NVIDIA/NeMo-Relay 0.11.0

0.11.0

Repository: NVIDIA/NeMo-Relay

Tag: 0.11.0

Published: 2026-10-28T01:34:03Z

Prerelease: no

Release notes: We are proud to release NeMo Relay 0.11.0!

This release includes changes since 0.10.0 and fixes shipped in the 0.10 patch releases. Highlights focus on diagnostics, framework integrations, and more reliable request handling.

Notable New Features

• 📊 Diagnostics: Inspect runtime counters, recent events, workers, and diagnostic memory across CLI, daemon, bindings, and plugin SDKs. Collection is opt-in with bounded storage.
• 🔗 Integrations: Add Python Google ADK support, including supported LiteLLM models, and configurable middleware placement for Deep Agents.
• 🔭 Observability: Configure OTLP custom certificate authorities and mutual TLS, Python trace-file sinks, and smaller ATIF exports. Doctor checks more exporter and privacy settings.
• 🔌 CLI Improvements Run full LLM middleware in daemon workers, supply API keys per request attempt, and use request codecs in manual LLM calls.
• 🧭 Scopes and coding agents: Inspect visible scope snapshots across bindings and plugin SDKs, and continue launcher traces through Codex and Claude Code hooks.
• ⚙️ Configuration: Configure worker callback deadlines, message-size limits, and optional operational log files. Expanded quick starts and Kubernetes guidance simplify deployment.

Breaking Changes

• Coding-agent hooks: Refresh installed integrations and restart affected processes. Local hooks verify TLS; tool checks default to fail-closed, and delivery failures return exit code 2. Generated hooks use PreToolUse without a separate PermissionRequest hook.
• Provider access and policy ordering: Relay checks client access before lending provider credentials. Workers now load layered user and system configuration. Response caches default to priority 150; review custom priorities so policy checks run first.
• Python and framework results: Object reconstruction requires explicit allowed_types. LangChain and LangGraph tool results use the typed nemo.relay.tool_result@1 wrapper; update middleware and subscribers that inspect results.
• Streaming and Node.js: Node.js requires version 24.13.1 or later. Streams expose provider, collector, and finalizer failures; handle errors and await stream.close() during early cleanup.
• Observability and request output: Automatic remote OTLP logs and metrics require HTTPS unless explicitly allowed. Fatal TLS errors stop HTTP trace exporters. CLI turn names inherit agent identity, and adaptive hints now honor configured injection settings.
• Rust and plugin SDKs: Update direct streaming and logging struct construction and exhaustive worker-error matches. Scope snapshots require native ABI v8 or the worker extension; existing v7 callback layouts remain compatible. Failed plugin integrity or signature checks prevent activation.
• Worker transport: Messages default to a 4 MiB limit each. Upgrade hosts and worker SDKs together before using callback-specific limits.

Non-Exhaustive Fixes

• Streaming and codecs: Preserve fragmented Unicode, Anthropic tool-input shapes, Gemini function arguments, Azure stream metadata, and OpenAI Responses usage. Retain failures through stream cleanup.
• Frameworks and coding agents: Preserve tool schemas, typed results, handled errors, and cache-replay content. Recover missing Codex MCP tool completions when matching results reach Relay.
• Observability and pricing: Correct ATIF content, trace parentage, exporter diagnostics, and token-cost accounting, including Gemini thinking and tool-use tokens.
• Daemon and plugins: Reduce startup memory and routing overhead; improve cancellation, session cleanup, configuration loading, and HTTP error propagation.
• Bindings and configuration: Preserve Python exceptions, isolate throwing Node.js codec callbacks, correct Go scope ownership, retain configured log sinks, and expose missing editor settings.

What's Changed

💥 Breaking Changes

• fix(python)!: require explicit codec reconstruction types by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/1352
• fix(node)!: surface stream provider, collector, finalizer errors by @greysonlalonde in https://github.com/NVIDIA/NeMo-Relay/pull/1287
• fix(cli)!: remove redundant PermissionRequest hooks by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/1388
• feat!: integrate bounded dynamic plugin diagnostics by @bbednarski9 in https://github.com/NVIDIA/NeMo-Relay/pull/1450

✨ New Features

• feat: support full LLM middleware in daemon workers by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/1281
• feat: make the Codex ChatGPT upstream configurable by @samratjha96 in https://github.com/NVIDIA/NeMo-Relay/pull/1319
• feat: add per-request upstream API key helpers by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/1342
• feat: expose consistent scope stack snapshots across bindings and plugin SDKs by @ericevans-nv in https://github.com/NVIDIA/NeMo-Relay/pull/1366
• feat: make short worker callback deadlines configurable by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/1371
• feat(observability): add insecure_http_hosts allowlist by @marcusds in https://github.com/NVIDIA/NeMo-Relay/pull/1367
• feat(core): recognize load_skill as a skill tool by @greysonlalonde in https://github.com/NVIDIA/NeMo-Relay/pull/1432
• feat(observability): add OTLP TLS and mTLS configuration by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/1422
• feat: add request codecs to manual LLM calls by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/1441
• feat: allow credential-free HTTP model upstreams on trusted hosts by @AjayThorve in https://github.com/NVIDIA/NeMo-Relay/pull/1429
• feat: add bounded runtime diagnostics foundation by @bbednarski9 in https://github.com/NVIDIA/NeMo-Relay/pull/1448
• feat: integrate bounded CLI diagnostics by @bbednarski9 in https://github.com/NVIDIA/NeMo-Relay/pull/1449
• feat: integrate bounded daemon diagnostics by @bbednarski9 in https://github.com/NVIDIA/NeMo-Relay/pull/1453
• feat: expose bounded diagnostics in language bindings by @bbednarski9 in https://github.com/NVIDIA/NeMo-Relay/pull/1455
• feat: unify diagnostics read contracts and conformance by @bbednarski9 in https://github.com/NVIDIA/NeMo-Relay/pull/1456
• feat: inspect diagnostic memory and correlate resource samples by @bbednarski9 in...