NeMo Relay Observability Plugin
Configure NeMo Relay event capture and telemetry exports.
The skill explicitly recommends synthetic non-sensitive payloads, sanitization before export, runtime credential injection, TLS verification, and redacted diagnostics; it also describes subscriber scope and lifecycle. Six points are deducted because external transmission does not require explicit user confirmation, rollback or isolation is not specified, and remote sinks are supported by default.
It documents a shared lifecycle, flush/shutdown behavior, transport constraints, common failures, and troubleshooting checks. Because this was not executed and the skill lacks dedicated reproducible tests, complete binding examples, and detailed abnormal-input feedback, the static score is conservatively limited to 9.
The audience, five output paths, language bindings, and adjacent-skill boundaries are relatively clear. Five points are deducted because the complete input/output contract, non-fit ranges, Chinese-language support, and mainland-China network reachability are not addressed.
Information is layered across the main file and four on-demand references, with license, author, version, evaluation report, related skills, and some risk guidance. Five points are deducted for missing recommended Instructions and Examples sections, the author-format finding, and an insufficiently explicit maintenance owner/update path.
The material covers output selection, registration, execution, export, and verification, so it can guide the core observability task. The score is capped at 7 because API details outside the selected references are incomplete and static review cannot establish directly runnable results.
The fixed revision, reference documents, benchmark report, and 20-task evaluation summary provide some auditability. The score is capped at 5 because no independently reproducible execution artifacts or skill-specific test suite/CI coverage are supplied.
- Remote stream, OTLP, or OpenInference export may contain business inputs and outputs; confirm recipients, redaction scope, and retention policy before production use.
- Complete installation guidance, compatibility matrices, and runnable per-language examples are absent; verify the actual NeMo Relay binding APIs and transport support.
- The benchmark reports a pass, but its results cannot be independently reproduced from the supplied static material and do not replace environment validation.
- Mainland-China network reachability is not documented; remote observability backends require separate validation of networking, TLS, and authentication.
What does this skill do, and when should you use it?
This skill guides users through NeMo Relay observability with the built-in Observability plugin, subscribers, and exporters. It covers raw ATOF JSONL events, ATIF trajectories, OpenTelemetry traces, OpenInference export, and custom event handling. It recommends verifying one output before adding more telemetry destinations and calls for sanitization before production export. It fits developers who need to inspect, record, analyze, or export Relay’s canonical event stream.
Guides creation, registration, flushing, deregistration, and shutdown of subscribers and exporters; helps choose among ATOF, ATIF, OpenTelemetry, OpenInference, and manual subscription; explains global, scope-local, and plugin-installed subscriber lifecycles; documents the event stream, LLM annotation behavior, skill-load marks, ATOF data fields, and typed profile data; and directs the user to references/atof.md, references/atif.md, references/opentelemetry.md, or references/openinference.md for the selected output.
- A NeMo Relay developer needs short-lived, in-process console inspection or custom event handling and uses a manual subscriber.
- A debugger needs raw canonical lifecycle events for local or offline inspection and chooses ATOF JSONL.
- A team building portable execution trajectories chooses ATIF.
- An engineer with OTLP tracing infrastructure needs general-purpose trace export and chooses OpenTelemetry.
- A team using an OpenInference-aware backend needs semantic trace export and chooses OpenInference.
What are this skill's strengths and limitations?
- Supports ATOF, ATIF, OpenTelemetry, and OpenInference outputs.
- Explains lifecycle behavior for plugins, global subscribers, and scope-local subscribers.
- Recommends synthetic, non-sensitive validation and sanitization before production export.
- Lists relevant binding names for Python, Node.js, Rust, and Go.
- The supplied source does not include concrete configuration examples or full code from the reference files.
- Supported NeMo Relay versions, runtime versions, and platform coverage are not specified.
- The supplied material does not show a test suite or test results for this individual skill.
- It configures observability only; uninstrumented calls require other NeMo Relay skills.
How do you install this skill?
Install the target skill with the skills CLI command supported by the repository README: npx skills add nvidia/skills --skill nemo-relay-plugin-observability --yes. The README does not document a standalone package installation and says users do not need to clone the repository or copy skill folders manually.
How do you use this skill?
Ask a compatible Agent Skills client something like: “Configure NeMo Relay OpenTelemetry observability and first verify it with a synthetic, non-sensitive payload.” The skill guides output selection, registration under a unique name before scoped work, execution of instrumented Relay work, flushing when deterministic delivery is needed, and final deregistration and shutdown. It lists binding names for Python, Node.js, Rust, and Go.
How does this skill compare with similar options?
The skill explicitly contrasts the built-in Observability plugin with manual subscriber or exporter APIs: the plugin is intended for reusable, configuration-driven process setup, while manual APIs suit tests, scripts, or applications requiring direct control over registration names, collection windows, or flush timing.