NeMo Relay Adaptive Tuning
Tune NeMo Relay plugin behavior from runtime signals with measured, reversible rollouts.
The material requires observe-first rollout, one change at a time, validation, idempotency and race checks, rollback, and sanitized diagnostics, while keeping hints advisory; however, telemetry data flows, sensitive-data handling, user-confirmation boundaries, dependency security, and exact runtime permissions are not specified. The repository license metadata is also NOASSERTION, so points are deducted.
SKILL.md, config.md, and hints.md are broadly consistent about prerequisites, staged rollout, validation failures, and rollback, and instruct the agent to stop on abnormal conditions; however, there are no reproducible configuration examples, pinned dependency versions, key-path test suites, or verifiable error outputs. Static calibration therefore limits this dimension to 9.
The trigger and non-fit conditions are clear, covering telemetry, hints, tool parallelism, and ACG, with Python, Node.js, and Rust entry points; however, finer capability boundaries, input/output contracts, Chinese-language support, and mainland-China reachability are not established, so points are deducted.
The information architecture, reference files, related skills, license, and repository synchronization/publication path are reasonably clear, with version and evaluation records; however, recommended Instructions and Examples sections are missing, the static report flags the author format, and installation dependencies, changelog, API compatibility policy, and maintenance ownership are incomplete.
For applications with existing NeMo Relay instrumentation, the skill gives a directly usable measured-tuning workflow covering baselines, in-memory state, single behavior changes, and rollback; however, it lacks complete runnable examples, representative direct outputs, and strong comparison against alternatives. Static evidence cannot establish broad effectiveness, so the score remains below the static cap.
The files provide a pinned revision context, evaluation tasks, benchmark report, static findings, and documented configuration/hint semantics, giving useful audit trails; however, benchmark claims are primarily author-reported and there is no independently reproducible test suite, CI coverage, or cross-source corroboration, so the score is limited.
- Before enabling tool_parallelism schedule mode or ACG/request rewriting, add verified checks for tool idempotency, race behavior, payload stability, and rollback.
- Ensure telemetry, hints, and diagnostics cannot expose secrets, personal data, or full request contents, and document retention, access control, and cross-border data-flow boundaries.
- Before publication, add runnable Python/Node.js/Rust configuration examples, dependency and compatibility information, and tests covering key paths.
- For mainland-China users, separately verify network reachability of NeMo Relay, Redis, package dependencies, and relevant documentation services.
What does this skill do, and when should you use it?
This skill configures or evaluates adaptive plugin behavior in applications that already have baseline NeMo Relay instrumentation. It covers telemetry, state, adaptive hints, tool parallelism, Adaptive Cache Governor (ACG), and rollout policy. The application must emit scope events and relevant managed-tool or LLM lifecycle events. The workflow emphasizes baseline measurement, one change at a time, and rollback when latency, correctness, or failure rates regress.
It guides the user through confirming scope and tool or LLM event streams, capturing a workflow baseline, and enabling telemetry through the first-party plugin component with kind adaptive. It documents the in_memory and redis state backends, observe_only, inject_hints, and schedule tool-parallelism modes, and passthrough, anthropic, and openai ACG providers. It directs the user to run representative traffic, inspect reports or runtime events, return configuration diagnostics, read references/config.md for exact fields, and read references/hints.md when application logic consumes adaptive hints or scheduling guidance.
- An engineering team with NeMo Relay instrumentation wants to reduce workflow latency using runtime signals.
- A developer evaluating tool parallelism wants to observe adaptive guidance before enabling hint injection or scheduling.
- A team wants prompt-cache planning after stabilizing provider request payloads.
- An operations team needs learned adaptive signals to survive restarts or be shared across workers.
- A developer building locally wants to begin with in-memory state and telemetry-only behavior.
What are this skill's strengths and limitations?
- Defines a clear workflow from baseline capture and telemetry to incremental rollout.
- Covers state backends, tool parallelism, ACG, hint consumption, and rollback conditions.
- Explicitly calls for checking tool idempotency, race behavior, and provider payload stability.
- Provides adaptive helper API surfaces for Rust, Python, and Node.js.
- It does not replace initial NeMo Relay instrumentation; uninstrumented applications need another skill first.
- The complete configuration schema is not included in SKILL.md and requires references/config.md.
- No quantitative performance results or test evidence are provided.
- Scheduling and prompt-cache planning require application-side safety and payload-stability validation.
How do you install this skill?
Install the specific skill with the CLI flow documented in the repository README:
npx skills add nvidia/skills --skill nemo-relay-plugin-adaptive-tuning --yes
The CLI handles the installation destination; cloning the repository or manually copying the skill directory is not required. The source does not document a more specific destination.
How do you use this skill?
First ensure that the application has baseline NeMo Relay instrumentation and the relevant event streams. Then use a request such as: "Evaluate and gradually tune NeMo Relay adaptive plugin behavior. Start with in-memory state and telemetry, establish a baseline, and only then consider tool parallelism or ACG." Read references/config.md for exact configuration fields and references/hints.md when application logic consumes adaptive hints or scheduling guidance. If validation fails or expected events are missing, keep the last known working configuration active and stop the rollout.
How does this skill compare with similar options?
The source explicitly directs users to nemo-relay-plugin-build when they need to build reusable plugin behavior. For initial instrumentation or onboarding, it points to nemo-relay-instrument-calls or nemo-relay-get-started.