Operational memory for AI-infra agents, in plain markdown

What it does It packages 11 skills as plain SKILL.md directories that any skill-capable agent runtime (Claude Code, Codex, Kimi) can load. The skills cover cross-framework serving benchmarks (SGLang, vLLM, TensorRT-LLM, TokenSpeed), torch profiler analysis, capacity planning, model Day-0 support, PR review, and incident triage. Alongside them sits a queryable knowledge base of 118 bilingual model dossiers — SGLang 45, vLLM 44, TensorRT-LLM 15, TokenSpeed 14 — mapping each model family to the PRs that touched it.

The interesting bit The PR history database is the standout: before you patch a model path or claim an optimization is new, you can check what upstream already did, with per-PR evidence cards and audit dates. The repo is also honest about provenance — generated entries are marked as pending manual diff review, and a source refresh is explicitly not a GPU rerun.

Key highlights

• Fair-benchmark skill pins model, framework commit, GPUs, workload, and SLA for every comparison row
• Profiler skills keep prefill and decode separate and check findings against a catalog of known optimizations
• Day-0 support skill turns a new architecture into an SGLang PR DAG with a validation matrix
• Explainer-video skill converts SGLang PRs or articles into narrated Chinese videos with bilingual captions
• Kernel campaigns live in the sibling KDA-Pilot repo, keeping this one focused on procedures and history

Caveats

• The PR dossiers are regenerated from git history, and new generated entries await manual diff review — treat them as leads, not gospel
• Strongly SGLang-centric in the review, incident, and Day-0 skills; the other frameworks mostly appear in benchmarks and histories

Verdict Worth a look if you run inference infrastructure and want your agent to stop rediscovering known optimizations. If you’re not in the SGLang/vLLM serving trenches, it’s a well-organized curiosity.