Skip to content

Prepare the agent skills for the NVIDIA skills catalog (v2-candidate) - #110

Open
kumo-isaac wants to merge 1 commit into
v2-candidatefrom
igutierrez/nvskills-onboarding-v2
Open

kumo-isaac wants to merge 1 commit into
v2-candidatefrom
igutierrez/nvskills-onboarding-v2

Conversation

@kumo-isaac

Copy link
Copy Markdown
Collaborator

This is #82 rebuilt on v2-candidate, as asked in review there. It prepares skills/ for NVSkills CI and the NVIDIA skills catalog.

What changes

  • One directory per skill, named for the skill. skills/tabular → skills/sdm-tabular and skills/relational → skills/sdm-relational, matching the frontmatter names these skills already use on this branch. The catalog copies each skill directory to one named for the skill, so sibling links (../sdm-tabular/...) only resolve when the two match. tests/test_skills.py, scripts/check_repo_hygiene.py, skills/README.md and docs/about/architecture.md follow the rename.
  • No relative link leaves a skill. The catalog copies the skill directory and nothing around it. Three links into docs/ are now GitHub URLs (blob/main, the same form the repo's docs already use). The new test_a_skill_links_only_to_skills holds every skill to this.
  • A skill card and evals per skill. skill-card.md is the catalog's governance card. evals/evals.json is what NVSkills CI's Tier 3 runs. The evals are written against these skills as they stand on v2-candidate, not main:
    • kumo-relational needs no extra.
    • The client imposes no size limits of its own.
    • Explanation keys are unique IDs, otherwise row_index.
    • New: one eval per skill for SDMClient.in_process().
  • skill-card.md and BENCHMARK.md are excluded from the markdown checks. They follow the catalog's template, and NVSkills CI writes BENCHMARK.md itself.
  • .github/workflows/request-nvskills-ci.yml. The catalog's request workflow, pinned to a SHA, identical to Add the NVSkills CI request workflow #109. The require-nvskills-ci gate runs from this copy on PRs into this branch. /nvskills-ci comments only work once the same file is on main (the default branch), which is what Add the NVSkills CI request workflow #109 is for.

Before NVSkills CI can sign this

  1. Add the NVSkills CI request workflow #109 merges to main.
  2. Repo secret NVSKILLS_CI_DISPATCH_TOKEN is set.
  3. NVIDIA/nvskills-ci lists this repo with allowed_base_refs: ["main", "v2-candidate"].
  4. A maintainer comments /nvskills-ci here.

Until then the require-nvskills-ci check fails, as it does on #82.

Open question for the skill authors

sdm-tabular/SKILL.md says a Request holds the context to no size of its own (#93), while sdm-tabular/references/in-process.md says "this client's own [limits] still are: a context above 10,000 rows is refused". I found no such check in the client, so the in-process sentence looks stale. I left both as written. Eval t03 doesn't depend on which one is right.

Replaces #82.

🤖 Generated with Claude Code

The catalog copies each skill directory to one named for the skill and
nothing around it, so each skill now lives in a directory named as its
frontmatter names it (sdm-tabular, sdm-relational), links to its sibling
by that name, and reaches the repository's docs by URL rather than by a
relative path that leaves skills/. A test holds every skill to that.

Each skill gains the catalog's governance card and evals for NVSkills CI.
The evals follow these skills as they stand on v2-candidate: no extra
for kumo-relational, no client-side size limits, unique-id keys for
explanations, and one eval each for running the model in this process.

The request workflow is the catalog's, pinned to a SHA, the same file
as on main's PR, so the NVSkills CI gate runs on PRs into this branch.

This is the v2-candidate port of #82.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: igutierrez <igutierrez@nvidia.com>
@kumo-isaac

Copy link
Copy Markdown
Collaborator Author

Test results: a0fe6e187

Local build of this branch (macOS, Python 3.12, Rust 1.97.0, editable installs of the api, client and connectors packages per CONTRIBUTING.md):

Check Result
pytest tests/test_skills.py 232 passed, 18 skipped
test_a_skill_links_only_to_skills, test_a_skill_announces_itself 10 passed
ruff check . clean
ruff format --check . clean (557 files)
scripts/check_repo_hygiene.py all clean
evals/evals.json (both skills) valid JSON

The 18 skips are the live checks, which need SDM_RELATIONAL_URL, SDM_TABULAR_URL or SDM_IN_PROCESS=1. They weren't run.

NVSkills CI itself hasn't run. It needs #109 on main and the dispatch secret first.

@coderabbitai

coderabbitai Bot commented Nov 18, 2026

Copy link
Copy Markdown

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

Run configuration

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • Trigger review
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant