skill-reach

An evaluation suite for AI agent skill routing, collision detection, multi-step trajectory evaluation, and description optimization.

While standard evaluation suites measure skill execution after invocation, skill-reach measures the discovery and selection phase: determining whether incoming user queries route to the appropriate skill, execute required multi-step tool handoffs, and avoid redundant calls in the presence of competing descriptions.

Contents

• Installation
• Authentication
• Quickstart
• Multi-Step Trajectory Evaluation
• Commands
• Supported Agents
• Configuration
• Security and Sandboxing
• Reproducibility and Provenance
• Development and Testing
• Repository Layout
• Support and Feedback

Installation

Requires Python 3.12+ and uv.

git clone https://github.com/google/skill-reach
cd skill-reach
uv sync

# Optional: initialize project configuration from template
cp reach.example.toml reach.toml

To enable dense semantic embeddings or the Google Antigravity SDK, install the optional extras:

# Install individual extras
uv sync --extra antigravity-sdk
uv sync --extra semantic

# Or sync with all extras
uv sync --all-extras

To install reach as an isolated standalone CLI tool from the local repository:

# Install base CLI tool
uv tool install .

# Or install with optional extras
uv tool install ".[semantic,antigravity-sdk]"

Authentication

Configure credentials for your target model provider:

Google Gemini API (AI Studio / Developer API)

When evaluating with Gemini models via API key authentication (antigravity-sdk or antigravity-cli):

export GEMINI_API_KEY="your-api-key"
# or
export GOOGLE_API_KEY="your-api-key"

Google Cloud Agent Platform (Vertex AI / ADC)

When using Agent Platform enterprise infrastructure (antigravity-cli or antigravity-sdk), authenticate with Application Default Credentials (ADC) and configure your project:

# 1. Authenticate with Google Cloud and bind quota project
gcloud auth application-default login --project your-project-id

# 2. Configure Agent Platform environment variables
export GOOGLE_CLOUD_PROJECT="your-project-id"