MCP HubMCP Hub
Вернуться к навыкам

harness:setup

raphaelchristi
Обновлено 5 days ago
27
4
27
Посмотреть на GitHub
Другоеai

О программе

Этот навык настраивает Harness Evolver v3 в проекте, анализируя код и конфигурируя LangSmith для оптимизации агента. Он запускает базовую оценку для улучшения производительности LLM-агента и активируется, когда пользователи впервые упоминают evolver в проекте без существующего файла .evolver.json. Настройка включает проверку API-ключей LangSmith и подготовку среды для эволюции агента.

Быстрая установка

Claude Code

Рекомендуется
Основной
npx skills add raphaelchristi/harness-evolver -a claude-code
Команда плагинаАльтернативный
/plugin add https://github.com/raphaelchristi/harness-evolver
Git клонированиеАльтернативный
git clone https://github.com/raphaelchristi/harness-evolver.git ~/.claude/skills/harness:setup

Скопируйте и вставьте эту команду в Claude Code для установки этого навыка

Документация

/harness:setup

Set up the Harness Evolver v3 in a project. Explores the codebase, configures LangSmith, runs baseline evaluation.

Prerequisites

Check for LangSmith API key — it can be in the environment, the credentials file, or .env:

python3 -c "
import os, platform
key = os.environ.get('LANGSMITH_API_KEY', '')
if not key:
    creds = os.path.expanduser('~/Library/Application Support/langsmith-cli/credentials') if platform.system() == 'Darwin' else os.path.expanduser('~/.config/langsmith-cli/credentials')
    if os.path.exists(creds):
        for line in open(creds):
            if line.strip().startswith('LANGSMITH_API_KEY='):
                key = line.strip().split('=',1)[1].strip()
    if not key and os.path.exists('.env'):
        for line in open('.env'):
            if line.strip().startswith('LANGSMITH_API_KEY=') and not line.strip().startswith('#'):
                key = line.strip().split('=',1)[1].strip().strip('\"').strip(\"'\")
print('OK' if key else 'MISSING')
"

If MISSING: "Set your LangSmith API key: export LANGSMITH_API_KEY=lsv2_pt_... or run npx harness-evolver@latest to configure."

The tools auto-load the key from the credentials file, but the env var takes precedence.

Resolve Tool Path and Python

# Prefer env vars set by plugin hook; fallback to legacy npx paths
TOOLS="${EVOLVER_TOOLS:-$([ -d ".evolver/tools" ] && echo ".evolver/tools" || echo "$HOME/.evolver/tools")}"
EVOLVER_PY="${EVOLVER_PY:-$([ -f "$HOME/.evolver/venv/bin/python" ] && echo "$HOME/.evolver/venv/bin/python" || echo "python3")}"

Use $EVOLVER_PY instead of python3 for ALL tool invocations. This ensures the venv with langsmith is used.

IMPORTANT: Never pass LANGSMITH_API_KEY inline in Bash commands. The key is loaded automatically by the SessionStart hook (from credentials file or environment) and by each Python tool's ensure_langsmith_api_key(). Passing it inline exposes it in the output. If the key is missing, tell the user to run export LANGSMITH_API_KEY=lsv2_pt_... instead.

Phase 1: Explore Project (automatic)

find . -maxdepth 3 -type f -name "*.py" -not -path "*/.venv/*" -not -path "*/node_modules/*" -not -path "*/__pycache__/*" | head -30

Monorepo detection: if the project root has multiple subdirectories with their own main.py or pyproject.toml, it's a monorepo. Use AskUserQuestion to ask WHICH app to optimize before proceeding — do NOT scan everything.

Look for:

  • Entry points: files with if __name__, or named main.py, app.py, agent.py, graph.py, pipeline.py
  • Existing LangSmith config: LANGCHAIN_PROJECT / LANGSMITH_PROJECT in env or .env
  • Existing test data: JSON files with inputs, CSV files, etc.
  • Dependencies: requirements.txt, pyproject.toml

To identify the framework, read the entry point file and its immediate imports. The proposer agents will use Context7 MCP for detailed documentation lookup — you don't need to detect every library, just identify the main framework (LangGraph, CrewAI, OpenAI Agents SDK, etc.) from the imports you see.

Detect virtual environments — check for venvs in the project or parent directories:

# Check common venv locations
for venv_dir in .venv venv ../.venv ../venv; do
    if [ -f "$venv_dir/bin/python" ]; then
        echo "VENV_FOUND: $venv_dir/bin/python"
        break
    fi
done

If a venv is found, use it for the entry point instead of bare python. The agent's dependencies are likely installed there, not in the system Python. For example: ../.venv/bin/python agent.py {input} instead of python agent.py {input}.

Identify the run command — how to execute the agent. Use {input} as a placeholder for the JSON file path:

  • .venv/bin/python main.py {input} — if venv detected (preferred)
  • python main.py {input} — agent reads JSON file from positional arg
  • python main.py --input {input} — agent reads JSON file from --input flag
  • python main.py --query {input_json} — agent receives inline JSON string

The runner writes {"input": "user question..."} to a temp .json file and replaces {input} with the file path. If the entry point already contains --input (without placeholder), the runner appends the file path as the next argument.

If no placeholder and no --input flag detected, the runner appends --input <path> --output <path>.

Phase 2: Confirm Configuration (interactive)

Present all detected configuration in one view with smart defaults and ask for confirmation.

Use AskUserQuestion:

{
  "questions": [{
    "question": "Here's the configuration for your project:\n\n**Entry point**: {command}\n**Framework**: {framework}\n**Python**: {venv_path or 'system python3'}\n**Optimization goals**: accuracy (correctness evaluator)\n**Test data**: generate 30 examples with AI\n\nDoes this look good?",
    "header": "Setup Configuration",
    "multiSelect": false,
    "options": [
      {"label": "Looks good, proceed", "description": "Use these settings and start setup"},
      {"label": "Customize goals", "description": "Choose different optimization goals"},
      {"label": "I have test data", "description": "Use existing JSON file or LangSmith project"},
      {"label": "Let me adjust everything", "description": "Change entry point, framework, goals, and data source"}
    ]
  }]
}

If "Looks good, proceed": Use defaults — goals=accuracy, data=generate 30 with testgen. Skip straight to Phase 3.

If "Customize goals": Ask the goals question, then proceed to Phase 3 with testgen as default data source.

Use AskUserQuestion:

{
  "questions": [{
    "question": "What do you want to optimize?",
    "header": "Goals",
    "multiSelect": true,
    "options": [
      {"label": "Accuracy", "description": "Correctness of outputs — LLM-as-judge evaluator"},
      {"label": "Latency", "description": "Response time — track and minimize"},
      {"label": "Token efficiency", "description": "Fewer tokens for same quality"},
      {"label": "Error handling", "description": "Reduce failures, timeouts, crashes"}
    ]
  }]
}

Map selections to evaluator configuration for setup.py.

Phase 2.5: Mode Selection

{
  "questions": [{
    "question": "Evolution mode?",
    "header": "Mode",
    "multiSelect": false,
    "options": [
      {"label": "light", "description": "20 examples, 2 proposers, ~2 min/iter. Good for testing."},
      {"label": "balanced (Recommended)", "description": "30 examples, 3 proposers, ~8 min/iter. Best trade-off."},
      {"label": "heavy", "description": "50 examples, 5 proposers, ~25 min/iter. Maximum quality."}
    ]
  }]
}

Pass selection to setup.py as --mode light|balanced|heavy.

The mode determines testgen count:

  • light: generate 20 examples
  • balanced: generate 30 examples (default, current behavior)
  • heavy: generate 50 examples

If "I have test data": Ask the data source question, then proceed to Phase 3 with accuracy as default goal.

Use AskUserQuestion with preview:

{
  "questions": [{
    "question": "Where should test inputs come from?",
    "header": "Test data",
    "multiSelect": false,
    "options": [
      {
        "label": "Import from LangSmith",
        "description": "Use real production traces as test inputs",
        "preview": "## Import from LangSmith\n\nFetches up to 100 recent traces from your production project.\nPrioritizes traces with negative feedback.\nCreates a LangSmith Dataset with real user inputs.\n\nRequires: an existing LangSmith project with traces."
      },
      {
        "label": "I have a file",
        "description": "Point to an existing file with test inputs",
        "preview": "## Provide Test Data\n\nSupported formats:\n- JSON array of inputs\n- JSON with {\"inputs\": {...}} objects\n- CSV with input columns\n\nExample:\n```json\n[\n  {\"input\": \"What is Python?\"},\n  {\"input\": \"Explain quantum computing\"}\n]\n```"
      }
    ]
  }]
}

If "Import from LangSmith": discover projects and ask which one (same as v2 Phase 1.9). If "I have a file": ask for file path.

If "Let me adjust everything": Ask all three original questions in sequence — confirm detection (entry point, framework, run command), then goals, then data source — using the question formats above.

Phase 3: Run Setup

Build the setup.py command based on all gathered information:

$EVOLVER_PY $TOOLS/setup.py \
    --project-name "{project_name}" \
    --entry-point "{run_command}" \
    --framework "{framework}" \
    --goals "{goals_csv}" \
    ${DATASET_FROM_FILE:+--dataset-from-file "$DATASET_FROM_FILE"} \
    ${DATASET_FROM_LANGSMITH:+--dataset-from-langsmith "$DATASET_FROM_LANGSMITH"} \
    ${PRODUCTION_PROJECT:+--production-project "$PRODUCTION_PROJECT"}

If "Generate from code" was selected AND no test data file exists, first spawn the testgen agent to generate inputs, then pass the generated file to setup.py.

Phase 4: Generate Test Data (if needed)

If testgen is needed, spawn it:

Agent(
  subagent_type: "harness-testgen",
  description: "TestGen: generate test inputs",
  prompt: |
    <objective>
    Generate 30 diverse test inputs for this project.
    Write them as a JSON array to test_inputs.json.
    </objective>

    <files_to_read>
    {all .py files discovered in Phase 1}
    </files_to_read>

    <output>
    Create test_inputs.json with format:
    [{"input": "..."}, {"input": "..."}, ...]
    </output>
)

Then pass --dataset-from-file test_inputs.json to setup.py.

Phase 5: Report

Setup complete!
  Project: evolver-{name}
  Dataset: {name}-eval-v1 ({N} examples)
  Evaluators: {list}
  Baseline score: {score}
  Config: .evolver.json

Next: run /harness:evolve to start optimizing.

Gotchas

  • If .evolver.json already exists, ask before overwriting.
  • If the agent needs a venv, the run command should activate it: cd {dir} && .venv/bin/python main.py
  • No project venv detected: Before building the entry_point, check if .venv/bin/python or venv/bin/python exists in the project directory. If NOT, warn the user:
    WARNING: No Python venv found in this project (.venv/ or venv/).
    The entry_point should use the project's own Python, not ~/.evolver/venv/bin/python.
    Create one first: python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
    
    Do NOT use ~/.evolver/venv/bin/python as the entry_point — that's for evolver tools only. Using it pollutes the evolver venv with agent dependencies.
  • If LangSmith connection fails, check API key and network. The ensure_langsmith_api_key() function validates key format and rejects dummy/test keys with a warning.
  • Eval concurrency defaults to 3 (runs 3 examples in parallel). If the agent can't handle parallel execution (writes to shared files, uses a fixed port, holds a DB lock), set eval_concurrency: 1 in .evolver.json after setup.
  • Companion plugin: For full proposer observability, recommend installing langsmith-tracing from langchain-ai/langsmith-claude-code-plugins. Each proposer's file reads, edits, and commits become visible in LangSmith.

GitHub репозиторий

raphaelchristi/harness-evolver
Путь: skills/setup
0
agent-evolutionclaude-code-plugincodex-skillsharness-engineeringmeta-harness

Похожие навыки

llamaguard

Другое

LlamaGuard — это модель от Meta с 7–8 миллиардами параметров для модерации входных и выходных данных больших языковых моделей по шести категориям безопасности, таким как насилие и разжигание ненависти. Она обеспечивает точность 94–95% и может быть развернута с помощью vLLM, Hugging Face или Amazon SageMaker. Используйте этот навык, чтобы легко интегрировать фильтрацию контента и защитные механизмы в ваши ИИ-приложения.

Просмотреть навык

cost-optimization

Другое

Этот навык Claude помогает разработчикам оптимизировать облачные расходы за счет правильного подбора ресурсов, стратегий тегирования и анализа затрат. Он предоставляет framework для сокращения облачных расходов и внедрения управления затратами в AWS, Azure и GCP. Используйте его, когда вам нужно проанализировать расходы на инфраструктуру, оптимизировать ресурсы или уложиться в бюджетные ограничения.

Просмотреть навык

quantizing-models-bitsandbytes

Другое

Этот навык выполняет квантизацию LLM до 8-битной или 4-битной точности с использованием библиотеки bitsandbytes, обеспечивая сокращение использования памяти на 50-75% при минимальной потере точности. Он идеально подходит для запуска больших моделей при ограниченной памяти GPU или для ускорения вывода, поддерживая форматы INT8, NF4 и FP4. Навык интегрируется с HuggingFace Transformers и позволяет использовать обучение QLoRA и 8-битные оптимизаторы.

Просмотреть навык

dispatching-parallel-agents

Другое

Этот навык Claude распределяет нескольких агентов для исследования и устранения трёх и более независимых проблем параллельно. Он предназначен для сценариев с несвязанными сбоями, которые можно устранить без общего состояния или зависимостей. Ключевая возможность — параллельное решение проблем, где за каждую независимую предметную область назначается отдельный агент для максимальной эффективности.

Просмотреть навык