SMILE 7.0.0 Release Notes

Welcome to SMILE 7.0.0! This major release unifies high-performance JVM machine learning with the next generation of agentic data science. Version 7.0.0 introduces autonomous desktop automation, multi-agent orchestration, high-throughput LLM serving with batched Multi-Token Prediction (MTP) speculative decoding, native Kotlin and Scala 3 notebook kernels, an extensive productivity and research skills suite, time series forecasting (ARIMA/ARMA), and comprehensive stability and security hardening across Studio and the Agent harness.

✨ New Features

🤖 AI Agents & Desktop Automation

• 🖥️ Desktop Operator Agent (@chuck):
• Introduced Chuck, a specialized desktop automation agent capable of interacting directly with GUI applications via screenshot analysis, mouse navigation, and keyboard control.
• Safe desktop automation pipeline with coordinate mapping, interactive user confirmation, and automatic session bounds.
• 👔 Chief of Staff Agent (@frank):
• Added Frank to coordinate multi-agent teams, synthesize status reports across active agents, track project priorities, and lead cross-agent briefings.
• Dynamic Agent Catalog:
• Studio now discovers and loads agent personas dynamically from configuration specifications, with clean peer-to-peer delegation and request queuing.

⚡ High-Performance LLM Serving & Speculative Decoding

• Batched Multi-Token Prediction (MTP) Speculative Decoding:
• Implemented high-throughput batched MTP speculative decoding with persistent per-request draft histories, enabled by default in smile.serve.
• Accelerated generation with fused DeltaNet window operations, eager FlashInfer attention, and CUDA graph verification for low-latency token generation.
• SHAP Explanations in Model Serving:
• Native SHAP explanation payloads in REST inference responses with clean 3-decimal formatting and zero overhead when unrequested.
• Dynamic Model Loading & Control APIs:
• Dynamically load and unload models on the fly via POST /api/v1/models/load.
• Added real-time telemetry, model control, and health endpoints (/smile/{id}, /onnx/{id}, and /q/health).
• Native Prometheus metrics scraping integration.
• Support for ONNX Runtime 1.30, ONNX Runtime GenAI models (including Phi-4 mini), LibTorch 2.14.1, and automatic KV-cache precision selection (auto).
• Centralized Prometheus metrics API:
• GET /q/metrics: Single scrape target formatted according to standard Prometheus text specifications.
• Multidimensional Labeling: Every time-series metric carries a model_id label, allowing unified visualization dashboards (e.g., Grafana) with dynamic multi-tenant filtering.
• Microsecond Precision: Latency observations, throughput gauges, and cache counters are recorded with sub-millisecond fidelity via Micrometer.
• Non-blocking Execution: Metric tracking uses atomic primitives and lightweight histograms without introducing latency penalties to inference pipelines.

📚 Expanded Skills Suite

• Comprehensive Document Processing:
• Excel (excel-processing): Read, analyze, create, and format spreadsheets (.xlsx, .csv) with support for formulas, styles, and embedded charts.
• Word (word-processing): Parse, edit, and generate Microsoft Word documents (.docx) with precise layout and typography control.
• PDF (pdf-processing): Extract, analyze, and render publication-quality PDF reports and slide decks.
• Research & Development Skills:
• Deep Research (research): Multi-source web search and structured synthesis for market research and competitive analysis.
• Weekly Briefing (weekly-briefing): Automated status roll-ups and progress tracking across all agent session logs.
• Scientific Posters (posters): Design and generate conference posters and presentation materials.
• Git Commit (git-commit): Intelligent, conventional commit generation from working directory diffs.
• Skill Creator (skill-creator): Interactive creation, evaluation, and optimization of custom agent skills.

📓 Interactive Notebooks & Polyglot Data Science

• 🚀 Native Kotlin Scripting Kernel:
• Interactive Kotlin notebooks (.kts) powered by the Kotlin Custom Scripting Host API with full access to SMILE algorithms.
• ⚡ AST-Aware Scala 3 Evaluation:
• ScalaKernel now parses code using the Scala 3 compiler to evaluate complete top-level classes, multi-line blocks, and closures without timeouts or truncation.
• Interactive Welcome & Release Notes:
• Replaced the blank starting file with an interactive Welcome tutorial and dedicated Release Notes viewer.

📈 Machine Learning, Analytics & Core Engine

• Time Series Analysis (ARIMA & ARMA):
• Added smile.timeseries.ARIMA supporting arbitrary differencing, stationary transformation, and recursive multi-step forecasting.
• Added model selection diagnostics: Gaussian log-likelihood, AIC, BIC, RSS, and coefficient $t$-test tables.
• Added pure MA fitting via Hannan-Rissanen stage-2 least squares in smile.timeseries.ARMA.
• Unit-Hypercube Hyperparameter Optimization (HPO):
• Added smile.hpo.ParameterSpace supporting continuous, integer, and discrete search spaces normalized into a unit hypercube $[0, 1]^d$, with uniform candidate sampling and elite local perturbations.
• Local Outlier Factor (LOF) Anomaly Detection:
• Density-based anomaly detection via smile.anomaly.LOF<T> across arbitrary metric spaces.
• Tree Ensemble Feature Selection:
• Introduced smile.feature.selection.EnsembleSelection to easily score, rank, and threshold feature importance from Random Forests, Gradient Boosted Trees, and Decision Trees.
• Java 25 Spark Module:
• Modernized spark module rewritten in pure Java 25.

🔌 Extensibility & Marketplace

• Claude Code-Style Plugin Marketplace:
• Integrated plugin marketplace dialog in Studio, seeded with official plugin catalogs and discovery tools.
• Scoped MCP Integration:
• Support for Model Context Protocol (MCP) servers with per-agent scoping via x-smile-agents and automatic omission of redundant provider tools via x-smile-provides.

🔧 Improvements

• Multi-Model LLM Management & Reasoning:
• Unified model catalog supporting OpenAI, Anthropic Claude, Google Gemini, and local Ollama models.
• Per-agent model selection with dynamic reasoning effort retry escalation (Default → High → Low).
• Clean separation of reasoning/thinking tokens to keep output panels concise.
• Chunked Conversation Compaction:
• Multi-stage conversation compaction summarizes lengthy chat histories without losing working context or interrupting execution.
• Agent Session Management & UX:
• /resume command and visual session picker to restore and continue past Studio conversations.
• Responsive cancellation: stop buttons provide immediate feedback and re-arm cleanly.
• Real-time queuing indicators (Queued vs. Thinking) with background thread dispatching.
• Notification badges when background agents request user input.
• Editor & Studio Usability:
• Unified Find/Replace dialog for both Notebooks and Notepads.
• External file change detection: cleanly reloads files modified on disk if not dirty in memory.
• Debounced background autosave for project files.
• Atomic Notebook Persistence:
• Safe .ipynb serialization writes to temporary files before atomically replacing target files, preventing corruption during system interruptions.
• Polyglot API Surface:
• Expanded idiomatic Kotlin and Scala wrappers covering smile.timeseries, smile.anomaly, smile.hpo, smile.data, and plot.swing.

🐛 Bug Fixes & Reliability

• Agent Tool Hardening:
• Added strict input validation and null guards across all built-in tools.
• Security & SSRF Protection:
• Blocked link-local and cloud-metadata addresses in WebFetch to prevent SSRF vulnerabilities.
• Memory & Visual Artifact Management:
• Fixed HTTP 413 payload errors during multi-turn vision tasks by replacing in-memory screenshot bitmaps with compact metadata placeholders between tool loops.
• Windows Platform Stability:
• Handled Windows user-mapped memory sections with automatic write-lock retry logic.
• Corrected Windows virtual key code mapping (VK_WINDOWS) and input chord handling in desktop automation.
• macOS Metal Rendering Pipeline:
• Resolved display sleep/wake crash issues on macOS Metal displays (CVDisplayLink).
• Jackson Polymorphic Serialization:
• Resolved type ID conflicts on Jupyter cell and output definitions in base/smile.util.ipynb.
• Stream Output Synchronization:
• Fixed race condition where REPL output streaming could lag behind prompt notifications during rapid resets.
• ONNX Model Crash Prevention:
• Prevented process termination on SIGINT/Ctrl+C during native ONNX and Torch model execution.

Getting Help & Feedback

• Read the Welcome tab in SMILE Studio for a quick tour of features and keyboard shortcuts.
• Explore documentation and guides: https://www.aihalo.dev/
• Join the community and report issues: https://github.com/haifengl/smile

Full Changelog: https://github.com/haifengl/smile/compare/v6.3.0...v7.0.0