What problem does it solve? Turning a failed mlx-vlm model run into a clear, reproducible upstream bug report is tedious: you must gather environment facts, minimize the repro, and format everything so a maintainer can act on it. This Skill converts check_models failures and native repros into paste-ready GitHub issue drafts without manual assembly. ## Core Features & Use Cases - Artifact-driven drafting: Builds issues from existing src/output/issues/issue_*.md, diagnostics.md, run_summary.md, results.jsonl, and environment.log instead of re-deriving facts. - Repro minimization guidance: Reduces failures to a single native python -m mlx_vlm.generate command or a server start plus one curl request, with conda/pip conventions (never uv). - Classification and quality bar: Distinguishes model/config, runtime, harness preflight, environment, and connectivity failures, and enforces a maintainer-ready template with environment, model revision, expected vs actual behavior, and trimmed tracebacks. - Use Case: After a benchmark sweep crashes one model, ask for an upstream issue draft and receive a complete Markdown report with exact versions, resolved model revision, minimal repro command, and root exception—ready to paste into GitHub only if you explicitly ask. ## Quick Start Draft an upstream mlx-vlm issue from the latest crash draft in src/output/issues without filing it on GitHub.
