Skip to content

README.md

Latest commit

 

History

History
95 lines (64 loc) · 4.36 KB

README.md

File metadata and controls

95 lines (64 loc) · 4.36 KB
Raw

UXL Agent Skills

Public catalog for agent skills that teach AI coding agents how to use UXL Foundation oneAPI libraries correctly.

This repository is the central distribution catalog for UXL skills. Skill status is still explicit: the current catalog release is public, while individual skills remain incubating or pilot until reviewed by the owning projects. The long-term source-of-truth model should be hybrid:

  • Library-owned skills live first in the owning project repositories.
  • This catalog mirrors reviewed skills for discovery, installation, validation, and release governance.
  • Cross-cutting skills that span multiple projects live directly here.

Skills

Skill Purpose
uxl-onednn oneDNN primitive selection, memory layout, graph/fusion, backend, and benchdnn workflows
uxl-onemath oneMath domain selection, runtime/compile-time dispatch, backend setup, and build guidance
uxl-onedal oneDAL native C++ and scikit-learn acceleration planning, data tables, and validation
uxl-onetbb oneTBB parallel algorithm, flow graph, task arena, and concurrency pattern selection
uxl-onedpl oneDPL host/device execution policy usage and SYCL algorithm migration
uxl-oneccl oneCCL collectives, launch setup, plugins, distributed training integration, and hang triage
uxl-sycl-build-debug Cross-project SYCL compiler, linker, device discovery, and runtime setup diagnosis
uxl-performance-validation Cross-project correctness, benchmark, profiling, and speedup evidence workflow

The catalog manifest is skills.yaml. Human-readable review records live in skill-cards.

Placement Model

Use this repo as an incubator first. Once project maintainers review the skill contents, copy each project skill into the matching repository under skills/<skill-name>/, then mirror it back here with provenance in release metadata.

Recommended product ownership:

  • uxl-onednn -> uxlfoundation/oneDNN
  • uxl-onemath -> uxlfoundation/oneMath
  • uxl-onedal -> uxlfoundation/oneDAL
  • uxl-onetbb -> uxlfoundation/oneTBB
  • uxl-onedpl -> uxlfoundation/oneDPL
  • uxl-oneccl -> uxlfoundation/oneCCL
  • uxl-sycl-build-debug and uxl-performance-validation -> central uxlfoundation/skills

Development

Validate the catalog locally:

python scripts/validate_catalog.py
python scripts/run_evals.py --validate
python scripts/generate_agent_wrappers.py --check
python scripts/check_links.py --timeout 15 --retries 1
npx agnix@0.36.0 .

Validate an individual skill with Codex's skill creator helper:

python "$env:USERPROFILE\.codex\skills\.system\skill-creator\scripts\quick_validate.py" .\skills\uxl-onemath

Emit eval prompts for forward testing:

python scripts/run_evals.py --write-prompts eval-prompts

Score saved answers:

python scripts/run_evals.py --answers-dir eval-answers

Accepted answer paths are eval-answers/<skill>/<eval-id>.md or eval-answers/<skill>--<eval-id>.md.

Agent Tools

This repo includes instruction files for Codex/AGENTS-compatible agents, Claude Code, Gemini CLI, GitHub Copilot, Cursor, Continue, Aider, Windsurf, Cline, Devin-style workflows, and OpenHands-style repository skills. See agent-tooling.md.

The CI workflow also runs agent-sh/agnix against the repository agent configuration files using .agnix.toml.

Install and usage guidance lives in install-and-use.md. Forward-test workflow lives in forward-testing.md.

Regenerate tool wrappers after changing canonical guidance:

python scripts/generate_agent_wrappers.py

Then review the generated files and run python scripts/generate_agent_wrappers.py --check.

Release and Promotion

The repository is already public and released. Promotion criteria remain because individual skills are still incubating or pilot-quality until maintainers review them.

  • Catalog releases make a pin-able snapshot of this repository.
  • Skill promotion requires maintainer review, current source verification, forward-test evidence, and updated skill-card evidence before a skill can be marked reviewed or project-owned.

See release-and-promotion-policy.md and the release ledger in docs/releases.