llm-d

Generate weekly nightly E2E trend report with failure patterns and success rates

Analyze nightly E2E job failures across llm-d repos. Smart router to sub-skills based on failure type.

Diagnose deployment failures in nightly E2E runs (helmfile, helm, install.sh)

Diagnose GPU contention failures in nightly E2E runs

Diagnose infrastructure and authentication failures in nightly E2E runs (GCP, AWS, runner issues)

Diagnose pod readiness timeout failures in nightly E2E runs

Full root cause analysis for nightly E2E failures that don't fit simple categories

Diagnose E2E test failures in nightly runs (e2e-validate.sh, make test-e2e)

Enforce the Product -> UI/UX -> Engineering development loop for new features. Run when starting a feature, reviewing progress, or before merging to main.

Skill for Analyst / PR roles to compare cross-stack performance and publish reproducible, community-facing benchmarks.

Skill for helping Solutions Architects and Customer Engineers sweep configurations and find optimal end-to-end stack setups.

Skill for helping llm-d Feature Developers run, compare, and validate benchmarks for new optimization features.

Skill for helping Stack Operators monitor production stacks, perform regular stress tests, and detect regressions.

Convert llm-d guides to benchmark scenario and experiment files. Use when the user wants to convert an llm-d deployment guide (Helm values or kustomize files) into llm-d-benchmark scenario files. Triggers on /convert-guide command with a URL or local path to a guide.

Add stack metadata to existing benchmark results in Google Cloud Storage (GCS). Use when the user wants to backfill or enrich previous reports with stack metadata from a specific serving cluster endpoint.

Run networking tests for LLM-D to ensure the environment is properly set up before vLLM is started.

Run preflight checks for LLM-D to ensure the environment is properly set up before vLLM is started.

Review an open pull request for this project. Checks Go code quality (AGENTS.md), test coverage, and Kubernetes security. Posts findings as a GitHub PR comment (idempotent: creates on first run, updates with resolved/new diff on re-runs). Use when the user asks to review a PR. Invoke with /pr-review [PR-number] or /pr-review to auto-detect the current branch's PR.