batch-scaling

name	batch-scaling
description	Use when the user wants to run inference-time scaling on multiple prompts from a file (JSONL, CSV, or TXT). Applies to batch processing, evaluation runs, or dataset-level scaling.
allowed-tools	["Bash(${CLAUDE_PLUGIN_ROOT}/scripts/its_batch_scale.sh:*)","Bash(${CLAUDE_PLUGIN_ROOT}/scripts/its_detect.sh:*)"]

Batch Scaling

Run inference-time scaling on multiple prompts from a file.

Supported Input Formats

• JSONL — one JSON object per line with a prompt or messages field
• CSV — must have a prompt column
• TXT — one prompt per line

Step 1: Check Environment

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

If config=missing, tell the user to run the setup-guide skill first.

Step 2: Run Batch Scaling

Call the batch scaling script with the input file and any overrides:

"${CLAUDE_PLUGIN_ROOT}/scripts/its_batch_scale.sh" [--algorithm ALG] [--budget N] [--model KEY] [--output FILE] <input-file>

The script loads config once and processes all prompts in a single process. Default output path is results/<input_name>_scaled.jsonl.

Step 3: Report Summary

The script outputs a JSON summary with total, succeeded, failed, failures, and output_file.

Report: "N/M prompts completed successfully. K failed. Results written to <output_file>"

If there were failures, list the line numbers and error messages.