The Git Times

The google-gemini/cookbook has added fresh Jupyter Notebook tutorials that reflect the latest Gemini API capabilities, giving developers concrete code patterns for recently released models.

Lyria 3 now features dedicated quickstarts for turning text prompts into 30-second clips or full songs, including image-to-music conversion and fine-grained control over musical structure. Nano-Banana 2 and Nano-Banana Pro receive expanded coverage demonstrating native image generation at 512px for speed or 4K for quality, with new examples of consistent image editing and visual storytelling that incorporate thinking and search grounding.

A new inference tiers guide explains how to choose between Priority and Flex options to balance latency, cost and reliability. The File Search quickstart shows how to build a hosted retrieval-augmented generation system that grounds model output in an organization's own documents. Additional notebooks demonstrate factual grounding using Google Maps data for location-aware applications.

The repository retains its established structure of Quick Starts for individual features and Examples that combine multiple capabilities, with separate demo repositories illustrating end-to-end implementations. These additions arrive as teams move beyond basic prompting into production multimodal workflows that require precise control over audio, visual and retrieval components.

The transformers library has incorporated two new model architectures in version 5.4.0, extending its computer vision capabilities.

VidEoMT is a lightweight encoder-only model for online video segmentation built on a plain Vision Transformer. It eliminates dedicated tracking modules by using a query propagation mechanism that carries information across frames and a query fusion strategy combining propagated queries with temporally-agnostic learned queries. The model achieves competitive accuracy while running 5x to 10x faster than existing approaches, reaching up to 160 FPS with a ViT-L backbone.

UVDoc performs document image rectification through geometric transformations that correct distortion, inclination and perspective deformation. It supports both single-input and batched inference, enabling efficient processing of multiple documents.

These additions reinforce the library's role as the standardized model-definition framework. Definitions remain compatible with training frameworks such as Axolotl, Unsloth and DeepSpeed, inference engines including vLLM and TGI, and libraries like llama.cpp. The centralized approach allows teams to adopt new architectures without rewriting core logic for different runtimes.

Installation uses standard virtual environments with pip install "transformers[torch]" or equivalent uv commands. The updates address growing demand for real-time video analysis and automated document workflows while preserving the project's emphasis on simplicity and efficiency.

The Julia project has released v1.12.5, a maintenance update that resolves issues identified since v1.12.4. The patch refines runtime stability and compiler behavior for the high-performance dynamic language built for technical computing.

Julia lets developers write high-level code that achieves C-level speed through just-in-time compilation and multiple dispatch. This design eliminates the usual trade-off between productivity and performance, letting researchers express mathematical algorithms directly while the juliaup tool manages installation and version switching.

The language provides native support for multidimensional arrays, parallel computing primitives, and efficient linear algebra routines. These features make it a practical choice for compute-intensive workloads where both speed and expressiveness matter. The project continues to welcome contributions ranging from bug fixes to documentation, with clear guidelines for new participants.

As demands for efficient simulation and machine-learning pipelines increase, this incremental release keeps the toolchain reliable for production use in laboratories and industry. Binary downloads and community resources remain available through julialang.org.

(168 words)

The ros-mcp-server has released v3.0.1, incorporating updates to its continuous integration framework. This builds on the major v3.0 improvements that strengthened its ability to connect large language models with robots.

Operators no longer need to alter robot source code. Installation involves adding a rosbridge node to the existing setup, enabling immediate bidirectional links with models such as Claude, GPT and Gemini.

Full context awareness allows LLMs to explore available topics, services and actions. The server automatically provides type information for custom messages, allowing correct syntax usage without manual input.

Recent enhancements improve stability across multiple ROS distributions. The tool supports both ROS 1 and ROS 2 versions including Jazzy and Humble.

In industrial settings, engineers use the system to diagnose robot states. They query topics and call custom services through natural language instructions.

Simulation teams leverage it with NVIDIA Isaac Sim to control the MOCA mobile manipulator. Commands entered in Claude Desktop translate directly into robot actions.

The open MCP standard ensures compatibility with various clients including Cursor and ChatGPT interfaces.

These changes come as robotics applications demand more sophisticated AI integration for observation and control tasks.

The maintainers of roslibjs have released version 2.1.0, introducing Action.cancelAllGoals() to the standard ROS JavaScript library. The new method, contributed by ikwilnaarhuisman, allows developers to terminate all active action goals with a single call rather than managing each one individually.

This change addresses a practical pain point in complex web-based robotic systems where multiple concurrent actions are common. roslibjs provides the canonical TypeScript implementation for connecting browser applications to ROS via WebSockets. It supplies classes for topics, services, actions, parameters, and TF transforms, enabling real-time publish and subscribe operations.

The library has served as the foundational client for web-robotics integration since 2013. Version 2.1.0 also includes extensive maintenance: dependency bumps to typescript-eslint, vite, jsdom and fast-png, plus a correction to a typo in the sendGoal implementation. These updates keep the project compatible with current web tooling.

Packaged as a monorepo with dedicated examples, roslibjs continues to support teams building browser interfaces for physical robots. The improvements ensure more reliable action lifecycle management without altering existing APIs.

As web technologies assume greater importance in robotics operations, incremental enhancements like this help the 13-year-old library remain relevant for modern development stacks.

The SSG-48 adaptive electric gripper has received its first official release with version 1.1. Thoroughly tested alongside the PAROL6 robotic arm, the update confirms reliable performance in practical deployments.

The device employs Spectral micro BLDC drivers to modulate gripping force between 5 N and 80 N. This range supports both delicate components and rigid parts, making it suitable for assembly lines and human-robot collaboration. A 48 mm stroke and 400 g mass give the gripper a practical balance of reach and portability.

All mechanical files, firmware and control software remain fully open source. Users can modify the end effector or integrate the unit with other robotic platforms. The project supplies complete documentation including building instructions, BOM, Python API, ROS2 packages, plus URDF and MJCF simulation files.

Control uses straightforward CAN communication. Typical code calibrates the mechanism then issues position and force commands in a simple loop. The latest release removes earlier stability issues identified during PAROL6 integration, allowing immediate use in both industrial automation and research settings.

Why it matters now: verified compatibility with popular open-source arms lowers the barrier for teams building complete robotic cells without proprietary hardware.

Trickest's automated CVE aggregator has strengthened its data collection mechanisms to keep pace with accelerating vulnerability disclosures. The repository pulls details directly from cvelist and systematically hunts for PoCs using refined techniques.

The workflow splits entries by year and deploys dual approaches for discovery. It parses References fields with a regex pattern (?i)[^a-z0-9]+(poc|proof of concept|proof[-_]of[-_]concept)[^a-z0-9]+ via ffuf. Additionally, it queries GitHub for repositories referencing the CVE ID through find-gh-poc.

Integration with HackerOne data now pulls video PoCs from public reports, expanding coverage of real-world exploits. Automated merges preserve hand-curated content while applying filters from blacklist.txt to reduce noise.

Generated markdown files feature version badges from shields.io, making impact assessment straightforward. An Atom feed enables monitoring of CVEs affecting chosen products or vendors.

Security teams leverage this to maintain current exploit libraries without manual trawling across disparate sources. The HTML template included allows quick generation of searchable tables for internal use.

Infisical released v0.159.5 with concrete upgrades to its privileged access management features. The update adds SSH PAM exec and SFTP support, letting operators run commands and transfer files through the platform's PAM integration. Administrators can now terminate active sessions directly, providing immediate revocation when security events occur or access must be cut.

Backend changes optimize recursive secrets fetching, lowering database usage for deployments with deep project hierarchies. This performance tweak addresses real operational costs as teams scale secrets across development, staging, and production environments.

Interface improvements include a redesigned project search modal and adjusted rotation dialog height for better usability. These changes arrive alongside existing capabilities such as dynamic secrets for PostgreSQL, MySQL, and RabbitMQ, automated credential rotation for AWS IAM, and point-in-time recovery of project states.

Certificate management continues to support internal private CA hierarchies and external authorities including Let’s Encrypt. The Kubernetes Operator and lightweight Agent deliver secrets to workloads without application code changes, while built-in scanning blocks leaks before they reach Git.

The release reflects steady evolution of the three-year-old platform toward tighter integration between secrets, certificates, and access controls.

Lazygit has released version 0.60.0 with practical refinements to its terminal interface for everyday git operations. Seven years after its initial launch, the Go-based tool continues to address persistent pain points that make raw git commands frustrating for many developers.

The update allows users to remove lines from patches directly, eliminating the need to hand-edit arcane patch files when a hunk cannot be split further. File views now support proper filtering rather than simple search, speeding up navigation in large repositories. Worktree handling has been enhanced to show branch names and detached HEAD status in the worktrees tab, while the branches list displays worktree names alongside each branch.

Additional changes include backward cycling support in the log view using shift-A, clearer labelling of the abbreviated commit hash copy function, and several fixes for URL matching, directory creation, and panel sizing.

These improvements build on lazygit's established strengths: visual interactive rebasing that avoids manual TODO file editing, straightforward cherry-picking, and the ability to stage individual lines without complex command sequences. The project maintains a consistent focus on reducing context switches between terminal and editor.

For developers who spend significant time in the terminal, the release delivers incremental but meaningful gains in usability without altering the core experience.

Nanobrew is a package manager for macOS and Linux written in Zig. It achieves warm installs in 3 milliseconds for packages already present in its local store.

The tool uses Homebrew formulas and bottles under the hood, functioning as a drop-in replacement for common workflows. It adds native .deb support for Linux and Docker containers, where it performs up to 13 times faster than apt-get on warm installs.

Key features include parallel downloads of all dependencies and a single 1.2 MB static binary with no Ruby runtime. The nb install command performs installations without triggering updates, which users must invoke explicitly through nb update. Cask installations on macOS skip the com.apple.quarantine attribute, so applications launch without Gatekeeper prompts. Third-party taps work directly with syntax such as nb install user/tap/formula.

Version 0.1.083 adds the nb leaves [--tree] command for listing packages with no installed dependents and fixes self-dependency cycle detection for packages including r, pnpm and texlive. It also corrects path rewriting in installed scripts and improves shell completions.

The project deliberately omits post_install hooks, build-from-source options and Mac App Store integration to maintain speed on the pre-built binary path. nb bundle install returns immediately when a Brewfile is already satisfied.

Prometheus has released version 3.10.0, introducing a distroless Docker image variant alongside the existing busybox image. The new option uses a minimal base for stronger security, runs with UID/GID 65532 instead of nobody, and removes the VOLUME declaration. Images are available with -distroless and -busybox suffixes, preserving the original as the default tag for compatibility.

Migration from named volumes requires adjusting ownership. The recommended command runs an Alpine container to execute chown -R 65532:65532 /prometheus before launching the distroless image.

The release adds an alertmanager dimension to notification metrics including prometheus_notifications_dropped_total. The UI now hides expanded alert annotations by default on the /alerts page, improving information density. New functionality includes MSK role support in AWS service discovery and PromQL binop modifiers fill(), fill_left() and fill_right().

These changes strengthen an architecture built on a multi-dimensional data model, autonomous single-server nodes, HTTP pull collection, and service discovery. After more than 13 years of development under the Cloud Native Computing Foundation, Prometheus continues adapting its core monitoring capabilities to current container and cloud requirements.

The Glasgow Interface Explorer continues to serve as a powerful asset for hardware developers seeking adaptable interfacing solutions. Its combination of FPGA hardware and Python software allows for on-the-fly reconfiguration to match specific electronic protocols.

Recent activity shows the project gaining momentum, with updates pushed as recently as April 2026. Plans are underway to expand the current team of three maintainers, which should help address the backlog of feature requests and support for new users.

Connecting through USB, the device provides multiple I/O channels that can be programmed to act as virtually any digital interface. Users create applets in Python that generate the necessary FPGA bitstreams and handle data processing on the host side.

This architecture supports a broad range of tasks without specialized equipment. Engineers benefit from the ability to quickly iterate on interface designs and debug communication issues in real time.

Key technical advantages:

• Dynamic FPGA reconfiguration for different protocols
• High-speed operation suitable for modern interfaces
• Extensive Python API for automation and scripting

As hardware projects grow more sophisticated, the need for such versatile tools increases. Glasgow reduces barriers for those working with legacy or obscure components by providing a single platform capable of many roles.

The comprehensive documentation manual assists both newcomers and experienced users in leveraging its full potential, from initial hardware setup to developing custom extensions.

micrOS 3.0 focuses on core stability and modern communication paths while preserving its lightweight footprint for DIY embedded projects.

The update integrates async SSL/TLS support into the HTTP client, with a redesigned urequests module that requires MicroPython 1.22 or later. This enables secure notification flows through LM_telegram and listener-based chatbot patterns without external services.

ESP-NOW peer-to-peer capability has been added to the InterCon module, providing lower-latency device-to-device messaging as a fallback to WiFi. The new command syntax simplifies remote execution — rgb toggle >>RingLight.local replaces the previous verbose intercon sendcmd format. The system automatically selects ESP-NOW when available on the target.

A multi-level filesystem layout now organises content into dedicated directories: /lib, /logs, /web, /data, /config and /modules. These are created automatically at boot and over USB, improving package management and long-term maintainability.

The platform continues to expose MicroPython functions through ShellCli socket sessions and WebCli REST endpoints. Its async task manager supports timestamp-based scheduling, geolocation-derived sunrise and sunset timing, and simple periodic tasks. All operations remain confined to the local WiFi network.

(178 words)

Luanti has issued 5.15.1, a bugfix release aimed at enhancing the stability of its open-source voxel game-creation platform. The update follows the project's rebranding from Minetest and focuses on polishing the user experience without introducing major new features.

Built in C++ with Lua for scripting, the engine empowers users to design and modify voxel worlds effortlessly. Key controls allow for intuitive interaction: WASD for movement, left mouse to dig, right to place. Advanced options include fly, fast, and noclip modes activated via keyboard shortcuts.

The release notes carry an important reminder: the legacy Minetest Game is not meant to be packaged with Luanti, pushing the community toward creating original content.

Documentation covers compilation steps, configuration files, and Docker support for easy deployment.

This maintenance update reaffirms Luanti's position as a versatile tool for hobbyists, educators, and independent developers. By keeping the codebase clean and mod-friendly, it ensures continued relevance in a landscape dominated by commercial engines. Community contributions drive its evolution, with the GitHub repository serving as the hub for updates and discussions.

GDQuest has released version 1.5.2 of the learn-gdscript project, introducing several bug fixes that improve the presentation of code within its interactive lessons.

Previously, certain symbols in code examples would fail to display correctly, leaving learners confused about syntax details. The new version corrects this for operators including =, <, and >. It further refines the syntax highlighting system by adding support for strings in bbcode and fixing how numbers are presented.

The application provides a complete learning path for GDScript, Godot's native scripting language. It targets users with no prior programming experience, offering hands-on exercises that explain and demonstrate basic concepts. Students write small scripts, receive instant feedback, and gradually build their understanding of programming logic.

Because it operates in the browser, the tool removes common obstacles to getting started. No downloads or complex configurations are necessary to begin the first lesson. For those who prefer it, desktop versions deliver improved responsiveness and higher quality text display.

This update, while focused on maintenance, underscores the project's ongoing relevance for new Godot users seeking an accessible introduction to game scripting.

Tracy has received version 0.13.1, bringing targeted fixes that improve reliability for developers profiling demanding applications.

The update corrects parsing of extended model and family information from x86 CPUID instructions, ensuring accurate processor detection. It eliminates a memory corruption bug triggered by long user names on Android and fixes an incorrect function signature when TRACY_DEBUGINFOD is enabled. Mount lists are now read through the proper API instead of parsing /proc/mounts.

Other changes include proper shadow warning suppression on gcc, silent handling of lost ETW Vsync events, and workarounds for older macOS systems that lack full C++20 support. A race condition during profiler shutdown has been resolved, and memory free faults can now be suppressed with the TRACY_IGNORE_MEMORY_FAULTS option.

New capabilities include a truncated mean parameter for csvexport and an experimental viewer for the user manual. These additions complement Tracy’s existing real-time, nanosecond resolution profiling of CPU, GPU, memory allocations, locks and context switches across OpenGL, Vulkan, Direct3D, Metal and CUDA.

The release strengthens the tool’s suitability for production environments where stability across platforms and compilers is essential.

(168 words)