convez.ai
SkillAutomaçãoIntermediário
vasuag09

Harness Claude

vasuag09/harness-claude

Pipeline SDLC completo com orquestração de subagentes e memória persistente

Ver no GitHub
BomPluginScore · de 0 a 100

Framework completo de SDLC para Claude Code que orquestra um pipeline Plan → Implement → Verify → Maintain com suporte a subagentes, gates de execução e memória entre sessões. Instala como plugin isolado em JavaScript/TypeScript, ideal para projetos React, Next.js e Python.

Instalação

$npx convez-ai vasuag09/harness-claude

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
3
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automatizar ciclo completo de desenvolvimento com validações de segurança e qualidade
  • Orquestrar múltiplos agentes Claude em fases distintas do projeto
  • Manter contexto de projeto entre sessões com memória persistente

Veredito

Harness bem-arquitetado para equipes que querem estruturar fluxos SDLC repetiveis com Claude. Abordagem plugin isolado reduz risco; roadmap claro sugere maturação contínua. Exige Node 18+ e claudicação em SDLC para aproveitar plenamente.

README

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

harness-claude

A lean, full-SDLC Claude Code harness — personal, isolated, and testable. It runs a four-phase pipeline (Plan → Implement → Verify → Maintain) backed by scoped subagents, runtime hooks, and cross-session memory. Built to evolve toward eval loops, retrieval, long-running agents, multi-agent orchestration, and computer-use (see ROADMAP.md).

Stack focus: TypeScript/JS (React, Next, Vercel) and Python.


Documentation

GuideRead it for
docs/USAGE.mdInstall → full pipeline → reuse-an-existing-codebase → common scenarios
docs/SKILLS.mdEvery command + agent, what each does, what it delegates to
docs/HOOKS.mdExactly what runs on your machine, and how to disable any hook
docs/TROUBLESHOOTING.mdHooks not firing, slow tsc, MCP, git boundary — fixes
CONTRIBUTING.mdAdding skills/hooks, portability rules, PR checklist
ROADMAP.mdWhere this is going (eval loops → retrieval → multi-agent → computer-use)

New here? Start with docs/USAGE.md.


Why it's structured as a plugin

It lives in its own repo and installs as a Claude Code plugin/marketplace, so you can:

  • Test in isolation — enable it via /plugin, run a real task, judge the result. Toggle it off to fall back. It never overwrites your existing ~/.claude.
  • Promote when proven — once it gives production-grade results, make it your default.
  • Share later — a plugin repo is already the shareable format; push to GitHub.

What's inside

LayerWhereCountRole
Rulesrules/7always-on guidance (cited by skills; imported by CLAUDE.md)
Skillsskills/<name>/SKILL.md2415 atomic /spec … /ship drivers + 5 phase orchestrators + 4 opt-in eval (/eval /extract /benchmark /health)
Agentsagents/*.md7scoped subagents the skills delegate to
Hookshooks/hooks.json + scripts/hooks/—tmux, format, typecheck, quality/design gates, strategic compact, build analysis, memory persistence
MCPs.mcp.json3memory · sequential-thinking · magic (load only when enabled)

The pipeline

PLAN ─────────────► IMPLEMENT ─────► VERIFY ─────────────► MAINTAIN
/spec  /research          /implement     /review  /security-review   /refactor-clean
/plan  /architect         /build-fix     /test    /verify  /ship      /onboard
                  memory: /save-session · /resume-session (cross-cutting)

Phase orchestrators run a whole phase in one command — thin wrappers that sequence the atomic skills, delegate to subagents, and pause for your input at each gate:

CommandRuns
/harness-planspec → research → plan → architect*
/harness-implementimplement (TDD) → build-fix
/harness-verifyreview → security-review → test → verify → ship
/harness-maintainonboard* → refactor-clean
/harnessthe full pipeline, with a checkpoint between phases

* conditional. Use the atomic skills (/plan, /review, …) when you only want one step.


Prerequisites

RequirementWhyWithout it
Claude Code ≥ 2.xHost for the plugin—
Node.js ≥ 18All runtime hooks are Node scriptsHooks no-op; rest of the harness works
gitPhase gates, /ship, session snapshotsMost value is inside a repo
bash + jq (optional)Status line onlyStatus line degrades gracefully

Project tools (prettier, eslint, tsc, ruff, pytest) are not required — the hooks that use them skip silently when absent. See docs/HOOKS.md for exactly what runs on your machine.


Install & test (isolated)

# 1) Add this repo as a marketplace — from GitHub, or via a local clone path
claude plugin marketplace add vasuag09/harness-claude        # straight from GitHub
# …or, if you've cloned it locally:
claude plugin marketplace add /path/to/harness-claude

# 2) In Claude Code, enable the plugin
/plugin            # find "harness-claude", enable it

# 3) Verify the pieces loaded
/plugin            # confirm skills/agents/hooks registered
/mcp               # confirm the 3 MCP servers (enable only what you need)

Then run a real task through the pipeline:

/spec    describe the feature you want
/research  /plan  /architect
/implement
/review  /security-review  /test  /verify  /ship

Toggle the plugin off in /plugin to return to your previous setup at any time.


Opt-in extras (global — not auto-applied)

These can't live inside an isolated plugin, so they're shipped as configs you apply when ready. Your real ~/.claude stays untouched until you do.

Companion plugins

./install-companions.sh          # prints the plan, asks before changing anything
./install-companions.sh --yes    # non-interactive

Installs (if missing): ralph-loop, frontend-design, security-guidance, feature-dev, explanatory-output-style, code-review, typescript-lsp, pyright-lsp, code-simplifier, context7, mgrep. The script needs a local clone (it's a .sh file); for the no-clone manual route and what each companion does, see docs/USAGE.md. Keep <10 plugins/MCPs enabled at once to protect your context window.

What's bundled vs. what's optional

The harness is self-contained: every skill and agent it ships references its own components by their namespaced names (harness-claude:planner, /harness-claude:plan), so they resolve deterministically no matter what else is installed in your ~/.claude. Companions are enhancers, not requirements — when one is absent the skills fall back to built-in tools and say so.

ItemStatusFallback when absent
7 agents (harness-claude:*)Bundled— (always present)
24 skills (/harness-claude:*)Bundled— (always present)
3 MCPs (memory · sequential-thinking · magic)Bundled, load only when enabledharness works without them
mgrepOptional companionGrep / Glob
context7Optional companionlibrary's primary docs / web
knowledge graphOptional companionGrep / Glob
LSP / analyzer plugins (ts/pyright/knip/…)Optional companionhooks skip silently

A reference-integrity check (scripts/check-reference-integrity.sh, run in CI) fails the build if any skill/agent ever reintroduces a bare, collision-prone reference.

Status line + output style + permissions

Merge the keys from settings.snippet.json into your ~/.claude/settings.json (adjust the absolute path). The status line shows: user dir ⎇branch✱ ctx:NN% model HH:MM ☑todos.

Make the rules global (opt-in)

The rules are active inside this repo via CLAUDE.md imports. To apply them on other projects, add to your ~/.claude/CLAUDE.md:

@/path/to/harness-claude/rules/workflow.md
@/path/to/harness-claude/rules/agents.md
@/path/to/harness-claude/rules/engineering.md
@/path/to/harness-claude/rules/security.md
@/path/to/harness-claude/rules/testing.md

Memory persistence

  • /save-session writes an evidence-based summary to .claude/sessions/<date>.md (project) or ~/.claude/harness-claude/sessions/ (outside a repo).
  • /resume-session loads the latest and re-establishes context.
  • Hooks back this up automatically: SessionStart surfaces the last session, PreCompact and Stop snapshot mechanical state so nothing is lost.

Tuning the hooks

Hooks are non-blocking nudges/gates — they never fail a tool call. To disable one, remove its block from hooks/hooks.json (match by id):

  • post:edit:tsc — runs tsc --noEmit on .ts/.tsx edits. Project-wide tsc can be slow; disable this if your edit loop feels sluggish.
  • post:edit:prettier, post:edit:quality-gate — need the project's local prettier/eslint/ruff; they skip silently if absent.
  • post:edit:design-quality — heuristic frontend "template drift" warnings.
  • post:edit:strategic-compact — suggests /compact ~every 50 edits.
  • pre:bash:tmux-reminder, pre:search:mgrep-nudge — reminders only.

Requirements: Node.js (hook scripts) and, for the status line, bash + ideally jq.


Roadmap

Phase 1 (subagents) is this release. Next: eval loops → retrieval → long-running agents → multi-agent orchestration → computer-use. See ROADMAP.md.

License

MIT.

Ver o original no GitHub

Bom para

Relacionados