What problem does it solve? When a check_models harness finding turns out to be a bug in upstream mlx-vlm, this Skill guides turning that finding into a clean, mergeable upstream fix instead of just an issue report, covering code placement, tests, formatting, and PR expectations. ## Core Features & Use Cases - Code Placement Guidance: Directs changes to the right locations in the editable mlx-vlm checkout, including model code under mlx_vlm/models/<model_type>/, processor clean_output hooks, and backward-compatible ModelConfig arguments. - Focused Test Execution: Runs targeted pytest classes in test_models.py and test_processors.py using conda + pip (never uv), then validates the fix on the real checkpoint with a native repro. - Formatting Alignment: Matches upstream's black/isort/autoflake hooks (88-column layout) using the ruff already installed in this repository, without installing extra tools. - Use Case: A check_models run shows a model emitting leftover special tokens; use this Skill to add a clean_output hook in the model's processor, write a TestOutputControlTokens-style test, format the diff to upstream's 88-column standard, and prepare the PR branch. ## Quick Start Turn my confirmed check_models finding about the glm4v processor output into an upstream mlx-vlm fix with tests and a prepared PR branch.
