What problem does it solve?
Generic AI code review misses the specific correctness, GPU runtime, API compatibility, and testing concerns that SGLang maintainers actually raise, producing feedback that does not match how the project is really reviewed.
Core Features & Use Cases
- Corpus-grounded review: Consults a gzip JSONL corpus of 40,110 review threads and 93,774 human comments from sgl-project/sglang PRs (project start through 2026-07-27), covering inline review threads, PR conversations, and review submissions with original multilingual text preserved.
- Exhaustive sweep and synthesis: A mandatory segmented sweep script aggregates every relevant historical thread by touched path and risk keyword, then a first-N query tool supports follow-up drill-downs by category, reviewer, kind, or PR number.
- Structured review output: Every review opens with a PR comprehension block (change summary plus Mermaid execution flowchart with changed nodes marked), followed by a historical synthesis, severity-ordered findings with file/line references, and test or benchmark gaps.
- Use Case: Before opening a PR that touches the SGLang scheduler and CUDA graph runner, run the sweep with those paths and keywords like "cuda graph" to surface recurring maintainer concerns, then produce a review-prep pass listing likely reviewer objections and missing benchmark evidence.
Quick Start
Review my current SGLang diff the way a human maintainer would, grounding the findings in the historical review corpus.
