Skip to content

Prepare ModelOpt catalog onboarding: original six skills plus QAD - #858

Draft
chadvoegele wants to merge 4 commits into
NVIDIA:mainfrom
chadvoegele:skills/add-modelopt-qad
Draft

chadvoegele wants to merge 4 commits into
NVIDIA:mainfrom
chadvoegele:skills/add-modelopt-qad

Conversation

@chadvoegele

@chadvoegele chadvoegele commented Nov 10, 2026 •

Copy link
Copy Markdown

Chad's Agent

Onboarding type

  • New product onboarding (new components.d/model-optimizer.yml file)
  • Other

Continues #843 under Chad's fork, with the requested root paths and product-prefixed names. Source changes: NVIDIA/Model-Optimizer#3281.

Registers Zhiyu's original six workflows (PTQ, deployment, evaluation, comparison, recipe search, and speculative decoding) plus QAD. No additional companion skills are registered. Every catalog directory matches the source directory and name: frontmatter under skills/modelopt-*.

For new product onboarding — author affirmations

  • Skills cleared for open source release by the owning team, with all required approvals in place
  • License selected: Apache 2.0, as declared by the source repository and skill cards
  • No new license or new third-party component introduced beyond what the source repo already carries — owning-team confirmation pending
  • Source repo is public and under an NVIDIA-owned GitHub org
  • Root skills/ paths used for new entries

Reviewer checklist (OSS Skills PIC)

  • Author confirmations above are checked
  • Component entry valid and paths exist on the merged source branch
  • Skill frontmatter spec-compliant
  • No new license or third-party dependency introduced

All PRs

  • All commits signed off with DCO — final verification before review

Validation and readiness

All seven mappings were checked against the source PR checkout for matching names, eval datasets, draft cards, and standalone loading guidance. This is a replacement draft for #843; retain the original PR for context. Source changes must merge first. NVSkills CI enablement, live evaluation evidence, generated cards/signatures, and owning-team publication confirmations remain pending.

Edwardf0t1 and others added 3 commits November 10, 2026 19:47
Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>
Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>
Signed-off-by: Chad Voegele <cvoegele@nvidia.com>
Signed-off-by: Chad Voegele <cvoegele@nvidia.com>
@chadvoegele chadvoegele changed the title Onboard ten product-prefixed ModelOpt skills including QAD Prepare ModelOpt catalog onboarding: original six skills plus QAD Nov 10, 2026
chadvoegele added a commit to NVIDIA/Model-Optimizer that referenced this pull request Nov 12, 2026
Chad's Agent

### What does this PR do?

Type of change: documentation.

Moves canonical skills to root `skills/` and makes the repository root
the Claude and Codex plugin source. Moves plugin manifests and MCP
configuration to root, makes the six `.claude/agents/*.md` files
canonical, and explicitly lists those six individual files in the Claude
manifest. Removes the obsolete `plugins/modelopt` wrapper; preserves
`.agents/skills`, `.claude/skills`, and `.codex/agents` checkout
discovery. Skill names and workflow behavior are unchanged.

The original external `plugins/modelopt/skills` symlink resolved in a
checkout but was skipped by Codex's installer, producing zero installed
skills. A real root `skills/` directory inside the plugin root fixes the
packaging regression.

Reference: [NVIDIA/skills — Recommended Skill Directory
Path](https://github.com/NVIDIA/skills/blob/main/CONTRIBUTING.md#recommended-skill-directory-path):

> When publishing skills in a product repo, keep the source of truth in
`skills/` at the repo root.

The guide recommends copying or symlinking from this canonical location
into agent discovery paths.

### Usage

Marketplace installation remains `modelopt@modelopt`. Both marketplace
entries select `./`; plugin manifests are `.claude-plugin/plugin.json`
and `.codex-plugin/plugin.json`. Repository skill discovery paths remain
available. The old `plugins/modelopt` paths are removed.

### Testing

- Skill tests at this stage: 111 passed, 3 skipped for unavailable
optional dependencies.
- Pre-commit passed.
- Offline marketplace regression check: codex-cli 0.163.0
installs/discovers 16 skills on base `cf5645153a`, zero on the original
PR head `8303b52bcd`, and 16 with this fix.
- Real isolated marketplace installs and runtime discovery: Codex 16
cached skills; Claude 16 cached skills and six agents. MCP configuration
is retained; servers were not launched. Claude Code 2.1.228 and 2.1.322
were checked. Runtime discovery uses the installed cache from outside
the checkout.
- Installer smoke checks remain outside the PR; no new test files were
added. The existing agent synchronization test is updated for the moved
canonical files.
- No live agent evaluations, GPU training, or container builds were run.

### Before your PR is "Ready for review"

- Compatibility: marketplace name and checkout discovery remain
available. Plugin component paths move to the repository root.
- New PIP dependencies: none.
- Changelog: N/A, onboarding preparation.
- Claude review: pending.

### Stack / merge order

Part 1/5. Base: `main`. Merge bottom-up.
This five-PR stack supersedes #3264 and #3281. 

1. [Move canonical ModelOpt skills to repository
root](#3282)
2. [Prefix ModelOpt catalog skill
names](#3283)
3. [Support independently installed ModelOpt
skills](#3284)
4. [Move skill evaluation datasets to canonical
paths](#3285)
5. [Add draft skill cards and catalog eval
cases](#3286)

This is preparation only. NVSkills CI wiring, enablement, live
evaluation, and signing remain follow-up work. Catalog NVIDIA/skills#858
remains limited to the original six entries plus QAD.




<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
<!-- coderabbit-ai-generated:v1 -->

## Summary by CodeRabbit

* **New Features**
* Added six ModelOpt agents for model acquisition, evaluation,
performance benchmarking, quantization, recipe search, and distillation.
  * Added a ModelOpt MCP server connection to the plugin.
* **Chores**
* Consolidated agent skills in a shared location while keeping existing
skill paths connected.
* Updated plugin configuration and repository guidance to reflect the
shared skills and agent definitions.
* **Tests**
* Skill validation now runs against the shared skills, and
agent-definition checks reflect their updated location.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: Chad Voegele <cvoegele@nvidia.com>
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.

2 participants