nemo-relay-plugin-observability

Files (9)

• BENCHMARK.md
• SKILL.md
• evals.json · evals/
• atif.md · references/
• atof.md · references/
• openinference.md · references/
• opentelemetry.md · references/
• skill-card.md
• skill.oms.sig

SKILL.md (first lines)

# Configure Observability Plugins

Start with one exporter managed by the built-in Observability plugin. This is
the default for reusable process configuration and the best first plugin for
most users because it makes Relay's captured activity visible.
Choose one proof output before layering additional telemetry destinations.

Use manual subscriber or exporter APIs only when a test, script, or application
needs direct control over registration names, collection windows, or flush
timing. Both paths consume the same canonical event stream.

## Select The Relay Version

Determine whether the application uses NeMo Relay 0.6 or 0.7 before proposing
configuration or binding APIs. Prefer the installed package version, lockfile,
or manifest. Ask the user when the version cannot be established; do not mix
the two surfaces in one example.

- **0.6** uses observability configuration version 2. OpenTelemetry and
  OpenInference are separate exporters with `OpenTelemetryConfig` /
  `OpenTelemetrySubscriber` and `OpenInferenceConfig` /
  `OpenInferenceSubscriber`.
- **0.7** uses observability configuration version 3. One typed OpenTelemetry
  exporter provides `full`, `gen_ai`, and `openinference` projections.

## Choose The Output

Select the output that best matches the user's immediate inspection target:

- **Console or custom event handling**
  Use a manual subscriber for short-lived in-process inspection.
- **Raw canonical lifecycle events**
  Use ATOF JSONL; read `references/atof.md`.
- **Portable execution trajectories**
  Use ATIF; read `references/atif.md`.
- **OTLP tracing**
  For 0.6, choose the separate OpenTelemetry or OpenInference exporter. For
  0.7, choose a typed OpenTelemetry endpoint (`full`, `gen_ai`, or
  `openinference`). Read `references/opentelemetry.md` and, for an
  OpenInference-aware backend, `references/openinference.md`.

Choose one output first and verify it before adding another. ATOF is the
default local proof because it preserves the raw event stream with the least
translation. Use synthetic, non-sensitive payloads for the first proof. Add and
verify sanitization before exporters receive production payloads, and never
display complete event records while validating an exporter.

## Embedded Event And Subscriber Model

Use this model when explaining how capture and export relate:

- NeMo Relay emits one canonical event stream from scopes, marks, managed tool
  calls, managed LLM calls, middleware, and manual lifecycle APIs.
- Subscribers consume events without defining the event model. Multiple
  subscribers can observe the same stream for logging, export, analytics, or
  diagnostics.
- Global subscribers remain active process-wide until removed.
- Scope-local subscribers are owned by one active scope and disappear when that
  scope closes.
- Plugin-installed subscribers are reusable, configuration-driven runtime
  components.
- Exporter-oriented subscribers preserve raw ATOF or translate the event stream
  into ATIF or the version-appropriate OpenTelemetry/OpenInference form.
- Event payloads reflect sanitized post-guardrail input and output when calls
  use managed helpers or manual lifecycle params provide those fields.
- LLM annotations follow the freshness rules:
  - Each owning agent scope starts fresh, and a `compaction` mark refreshes it.
  - The first subsequent LLM start retains complete annotation history. Later
    starts retain system instructions, the latest user message, and every
    following assistant or tool message.
  - When a request codec supplies an annotation, Relay applies the same
    event-only projection to provider-shaped event input without changing
    provider execution.
- Event fields include semantic input/output through the ATOF `data` field,
  typed profile data such as `model_name` and `tool_call_id`, and codec-provided
  annotated LLM request/response data for in-process subscribers and exporters.
- First-class skill tools and the requests to read a complete `SKILL.md`
  automatically emit `skill.load` marks under the tool span. The payload
  contains only `skill_name`; metadata records the load source and tool name.
  Partial reads do not count, and ambiguous slash-command expansions use the
  separate `skill.load.inferred` name. The eager mark remains present if tool
  execution later fails.

## Shared Lifecycle

1. Create the exporter or subscriber.
2. Register it with a unique name before the relevant scoped work.
3. Run NeMo Relay-instrumented work inside scopes.
4. Flush and deregister in the exporter-specific reference's documented order.
5. Shut it down when the process or subsystem is done.

## Binding Names

Use the names exported by the selected language binding and Relay version:

- Python 0.6: `nemo_relay.subscribers.register(...)`, `AtofExporter`,
  `AtifExporter`, `OpenTelemetrySubscriber`, and `OpenInferenceSubscriber`
- Python 0.7: the same registration and file exporters, plus
  `OpenTelemetrySubscriber` for all three typed projections
- Node.js follows the same 0.6/0.7 split through its root exports
- Rust: `nemo_relay::api::subscriber` and `nemo_relay::observability::*`
- Go: source-first wrappers expose equivalent register, exporter, and subscriber
  lifecycle methods

## Load A Reference When

Load only the reference required by the selected output:

- Load `references/atof.md` for raw JSONL events used in local debugging or
  offline inspection.
- Load `references/atif.md` for ATIF trajectories.
- Load `references/opentelemetry.md` for OTLP/OpenTelemetry traces.
- Load `references/openinference.md` for the standalone 0.6 OpenInference
  exporter or the 0.7 `openinference` OpenTelemetry projection.

## Use Another Skill When

Choose another skill when the task belongs to an adjacent workflow:

From NVIDIA/skills, Apache-2.0.

More from NVIDIA/skills