convez.ai
AgenteDesenvolvimento WebAvançado
waynehacking8

Claude Code Bare Harness

waynehacking8/claude-code-bare-harness

Strip Claude Code harness para testes A/B de comportamento do modelo

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent para testes A/B que remove camadas da instrução do Claude Code, permitindo observar o comportamento do modelo com injeção mínima de contexto. Útil para pesquisa de como o harness afeta estilo e resposta, mantendo os limites de segurança do modelo intactos.

Instalação

$npx convez-ai waynehacking8/claude-code-bare-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
1
Forks
1
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • A/B testing: comparar respostas com e sem agent harness
  • Pesquisa de comportamento: estudar como instruções afetam saída do modelo
  • Transparency research: entender camadas de injeção no Claude Code

Veredito

Ferramenta honesta e bem documentada para pesquisa reproducível, com posicionamento claro sobre o que pode e não pode remover. Ideal para engenheiros que estudam behavior modeling e querem isolar variáveis no Claude Code.

README

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

Minimizing the Claude Code Harness for Model-Behavior Testing

Purpose: When observing model behavior or running A/B tests, strip as much of the Claude Code agent harness as possible so that what reaches the model is closer to "just the user message."

⚠️ Read this first This documents removing the agent-instruction layer, NOT "disabling safety."

  • The model's safety behavior comes from training (weights), not the system prompt. No CLI flag can remove it.
  • The server-side content safety classifier cannot be disabled via local config.
  • After --system-prompt "" you still get the same Claude with the same safety boundaries — just without the tool definitions, CLAUDE.md, and behavior rules. Treat this as a "minimal harness," not a "bare/unrestricted model."

1. Background: what actually gets injected per request

Before sending your message, Claude Code stacks several context layers:

  1. Built-in system prompt — agent identity, behavior rules, output style
  2. Tool definitions (schema) — JSON schemas for Bash / Read / Edit / …
  3. Setting sources — ~/.claude/settings.json, CLAUDE.md, auto-memory
  4. Dynamic blocks — cwd, git status, OS, date (system-reminder)
  5. Model training — baked into weights, NOT visible/removable here

2. What each flag removes

FlagRemovesLayer
--system-prompt ""Built-in Claude Code agent instructionssystem prompt
--setting-sources ""CLAUDE.md / memory / settings injectionconversation injection
--tools ""Tool schemastool definitions
clean cwdAvoids triggering auto-memoryconversation injection
--append-system-prompt(adds instead of replacing — for reference)system prompt

3. Minimal-harness invocation

mkdir -p /tmp/bare-test && cd /tmp/bare-test  # clean cwd -> no memory
claude --system-prompt "" --tools "" --setting-sources "" "your prompt"
# one-shot:
claude -p --system-prompt "" --tools "" --setting-sources "" "your prompt"

4. Residuals you CANNOT remove (subscription/OAuth CLI)

  1. One SDK identity line: You are a Claude agent, built on Anthropic's Claude Agent SDK.
  2. A system-reminder containing user email + current date
  3. The model's safety training — not in the system prompt, not removable by any flag
  4. The server-side safety classifier — not disableable via local config

5. Scope / honest positioning

  • Good for: A/B tests (harness vs. no-harness response diffs), studying how the harness affects style and behavior, transparency/research.
  • Not for: "removing safety limits" — it does not and cannot do that.
  • A truly zero-injection setup (no SDK identity line) only exists via the raw API with no system parameter; that's a separate path and still subject to the same model-level safety.
Ver o original no GitHub

Bom para

Relacionados