setup-guide

its_hub Setup Guide

You are helping the user set up inference-time scaling.

Step 1: Detect Environment

"${CLAUDE_PLUGIN_ROOT}/scripts/its_detect.sh"

Also check for API keys in the environment:

echo "openai_key=${OPENAI_API_KEY:+found}" "anthropic_key=${ANTHROPIC_API_KEY:+found}" "endpoint=${OPENAI_BASE_URL:-none}"

Step 2: Install if Needed

If library=missing:

• Explain: "its_hub is a library for inference-time scaling — it generates multiple LLM responses and selects the best one using voting or scoring algorithms."
• Ask permission: "I can install it for you. Want me to proceed?"
• If yes and installer=uv: run uv pip install "its_hub[lm]"
• If yes and installer=pip: run pip install "its_hub[lm]"
• If installer=none: tell the user they need Python 3.11+ and pip/uv installed first

Step 3: Quick Setup or Custom

If an API key was detected, offer a one-question fast path:

"I detected your OpenAI API key. I can set up with these defaults:

• Model: gpt-4o-mini
• Algorithm: self-consistency
• Budget: 8

Accept these defaults, or would you like to customize?"

If the user accepts, skip to Step 5 using:

• endpoint: https://api.openai.com/v1 (or OPENAI_BASE_URL if set)
• model: gpt-4o-mini
• algorithm: self-consistency
• budget: 8

If no API key was detected, or the user wants to customize, proceed to Step 4.

Step 4: Collect Configuration

Ask these questions one at a time:

1. Endpoint: "What's your model endpoint URL?" — e.g., http://localhost:8000/v1 for vLLM, https://api.openai.com/v1 for OpenAI
2. Model name: "What's the model identifier?" — e.g., gpt-4o, Qwen/Qwen2.5-32B-Instruct
3. Algorithm: "Which scaling algorithm do you want to use?" — consult the inference-scaling-guide skill for guidance if the user is unsure.
   • Self-consistency — Votes on the most common answer. No extra setup needed.
   • Best-of-N — Scores each with an LLM judge. Requires a judge model.
4. Budget: "Which budget do you want to use?" — consult the inference-scaling-guide skill for guidance if the user is unsure.

Algorithm-Specific Config

Self-consistency:

• Ask: "Do you need regex patterns for answer extraction? (e.g., \boxed{...} for math). If unsure, skip — default voting works for most cases."
• Ask: "Will your prompts involve tool/function calls?" If yes, ask which voting strategy: tool_name, tool_args, or tool_hierarchical

Best-of-N:

• Ask: "Which model should be the judge? (This can be the same model or a different one)"
• Collect: judge model name, judge endpoint (default: same as generation model)

Step 5: Ensure API Key

API keys are read from environment variables — never store them in the config file.

If no API key was detected in Step 1, tell the user to set the appropriate environment variable:

"Set your API key as an environment variable before running scaling:

export OPENAI_API_KEY="sk-..."        # OpenAI or OpenAI-compatible endpoints
export ANTHROPIC_API_KEY="sk-ant-..."  # Anthropic endpoints

For local vLLM endpoints that don't require authentication, no API key is needed."

Step 6: Save Config

Write the config to .its-hub/config.json:

{
  "models": {
    "default": {
      "endpoint": "<endpoint>",
      "model": "<model_name>"
    }
  },
  "algorithm": "<algorithm>",
  "budget": "<budget>",
  "algorithm_config": {}
}

Add .its-hub/ to .gitignore if not already present.

Confirm the config file was written, then report success:

"Setup complete! To run scaling, use the inference-scaling skill.

API keys are read from environment variables, not the config file. Make sure the appropriate variable is set in your shell:

export OPENAI_API_KEY="sk-..."        # OpenAI or OpenAI-compatible endpoints (including vLLM with auth)
export ANTHROPIC_API_KEY="sk-ant-..."  # Anthropic endpoints

Local endpoints (e.g., vLLM without auth) don't need an API key."

Adding More Models

If this skill is invoked again and a config already exists, ask: "You already have a configuration. Do you want to update it or add another model?"

If adding a model: collect endpoint and model name. Add a new entry to the models dict using the model name as key.