klaach_harness
klaachkinggit/klaach_harness
Base Claude Code harness for software development — MCP servers, slash commands, skills
Instalação
npx convez-ai klaachkinggit/klaach_harness1 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
README
A tradução para o português deste README ainda está na fila.
klaach_harness
Model-agnostic harness for AI-assisted software dev. Works with Claude Code, Codex CLI, Cursor, Windsurf, Gemini CLI, Copilot, Cline, and any future tool.
Use: point any agent at this repo; it self-installs. Setup → APPLY.md. Per-tool capability map → HARNESS.md. Per-project skill/MCP/plugin packs → PROFILES.md.
Why it's model-agnostic: rules + prompts are universal; MCP config is generated per-tool; all enforcement (no secrets, formatted code, passing tests) lives at the git + CI layer, so guarantees hold under any agent or human. Runtime hooks (Claude Code + Codex) are a faster copy of the same checks.
Quick apply
# TOOL = claude | codex | cursor | windsurf | gemini | copilot | cline | all
TOOL=claude bash <(curl -fsSL https://raw.githubusercontent.com/klaachkinggit/klaach_harness/main/apply.sh)
What's included
Rules — behavioral contract
RULES.md is the single source, synced to every tool's file by sync-rules.sh.
| Tool | File | Tool | File | |
|---|---|---|---|---|
| Claude Code | CLAUDE.md | Gemini | GEMINI.md | |
| Codex | AGENTS.md | Copilot | .github/copilot-instructions.md | |
| Cursor | .cursorrules + .cursor/rules/harness.mdc | Cline | .clinerules | |
| Windsurf | .windsurfules | Other | copy RULES.md into system prompt |
Universal enforcement (git + CI — any tool or human)
.githooks/pre-commit— block secrets, auto-format staged files.githooks/pre-push— run tests, block on failure.github/workflows/ci.yml— secret scan + lint + test on push/PR- Activated by
apply.sh:git config core.hooksPath .githooks
MCP servers (per-tool config via tools/gen-mcp.py)
Base set: github, filesystem, git, playwright, db (Postgres, set DATABASE_URL). Emits the right format/path per tool (.mcp.json / .cursor/mcp.json / .codex/config.toml). Stack-specific servers → PROFILES.md.
Plugins (Claude Code — installed by apply.sh)
obra/superpowers (workflow), anthropics/skills → claude-api + document-skills (Anthropic SDK + PDF/docx/xlsx/pptx). The trailofbits/skills marketplace is also registered so you can claude plugin install <name>@trailofbits on demand (it ships ~19 security plugins — opt-in to avoid bloat).
Hooks (Claude Code & Codex — .claude/hooks/)
| Script | Trigger | Does |
|---|---|---|
protect-secrets.sh | Read/Write/Edit | block .env/.pem/.key/credentials |
block-dangerous.sh | Bash | block recursive-rm of root/home/cwd, curl|sh, force-push main, fork bomb |
log-bash.sh | Bash | log commands to .claude/bash.log |
auto-format.sh | Edit/Write | prettier/black/ruff/gofmt if installed |
pre-pr-gate.sh | create_pull_request | block PR if tests fail |
| done-notify | Stop | macOS notification / terminal bell |
Skills (lean by design)
Base bundles only find-skills (discovers/installs any other skill on demand). Niche skills are per-project — see PROFILES.md. Skill descriptions tax every session's context, so the base stays minimal. Token discipline comes from model-tier routing (prompts/subagent.md), /compact, and /cost-review — not from output-compression skills.
Prompts (prompts/ — paste into any tool)
grill-me (requirements interview), sparc (5-phase build), tdd (RED-GREEN-REFACTOR), diagnose (8-step debug), to-issues (plan → issues), zoom-out (system map), security-scan (OWASP + secrets + PII), risk-review (diff risk classifier), preflight (pre-ship checklist), audit (periodic repo health), adr (architecture decision record), memorize (append to MEMORY.md), learn (append to LESSONS.md), subagent (when to delegate + Haiku/Sonnet/Opus tier routing), cost-review (token/spend check), assess-capabilities (acquire skills/MCP/plugins for a project/feature), adopt-harness (adopt into existing project + clean up). Claude/Codex get the same as /slash commands.
Memory, decisions & learning
MEMORY.md— append-only cross-session memory (state of the world). Seeprompts/memorize.md.LESSONS.md— append-only heuristics learned (what to do next time). Seeprompts/learn.md. Together these are the portable "self-learning" loop: read at session start, append when something non-obvious happens. No daemon, no vector DB.docs/adr/— architecture decision records (template at0000-template.md, write viaprompts/adr.md).
Per-project customization
Append below the <!-- Add project-specific rules --> line in your rules file:
## Project: my-app
- Stack: Next.js 14, Prisma, PostgreSQL
- Test: vitest — `npm test` · Lint: `npm run lint`
- Deploy: Vercel, CI only — never from local
Maintaining (harness repo only)
Edit RULES.md, then ./sync-rules.sh regenerates all tool files (CLAUDE.md = RULES.md + .claude-extra.md). Run it only here — in a target project you own the rules files. apply.sh backs up overwrites to <file>.bak.
Env
Copy .env.example → .env: GITHUB_TOKEN (required for github MCP), DATABASE_URL (optional, enables db MCP).