convez.ai
AgenteAutomaçãoAvançado
SynapSync

Kyro AI

SynapSync/kyro-ai

Harness de sprint multi-agente com orquestração portável para fluxos de código

Ver no GitHub
InicialPluginScore · de 0 a 100

Sistema completo de orquestração de sprints para agentes de IA em projetos TypeScript. Instala um harness de workflow portável que permite múltiplos agentes trabalhar em ciclos de sprint estruturados, com rastreamento de estado local, comandos de skill e aprendizado adaptativo sem perder contexto entre trocas de agentes.

Instalação

$npx convez-ai SynapSync/kyro-ai

1 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

Estrelas
5
Forks
0
Acompanhando
0
Saúde
37%
Último commit
há 1 dia
Versão
v6.6.1
Issues abertas
0
Contribuidores
1

Casos de uso

  • Coordenar múltiplos agentes IA em um mesmo projeto sem reaprendizado de workflow
  • Estruturar ciclos de sprint com análise, planejamento, execução e revisão formalizados
  • Rastrear débito técnico e contexto entre sessões de diferentes agentes

Veredito

Kyro AI resolve um problema legítimo: manter consistência de workflow quando múltiplos agentes IA trabalham no mesmo projeto. A implementação é sólida com CI/CD, npm registry e boa separação entre runtime global e estado local. Ferramenta madura para equipes que precisam de rigor processual em fluxos de IA.

README

A tradução para o português deste README ainda está na fila.


What Kyro AI Does

Kyro AI installs a project-local workflow harness that helps AI coding agents analyze, plan, execute, review, and close software work through repeatable sprint cycles.

It is designed for teams that switch between agents and do not want every agent to relearn the workflow from a giant prompt.

Kyro gives agents:

  • a managed core with orchestrator, command, skill, and template instructions
  • command-like skills such as kyro-forge, kyro-status, and kyro-wrap-up
  • workspace state that tracks installed adapters and future scopes
  • human-readable artifacts for roadmaps, phase documents, debt, retros, and re-entry context
  • doctor/sync/uninstall commands so setup can be validated instead of guessed

Quick Start

Install Kyro AI into the current project:

npx kyro-ai install --agent opencode,codex --scope workspace --yes
npx kyro-ai doctor

Use one agent only if needed:

npx kyro-ai install --agent opencode --scope workspace --yes
npx kyro-ai install --agent codex --scope workspace --yes

After install, open your agent and invoke the installed Kyro command skill:

kyro-forge auth-refactor

Other installed command skills:

kyro-status
kyro-wrap-up

If the host exposes slash commands, the equivalent public command namespace is:

/kyro:forge
/kyro:status
/kyro:wrap-up

Installed Layout

Kyro separates global runtime from project state. Runtime files live in your user-level agents directory; project files keep only state and artifacts.

Global runtime:

~/.agents/
├── kyro/
│   ├── versions/
│   │   └── {version}/
│   │       ├── core/
│   │       │   ├── agents/
│   │       │   ├── config.json
│   │       │   └── WORKFLOW.yaml
│   │       ├── commands/
│   │       ├── skills/
│   │       │   ├── sprint-forge/
│   │       │   └── qa-review/
│   │       ├── KYRO.md
│   │       └── manifest.json
│   └── current -> versions/{version}
└── skills/
    ├── kyro-forge/SKILL.md
    ├── kyro-status/SKILL.md
    └── kyro-wrap-up/SKILL.md

Project state and artifacts:

<project>/
└── .agents/
    └── kyro/
        ├── kyro.json
        └── scopes/
            └── {scope}/
                ├── state.json
                ├── ROADMAP.md
                └── phases/

Important invariant: kyro install creates the root project state at .agents/kyro/kyro.json. It does not create .agents/kyro/scopes/{scope}/state.json; scoped state belongs to future scope/forge creation.

Who Invokes Whom

User
  ↓
kyro / kyro-ai CLI
  ↓
installs managed core + adapter projections
  ↓
agent opens the project
  ↓
agent discovers AGENTS.md and/or ~/.agents/skills
  ↓
user invokes kyro-forge / kyro-status / kyro-wrap-up
  ↓
projected skill reads ~/.agents/kyro/current/commands/*.md
  ↓
orchestrator reads ~/.agents/kyro/current/core/agents/orchestrator.md
  ↓
sprint-forge skill assets guide the workflow
  ↓
artifacts are written under <project>/.agents/kyro/scopes/{scope}/

The user should not have to explain the workflow in natural language. If an agent cannot discover the installed command skills, that agent needs a better adapter.


CLI Reference

Kyro AI exposes two equivalent bins:

kyro
kyro-ai

Commands:

CommandPurpose
kyroOpen the basic interactive configuration TUI
kyro installInstall managed core and selected agent adapters
kyro doctorValidate package health, workspace state, core files, and adapters
kyro syncRefresh managed assets without rewriting unmanaged files
kyro uninstallRemove managed workspace assets

Common usage:

kyro install --agent opencode,codex --scope workspace --yes
kyro doctor --tokens --artifacts
kyro sync --agent codex --dry-run
kyro uninstall --dry-run

Supported install adapters today:

AdapterStatusWhat it installs
opencodeImplemented~/.agents/skills/kyro-* command skill projections
codexImplemented~/.agents/skills/kyro-* plus a managed Kyro block in root AGENTS.md
claudePlanned for CLI workspace installClaude plugin remains first-class through .claude-plugin/
cursorPlannedNot installed by the CLI yet

There is intentionally no generic adapter. Cross-agent instructions belong in root AGENTS.md; install adapters should target concrete agent capabilities.


Artifact Integrity

Kyro treats the repo as the system of record. Validate and repair the knowledge contract with:

kyro doctor --artifacts
kyro repair --kyro-scope <scope> --yes
kyro scope inspect <scope>

repair only rebuilds JSON routing/summaries from existing Markdown evidence; it does not rewrite user-authored Markdown.

Core Workflow

Kyro has three stable workflow intents:

IntentSkill / commandWhat it does
Forgekyro-forge / /kyro:forgeAnalyze, plan, execute, review, and close a sprint cycle
Statuskyro-status / /kyro:statusReport project progress, roadmap health, and technical debt
Wrap-upkyro-wrap-up / /kyro:wrap-upClose a session, update handoff context, and preserve learnings

Forge is routed progressively:

read kyro.json/state.json/index.json
  → choose INIT, plan, execute, review, close, or recover
  → load only the selected mode/helper/template
  → update Markdown evidence plus JSON summaries

Kyro is intentionally sprint-by-sprint and summary-first. It should not pre-load every roadmap, sprint, helper, and template just to decide the next action.


Architecture

The source package is organized around portable markdown instructions plus a small deterministic CLI:

kyro-ai/
├── src/cli/                 # installer, doctor, sync, uninstall, adapters
├── agents/                  # orchestrator instruction
├── commands/                # forge, status, wrap-up command definitions
├── skills/
│   ├── sprint-forge/        # main workflow skill and assets
│   └── qa-review/           # senior QA review skill
├── docs/                    # architecture, CLI, adapter, and command docs
├── rules/                   # reusable operating rules
├── templates/               # context templates
├── .claude-plugin/          # Claude plugin adapter packaging
├── WORKFLOW.yaml
└── config.json

Runtime contract:

LayerResponsibility
CLIInstall, sync, uninstall, validate, and project managed files
AdapterTranslate Kyro into each agent's native/compatible instruction surface
Projected skillGive the agent a short command entrypoint without duplicating lifecycle prose
OrchestratorCoordinate phases, gates, review, debugging, and handoff
Sprint Forge skillDefine workflow modes, helpers, templates, and discipline rules
ArtifactsPersist roadmap, phases, debt, retros, and re-entry context in markdown

Claude Plugin Support

Claude remains first-class.

The Claude plugin adapter lives in .claude-plugin/ and is shipped with the package. The CLI workspace installer does not replace the Claude plugin; it complements it for agents that need workspace-local command skills and AGENTS.md instructions.

Claude plugin install:

/plugin marketplace add SynapSync/kyro-ai
/plugin install kyro-ai@kyro-ai

Local plugin development:

git clone https://github.com/SynapSync/kyro-ai.git
cd kyro-ai
npm install
npm run build
claude --plugin-dir /path/to/kyro-ai

Development

Requirements:

  • Node.js 18+
  • npm

Useful commands:

npm ci
npm run check
npm run build
npm pack --dry-run

npm run check runs:

  • TypeScript typecheck
  • package/plugin/workflow version validation
  • relative markdown link validation

Release tags publish to npm through GitHub Actions when the tag matches package.json.version:

git tag v3.2.2
git push origin v3.2.2

The release workflow expects the repository secret NPM_TOKEN.


Documentation

GuideDescription
CLIInstaller, doctor, sync, uninstall, and adapter commands
Commands Reference/kyro:* command semantics
ArchitectureSystem architecture and data flow
Agent AdaptersAdapter setup and host-specific notes
Harness MigrationDirection for the multi-agent runtime
Getting StartedIntroductory workflow guide
Rules GuidePersistent learning rules
Context ManagementRe-entry prompts and continuity strategy
Programmatic UsageUsing Kyro instructions from custom LLM apps

Philosophy

  1. Commands over prose — users should invoke workflows, not explain them repeatedly.
  2. Markdown remains the collaboration layer — humans and agents can inspect the same artifacts.
  3. The CLI owns deterministic checks — package health should not depend on prompt discipline.
  4. Adapters are concrete — each supported agent gets the files it actually knows how to use.
  5. One sprint at a time — each cycle adapts from evidence, retro, and technical debt.
  6. Claude stays first-class — multi-agent support does not mean retiring the Claude plugin.

Ver o original no GitHub

Bom para

Relacionados