Skip to navigation
About NeMo Fabric

Release Notes

View as Markdown

This page summarizes the current documentation-visible release state. GitHub Releases remain the source of truth for complete release history and tag-specific notes.

Current Release

NVIDIA NeMo Fabric 0.4 adds adapters for Cline, OpenClaw, OpenCode, and Qwen Code. It also adds a shared-service lifecycle for adapters that support it, extends Pi observability with NeMo Relay telemetry and streaming, and supports normalized model sampling and token-limit settings.

For release artifacts and the complete change history, refer to GitHub Releases.

Upgrade Notes

Relay-enabled integrations in NeMo Fabric 0.4 require NeMo Relay >=0.9,<0.10. Upgrade separately installed Relay Python packages and nemo-relay CLI installations before using Relay with 0.4 adapters. The published Hermes Agent v2026.9.24 release requires Relay 0.8, so Hermes Relay telemetry and streaming currently require the compatible, merged Hermes Relay 0.9 revision in a separate Python 3.14 environment. Ordinary Hermes Agent invocations do not require that development checkout.

The public Rust ModelConfig and AgentModelConfig structs now include optional top_p and max_tokens fields, and RuntimeHandle includes an optional service_id. Applications that construct these structs with explicit Rust struct literals must add the new fields or use a constructor or default that supplies them. Existing serialized configurations can omit the new fields.

Release Highlights

More Harness Integrations

  • Cline runs a local Cline SDK session across ordered invocations and maps supported model, tool, skill, and MCP configuration.
  • OpenClaw can run a Gateway for one runtime or connect multiple runtimes to a shared Gateway.
  • OpenCode runs an embedded OpenCode host and persistent session. The adapter can return a patch artifact for files changed during an invocation.
  • Qwen Code runs the Qwen Code SDK and its bundled CLI in an isolated child process, reusing one session across ordered invocations.

Shared Adapter Services

The Python SDK now exposes Fabric.prepare_service() for a NeMo Fabric-managed service and Fabric.attach_service() for a service managed by another system. Multiple runtimes can use one service while keeping separate runtime IDs and harness sessions. OpenClaw supports both service paths. Service handles are local to the creating process and cannot be used for cold resume.

Relay and Model Configuration

The Pi adapter now emits Relay telemetry and supports live ATOF records through Runtime.invoke_stream(). Pi streaming uses the embedded collector; its Relay 0.9 extension and CLI must be installed separately. Pi model telemetry also depends on a supported model API and a matching Relay gateway upstream.

The normalized model configuration now accepts models.<role>.top_p and models.<role>.max_tokens. Adapters use these fields when their descriptors declare support; unsupported settings fail during planning. See the Python SDK guide and the selected adapter guide for the exact mapping and limits.

The Deep Agents adapter also adds an opt-in local shell backend for supervised demonstrations. This backend runs commands with the adapter host user’s permissions and provides no process isolation.

Current Limitations

The current release has the following limitations:

  • Provider-backed third-party adapter registries are not included.
  • An automated adapter conformance suite is not included.
  • Cancellation, live updates, and cold resume are not exposed as supported adapter lifecycle operations. Shared service handles cannot be persisted across processes.
  • The Remote Agent adapter does not normalize the remote service’s MCP, skills, tool policy, or subagents. Relay streaming requires a separately instrumented remote service and a collector reachable by that service and NeMo Fabric.
  • The new Cline, OpenCode, and Qwen Code adapters do not support Relay telemetry or native streaming. Consult each adapter guide for other capability limits.