Chaos Theory
whfreelove/chaos-theory
Plugins de agentes com comportamento estocástico e workflows estruturados
Ver no GitHubConjunto de plugins para Claude que implementam comportamento estocástico em agentes de IA. Inclui Tokamak (reator de especificações), finite-skill-machine (orquestração de workflows) e ferramentas de isolamento e testes. Designed para melhorar a iteração de design e a execução de tarefas complexas com agentes.
Instalação
npx convez-ai whfreelove/chaos-theory1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automatizar workflows de desenvolvimento com task lists baseadas em dependências
- Implementar ciclos test-driven development forçados por hooks
- Gerar e iterar especificações de software através de feedback de agentes
Veredito

Projeto experimental com abordagem criativa para orquestração de agentes, mas a documentação é densa e a proposta central ('plasma thrashing') é pouco clara. Recomendado para quem explora limites técnicos de agentes Claude, não para uso em produção sem validação profunda.
README
A tradução para o português deste README ainda está na fila.
Chaos Theory Plugin Marketplace
Claude Code plugins leveraging stochastic agent behavior. "Agents, uhh, find a way."
Installation
claude plugin marketplace add whfreelove/chaos-theory
Dependencies
- jq (standard in MacOS Sequoia 15.0+)
Tokamak
pip install ruamel.yaml- OpenSpec (warning: confirm your telemetry privacy settings)
A SessionStart hook will install chaos-theory OpenSpec schemas for you (if openspec/ already exists), and if a Tokamak update includes a new schema version, use /init-schemas upgrade.
Available Plugins
Tokamak
A spec-as-code reactor that harnesses power from AI agents' thrashing plasma.
An OpenSpec-based, iterative, structured specification workflow based on turning software engineering best practices overhead into background noise. Agents push you to understand your design through repeated rounds of finding gaps, clarifying your solutions, documenting your decisions, then updating specifications.
pip install ruamel.yaml
claude plugin install tokamak
See plugins/tokamak/README.md for the full change and brownfield documentation workflows.
finite-skill-machine
Drive complex agent workflows with minimal context and no task transcription mistakes.
Auto-populates TaskList from skill fsm.json companion files. Skills can define multi-step workflows that automatically populate the task list when invoked.
claude plugin install finite-skill-machine
Create a skill with an fsm.json alongside SKILL.md:
my-skill/
├── SKILL.md # Agent instructions
└── fsm.json # Task definitions
Example fsm.json:
[
{"id": 1, "subject": "Set up environment", "activeForm": "Setting up"},
{"id": 2, "subject": "Implement feature", "blockedBy": [1]}
]
When the skill is invoked, tasks appear automatically in the task list with proper dependencies.
worktree-isolation
Restricts Claude Code to git worktree or Jujutsu workspace boundaries. Prevents access to project root or sibling worktrees.
test-driven-everything
Enforces RED→GREEN→REFACTOR as a workflow constraint. Hooks block code writes until tests fail, then block completion until tests pass.
claude plugin install test-driven-everything
Invoke tdd-begin to start a cycle. The skill guides you through intent declaration, test definition, and implementation. Hooks enforce ordering: writes are blocked in RED state, and the Stop hook prevents session end until the cycle reaches READY.
No dependencies beyond jq.
Update
claude plugin marketplace update
claude plugin update finite-skill-machine
Test
python -m venv .venv
source .venv/bin/activate
pip install ".[test]"
Run all plugin tests:
make test
Run tests for a specific plugin (target name matches directory name under tests/plugins/):
make test-finite-skill-machine
List all available targets:
make help
Research
Launch a JupyterLab environment for exploring plugin behaviors:
make research
The research/ directory holds notebooks for static analysis (hook outputs, JSON configs, transcript logs) and live CLI execution via subprocess.