Refresh Data Designer Skill

The user-facing skill at skills/nemo-data-designer-plugin/ is derived from the upstream skill bundle at https://github.com/NVIDIA-NeMo/DataDesigner (path skills/data-designer/).

This skill is for plugin contributors. End users do not run it — they get the already-vendored output.

The governing principle

The shipped skill describes NeMo Helix execution only. It is not a copy of the upstream skill with caveats bolted on.

Upstream's skill documents the standalone library, where models come from a local YAML registry, seed data is a file on disk, and persona locales are downloaded to ~/.data-designer/managed-assets/. None of that is true here. Data Designer on NeMo Helix resolves inference through Inference Gateway, seed data through HuggingFace or the Files service, and persona data through system workspace filesets.

Earlier versions of this playbook copied upstream wholesale and tried to teach agents the delta in a separate additions file. That failed: agents followed the standalone-library instructions that were still sitting in the bundle. Do not reintroduce that pattern. When upstream content contradicts platform behavior, replace it, don't annotate it.

When to use

• After bumping the data-designer==X.Y.Z pin. The data-designer-version-bump skill (.agents/skills/data-designer-version-bump/) runs this as its Step 8.
• When upstream ships skill changes that should land here.
• When auditing drift between the vendored copy and upstream.

File ownership

Refreshing is not a whole-tree overwrite. Each file is either tracked from upstream or owned by this plugin:

For plugin-owned files, "review only" means: read upstream's version to see whether it documents a new capability worth covering, then write that capability in platform terms yourself. Never paste upstream's text in.

Procedure

Step 1 — Determine the pinned version

grep -E '"data-designer==' plugins/nemo-data-designer/pyproject.toml | head -1

Capture the version string (e.g., X.Y.Z); this is $VERSION for the rest of the procedure.

Step 2 — Fetch the upstream bundle to a temp directory

gh api "repos/NVIDIA-NeMo/DataDesigner/git/trees/v$VERSION?recursive=1" \
  | jq -r '.tree[] | select(.path | startswith("skills/data-designer/")) | select(.type=="blob") | .path'

Fetch each path from https://raw.githubusercontent.com/NVIDIA-NeMo/DataDesigner/v$VERSION/<path> into a scratch directory. Stage it — do not write into skills/nemo-data-designer-plugin/ yet. You are diffing against the shipped bundle, not replacing it.

Step 3 — Diff and triage

Diff each staged upstream file against its shipped counterpart. For every difference, decide:

• A new or changed capability (new column type, new sampler, new flag) → adopt it, translated to platform terms.
• Standalone-library mechanics (local paths, ~/.data-designer/, the local model registry, artifact folders on disk) → drop it. It does not apply here.
• Cosmetic upstream churn → take it if it costs nothing.

If upstream adds a file with no platform meaning, do not vendor it. If upstream deletes a file the plugin still owns, keep the plugin's.

Step 4 — Apply adaptations to upstream-tracked files

Prefix CLI invocations. Every shell-shaped invocation of the data-designer binary becomes nemo data-designer …. Subcommand names, flags, and positional args are otherwise identical. Do not rewrite prose mentions of the project name ("the Data Designer library", "Data Designer columns") — only shell-shaped invocations.

create takes no --dataset-name or --artifact-path. Artifacts live at a Jobs-service-managed path, so there is no local folder to name or relocate. Strip those flags and their values; keep others like --num-records.

Resolve-CLI-command step. Upstream's workflows open by detecting the standalone data-designer binary. Replace with nemo detection:

1. **Resolve CLI command** — Run `command -v nemo 2>/dev/null || (test -x .venv/bin/nemo && realpath .venv/bin/nemo) || echo CLI_NOT_FOUND`.
  - If the output is a path, use it in place of `nemo` in every `nemo …` invocation in this workflow — including commands outside the `data-designer` group, such as `<path> inference providers list`.
  - If the output is `CLI_NOT_FOUND`, STOP and follow the Troubleshooting section in SKILL.md. Do not continue to the next step.

Learn step. Upstream tells the agent to stop when agent context lists no model aliases, and to trust its output generally. Both are wrong here. agent context prints three sections that describe a standalone local install — Model Aliases (a local YAML registry), Persona Datasets (an installed column reflecting local downloads), and Commands (unprefixed data-designer …). Its config_root path and type tables are correct and are the reason to run it at all. Replace the stop-on-no-aliases line with an instruction to read schemas from config_root and ignore those three sections, pointing at references/platform-execution.md.

Clarify / Infer step. Where upstream picks among locally configured aliases, have the agent run nemo inference providers list and choose an Inference Gateway provider instead.

Output Template. Preserve the plugin's template shape:

• model_configs=[dd.ModelConfig(...)] declared inline in the DataDesignerConfigBuilder(...) constructor, with a comment pointing at nemo inference providers list. Upstream's bare DataDesignerConfigBuilder() produces configs that cannot run here.
• The commented seed line uses FilesetFileSeedSource, imported from data_designer_nemo.fileset_file_seed_source, not dd.LocalFileSeedSource.
• The PEP 723 dependency list includes data-designer-nemo for the seed source import.

Rules section. Keep the bullet pointing at references/platform-execution.md, and the one routing retrieval SDG to references/retrieval-sdg.md.

Step 5 — Write the bundle

Write adapted upstream-tracked files to skills/nemo-data-designer-plugin/, mirroring the upstream tree (workflows/, references/, scripts/). Leave plugin-owned files alone unless Step 3 turned up a capability change.

Step 6 — Verify

Resolve every failure before considering the refresh complete.

No untranslated CLI invocations.

SKILLS=skills/nemo-data-designer-plugin
grep -RnE '(\s|`)data-designer\s' "$SKILLS" | grep -v 'nemo data-designer'

Clean run = no output. Prose mentions are fine; CLI-shaped invocations are not.

Every referenced subcommand resolves.

grep -RhoE 'nemo data-designer( [a-z][a-z-]*)+' "$SKILLS" \
  | sort -u \
  | while read -r cmd; do
      eval "$cmd --help" > /dev/null 2>&1 && echo "OK: $cmd" || echo "MISSING: $cmd"
    done

Any MISSING: means upstream changed the command surface or nemo data-designer drifted — investigate before merging. In particular, personas exposes only make-fileset; a personas download reference is always a bug.

No standalone-library guidance leaked back in.

grep -RniE '~/\.data-designer|managed-assets|LocalFileSeedSource|personas download|data-designer config' "$SKILLS"

Every hit must be either absent or an explicit statement that the thing does not apply on NeMo Helix. A hit that reads as instruction is a failure.

Non-obvious claims are checked against the source, not against upstream's docs:

• Supported seed types: packages/data_designer_nemo/src/data_designer_nemo/seed.py
• Persona locales and fileset naming: packages/data_designer_nemo/src/data_designer_nemo/nemotron_personas.py
• CLI surface: plugins/nemo-data-designer/src/nemo_data_designer_plugin/cli/

Notes

• Upstream may add files under skills/data-designer/. Pick them up via the Step 2 tree enumeration, then triage per Step 3 — enumeration is not adoption.
• Don't edit references/preview-review.md beyond CLI prefixing; it is genuinely upstream's.
• scripts/get_person_object_schema.py diverges from upstream on purpose: it reads the locale's system fileset through the plugin's FilesetFileSystem and fails with instructions to run personas make-fileset when that fileset is missing. Upstream's version reads a local download path that never exists here.