What problem does it solve? Integrating a local MLX Swift VLM provider into the TritonKit CLI requires strict separation between the main CLI and heavy ML dependencies, plus disciplined model cache, evidence, and grounding-failure handling that is easy to get wrong. ## Core Features & Use Cases - Helper Contract Enforcement: Delegates real MLX execution to an external triton-mlx-provider helper resolved via --helper, TRITON_MLX_HELPER, or TRITON_MLX_SWIFT_LM_HELPER, keeping mlx-swift-lm out of the main CLI package. - Model Cache Management: Covers triton vlm model list/inspect/preflight/download/prune/remove with idempotent ready entries, fail-closed missing-helper behavior, and explicit removal semantics. - Evidence-Gated Grounding: Ensures VLM grounding writes request, raw output, parsed point, transform, overlay, and policy artifacts, converting failures into vlm_grounding_failed recovery proposals instead of executing actions. - Use Case: When adding Qwen2-VL local smoke validation, use this Skill to build the helper, copy the Metal metallib runtime, run focused swift test --filter VLMMlxSwiftLM tests, and capture compact evidence without committing model weights. ## Quick Start Ask the agent to implement and validate a local mlx-swift-lm VLM grounding flow for the TritonKit CLI, including helper resolution, model cache preflight, and evidence artifacts for a fixture screenshot.
