[andreatnvidia]

Summary

Invocation guidance in argument-hint is missing from the standard skill metadata that other harnesses load. Move it into natural-language descriptions across the seven affected skills, following the Agent Skills specification.

Carries forward @janEbert's original contribution in #1006, based on his signed commit 963a7145. The maintainer commit credits him as co-author, retains his existing DCO sign-off and adds my own.

Related Issue

Closes #1010

Changes

• Move argument guidance into descriptions for commit, create-pr, review-code, search-docs, search-github, update-pr and data-designer.

• Preserve quoted literals such as "--title", all skill bodies and current main's invocation controls, including model invocation for review-code.

• Keep the useful Claude Code extensions in place. Claude Code documents argument-hint as an autocomplete hint; the descriptions now carry the guidance across harnesses.

• Add evaluation-only CLI installation and an offline model alias through evals/config.yml, so the script-writing test has its prerequisites. The fixture uses a loopback endpoint and no inference credentials.

Testing

• Safely parsed all seven skill frontmatters and verified that bodies and remaining metadata match current main, including quoted "--title" and invocation controls.
• Ran Ruff lint/format and applicable pre-commit hooks; restored the unrelated generated notebook touched by Ruff. git diff --check passed. Validation used current main's existing environment because local uv is below the workspace's required version.
• Cold-tested the full evaluation setup and CLI agent context as uid 1000 with a login shell, empty Python loader variables, and /tmp mounted noexec. The setup uses the image's Python/pip to create an isolated environment without ensurepip; it completed in 31 seconds within Harbor's 120-second setup limit.
• GitHub unit/E2E tests and lint checks passed on f023dce0. NVSkills Tier 3 also passed: all four agent trials completed, and both Claude Code and Codex passed every evaluation dimension. The content gate passed. The sandbox image must provide a compatible Python, the venv module and pip 22.3+. Local unit/E2E tests were not rerun for this skill metadata and evaluation setup change.

Checklist

• Follows commit message conventions
• Commits are signed off (DCO)
• Linked issue is open and labeled triaged
• Architecture docs: not applicable

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge.

Summary

Moves invocation guidance into descriptions across seven skills while preserving their bodies and invocation controls.

• Skill descriptions tell agents what invocation arguments mean.
• The data-designer evaluation has a local CLI setup and refreshed results.

_{Reviews (10) · Last reviewed commit: "Attach NVSkills validation signatures" · Reviewed by Greptile}

[svcnvidia-nemo-ci]

No merge-blocking defects found across the seven changed files. The descriptions match the existing argument contracts, including PR-number and branch review selection, the optional title update, and the dataset-description argument. Skill bodies and remaining invocation controls are preserved.

Validation was static only; no YAML parser, tests, or skill workflows were executed. Two delegated reports were unavailable despite completed statuses; their assigned files and directly affected contracts were independently inspected.

Note: formal approval is not available for this request, so this review is posted as a comment only.

Review generated by codex. This was a full review. Request a full review with /review reset=true.

[andreatnvidia]

/nvskills-ci

[andreatnvidia]

/nvskills-ci

[andreatnvidia]

/nvskills-ci

[andreatnvidia]

/nvskills-ci

[andreatnvidia]

/nvskills-ci

[andreatnvidia]

/nvskills-ci

[andreatnvidia]

/nvskills-ci

[andreatnvidia]

/nvskills-ci