Repository navigation
Add the NVSkills CI request workflow - #109
kumo-isaac wants to merge 1 commit into
Conversation
/nvskills-ci is an issue_comment trigger, so GitHub runs it from the copy of this workflow on main, never from a pull request's. #82 adds this file beside the skill changes it is meant to sign, so the first /nvskills-ci on #82 cannot start until the file is on main. This lands it on its own, byte for byte as #82 carries it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: igutierrez <igutierrez@nvidia.com>
|
No actionable comments were generated in the recent review. Walkthrough
Change: Other Merge Risk:
|
| # NVSkills CI evaluates and signs changes under skills/. The NVIDIA skills | ||
| # catalog (github.com/NVIDIA/skills) syncs these skills from main every hour, | ||
| # and accepts only files that match the signature (skill.oms.sig) NVSkills CI | ||
| # writes into the PR. This is the catalog's request workflow, unchanged except |
There was a problem hiding this comment.
P2: Generated signing documents fail the skills inventory test
When signing adds skill-card.md or BENCHMARK.md under a skill folder, normal CI raises KeyError: the skills tests collect every Markdown file but require each filename in their table inventory. Adding a plain BENCHMARK.md reproduces this. Could we exclude the known generated documents from authored-skill checks, or validate them separately, before running the signing flow on #82?
There was a problem hiding this comment.
Fixed in #110 (a0fe6e187, on v2-candidate). tests/test_skills.py now leaves out the catalog's own files (CATALOG_FILES = {'skill-card.md', 'BENCHMARK.md'}) when it collects the documents it checks. They follow the catalog's template, NVSkills CI writes BENCHMARK.md itself, and no skill links to either. skill.oms.sig isn't Markdown, so nothing collected it.
To reproduce your scenario, I took the files a real signing commit added on kumo-engine (NVIDIA/kumo-engine@9680c1ac: BENCHMARK.md, the rewritten skill-card.md and skill.oms.sig), dropped them into both skill directories on #110's branch and staged them:
pytest tests/test_skills.py: 232 passed, 18 skipped (the live checks). NoKeyError.scripts/check_repo_hygiene.py all: clean, once the borrowed card names this repo (a real signing run here writes this repo's URL).ruff check .: clean.
| # SPDX-License-Identifier: Apache-2.0 | ||
|
|
||
| # NVSkills CI evaluates and signs changes under skills/. The NVIDIA skills | ||
| # catalog (github.com/NVIDIA/skills) syncs these skills from main every hour, |
There was a problem hiding this comment.
P2: Question: preserve the documentation reference when publishing skills
Question: before catalog onboarding, could we make the tabular skill’s ../../docs/reference/prediction-output.md link portable? The catalog copies each skill directory, leaving this explanation-schema document behind; an isolated copy reproduces the missing target. Assuming onboarding uses that layout, an absolute repository URL or an included reference would preserve access for readers of the published skill.
There was a problem hiding this comment.
Done in #110 (a0fe6e187, on v2-candidate). The three links that left a skill now use the repository URL:
sdm-tabular/SKILL.mdandsdm-relational/references/beyond-one-answer.md→https://github.com/NVIDIA/structured-data-models-client/blob/main/docs/reference/prediction-output.mdsdm-relational/references/in-process.md→.../blob/main/docs/get-started/quickstart.md
A new test, test_a_skill_links_only_to_skills, fails any relative link that leads out of a skill directory, so this can't come back. I also copied both skill directories out on their own, the way the catalog lays them out: all 29 remaining relative links resolve.
One thing to decide: I pointed at blob/main because that's where these docs will live after 2.0 and it's the form the docs already use. Until v2-candidate merges, though, main's prediction-output.md is the 1.x version: it documents gradients but not result.explanations.columns or the sdm_gradient_v1 details. Linking blob/v2-candidate would be accurate today but would break when the branch is retired. Happy to switch if you'd rather.
Problem
/nvskills-cistarts an NVSkills CI signing run. It is anissue_commenttrigger, and GitHub runsissue_commentworkflows only from the copy onmain, never from a pull request's.#82 adds this workflow together with the skill files it is meant to sign. Until the workflow is on
main,/nvskills-cion #82 starts nothing, and the skills cannot be signed or published to the NVIDIA skills catalog. Kumo Engine split it out the same way in NVIDIA/kumo-engine#371.Change
Adds
.github/workflows/request-nvskills-ci.ymland nothing else. It is byte-identical to the copy on #82, so #82 merges cleanly over it.team-request.ymltob707bbeinstead of@main, the same pin Kumo Engine uses.team-request.ymlis unchanged between that commit and the catalog's currentmain.CODEOWNERS(*) already requires owner review for workflow changes, which NVSkills CI onboarding asks for.skills/the workflow skips. On one that does, itsrequire-nvskills-cijob fails until someone comments/nvskills-ciand the run reports back. Prepare the agent skills for the NVIDIA skills catalog #82 shows that today.What it needs before it does anything
nv-skills-ciGitHub App on this repository and saves the repository secretNVSKILLS_CI_DISPATCH_TOKEN.config/onboarded-repositories.jsonin NVIDIA/nvskills-ci, as they did for Kumo Engine in NVIDIA/nvskills-ci#112./nvskills-cineedsmaintainoradminon this repository.Validation
cmpagainst Prepare the agent skills for the NVIDIA skills catalog #82's copy: identicalrequest-nvskills-ci.ymlonmain, apart from the pinteam-request.ymlatb707bbeis identical to the catalog'smainpython scripts/check_repo_hygiene.py all: passes/nvskills-cion Prepare the agent skills for the NVIDIA skills catalog #82, once the App, the secret and the onboarding entry are in place🤖 Generated with Claude Code