Harness Kit
HoangTechCS-AIE/harness-kit
Setup de repo para coding agents em 5 níveis de maturidade — um comando
Ver no GitHubKit de inicialização para preparar repositórios como agents de IA coding (Claude Code) em 5 níveis de maturidade. Instala artefatos, configurações de guardrails, context management e observabilidade em um único comando npx.
Instalação
npx convez-ai HoangTechCS-AIE/harness-kit1 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
Casos de uso
- Configurar repositórios para agents Claude Code com boas práticas
- Implementar guardrails e constraints de segurança em agents
- Escalar trabalho de longa duração de agents com observabilidade
Veredito

Ferramenta sólida para engenheiros que querem estruturar repos para agents com disciplina. Os 5 níveis (foundation, context, guardrails, long-running, evals) cobrem falhas reais; instalação idempotente reduz atrito.
README
A tradução para o português deste README ainda está na fila.
harness-kit
Set up any repo for a coding agent (Claude Code) across the 5 maturity levels of harness engineering — in one command.
Vietnamese · English
Harness = everything around the model (context, instructions, tools, safety constraints, orchestration, measurement) that makes an agent reliable. This kit packages it into installable artifacts plus a tutorial that teaches when to use each. The kit ships files; the tutorial teaches the discipline.
1. How it works
flowchart LR
D([You]) -->|"npx @htechcs/harness-kit"| I[["harness-kit<br/>installer"]]
I --> R1[".claude/<br/>settings.json · agents/"]
I --> R2["repo root<br/>setup.sh · TASK.md"]
I --> R3["~/.claude/skills/<br/>init-harness"]
I --> R4["docs/harness/<br/>tutorial + guides"]
R1 --> A([Claude Code<br/>reads every session])
R2 --> A
R3 --> A
R4 --> A
One npx command drops artifacts in the right places; from there Claude Code reads them every session.
2. Quick install
npx @htechcs/harness-kit # pick levels interactively, then install
npx @htechcs/harness-kit --all # install all 5 levels
npx @htechcs/harness-kit --levels=1,3 # specific levels only
Requires Node ≥18. The command saves all docs into docs/harness/ so your team keeps them.
Idempotent — safe to re-run (--force to overwrite).
3. The 5 levels — each prevents a failure mode
flowchart TD
L1["Level 1 · Foundation — CLAUDE.md"] --> L2["Level 2 · Clean context"]
L2 --> L3["Level 3 · Guardrails"]
L3 --> L4["Level 4 · Long-running"]
L4 --> L5["Level 5 · Evals & Observability"]
L5 -.->|"surface a failure → reinforce the harness"| L1
| Level | Without it | Artifact |
|---|---|---|
| 1 — Foundation | the agent has no durable repo guidance | /init-harness skill → generates CLAUDE.md |
| 2 — Clean context | context floods, the agent's attention dilutes | sample subagent + MCP audit checklist |
| 3 — Guardrails | agent deletes/pushes by mistake, asks permission nonstop | settings.json (deny/ask/allow) |
| 4 — Long-running | long tasks break mid-way, can't resume | setup.sh, new-worktree.sh, TASK.md |
| 5 — Evals & Obs | no idea whether the agent does well or badly | golden-task template + observability guide |
4. Manual install per level
The installer just automates the
cpcommands below — expand them to understand/do it by hand.
cp -r skills/init-harness ~/.claude/skills/ # then run /init-harness in the target repo
mkdir -p .claude/agents && cp templates/agents/repo-explorer.md .claude/agents/
Read templates/agents/README.md + templates/mcp-audit.md.
mkdir -p .claude && cp templates/settings.json .claude/settings.json
First thing to do: add your repo's test/lint commands to allow (see templates/guardrails/README.md).
cp templates/setup.sh templates/new-worktree.sh . && chmod +x setup.sh new-worktree.sh
cp templates/long-running/TASK.md . # when starting a long task
mkdir -p evals/cases && cp templates/evals/cases/example-task.md evals/cases/
Read templates/evals/README.md (includes a no-harness baseline step) + observability.md.
mkdir -p docs/specs && cp templates/spec/FEATURE.md docs/specs/<feature>.md
5. Level dependencies
- Level 1 first — it's the backbone; later levels reference
CLAUDE.md. - Levels 3 & 4 both "point
CLAUDE.mdto" their artifacts → require Level 1 done. - Level 5 needs at least one level applied to have a change to measure (see the feedback loop above).
6. Guardrails benchmark
The Level 3 guardrails are measured, not asserted. A reproducible benchmark runs a 25-op risk suite (catastrophic commands, secret-file reads, safe dev ops) through the permission rules and scores what is denied / gated / allowed.
The shipped defaults gate 68% of the suite out of the box; a hardened profile raises it to 92% — blocking 6/8 catastrophic ops and 7/7 secret-file reads. This measures guardrail policy coverage, not task success.
node bench/guardrails-bench.js # default
node bench/guardrails-bench.js templates/settings.json bench/settings.hardened.json # before / after
See bench/RESULTS.md for methodology and the two residual gaps.
7. Docs
docs/harness-engineering-tutorial.en.md (Vietnamese) — the
why + when to use each piece (after install: docs/harness/). Full industry-wide source catalog:
Awesome Harness Engineering.
8. License
MIT.