Agent skill · NVIDIA
nemo-relay-plugin-adaptive-tuning
Use this skill when baseline NeMo Relay instrumentation exists and the user wants to configure or evaluate adaptive plugin behavior, including telemetry, state, adaptive_hints, tool_parallelism, ACG, hint consumption, or measured rollout.
What it needs
About 3k tokens when loaded.
What this skill does
Tune Adaptive Plugin Behavior Use This When Use this skill when a user has baseline NeMo Relay instrumentation and wants to improve latency, parallelism, prompt-cache behavior, or model-request behavior from runtime signals. Keep adaptive behavior measured against a known baseline. Do Not Use This When Do not use this skill when the application is not instrumented yet. Start with nemo-relay-instrument-calls or nemo-relay-get-started first. Default Guidance Observe first, compare against a baseline, then enable one behavior change at a time. Use the adaptive plugin component rather than inventing separate tuning logic or hand-registering adaptive behavior at every call site. Start with in-memory state and telemetry-only behavior for local development. Move to persistent state only when learned signals must survive restarts or be shared across workers. Add active behavior only after representative runtime events show what should change. Embedded Adaptive Model Adaptive behavior is configured through the first-party plugin component with kind adaptive. Adaptive requires existing NeMo Relay scopes and at least one relevant managed tool or LLM lifecycle event stream because it learns from runtime signals. Main configuration areas are state, telemetry, adaptive hints, tool parallelism, Adaptive Cache Governor (ACG), and rollout policy. State backends are inmemory and redis. Tool-parallelism modes are observeonly, injecthints, and schedule. Adaptive Cache Governor providers are passthrough, anthropic, and openai; omit ACG until prompt-cache planning is needed. Helper APIs exist in Rust nemorelayadaptive, Python nemorelay.adaptive, and Node.js nemo-relay-node/adaptive. Go and raw FFI are source-first or advanced surfaces. Default Path Use this rollout sequence: 1. Confirm the app emits scope events and the managed tool or LLM events needed for the behavior being evaluated. Do not require both call types when the workflow uses only one. 2. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills NVIDIA/nemo-relay-plugin-adaptive-tuning