convez.ai
AgenteAutomaçãoAvançado
jstuart0

Mozart Orchestration

jstuart0/mozart-orchestration

Pipeline orquestrado com 14 subagentes para entrega estruturada de código

Ver no GitHub
BomPluginScore · de 0 a 100

Agent orquestrador multi-estágio para Claude Code que transforma uma solicitação em um pipeline estruturado com 14 subagentes especializados. Realiza fluxos de DELIVER (construção), AUDIT (revisão) e DIAGNOSE (investigação) com pesquisa, planejamento, implementação, verificação e documentação narrados em tempo real.

Instalação

$npx convez-ai jstuart0/mozart-orchestration

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
1
Forks
0
Acompanhando
1
Saúde
100%
Último commit
há 2 semanas
Versão
v0.3.0
Issues abertas
13
Contribuidores
1

Casos de uso

  • Implementar features complexas com revisão automática em múltiplas etapas
  • Auditar código e arquitetura contra objetivos de qualidade
  • Investigar falhas em produção com rastreamento de causa-raiz

Veredito

Ferramenta sólida para times que precisam de rigor em pipelines de entrega de código. O modelo de orquestração com subagentes nomeados e tiers ajustáveis é bem pensado, mas requer compreensão clara da arquitetura e workflow para máxima efetividade.

README

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

mozart-orchestration

A Claude Code plugin that turns one request into an orchestrated, narrated, multi-agent pipeline.

License Version Claude Code Plugin

Why this exists

One-shot Claude Code requests either over-fire (one giant context doing everything) or under-deliver (no review, no plan, no validation). Mozart threads the needle: you describe what you want, and he routes it through a real delivery pipeline — research, plan, specialist review, implement, verify, document — using named subagents running in their own contexts. Each stage has a defined scope and a clear handoff. You see every move as it happens.

Mozart handles three shapes of work: DELIVER (build or change something), AUDIT (review against a goal), and DIAGNOSE (investigate a failure). He tiers tasks — TINY / STANDARD / HEAVY — to right-size the gates, classifies project context (GREENFIELD / BROWNFIELD) to decide when duplicate-check agents run, and narrates every Task spawn so you always know who is working and why.

Quickstart

Install from the Claude Code plugin marketplace:

/plugin marketplace add jstuart0/mozart-orchestration
/plugin install mozart-orchestration

Then invoke mozart with your request:

/mozart <your request>

Example invocations:

  • /mozart add OAuth login to the admin panel — DELIVER, STANDARD tier
  • /mozart audit our auth middleware — AUDIT shape
  • /mozart investigate why pgvector queries are slow on staging — DIAGNOSE shape
  • /mozart resume thoughts/shared/plans/<slug>.state.md — resume a stopped campaign

Without an argument, mozart asks what to orchestrate.

The pipeline at a glance

flowchart LR
    A[1 · Intake] -.-> B[2 · Research<br/><i>STANDARD/HEAVY</i>]
    B -.-> C[3 · Plan<br/>harry]
    A --> C
    C -.-> D[4 · Plan review<br/><i>STANDARD/HEAVY</i>]
    D -.-> E[5 · Codex r1]
    C --> E
    E --> F[6 · Iterate]
    F --> G[7 · Implement<br/>jackson]
    G -.-> H[8 · Mid-build<br/><i>conditional</i>]
    H -.-> I[9 · Codex r2<br/><i>STANDARD opt /<br/>HEAVY req</i>]
    G -.-> I
    I -.-> J[10 · Validate<br/>valerie]
    G --> J
    J --> K[11 · Reconcile<br/>jackson]
    K --> L[12 · Document<br/>scott]
    L --> M[13 · Report]

Solid edges (-->) run on every tier. Dashed edges (-.->) mark conditional stages: Research runs on STANDARD/HEAVY; Plan review fan-out runs on STANDARD/HEAVY; Mid-build specialists trigger per-phase when conditions match; Codex r2 is optional on STANDARD and mandatory on HEAVY.

AUDIT and DIAGNOSE flows are shorter — see PIPELINE.md for the full reference.

Three orchestration shapes

ShapeWhenOutput
DELIVER"build X", "ship Y", "fix Z"Working code, verified against the plan, documented
AUDIT"review X", "audit Y for Z"Findings document; optional remediation flow
DIAGNOSE"why is X broken?", "investigate Y"Findings document with symptom / repro / root cause / remediation options

Bug-shaped DELIVER on STANDARD/HEAVY auto-promotes to DIAGNOSE first.

Task tiers

TierWhat it adjusts
TINYSkip research, plan-review fan-out, mid-build specialists. Brief jackson directly → verify → commit
STANDARDDefault — full DELIVER pipeline
HEAVYSTANDARD + mandatory ian on every phase + mandatory xander mid-build + mandatory codex r2 on the final diff

Mozart classifies tier at intake based on surface area (auth, schema, migrations, infra, security-critical → HEAVY).

Project context

ContextWhat it controls
GREENFIELDSkip librarian. Nothing meaningful to search against.
BROWNFIELDLibrarian runs at plan review and mid-build for new functions, classes, and services

When in doubt, mozart classifies BROWNFIELD; librarian short-circuits if the work turns out to be greenfield-shaped in practice.

What's in the box

mozart-orchestration/
├── .claude-plugin/
│   ├── marketplace.json
│   └── plugin.json
├── agents/
│   ├── LEARNINGS.md             # cross-project field notes (append-only)
│   ├── PIPELINE.md              # full pipeline reference
│   ├── bob.md
│   ├── codebase-analyzer.md
│   ├── codebase-locator.md
│   ├── codebase-pattern-finder.md
│   ├── dexter.md
│   ├── dick.md
│   ├── harry.md
│   ├── ian.md
│   ├── jackson.md
│   ├── librarian.md
│   ├── mozart.md                # the conductor
│   ├── otto.md
│   ├── ruby.md
│   ├── sarah.md
│   ├── scott.md
│   ├── valerie.md
│   ├── web-search-researcher.md
│   └── xander.md
├── commands/
│   └── mozart.md                # the /mozart slash command
├── docs/                        # created in v0.1.0 release
│   └── README.md
├── .github/                     # created in v0.1.0 release
│   ├── ISSUE_TEMPLATE/
│   └── workflows/
├── CHANGELOG.md
├── CODE_OF_CONDUCT.md
├── CONTRIBUTING.md
├── INTEGRATION.md               # ticketing + docs configuration
├── LICENSE
├── README.md
└── SECURITY.md

thoughts/ (gitignored) — per-campaign state files, plans, flow sketches, and investigation notes. Never committed.

Live narration

Mozart announces every Task invocation before it starts (one line) and summarizes every return when it comes back (one line), each prefixed TASK [<stage>] for scannability. You always know which agent is running, on which campaign, and why.

Integration

Mozart is pluggable for two surfaces that vary by team:

  1. Ticketing — Plane, Linear, Jira, GitHub Issues, or none.
  2. Documentation surfaces — GitHub wiki, in-repo docs, an external wiki (Wiki.js, Notion, Confluence), or a custom mix.

Configure both by adding stanzas to your repo's CLAUDE.md. See INTEGRATION.md for templates and the contract mozart follows.

If you don't configure ticketing, mozart skips ticket steps entirely — research, planning, implementation, and verification still work. If you don't configure docs, scott will publish to in-repo README.md / CHANGELOG.md / docs/ only.

Optional: codex CLI

Mozart's pipeline calls an external codex CLI at stages 5 (codex-r1-plan, plan review) and 9 (codex-r2-diff, diff review) for fresh-context, second-opinion review. The value is that codex runs with no plan-iteration history, which surfaces issues that in-context agents sometimes miss. The plugin works without codex — those stages skip with a logged note and the pipeline continues.

If you want codex's input, install it from https://github.com/openai/codex. On HEAVY tier, codex r2 is mandatory; on STANDARD it's optional; TINY skips both rounds entirely.

Authority

The mozart agent persona (agents/mozart.md) is authoritative for orchestration behavior. The /mozart slash command is a thin wrapper that hands control to the persona at the top level of a session.

See CONTRIBUTING.md for the persona-authoring contract and instructions for adding new agents.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados