sglang-prod-incident-triage

Replay failing SGLang requests on a clean target for incident triage.

Updated Jun 29, 2026
One-click install
npx skills add https://github.com/hey-kong/sglang --skill sglang-prod-incident-triage-hey-kong
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: sglang-prod-incident-triage
Source: https://github.com/hey-kong/sglang/tree/main/.claude/skills/sglang-prod-incident-triage
Command: npx skills add https://github.com/hey-kong/sglang --skill sglang-prod-incident-triage-hey-kong

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires requests, and includes scripts (resource) and references (resource) components.

What problem does it solve?

Replay-first debugging flow enables rapid incident resolution for SGLang serving problems. It guides operators from baseline collection to failing-request capture, replay on a clean target, and then targeted tool use for CUDA crashes, distributed hangs, or profiling scenarios.

Core Features & Use Cases

  • Replay-first workflow turns live serving problems into reproducible debug steps.
  • It links to specialized tools for CUDA crash, distributed hang, and torch profiling when narrowed.
  • Use cases include health degradation, latency regressions, queue growth, or post-deploy behavioral changes.

Quick Start

Collect a live bundle, save the failing request or crash dump, then replay on a clean target to identify the root cause.

Frequently Asked Questions about sglang-prod-incident-triage

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I debug SGLang serving incidents like latency regressions or queue growth?

Debug SGLang serving incidents using a replay-first workflow that collects baseline data and captures failing requests to replay on a clean target. This approach isolates the root cause of health degradation or post-deploy behavioral changes.

What is the best way to reproduce a CUDA crash or distributed hang from an SGLang server?

Reproduce a CUDA crash or distributed hang by saving the failing request or crash dump from the live server. Replay this bundle on a clean target to trigger specialized tools for deterministic debugging.

How do I triage an SGLang incident step by step?

Triage an SGLang incident by collecting a live bundle, saving the failing request, replaying it on a clean target, and then applying targeted tools for CUDA crashes or profiling to identify the root cause.

Do I need specific dependencies to run the SGLang incident triage workflow?

You need the standard Python requests library to run the SGLang incident triage workflow. The process relies on standard Python tooling and integrates optional scripts and references for end-to-end support.

Can I use this replay debugging flow for post-deploy behavioral changes in SGLang?

Yes, you can use this replay debugging flow for post-deploy behavioral changes in SGLang. It captures the failing request from the live environment and replays it to resolve serving problems rapidly.

Why does replaying a failing request on a clean target help troubleshoot SGLang serving problems?

Replaying a failing request on a clean target helps troubleshoot SGLang serving problems by isolating the issue from live environment noise. This deterministic step pinpoints whether CUDA crashes or distributed hangs are caused by specific inputs.