NVIDIA/NeMo-Relay 0.4.0

NeMo Relay 0.4.0

Repository: NVIDIA/NeMo-Relay

Tag: 0.4.0

Published: 2026-06-13T01:06:33Z

Prerelease: no

Release notes: We are proud to release NeMo Relay 0.4.0!

Key Features

• 🛡️ First-party PII redaction plugin with deterministic local backend support
• 🚦 Built-in NeMo Guardrails local backend support, so the nemo_guardrails component can run against either remote services or a Python-backed local worker.
• ⚙️ Code-driven plugin configuration layered over materialized global, project, and user plugin files while preserving documented precedence.
• 📡 Streaming ATOF export endpoints for HTTP POST, WebSocket, and long-lived NDJSON collectors.
• 🗄️ ATIF HTTP storage export support alongside S3-compatible trajectory storage.
• 💸 Pricing-aware LLM observability with model-pricing lookup, cost layering, and cost metadata propagated into ATIF, OpenInference, and OpenTelemetry output.
• 🧭 Stronger coding-agent trace fidelity across Hermes, OpenClaw, Claude Code, and Codex, including nested subagent lineage, routed-provider spans, and error-path consistency.
• 🧩 Improved framework integration behavior for LangChain, LangGraph, and Deep Agents, including serialization fixes, callback coverage, and model-response annotations.
• 🔌 Host plugin installation for Claude Code and Codex, including native plugin manifests, and install/uninstall/doctor --plugin subcommands.

What's Changed

💥 Breaking Changes

• chore(guardrails)!: remove python-only guardrails example plugin by @afourniernv in https://github.com/NVIDIA/NeMo-Relay/pull/198

✨ New Features

• feat: Improve Hermes Relay hook injection and event alignment by @yczhang-nv in https://github.com/NVIDIA/NeMo-Relay/pull/205
• feat: add nested subagent session lineage for OpenClaw by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/210
• feat: add ATIF HTTP storage export by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/231
• feat: add streaming ATOF endpoints by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/232
• feat: add LLM cost layering and pricing lookup by @AjayThorve in https://github.com/NVIDIA/NeMo-Relay/pull/236
• feat: add local backend for built-in nemo guardrails by @afourniernv in https://github.com/NVIDIA/NeMo-Relay/pull/197
• feat: layer code-driven plugin config over materialized file configs by @zhongxuanwang-nv in https://github.com/NVIDIA/NeMo-Relay/pull/211
• feat: add claude code and codex plugins by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/224
• feat: add PII redaction plugin crate with local backend by @afourniernv in https://github.com/NVIDIA/NeMo-Relay/pull/245

⚡ Improvements

• refactor: reduce SonarQube complexity findings by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/243

🐛 Bug Fixes

• fix: pass a real ScopeHandle to Node withScope callbacks by @zhongxuanwang-nv in https://github.com/NVIDIA/NeMo-Relay/pull/196
• fix: improve OpenClaw observability consistency by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/203
• fix: remove race condition in LC integration tests by @dagardner-nv in https://github.com/NVIDIA/NeMo-Relay/pull/202
• fix: improve OpenClaw observability cost consistency by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/206
• fix: add flattened OpenInference LLM attributes for annotations and replay by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/207
• fix: Fix LangChain input serialization in callback handler and fix deadlock in plugin context manager by @dagardner-nv in https://github.com/NVIDIA/NeMo-Relay/pull/208
• fix: tighten OpenClaw ATOF hook-backed provenance metadata by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/209
• fix: suppress Claude Code lifecycle noise by @AjayThorve in https://github.com/NVIDIA/NeMo-Relay/pull/213
• fix: preserve structured ATIF tool results by @dnandakumar-nv in https://github.com/NVIDIA/NeMo-Relay/pull/223
• fix: validate Hermes routed provider observability by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/222
• fix: preserve Hermes error-path export consistency by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/229
• fix: preserve Hermes tool-result ATIF observations by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/233
• fix: resolve sanitized LLM request from annotations by @fallintoplace in https://github.com/NVIDIA/NeMo-Relay/pull/241
• fix: avoid ATIF shutdown subscriber deadlock by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/242
• fix: annotate Deep Agents model responses by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/246
• fix: include pii redaction in version updates by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/251
• fix: configure gateway body limits by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/261

✅ Tests

• test: validate OpenClaw nested subagent exports by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/214
• test: validate OpenClaw placeholder replay exports by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/217
• test: tighten Hermes wrapped observability contract by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/215
• test: validate OpenClaw model timing diagnostic exports by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/218
• test: cover Hermes wrapped ATIF fidelity and cost by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/219
• test: validate OpenClaw hook-only fallback exports by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/220
• test: add NeMo Relay skill eval datasets by @abhisawa-Nvidia in https://github.com/NVIDIA/NeMo-Relay/pull/225
• test: validate Hermes routed provider OpenInference spans by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/235
• test: add Codex observability contract coverage by @yczhang-nv in https://github.com/NVIDIA/NeMo-Relay/pull/234
• test: regenerate NeMo Relay evals by @abhisawa-Nvidia in https://github.com/NVIDIA/NeMo-Relay/pull/226
• test: validate Hermes subagent lineage and orphan export coverage by @mnajafian-nv in https://github.com/NVIDIA/NeMo-Relay/pull/240
• test(wasm): drop --test-force-exit to avoid Windows libuv abort by @zhongxuanwang-nv in https://github.com/NVIDIA/NeMo-Relay/pull/247
• test: raise Rust and CLI coverage by @willkill07 in https://github.com/NVIDIA/NeMo-Relay/pull/259

🧰 Maintenance

• chore: prepare 0.3 code freeze by @willkill07 in...