2.1.748
馃殌 Arthur Engine Release
July 29, 2026
This release deepens dataset traceability and version management, refines the task list experience, and delivers a broad wave of security remediations across container images and dependencies.
Datasets & Traceability
Dataset Records & Versions
- Added the ability to link dataset records back to their originating traces, with a new SourceTraceLink component and navigation directly from the dataset UI (#1978)
- Introduced dataset version reinstatement, letting users restore a prior version as a new latest version鈥攕imilar to a git revert鈥攚ithout overwriting history (#1962)
Add-to-Dataset Flow
- Fixed transform auto-matching and auto-filling so transforms are recognized after columns are first added, removing the need for the "Fill from Object" workaround (#1961)
- Fixed the unsaved changes confirmation to guard all exit paths鈥攕idebar navigation, in-page Back button, and version switching鈥攙ia a unified NavigationGuardContext (#1990)
Dataset workflows are now more transparent and forgiving, giving users clearer lineage back to source traces, safer version recovery, and consistent protection against accidental data loss.
Task List & Filtering
- Moved sorting and activity filtering server-side and based the "Active in last N days" filter on real trace activity, so task lists remain accurate and complete regardless of size (#1975)
- Removed the sort dropdown from the task list and standardized ordering to newest-first by creation date (#2027)
Task list behavior is now reliable at scale, with accurate "Last active" labels and a simplified, consistent ordering model.
Formatting & UI Consistency
- Migrated to shared DurationCell and formatter components from the unify-shared library, ensuring latency and duration values render with consistent decimal precision (#1976)
Latency and duration values now display cleanly across all panels, free from trailing floating-point artifacts.
Deployment & Infrastructure
Helm Charts
- Exposed the ml-engine HorizontalPodAutoscaler in the umbrella chart values, letting operators enable and configure ML engine auto-scaling with sensible defaults (#1984)
- Fixed genaiEngineInternalHost Helm configuration to include the required
http://scheme, resolving ML Engine startup failures (#1942)
CI & Release Tooling
- Updated CI to rebuild model-upload images when their Dockerfile, dependencies, or entrypoint scripts change, ensuring security and dependency updates reliably reach production (#2019)
- Added a scheduled stale PR cleanup workflow that labels and eventually closes inactive pull requests while preserving active and kept-open ones (#1991)
- Improved Renovate configuration to automatically flow transitive and uv updates and to group langchain and openai package updates into coordinated PRs (#2012, #2015)
- Added sourcemap upload configuration for meticulous test coverage analysis, scoped to testing builds only (#2009)
These changes give operators more control over ML engine scaling while strengthening the reliability of the build and release pipeline.
Security & Vulnerability Remediation
Container Image CVEs
- Bumped torch to 2.13.0 and uv to 0.12.0 to clear HIGH-severity RCE, deserialization, and Rust crate CVEs across genai-engine and ml-engine images (#2034)
- Bumped hf-xet to 1.5.2 to patch a vendored quinn-proto vulnerability, eliminating a remote memory exhaustion risk (#2041)
- Triaged and resolved the full Trivy HIGH backlog, remediating findings via uv upgrades and formally justifying the remainder with OpenVEX statements (#2018)
- Added OpenVEX justifications for perl-base CVEs in the models-gcs image where the vulnerable code paths are unreachable (#2002)
- Dropped the remediated nltk CVE-2026-54293 VEX exception now that the fixed version is pinned (#2016)
Dependency Security Patches
- Updated pypdf to v6.14.2 to address two critical CVEs causing infinite loops when parsing malformed PDF inline images (#1985)
- Bumped vulnerable transitive Python dependencies (pyasn1, Pillow, mcp, joserfc), resolving 30 of 31 open Dependabot alerts (#2014)
- Updated frontend dependencies including tar, React Router, and shared-components to patch known CVEs in the frontend security scan (#2013, #2017)
Container images now pass HIGH-severity security scanning with all findings either remediated or documented, and users are protected against critical PDF-parsing and transitive dependency vulnerabilities.
Dependency Updates
ML & Data Libraries
- Updated sentence-transformers to v5.6.1, fixing silently degraded embeddings for RoBERTa-family models under flash attention (#2008)
- Updated GLiNER to v0.2.28 with improved inference reliability and Transformers v5 compatibility (#2004)
- Updated presidio-analyzer and presidio-anonymizer to v2.2.364, improving UK NINO and Indian vehicle registration detection and adding a CLI threshold override (#2006, #2007)
- Updated duckdb to v1.5.5 for numpy compatibility and a DuckDBPyRelation.query alias fix (#2003)
- Updated croniter to v6.2.4 with corrected cron range calculations and faster evaluation of comma-separated ranges (#2021)
- Updated pandas to v3.0.5 and google-cloud-bigquery to v3.42.2 for stability and upstream fixes (#2005, #2024)
- Updated boto3 across several patch releases to capture upstream AWS SDK improvements (#1989, #1996, #2001)
Frontend & UI Libraries
- Updated the React monorepo to v19.2.8 for improved stability (#2020)
- Updated monaco-editor to v0.56.0, adding new editor APIs, configuration options, and improved Markdown security handling (#2029)
- Updated framer-motion to v12.42.2, graduating animateView with selector support and improved view-transition animations (#2023)
- Updated nuqs to v2.9.2 with improved query-string state recovery and Next.js security patches (#2030)
- Fixed a @vitejs/plugin-react regression by updating to v6.0.4, resolving a
$RefreshSig$ is not definederror underNODE_ENV=production(#1994) - Updated Amplitude analytics and session replay browser SDKs to their latest patch versions (#1986, #1987, #1993)
Arthur Client
- Updated the arthur-client dependency across genai-engine and ml-engine through successive patch releases (#1988, #1995, #2000)
These updates keep the Arthur Engine's ML, data, and frontend stacks current, delivering improved accuracy, stability, and compatibility across the platform.
Release notes generated by Louisa