Skip to navigation

Release Notes

View as MarkdownOpen in Claude

This page summarizes NVIDIA NeMo Fabric 0.7.0. GitHub Releases remain the source of truth for complete release history and tag-specific notes.

Current Release: 0.7.0

NeMo Fabric 0.7.0 adds conversation resume, saved invocation results, richer harness configuration, and structured compatibility diagnostics. The release also improves runtime cleanup, deadline reporting, and recovery after a caller loses a response.

Upgrade Notes

inspect_adapter() now returns AdapterInspection instead of AdapterCapabilityProfile. Compatibility issues are returned by default. Check inspection.issues before reading inspection.capability_profile or admitting a deployment. Pass diagnostics=False for exception-based rejection. Malformed input raises FabricConfigError in both modes.

Rust inspect_adapter_config callers now pass deserialized descriptor values instead of file paths and explicitly choose diagnostic or exception-based behavior. Refer to Inspection API Migration for both language APIs.

Upgrade nemo-fabric, nemo-fabric-runtime, the adapters, the adapter contract, the adapter catalog, and nemo-fabric-evidence together. Keep the package versions aligned at 0.7.0. Rust callers using complete struct literals must account for the added runtime controls and result fields, or use defaults where available.

Invocation deadlines use error code timeout at stage invoke. Startup and shutdown deadlines retain their lifecycle stage. Applications that inspect timeout codes should review Invocation and Cleanup Deadlines.

Conversation resume requires runtime.keep_state, retained artifact storage, and an explicit SessionRef. Omitting the reference starts a fresh conversation. Hermes resume also requires the SDK interface to be enabled. An uncertain earlier turn can require acknowledge_recovery=True after the caller inspects native history and external effects. Refer to Requests and Runtimes.

Harness and Consumer Compatibility

Hermes Agent requires Python 3.14 in its own environment. Applications using another supported Python version can select that environment with ADAPTER_PYTHON. Keep the Relay version matched to the installed Hermes release: v0.22.7 uses Relay >=0.11,<0.12, while the supported legacy v0.22.5 pairing uses >=0.10,<0.11. Refer to the Hermes installation guide.

The NeMo Fabric Harbor extra continues to select harbor>=0.24,<0.25. Harbor’s phase-timeout ownership fix is a separate requirement for correct inner-deadline reporting in trial summaries. Upgrading NeMo Fabric alone does not correct those summaries in older Harbor versions. Refer to Harbor Execution Deadlines.

Release Highlights

Resume Hermes and OpenClaw Conversations

Retain native agent state across runtimes and explicitly reopen a selected Hermes or OpenClaw conversation with SessionRef. A replacement runtime receives a new NeMo Fabric runtime identifier while retaining the native conversation. Missing or incompatible sessions fail explicitly, and uncertain earlier work is not automatically replayed.

NeMo Fabric locks retained state against overlapping managed runtimes for the same agent storage scope. Failed cleanup can keep ownership held and block a replacement until cleanup succeeds. Refer to the Hermes resume guide and the OpenClaw integration guide.

Retrieve Results and Inspect Runtime Status

Completed invocation results are saved before being returned. After a caller timeout, use Fabric.invocation_result() or fabric-agent-ctl result with the original runtime and request identifiers to retrieve the saved result. Stable request identifiers let compatible buffered retries return the saved result instead of dispatching another invocation. A missing result does not establish that work is safe to repeat.

Use Fabric.runtime_status() or fabric-agent-ctl status to inspect recorded runtime state and available handles. Recorded state is separate from a fresh health observation; use Fabric.check() for the latter. Refer to Saved Invocation Results and Recorded Runtime Status.

Configure Native Interfaces, Skills, and Plugins

Hermes and OpenClaw expose native interfaces alongside SDK execution, with explicit ownership of the managed processes and credentials. Hermes supports its HTTP API, browser interface, and experimental terminal attachment. OpenClaw supports Gateway API, dashboard, and terminal access. Refer to the Hermes interface guides and OpenClaw interface guides.

Supported adapters can load skill directories, select bundled skills, and control duplicate names. Reports identify loaded, unavailable, or conflicting skills. The harness-specific rules and file-system permissions still apply. Refer to Configure Skills.

Hermes and OpenClaw support bundled and external plugins. OpenClaw can also retain plugins installed by the agent when state retention is enabled. Hermes requires external plugin dependencies to be installed in advance and rejects agent-installed plugin environments. Refer to Hermes Plugins and OpenClaw Plugins.

Extend Deep Agents and Report Unavailable MCP Servers

Deep Agents accepts installed Python tools and middleware through configuration. Tool definitions name LangChain tools, and middleware entries name LangChain middleware classes. The adapter imports the configured code at runtime; NeMo Fabric does not fetch or install those extensions. Deep Agents also loads skills through its configured backend and supports OAuth client-credentials authentication for eligible remote MCP servers.

Deep Agents, Claude, and Codex report unavailable MCP servers while allowing supported startup with the remaining tools. The Deep Agents adapter returns MCP tool and argument-validation failures as model-visible tool feedback where supported, so the model can correct a call. Invalid configuration still fails startup. Refer to the Deep Agents guide and each adapter’s MCP guidance.

Inspect Compatibility and Bound Returned Data

Inspect adapter descriptors in memory before scheduling execution. AdapterInspection reports configuration issues without importing the harness SDK or discovering executable adapters. The inspection validates supplied metadata; the task environment still resolves and verifies its installed descriptor. Refer to Inspect Metadata on a Separate Host.

Set runtime.response_limits to bound buffered response size, artifact count, total artifact bytes, and individual file size. Local execution enforces these budgets when configured explicitly; OpenShell retains default transfer budgets when limits are omitted. A breach returns response_limit_exceeded, preserves available execution evidence, and does not undo agent side effects. Refer to Response and Artifact Limits.

Reliability and Security Updates

  • Preserve invocation deadline ownership across local execution, OpenShell, and Harbor.
  • Improve interruption and cleanup of busy runtimes before replacement, and preserve more accurate health and process-exit observations.
  • Retain available execution evidence when Harbor’s runner fails.
  • Improve OpenClaw Gateway shutdown, custom-provider API selection, and native session tracking between calls.
  • Keep host ownership records under the configured artifact root.
  • Upgrade LiteLLM to 1.100.4 for the security fix tracked in PR #642.
  • Reorganize Getting Started and SDK guides around consumer workflows, harnesses, and execution environments.

Current Limitations

  • Conversation resume is adapter-specific and requires the original retained storage and explicit session reference. An OpenClaw native session reset can still race with final invocation dispatch; refer to issue #638.
  • Saved results remain until their storage is removed. Runtime status records belong to the owning process and do not survive server restart. Terminal results from completed streams can be retrieved, but automatic stream replay is unsupported.
  • Native terminal attachment and interactive approvals remain subject to the qualification limits in the Hermes and OpenClaw guides. Caller-owned OpenClaw Gateways reject retained-session selection and plugin configuration when NeMo Fabric cannot establish the required ownership boundary.
  • Metadata inspection does not verify image contents, installed harness software, credentials, endpoint access, or runtime readiness.
  • Prepared and attached service hosting remains a local SDK capability. The OpenShell command bridge returns buffered results and does not support native streaming; artifact transfer remains bounded.
  • Health observations are adapter-specific. Resource statistics exclude remote model and tool servers and GPU usage. Selected software evidence is not a complete dependency inventory or an attestation.
  • Kiro and AWS Strands do not provide Relay telemetry or native NeMo Fabric streaming. The adapter catalog provides metadata snapshots rather than executable provider registries or an automated conformance suite.