Skip to content

Agent Skill

Without a shared workflow, a coding agent must rediscover how each project distributes jobs, records progress, retries failures, and collects results. Labtasker's labtasker skill supplies those operating rules, so a request such as “run these cases in parallel across 8 GPUs with Labtasker” is enough to start the standard Labtasker workflow.

The skill covers v2 Task submission, Worker design, routing, inspection, updates, recovery, and adapting an existing experiment pipeline. It also explains how to compare pending Task demand with observed Worker activity through Python or CLI grouped counts, including pagination and delayed observations. It covers fuzzy Task name search, advisory Server version warnings, and the Worker’s consecutive-failure limit without confusing it with Task retry budgets. During a migration it asks about the project's entry points, reusable setup, retry units, resources, dependencies, and outputs, then owns the mapping into Labtasker instead of requiring a newcomer to design Tasks, Workers, routes, or Queues. It does not choose the experiment, allocate GPUs, or keep the agent inside the execution loop. Its short SKILL.md routes deployment, Worker, inspection, and recovery questions to focused references; the official installable package is skills/labtasker/.

LLM-readable documentation

The Agent Skill teaches an agent how to operate Labtasker. A built documentation site exposes llms.txt to help an agent find the right supporting material. It links directly to the raw Markdown for the main guides, API references, and full specification. The llms.txt source can also be read directly.

Claude Code marketplace

Add the repository as a marketplace, then install its plugin:

/plugin marketplace add luocfprime/labtasker
/plugin install labtasker-skill@labtasker

The installed skill is invoked as /labtasker-skill:labtasker. Claude Code can also select it automatically when a request matches its description.

Agent Skills CLI

The open Agent Skills installer supports Claude Code, Codex, OpenCode, Cursor, and other compatible agents:

npx skills add \
  https://github.com/luocfprime/labtasker/tree/main/skills/labtasker

The direct skill path keeps contributor-only repository workflows out of the installation. The installer prompts for target agents and project or global scope. For a non-interactive Codex installation, for example:

npx skills add \
  https://github.com/luocfprime/labtasker/tree/main/skills/labtasker \
  --agent codex --global --yes

Use --agent claude-code instead to install through the same open skill format without using the Claude Code marketplace.

Repository checkouts

This repository exposes the same skill content at .agents/skills/labtasker for repository-aware agents. It is a relative symlink, so every discovery method reads the same files.

The agent skill describes the public Labtasker product. Contributor-only workflows such as releases and cross-surface contract changes remain separate under .agents/skills/ and are documented in Development.

Maintaining the skill

The public skill and its bundled references are maintained together. Feature changes should update the relevant reference and add a realistic agent scenario to the skill regression suite. The suite separates candidate requests from examiner setup and checks. Candidates operate only in temporary test environments, using the skill and public interfaces. Test fixtures and grading rules are not part of the installed skill. Questions describe ordinary user goals without prescribing the API, commands, steps, or expected answer. The examiner checks whether the goal was met, whether the agent interpreted the results correctly, and whether it made unnecessary changes or requests for clarification. Low-level product invariants remain in the ordinary automated tests.

See Development for the examination and revision workflow. Executable setup/check validation and an independent agent pass are separate results; neither guarantees support on untested platforms.