Summary

• Added run_custom.py, a CLI entry point for running RecursiveMAS on user-provided prompts without modifying benchmark datasets.
• Supports custom inputs from repeated -q/--question flags or .txt, .json, .jsonl, and .ndjson files.
• Adds task routing for math, choice, and code, plus a new custom reasoning mode for open-ended prompts, mapping each mode onto the existing RecursiveMAS inference pipelines and prompt/evaluation
  families.
• Captures the internal runner JSONL output, normalizes it into cleaner per-sample records, prints readable outputs, and optionally writes results with --output_jsonl.
• Adds controls for sampling, rollouts, recursion depth, token limits, dtype/device settings, answer-format retries, pipeline log capture, and tool/code-generation options.
• Updated the README with run_custom.py in the project layout, a short component description, supported file formats, task guidance, and examples for direct prompts, file-based prompts, and collaboration-style comparisons.

Example usage

python run_custom.py --style sequential_scaled --task reasoning \
  -q "In materiomics, a material is studied across atomic, molecular, microstructural, and macroscopic scales. Explain how hierarchical structure can create toughness in bone while preserving stiffness." \
  --device cuda

Validation

• Checked the README custom prompt section against the implemented CLI behavior.
• Verified run_custom.py syntax with python -m py_compile run_custom.py

TODO: Replace the current compatibility-layer implementation with a first-class custom dataset and inference API. The present runner intentionally reuses the existing benchmark-oriented inference path so that custom prompts behave consistently with the current RecursiveMAS pipelines, but a dedicated custom inference endpoint would make this workflow cleaner, easier to maintain, and less dependent on runtime adaptation.