post-commit	After the commit, before the push/PR prompt

By default, only devils-advocate is enrolled (at post-plan). See HOOKS.md for the authoritative reference — the full frontmatter schema, the deterministic discovery routine, and tie-break/ordering rules.

Enrollment frontmatter:

An agent enrolls by adding three keys to its frontmatter:

---
name: devils-advocate
description: "..."
model: opus
tools: Read, Write, Edit, Glob, Grep
# --- hook enrollment ---
phase: post-plan   # one of the 8 hook points above
order: 10          # lower runs first within a hook; default 100
mode: single       # "single" | "batch"; default "single"
---

Customizing the pipeline:

Enable, add, remove, or reorder agents by editing frontmatter — never a central file.

• Enable a built-in agent. standards-enforcer ships with its enrollment commented out. Uncomment its phase (and optional order / mode) to turn it on:

  ---
  name: standards-enforcer
  description: "..."
  model: opus
  tools: Read, Edit, Glob, Grep
  # To enable code-standards enforcement after implementation, uncomment:
  phase: post-implementation
  order: 50
  mode: batch
  ---

• Add a custom gate. Drop an agent file into your project's .claude/agents/ directory with a phase, and it is enrolled automatically:

  ---
  name: doc-updater
  description: "Updates documentation when implementation changes."
  model: sonnet
  tools: Read, Edit, Glob, Grep
  phase: post-implementation
  order: 100
  mode: single
  ---
  
  You are a documentation updater. Your job is to...
  {define the agent's behavior, process, and output format}

• Disable an agent. Remove (or comment out) its phase key — there is no condition to toggle.

The agent's filename (without .md) should match its frontmatter name. Local agents in .claude/agents/ override plugin agents with the same name (the local file wins entirely).

Upgrading from pipeline.md: Earlier versions of HCF configured the pipeline in a central .claude/pipeline.md file. That file is no longer read — agent frontmatter is now the only source of pipeline configuration. If a leftover .claude/pipeline.md is present, HCF blocks plan-create and plan-orchestrate (you'll see a prompt on session start and when you try to plan) until you run /project-update, which migrates your configuration into agent frontmatter and removes the file. /project-update is never blocked, and the gate clears automatically once the file is gone.

Dependency Graph

After generating a plan, HCF visualizes the task dependency graph so you can verify parallelism and ordering before execution:

001 ─┬─► 002 ─┬─► 005
     │        │
     └─► 003 ─┘
     │
     └─► 004 ────► 006

This tells the orchestrator which tasks can run in parallel and which must wait. In this example:

• Batch 1: Task 001 (no dependencies)
• Batch 2: Tasks 002, 003, 004 (all depend only on 001)
• Batch 3: Tasks 005, 006 (005 depends on 002+003; 006 depends on 004)

Parallel Execution

Independent tasks within each batch run simultaneously:

Batch 1: Task 001 (no deps)          → 1 worker
Batch 2: Tasks 002, 003, 004         → 3 parallel workers
Batch 3: Tasks 005, 006              → 2 parallel workers

100 tasks might complete in 5-10 batches instead of 100 sequential runs.

TDD Methodology

Each task follows strict Red → Green → Refactor:

1. Write failing test for one requirement
2. Write minimum code to pass
3. Refactor while tests stay green
4. Repeat for next requirement
5. Commit when task complete

Reference

Files Created

Project Config (.claude/)

File	Purpose
testing.md	Test commands, coverage requirements
code-standards.md	Linting, formatting rules
architecture.md	Directory structure, patterns

Plans (.claude/plans/{name}/)

File	Purpose
_plan.md	Plan overview, task table
001-{task}.md	First task with requirements
002-{task}.md	Second task
...	More tasks

Task Format

# Task 001: Create User Model

**Status**: pending
**Depends on**: none
**Retry count**: 0

## Description
Create the User model with authentication fields.

## Requirements (Test Descriptions)
- [ ] `it creates a user with valid email and password`
- [ ] `it hashes the password before storing`
- [ ] `it validates email uniqueness`

Requirements become test names directly.

Skills

All skills can be invoked directly with /skill-name or triggered automatically by Claude when your request matches their description.

Skill	Invocation	Auto-triggers	Description
project-setup	/project-setup	No	Configure project (one-time)
project-update	/project-update	No	Sync config with latest plugin defaults
plan-create	/plan-create [description]	Yes — "Build a...", "Let's start building...", "I want an app that...", "Help me implement...", capability lists, etc.	Interactive planning with codebase discovery and grounded clarification
plan-orchestrate	/plan-orchestrate [plan-name]	Yes — "Run the plan", "Execute", "Start implementation"	Parallel TDD execution

Outputs

Output	Meaning
ALL_TASKS_COMPLETE	Plan finished successfully
TASKS_BLOCKED: [003, 007]	Some tasks failed after retries

Architecture

hcf/
├── .claude-plugin/
│   └── plugin.json           # Plugin manifest
├── agents/
│   ├── devils-advocate.md    # Plan reviewer - finds gaps before execution (opus)
│   ├── tdd-worker.md         # TDD implementation worker (sonnet)
│   └── standards-enforcer.md # Code standards enforcement (sonnet)
├── skills/
│   ├── project-setup/
│   │   └── SKILL.md          # One-time setup skill (manual invocation only)
│   ├── project-update/
│   │   └── SKILL.md          # Sync project config with plugin updates (manual)
│   ├── plan-create/
│   │   └── SKILL.md          # Interactive planning skill (auto-triggers)
│   └── plan-orchestrate/
│       └── SKILL.md          # Parallel TDD execution skill (auto-triggers)
├── HOOKS.md                  # Authoritative hook/frontmatter reference
├── CLAUDE.md                 # Portable CLAUDE.md template for projects
└── README.md

Design Principles

1. Pure TDD - No Gherkin/BDD, requirements map directly to test names
2. Parallel by default - Auto-detect independent tasks, maximize concurrency
3. 100% portable - CLAUDE.md identical across all projects
4. Explicit dependencies - Tasks declare what they depend on
5. Fail gracefully - Retry 3x, then block and continue with other tasks

Development

Requirements

• Claude Code CLI
• Git repository
• Test framework configured in project

Local Testing

Test the plugin locally without publishing using the --plugin-dir flag:

claude --plugin-dir /path/to/hcf

Skills are namespaced with the plugin name:

• /hcf:project-setup
• /hcf:plan-create [description]
• /hcf:plan-orchestrate [plan-name]

Skills with disable-model-invocation: true (like project-setup) require manual invocation. Others auto-trigger based on their descriptions.

Testing Workflow

1. Start Claude Code with the plugin:

   cd ~/Sites/your-test-project
   claude --plugin-dir ~/Sites/hcf

2. Verify plugin loads:

   /help
   

   Skills should appear under the hcf namespace.

3. Test each component:

   # Test project setup (direct invocation only)
   /hcf:project-setup
   
   # Test plan-create (auto-triggers on feature requests, or invoke directly)
   "Create a plan to implement user authentication"
   /hcf:plan-create user authentication with JWT
   
   # Test plan-orchestrate (auto-triggers on execution requests, or invoke directly)
   "Run the user-auth plan"
   /hcf:plan-orchestrate user-auth

Debugging

Run with debug output to see plugin loading details:

claude --plugin-dir ./hcf --debug

To load multiple plugins:

claude --plugin-dir ./hcf --plugin-dir ./other-plugin

Plugin Structure Notes:

• Only plugin.json goes in .claude-plugin/
• Skills, agents, and other components stay at the plugin root level
• Each skill is a directory with a SKILL.md entrypoint
• Skills auto-trigger based on their description frontmatter (unless disable-model-invocation: true)

Changelog

See CHANGELOG.md for the full release history. New user-visible changes are added to the [Unreleased] section as they land and moved into a versioned section at release time.

Contributing

1. Fork the repository
2. Create a feature branch
3. Test locally with --plugin-dir
4. Add an entry to the [Unreleased] section of CHANGELOG.md if your change is user-visible
5. Submit a pull request

Credits

Created by Mark Shust

License

HCF is open-source software licensed under the MIT License.