Skip to content

AGENTS.md

Latest commit

 

History

History
59 lines (48 loc) · 2.85 KB

AGENTS.md

File metadata and controls

59 lines (48 loc) · 2.85 KB
Raw

AGENTS.md — layer-python-ml

Standalone candy repo for the python-ml layer — a CUDA-ready Python 3.13 ML environment (PyTorch, transformers, vLLM, llama.cpp). The candy lives in charly.yml at the repo root: the require: on layer-cuda, the nested candy: composition of llama-cpp, the env: block, the vLLM pip step, the check: assertions, and the embedded skill: entity projected into the marketplace corpus as /charly-languages:python-ml.

Canonical files:

  • charly.yml — the python-ml: candy entity and the python-ml-skill: skill entity.
  • pixi.toml / pixi.lock — the core ML Python environment.
  • .github/workflows/tag-on-merge.yml — CalVer tag + CHANGELOG/ on merge.
  • README.md — user overview only; never agent guidance.

Load these skills first (R0)

  • /charly-languages:python-ml — the owning skill. The Tier 2 environment-owner meta-layer, the pixi.toml it owns, the llama-cpp sub-candy composition, and the vLLM wheel install. Load before editing or troubleshooting the candy.
  • /charly-image:layer — the candy authoring reference (charly.yml schema, plan: step verbs incl. check:, the nested candy: composition list, service declarations). Load before editing any entity field or plan step.

Build / validate / test

  • The merge gate is the org-wide charly/pr-validator (required check validate / validate, defined in opencharly/.github); this repo has no per-repo candy gate and ships only .github/workflows/tag-on-merge.yml.
  • charly box validate at the repo root checks the manifest parses and validates.
  • The candy's plan: check: steps are the functional evidence: the pixi interpreter exists and PyTorch, transformers, and vLLM each import and print a version.
  • Regenerate pixi.lock whenever pixi.toml changes — the build installs with pixi install --frozen and fails loudly on a stale lock.

Modify this repo

  • Edit the python-ml: candy entity AND the python-ml-skill: skill entity in charly.yml together. The skill is the projected usage source, so a version or path change not mirrored in the skill leaves the corpus stale.
  • Keep the build order (pixi environment → llama-cpp → vLLM wheel) intact; the vLLM wheel is pip-installed after the pixi environment exists.
  • New behaviour claims belong in the plan: as an observable check: step, and in the skill body.

Landing

  • PR-only. Every change lands through a pull request; the org-required charly/pr-validator validates the diff and body and arms native auto-merge on PASS. Direct pushes to main are blocked.
  • History lives in CHANGELOG/ (written by tag-on-merge at merge time); the PR body IS the changelog.
  • The authoritative rulebook is the umbrella AGENTS.md in opencharly/opencharly and charly/AGENTS.md in the charly repo. Do not restate its rules here.