About NVIDIA NeMo RelayRelease Notes

Highlights

View as Markdown

This page summarizes the notable capabilities in the current release documentation set.

NVIDIA NeMo Relay 0.5

NVIDIA NeMo Relay 0.5 adds dynamic plugin SDKs and controls, verified CLI installation and diagnostics, clearer runtime contracts, and improved lineage for managed LLM requests.

Compatibility and Migration

  • Move plugin configuration from config.toml into plugins.toml. The --plugin-config CLI option and hook-forwarding input have been removed.
  • Update every registered LLM request and tool execution intercept to return its canonical outcome. The outcomes can include ordered pending marks while Relay retains lifecycle ownership.
  • Rebuild development native plugins and grpc-v1 workers against 0.5 because the native ABI and worker outcome contracts are now finalized.
  • Migrate away from the removed experimental nemo-relay-wasm package, Cursor CLI support, and patch-based integrations.

Dynamic Plugins and Configuration

  • Added dynamic plugin discovery from plugins.toml, host policy and attestation gates, and lifecycle commands for inspecting, validating, enabling, and disabling plugins.
  • Added the Rust SDK for trusted in-process rust_dynamic plugins and the grpc-v1 worker protocol with Rust and Python worker SDKs.
  • Added schema-aware dynamic-plugin configuration, validation, secret-aware editing, and merged TOML preview in the CLI.
  • Added Python worker-environment management and worker invocation cancellation.

CLI Installation and Diagnostics

  • Added verified Unix and Windows installers that select the release asset, verify its SHA-256 checksum, and support version pinning.
  • Expanded nemo-relay doctor with agent-host readiness checks, persistent host-plugin diagnostics, and discovered plugins.toml source and validation reporting.
  • Added gateway status reporting for bind, exporter, and remote ATIF storage destinations.

Runtime and Observability

  • Added canonical LLM request and tool execution outcomes across Rust, Python, Node.js, Go, C FFI, native plugins, and worker plugins.
  • Added ordered pending marks to managed LLM and tool lifecycles without adding control data to application-visible requests or tool results.
  • Added automatic Dynamo session and parent-session headers to managed LLM requests, plus bounded CLI turn scopes for trace correlation.
  • Documented stable normalized token and cost semantics, including provider mapping and exporter projection behavior.
  • Hardened ATOF HTTP output and ensured gateway shutdown flushes queued events.

Documentation and Tooling

  • Added public contract references for LLM request and tool execution intercept outcomes.
  • Expanded dynamic plugin, configuration, runtime-contract, installation, and integration examples.
  • Updated package, build, test, and release guidance for the primary Rust, Python, and Node.js surfaces, plus experimental Go and raw C FFI notes.

The complete changelog and release artifacts can be viewed on GitHub Releases.