roofline

Roofline, a Lenialno product line. По-русски

Agent Skills for the Inference Engineer role: run vLLM or SGLang on your own GPUs behind an OpenAI-compatible endpoint, serve Claude Code and Qwen Code through a proxy, stay healthy on air-gapped hosts. 13 skills, one Claude Code subagent and a Taskfile toolbox; the competency map follows roadmap.sh/inference-engineering.

Quick start

As Agent Skills, for any agent that supports them (Claude Code, Codex, Cursor, OpenCode and others):

npx skills add lenialno/roofline --list                                      # see what is inside
npx skills add lenialno/roofline --skill roofline --skill roofline-toolbox   # pick skills
npx skills add lenialno/roofline --all                                       # everything, all agents
npx skills add lenialno/roofline -g -a claude-code --skill roofline-toolbox  # user scope, Claude Code only

As a Claude Code plugin, available in every project on the machine, from the Lenialno catalog lenialno/skills:

claude plugin marketplace add lenialno/skills
claude plugin install roofline@lenialno

[!TIP] Update with claude plugin marketplace update lenialno, then claude plugin update roofline@lenialno and /reload-plugins. To try a local clone for one session: claude --plugin-dir /path/to/roofline.

Then ask the agent what to do first with your local LLM server. The roofline skill routes the question to the right skill, task or subagent.

The toolbox needs python3 and curl, go-task for task, and nvidia-smi for the gpu task. The scripts also run without task.

Skills

Agents pick a skill by its description, so phrases like the ones in the last column are enough. The first six also ship in Russian under the same name with a -ru suffix: roofline-ru, roofline-runtime-ru, roofline-toolcall-ru, roofline-perf-ru, roofline-incident-ru, roofline-capacity-ru. The toolbox is language-neutral and shared by both sets.

[!NOTE] No MCP servers and no hooks, by design. Live state comes from Task: the agent reads GPU memory, health, metrics and logs from the real server instead of guessing.

Subagent

incident-investigator (agents/incident-investigator.md) is a Claude Code subagent that roofline-incident hands an investigation to when it needs long logs or hours of metrics. It only reads and measures (health, metrics, gpu, models, logs, smoke-toolcall), never runs serve, stop or restart, and returns a verdict of at most 25 lines: cause, evidence, rejected hypotheses, next action.

Toolbox

A Taskfile plus dependency-free bash and python3 scripts in skills/roofline-toolbox/; the root Taskfile.yml re-exports it without a namespace. Every task is idempotent and exits non-zero on failure. From any project:

task -t /path/to/roofline/Taskfile.yml health

Or include it in the project Taskfile and call task infer:health:

version: '3'
includes:
  infer:
    taskfile: /path/to/roofline/skills/roofline-toolbox/Taskfile.yml

Commands run in the directory you call task from, where inference.env (KEY=value lines) and .inference-state/ live. task with no arguments prints the task list under the Roofline mark.

[!NOTE] task serve starts only a named profile from profiles/ (or PROFILES_DIR). Before promoting a profile: task health && task smoke-toolcall && task bench.

Profiles and their variables: skills/roofline-toolbox/profiles/README.md.

Priority, highest first: task variable (task serve INFER_PROFILE=x), environment, inference.env in the current directory, built-in default.

Examples

Real output, cap