What problem does it solve?
Contributors to vllm-omni often open PRs that fail review for preventable reasons: missing title prefixes, new model-specific examples, dead code, unverified accuracy or benchmark claims, and code-quality regressions. This Skill runs a maintainer-grade self-review of your branch locally so you catch blocking issues before opening the PR.
Core Features & Use Cases
- Diff-scoped code-quality sweep: Detects five fragility patterns introduced by your diff only — kwargs string-lookup plumbing, broad exception swallows,
Any/wrong type hints, hot-path .clone()/deepcopy, and event-loop blocking — without flagging the pre-existing backlog.
- PR-type checklists in quick or full mode: Categorizes the change as new model, diffusion model, bug fix, performance, or general, then applies the matching checklist covering registry entries, accuracy tests, benchmark rigor, and PR title format.
- Policy enforcement: Blocks new model-specific Python examples, runs a simplification pass for redundant abstractions, and confirms local pre-commit gates (SPDX headers, forbidden imports, mypy, shellcheck, markdownlint).
- Use Case: Before pushing a branch that adds a new TTS model, run the full precheck to verify registry entries match
__init__.py exports, the PR title uses [Model] <Name> format, and no new broad except clauses were introduced.
Quick Start
Ask the assistant to run the precheck skill in full mode on my current branch before I open a PR against vllm-omni.
