Factory function to create the appropriate MCP connection. Args: transport: Connection type ("stdio", "sse", or "http") command:

Remove duplicate or near-duplicate consecutive frames. Args: threshold: Similarity threshold (0.0-1.0). Higher = more st

Bounce ease-in (bouncy start).

Read a file and return an embedded representation.

Evaluate a single QA pair with the given tools.

Generate complete benchmark.json from run results.

Generate human-readable benchmark.md from benchmark data.

Get easing function by name.

Wait for server to be ready by polling the port.

Retrieve available tools from the MCP server.

Load previous iteration's feedback and outputs. Returns a map of run_id -> {"feedback": str, "outputs": list[dict]}.

Load all run results from a benchmark directory. Returns dict keyed by config name (e.g. "with_skill"/"without_skill", or "new_skill"/"o

Reduce colors in all frames using quantization. Args: num_colors: Target number of colors (8-256) use_global

Package a skill folder into a .skill file. Args: skill_path: Path to the skill folder output_dir: Optional output directory

Parse environment variable strings in format 'KEY=VALUE' into a dictionary.

Parse XML evaluation file with qa_pair elements.

Parse header strings in format 'Key: Value' into a dictionary.

Run evaluation with MCP server tools.

Run the eval + improvement loop.

Check if a path should be excluded from packaging.

Split eval set into train and test sets, stratified by should_trigger.