v1.6.0

🚀 Notable Features and Improvements

• This release introduces ATIF-based evaluation, including a standalone evaluator class that can evaluate ATIF trajectories generated by any system

What's Changed

🚨 Breaking Changes

• NAT Eval Dependency Reduction by @AnuradhaKaruppiah in #1690
• Migrate nvidia_nat_ragas to Ragas v0.4 by @AnuradhaKaruppiah in #1780
• Add support for optionally disabling SSL/TLS verification by @dagardner-nv in #1640
• Add centralized user identity management and custom WebSocket authentication by @ericevans-nv in #1775
• Feat/optimizer package extraction by @bhuvan002 in #1637
• Decouple ATIF schemas from core and make nvidia-nat-eval lightweight by default by @AnuradhaKaruppiah in #1833

✨ New Features

• Add TimeoutMiddleware by @ericevans-nv in #1677
• feat(agent): stream tool call chunks by @MylesShannon in #1717
• Add ATIF adapter: IntermediateStep to Harbor ATIF v1.6 trajectory conversion by @yczhang-nv in #1663
• feat(control_flow): add built-in parallel_executor with example by @antoniomtz in #1733
• Restore TimeoutMiddleware reverted by #1710 by @ericevans-nv in #1753
• Add ATIF-native evaluator interface and runtime dual-lane dispatch by @AnuradhaKaruppiah in #1747
• Add LLM error recovery, retry logic, and diagnostic logging to Tool Calling Agent by @ericevans-nv in #1757
• Add ATIF-only custom evaluator example and docs by @AnuradhaKaruppiah in #1766
• feat(profiler): ATIF-native profiler by @afourniernv in #1751
• Add pre-tool verifier defense middleware for input instruction violation detection by @lidan-capsule in #1605
• Add MemMachine memory integration for NeMo Agent toolkit by @Charlie-Yi-2002 in #1460
• Nested ATIF lineage by @AnuradhaKaruppiah in #1814

🔧 Improvements

• Forward-merge release/1.5 into develop by @willkill07 in #1676
• Add user attribution to Weave traces and feedback by @thepatrickchin in #1668
• Move eval callbacks to nvidia_nat_eval and guard core shim path by @AnuradhaKaruppiah in #1748
• Refactor eval file output into FileEvalCallback for Python API support by @yczhang-nv in #1743
• Add AtifBaseEvaluator for reusable concurrent ATIF custom evaluators by @AnuradhaKaruppiah in #1769
• Use nvidia-nat-eval from PyPI in the custom evaluator notebook by @AnuradhaKaruppiah in #1778
• chore(pre-commit): any pyproject.toml changes should trigger uv lock checks by @willkill07 in #1790
• Remove unneeded warning filter by @dagardner-nv in #1789
• Remove old/unused compatibility code by @dagardner-nv in #1794
• fix: handle Anthropic-style tool schemas in LangchainProfilerHandler by @cluster2600 in #1770
• Add ATIF-native evaluators for nvidia-nat-profiler runtime metrics by @afourniernv in #1791
• Update old CI variable names by @dagardner-nv in #1813
• Explicitly set framework_wrappers for LangChain tools by @dagardner-nv in #1820
• Report installed packages when testing built wheels by @dagardner-nv in #1821
• chore: pin LiteLLM to known good version by @willkill07 in #1823
• Ensure file paths are properly quoted in gitutils.py by @dagardner-nv in #1832
• Update LICENSE-3rd-party.txt by @dagardner-nv in #1840
• Informational ATIF mapping docs and sample output colocation (no functional changes) by @AnuradhaKaruppiah in #1844
• fix: guard nim_langchain against ChatNVIDIA model lookup crash by @bbednarski9 in #1843
• Document that MemMachine integration is Experimental by @AnuradhaKaruppiah in #1854
• Restore profiler intermediate-step compatibility by rolling back selected ATIF-native changes by @AnuradhaKaruppiah in #1853

🐛 Bug Fixes

• Fix Simple Calculator MCP example to return consistent timezone results by @dagardner-nv in #1708
• fix: preserve nested Pydantic models when unpacking multi-arg functions by @MylesShannon in #1696
• Fix _running_workflow_task always being None due to method chaining by @AjayThorve in #1744
• Fix LoRA inference routing in OpenPipe ART cookbook by @aslanshi in #1662
• Add profiler callback to sequential executor for LLM observability by @thepatrickchin in #1786
• Enhance chat completion functionality to support OpenAI-style message history by @mdemoret-nv in #1674
• fix(crewai): CrewAI Profiler Handler logic >=1.1.0 by @willkill07 in #1803
• fix: add missing dependencies and repair logic for some tests by @willkill07 in #1815
• Fix urls in pypi markdown files by @dagardner-nv in #1828
• Fix default values for datetime fields in JobInfo objects by @dagardner-nv in #1829
• add targeted MCP OAuth2 diagnostics for redirect URI mismatches by @AnuradhaKaruppiah in #1830
• Ensure proper cancellation of in-flight workflow tasks by @thepatrickchin in #1819
• Ensure HTTP request attributes are propagated into eval and async requests by @dagardner-nv in #1838
• Update weave version to pickup fixes for tracing LangChain workflows by @dagardner-nv in #1841
• fix(middleware): improve pre-tool middleware guarding logic by @cparadis-nvidia in #1824

📝 Documentation Updates

• Add notebooks demonstrating standalone ATIF evaluation via Python API by @yczhang-nv in #1752
• Add notebook demonstrating end-to-end interoperability between third-party agent frameworks and the NeMo Agent toolkit evaluation harness by @yczhang-nv in #1799
• docs: add LangSmith to list of supported observability platforms by @abhisawa-Nvidia in #1825
• Fix license header in memmachine's README.md by @Charlie-Yi-2002 in #1831
• add memmachine into docs/ by @Charlie-Yi-2002 in #1839
• Document Otel endpoints for custom LangSmith deployments by @dagardner-nv in #1845

New Contributors

• @AjayThorve made their first contribution in #1744
• @cluster2600 made their first contribution in #1770
• @afourniernv made their first contribution in #1791
• @bhuvan002 made their first contribution in #1637
• @lidan-capsule made their first contribution in #1605
• @abhisawa-Nvidia made their first contribution in #1825
• @Charlie-Yi-2002 made their first contribution in #1460
• @cparadis-nvidia made their first contribution in #1824

Full Changelog: v1.5.0...v1.6.0