What problem does it solve? Research codebases rarely ship with complete, working dependency specs, so reproducing a paper often stalls on Python/CUDA/torch version conflicts and broken environments. This Skill turns scattered version evidence (lockfiles, READMEs, the paper itself, code-level API clues) into one working, verified conda environment with every pin recorded. ## Core Features & Use Cases - Tiered version resolution: Ranks evidence from lockfiles and Dockerfiles (tier 1) down through README instructions, the paper's text, and inference from imports, resolving conflicts deterministically and recording which tier decided each fact. - Three operating modes: Fresh setup, repair mode for fixing an existing env given a traceback (with the smallest change that resolves the error), and candidate-env reuse checks that never modify another repo's environment. - Verification and lockfile export: Import-tests every major package and entry-point module, checks torch.cuda.is_available() against host GPU reality, and exports a pip-freeze lockfile plus a structured markdown report. - Use Case: While reproducing a machine learning paper, stage 4 of your pipeline hits "ModuleNotFoundError: No module named 'scipy'". Invoke this Skill in repair mode to install a compatible scipy version, re-verify the import, re-export the lockfile, and append a dated entry to the repair log. ## Quick Start Set up the conda environment for this cloned research repo, pin all package versions into a lockfile, and verify the imports and CUDA setup work on this machine.
