agentleFS
NVIDIAKnown publisherNVIDIA/NeMo-Fabric/.agents/skills/contribute-adapter/SKILL.md

contribute-adapter

Skill73 starsChanged 59 days agoApache-2.0View on GitHub

Add or substantially change a first-party NVIDIA NeMo Fabric adapter, including repository packaging, discovery metadata, catalogs, CI wiring, and validation.

The file

We split this skill into the rules it gives an agent

11 rules. About 0 are common, 2 are distinctive, and 9 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 9

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Do not use this skill for a consumer that selects an existing adapter.

    line 16
  2. Very specific:

    Use karpathy-guidelines to keep the change scoped, python-tests for test design, maintain-packaging for package or dependency changes, …

    line 21
  3. Distinctive:

    Follow these repository-specific requirements after applying the public skill:

    line 28
  4. Very specific:

    Place a Python adapter under adapters/python/<name>/ with LICENSE -> ../../../LICENSE, README.md, <name>.fabric-adapter.json, Python …

    line 30
  5. Very specific:

    Add one canonical root extra that delegates to the matching leaf adapter and its harness extra.

    line 40
  6. Very specific:

    Add the package to the root adapter-test dependency group, [tool.uv.sources], python_projects in justfile, applicable catalogs, and CI …

    line 44
  7. Very specific:

    Update examples/code_review_agent/ and examples/harbor/calculator/ to support the new adapter.

    line 48
  8. Very specific:

    If the new adapter provides a coding harness, update examples/harbor/swebench/ to support it.

    line 50
  9. Distinctive:

    Keep descriptor claims, implementation, focused tests, public documentation, catalog entries, and packaged metadata synchronized.

    line 56
  10. Very specific:

    Use validate-change to select the complete matrix.

    line 87
  11. Very specific:

    Run python3 scripts/ci/check_npm_bootstrap.py when introducing a public npm package.

    line 113

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Contribute a First-Party NVIDIA NeMo Fabric Adapter

Use the public nemo-fabric-build-adapter skill for adapter-contract semantics, descriptor design, configuration mapping, lifecycle behavior, and conformance evidence. This maintainer skill adds only the repository integration required for adapters shipped by NVIDIA NeMo Fabric.

Do not use this skill for a consumer that selects an existing adapter. Use the consumer nemo-fabric-integrate skill instead.

Companion Guidance

Use karpathy-guidelines to keep the change scoped, python-tests for test design, maintain-packaging for package or dependency changes, contribute-docs and review-doc-style for public text, validate-change for the validation matrix, and prepare-pr for review handoff.

Repository Integration

Follow these repository-specific requirements after applying the public skill:

  1. Place a Python adapter under adapters/python/<name>/ with LICENSE -> ../../../LICENSE, README.md, <name>.fabric-adapter.json, Python package and lock files, a source entry point, and focused tests. Place TypeScript packages under adapters/typescript/<name>/ and wire them into that npm workspace.
  2. Give each Python leaf adapter a small base installation, a harness extra for package-installable target packages, and a full extra for package-installable integrations. The Hermes adapter is the sole source-only exception and omits harness. Add a relay extra only when the adapter imports NVIDIA NeMo Relay Python APIs.
  3. Add one canonical root extra that delegates to the matching leaf adapter and its harness extra. The Hermes root extra delegates to the bare adapter because users install Hermes Agent separately from source. Keep nemo-fabric-runtime an exact-version, unconditional root dependency.
  4. Add the package to the root adapter-test dependency group, [tool.uv.sources], python_projects in justfile, applicable catalogs, and CI enumerations. Ship its descriptor under share/nemo-fabric/adapters/<name>.
  5. Update examples/code_review_agent/ and examples/harbor/calculator/ to support the new adapter.
  6. If the new adapter provides a coding harness, update examples/harbor/swebench/ to support it. Skip this step for adapters with non-coding harnesses.
  7. Regenerate lockfiles and inspect the root and leaf wheel metadata. Verify root-to-leaf delegation and every published leaf extra.

Keep descriptor claims, implementation, focused tests, public documentation, catalog entries, and packaged metadata synchronized. Start with the narrowest truthful capability set.

For usage changes, verify invocation-local accounting through the adapter contract and consumer SDK, including warm-session deltas, unknown counters, cache semantics, and unsuccessful results. Keep generated schemas and language bindings in parity. Preserve the closed fabric.adapter/v1alpha2 AgentUsage shape; carry detailed cache-write, reasoning, peak request, and per-model data in the descriptor-declared usage.extensions["nemo_fabric.detailed_token_usage.v1"] object until a new adapter contract revision negotiates those fields. Consumers treat any detailed field for an input or output dimension as selecting the detailed counters for that dimension, so require complete, internally consistent counters instead of relying on the top-level view as a fallback.

For instructions.system, keep config.system_instruction_modes, planning behavior, direct adapter validation, and target-native composition synchronized. New descriptors must declare their exact replace and append support rather than relying on the legacy omitted-value behavior.

For software evidence, declare selected Python packages in the adapter's PROVENANCE_PACKAGES or configure the TypeScript lifecycle host's software selection. Keep this list specific to the executing adapter, harness and framework. Forward versions already observed at startup with the shared passive capture helper. Do not add version probes or infer a remote service's packages from its local client. Test disabled collection, original capture times and missing metadata without changing the invocation outcome.

Repository Evidence

In addition to the evidence required by the public skill, include:

  • A subprocess test of the packaged entry point.
  • Exact descriptor assertions for every claimed capability.
  • A credential-free fixture that exercises plan, doctor, and run.
  • Wheel inspection when package data, dependencies, or extras change.
  • Deterministic CI coverage; keep credentialed live-target tests opt-in.

Validation

Use validate-change to select the complete matrix. The common adapter checks are:

uv sync --group adapter-tests
uv run --no-sync pytest tests/adapters/test_<name>*.py
just test-python
just lock-python && just wheels
just schemas
cargo fmt --all -- --check && just test-rust
just docs
uv run pre-commit run --all-files --show-diff-on-failure
git diff --check

References

  • Public contract: docs/adapter-contract/ and schemas/adapter-contract/.
  • Descriptor schema: schemas/adapter-contract/adapter-descriptor.schema.json.
  • Repository packaging: root pyproject.toml, justfile, adapter catalogs, and CI workflows.
  • Shared first-party Python host patterns: adapters/python/common/ and the closest adapter with the same target boundary.

New Public npm Packages

Run python3 scripts/ci/check_npm_bootstrap.py when introducing a public npm package. The check discovers the contract and public adapter workspaces from manifests; add the adapter to the npm workspace. Separately wire it into release version stamping, build/test coverage, and the publication list before enabling uploads. Discovery alone does not enable publication. Complete the npm maintainer bootstrap and authenticated trusted-publisher readback in RELEASING.md before release. Public existence alone cannot verify trusted publishing. If the PR check reports HTTP 404, ask an authorized npm maintainer to bootstrap the exact manifest name, then rerun the check. Do not add CI write tokens or manually publish supported versions.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.