nemo-relay-plugin-build

Files (5)

• BENCHMARK.md
• SKILL.md
• evals.json · evals/
• skill-card.md
• skill.oms.sig

SKILL.md (first lines)

# Build a Plugin

Use this skill when a user wants to package reusable NeMo Relay runtime behavior
behind plugin configuration.
Keep reusable plugin behavior separate from one-off application startup code.

## Use This When

Use this skill when the behavior should be activated by shared config and reused
across applications, teams, or process startup paths.

Common cases:

- Register subscribers, guardrails, intercepts, or a small bundle of related
  runtime behavior.
- Validate operator-supplied config before changing runtime behavior.
- Give reusable behavior a stable plugin `kind` and activation lifecycle.
- Package behavior that should be enabled, disabled, or rolled out through
  plugin config rather than repeated application startup code.

## Do Not Use This When

Do not build a plugin when a narrower NeMo Relay surface is enough:

- One request or tenant needs temporary behavior -> use scope-local middleware.
- The user only needs first-time scopes, tool calls, or LLM calls ->
  `nemo-relay-instrument-calls`.
- The user only needs to choose an exporter path ->
  `nemo-relay-plugin-observability`.
- The behavior depends on live callables, provider clients, file handles,
  credentials, or framework objects inside config.

## Choose A Delivery Model

Choose the delivery model before designing config or registration:

- **Embedded component:** Use the shared plugin document when behavior ships
  with the Relay host application. Follow the embedded component model below.
- **Discoverable dynamic package:** Use a `relay-plugin.toml` manifest when the
  plugin ships independently of the host. Use `rust_dynamic` for a trusted
  in-process native Rust library, or `worker` for a local `grpc-v1` worker.
  Read the public native dynamic-plugin or gRPC worker guide before designing
  the package boundary. Do not reproduce native ABI or worker-protocol details
  from memory.

Native plugins are trusted C-ABI extensions that run in the Relay process and
are not sandboxed. Worker plugins provide process isolation, not a security
sandbox.

## Embedded Component Model

- Plugins package reusable process-level behavior.
- A plugin exposes a stable `kind` string and receives component-local config
  from a shared plugin document.
- Plugin config must be JSON-compatible across Rust, Python, Node.js, files,
  tests, and deployment systems.
- Validation is deterministic and side-effect free. It inspects config and
  returns structured diagnostics before runtime behavior changes.
- Registration runs after validation and installs real behavior through
  `PluginContext`, such as subscribers, guardrails, request intercepts,
  execution intercepts, or stream execution intercepts.
- `PluginContext` gives the plugin system enough ownership to qualify runtime
  names and roll back partial setup when activation fails.
- Disabled components should still validate when possible so operators can find
  config problems before rollout.

## Default Path

1. Decide whether a plugin is actually needed. Prefer direct instrumentation or
   scope-local behavior when the use case is not reusable process-level
   behavior.
2. Choose an embedded component or a discoverable dynamic package. For a
   dynamic package, establish the manifest-backed native or worker boundary
   before implementing behavior.
3. For native dynamic plugins, determine the target Relay version from the
   package constraint, lockfile, or deployment target before choosing callback
   APIs.
4. Pick one first runtime surface: subscriber-oriented export, sanitize
   guardrail, conditional guardrail, request intercept, execution intercept, or
   stream execution intercept.
5. Choose a stable plugin `kind` and the smallest JSON-compatible config shape.
6. Define diagnostics for missing fields, unsupported values, unknown fields,
   unsafe config, and invalid field combinations.
7. Validate config before initialization. Validation must not open network
   connections, create clients, register middleware, or mutate process state.
8. If validation returns error diagnostics, return them and stop without
   initialization or registration.
9. Register runtime behavior through `PluginContext`, not by hand-registering
   global behavior inside application startup.
10. Test activation, disabled components, validation failures, and registration
   failure rollback.
11. Document how to enable the plugin, what config fields are supported, and how
   to roll back the component.
12. For a dynamic plugin that should provide structured fields in
   `nemo-relay plugins edit`, declare the `config_schema` capability and
   reference a local Draft 7 or Draft 2020-12 JSON Schema file from
   `[config_schema].path` in `relay-plugin.toml`. Schema-less plugins remain
   editable as raw JSON objects.

## Config Shape

The top-level plugin document contains `version`, `components`, and `policy`.
Each component supplies the plugin `kind`, `enabled`, and component-local
`config`:

```json
{
  "version": 1,
  "components": [
    {
      "kind": "redaction-policy",
      "enabled": true,
      "config": {
        "preset": "strict"
      }
    }
  ],
  "policy": {
    "unknown_component": "warn",

From NVIDIA/skills, Apache-2.0.

More from NVIDIA/skills