convez.ai
AgenteAutomaçãoAvançado
ZenHive

Harness: Orquestrador de Agentes de Código Headless

ZenHive/harness

Orquestrador multi-agente de IA com gates de qualidade verificados

Ver no GitHub
BomPluginScore · de 0 a 100

Engine OTP nativo em Elixir para orquestrar agentes de codificação headless (Claude Code, Cursor, Codex, Grok, Antigravity) através de loops verificados de delegação→verificação→reparo. Gerencia tarefas em paralelo com fila persistente (Oban), executa cada uma em worktree isolado, e valida resultados com IA revisora cross-family antes de aprovar ou corrigir.

Instalação

$npx convez-ai ZenHive/harness

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
2
Forks
0
Acompanhando
1
Saúde
75%
Último commit
há 2 semanas
Versão
—
Issues abertas
0
Contribuidores
4

Casos de uso

  • Automação end-to-end de tarefas de codificação com validação cruzada
  • Gerenciamento de roadmaps de projetos acionados por agentes de IA
  • Deploy automático de PRs validadas com auditoria pós-merge

Veredito

Harness é uma infraestrutura robusta e production-ready para quem precisa orquestrar múltiplos agentes de código em escala, com gates de qualidade não negociáveis. Requer expertise em Elixir/OTP e DevOps, mas resolve um problema real em pipelines de automação de IA.

README

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

Harness

OTP-native task-execution engine an AI orchestrator drives end to end.

Harness pulls tasks from an rmap roadmap, dispatches each to a headless coding agent (Claude Code, Cursor, Codex, Grok, Antigravity, Pi) running in an isolated git worktree, then gates the result with a cross-family reviewer AI — the reviewer runs the target project's own checks itself, fixes what it can inline, and writes the verdict. The primary user is an AI orchestrator, not a human. The reviewer's verdict — not the implementer's self-report — is the source of truth for success/failure. Every adapter is held to the same AgentAdapter behaviour and a reusable conformance suite.

Status

Harness is a long-running multi-project OTP node. Harness.ProjectRegistry holds N first-class projects (Elixir, Rust, anything an agent can check); Oban (queue-per-project, Postgres-persisted) provides dispatch with restart resilience; six agent adapters (Claude Code, Codex, Cursor, Grok, Antigravity, Pi) drive runs; the Harness.Run gen_statem owns the per-run lifecycle (implement → commit → review → settle); approved runs can auto-land (rebase + ff-push) and get a post-merge audit agent; Oban.Plugins.Cron lets the roadmap drive itself unattended. Architecture spec: docs/agent-gate-workflow.md.

The cold-path consumer surface is the Phoenix LiveView dashboard + embedded Oban Web + a native MCP server (/harness/mcp, flat JSON-RPC tools) + a Tidewave MCP plug (/tidewave/mcp, project_eval), all served by one standalone Bandit endpoint on http://localhost:4018. The native MCP tools (dispatch__task, dispatch__status, dispatch__verdict_detail, roadmap__*, …) are the primary surface for any JSON/MCP orchestrator; Tidewave project_eval + IEx are the escape hatch for arbitrary eval and the struct-passing ops the flat tools deliberately omit.

See ROADMAP.md for the current task state (rendered from roadmap/tasks.toml by rmap), @~/.claude/includes/harness-workflow.md (via mix harness.install_includes) for the operator workflow (adopt in any repo), docs/dogfooding-workflow.md for harness-incubator specifics, and skills/harness-driver/SKILL.md for the AI-orchestrator contract.

Running the node

iex -S mix

Boots the OTP application, Postgres-backed Oban, and the standalone dashboard endpoint. Live surfaces:

URLWhat it is
http://localhost:4018/harnessLiveView dashboard — project switcher, per-bucket run counts, per-run drill-down with live transcript pane
http://localhost:4018/harness/obanOban Web — queue / job rows / retries / scheduled work
http://localhost:4018/harness/mcpNative MCP server — flat JSON-RPC tools (dispatch__*, roadmap__*, …); the primary surface for a JSON/MCP orchestrator
http://localhost:4018/tidewave/mcpTidewave MCP endpoint (dev only) — project_eval escape hatch for arbitrary eval + struct-surface ops

The standalone Bandit endpoint is gated by config :harness, :dashboard, enabled: true AND Bandit being in the dep stack. Mountable consumers (their own Phoenix endpoint) leave enabled: false and route live "/harness/*path", Harness.Dashboard.Live themselves.

Use harness from another repo

The common case: you have a project (myapp) and want harness — running as a long-lived iex -S mix BEAM in ~/_DATA/code/harness/ — to dispatch tasks from myapp's roadmap to headless coding agents, gate each result with a cross-family reviewer AI (which runs myapp's own checks itself), and report the reviewer's verdicts back to the AI agent driving from inside myapp.

Three setup steps:

1. Register myapp with harness. Add an entry alongside the self-registered "harness" project in config/dev.exs, then restart iex -S mix:

# ~/_DATA/code/harness/config/dev.exs
config :harness, :projects, [
  [
    name: "harness",
    source: {:local, Path.expand("..", __DIR__)},
    check_command: "mix precommit.full",
    roadmap_path: Path.expand("..", __DIR__)
  ],
  [
    name: "myapp",
    source: {:local, "/Users/you/_DATA/code/myapp"},
    check_command: "mix precommit",      # free-text hint for the reviewer AI
    roadmap_path: "/Users/you/_DATA/code/myapp",
    concurrency_cap: 2
  ]
]

check_command is a free-text hint handed to the reviewer AI — the reviewer runs the project's checks itself and judges the output; harness never executes the command. Point it at your project's mergeable bar (mix precommit, cargo test && cargo clippy, …); for a multi-language monorepo just describe each component's command in the hint. Omit it to let the reviewer discover the checks on its own.

2. Add harness's MCP endpoints to myapp/.mcp.json — alongside myapp's own Tidewave if it has one. The harness entry (native flat tools) is your primary surface; the optional harness_eval entry is the project_eval escape hatch into harness's BEAM:

{
  "mcpServers": {
    "tidewave": {
      "type": "http",
      "url": "http://localhost:4001/tidewave/mcp"
    },
    "harness": {
      "type": "http",
      "url": "http://localhost:4018/harness/mcp"
    },
    "harness_eval": {
      "type": "http",
      "url": "http://localhost:4018/tidewave/mcp"
    }
  }
}

Claude Code surfaces a server's tools as mcp__<server-name>__<tool>, giving three distinguishable surfaces: mcp__tidewave__project_eval (inspect myapp's state, port 4001), mcp__harness__dispatch__task & the rest of the flat driver tools (dispatch + observe + triage against harness's :4018 BEAM — the primary surface), and mcp__harness_eval__project_eval (escape hatch for arbitrary eval + struct-surface ops). Drop harness_eval if you only need the flat tools. No port collision — different BEAMs / paths.

3. Import the driver skill from myapp/CLAUDE.md so the AI agent in myapp knows how to use the surface:

@~/_DATA/code/harness/skills/harness-driver/SKILL.md

Restart the Claude Code session in myapp to pick up the new .mcp.json entries. After that, the agent dispatches via the flat mcp__harness__dispatch__task tool (and observes with mcp__harness__dispatch__status / dispatch__verdict_detail) against :4018; harness manages isolated worktrees of myapp, gates each run with the cross-family reviewer AI, and reports the reviewer's verdict back.

Full driver contract (entry points, two-eval pattern for ephemeral MCP eval processes, cross-checkout sharp edges, secret scrubbing): skills/harness-driver/SKILL.md § "Context A — Driving harness from another repo".

Development

# First time
mix deps.get
mix compile

# Fast local gate (hook-bound, ~180s)
mix check.fast

# Pre-commit gate (no dialyzer — dialyzer lives in precommit.full)
mix precommit

# Full hand-off gate — mirrors CI, includes dialyzer
mix precommit.full

# Focused checks
mix test
mix credo --strict        # includes TODO/FIXME debt visibility by design
mix sobelow --exit --skip
mix sobelow.baseline      # refresh Sobelow skip baseline intentionally

# AI-friendly output
mix test.json
mix dialyzer.json

All tooling is wired per the global Elixir setup conventions (Styler first, Reach for OTP analysis, etc.).

License

MIT (or your preferred license).

Ver o original no GitHub

Bom para

Relacionados