The Git Times

FaceSwap has shipped v3.0.0 with an installer that removes most manual configuration steps for its Python-based deep learning face swapping pipeline. The faceswap_setup_x64.exe downloads and provisions Git, MiniConda, PyTorch and the appropriate GPU libraries, then places a desktop shortcut that launches straight into the GUI.

Nvidia users receive a self-contained CUDA 11.8+ and cuDNN environment inside Conda. AMD users get the ROCm Torch build; on Windows this still requires WSL2 with ROCm 6.0-6.4 installed. The release notes stress that the process can appear to hang while large packages download, but the entire prerequisite chain completes without administrator rights.

The core workflow remains unchanged: extract faces, train a model (Phaze-A and Villain architectures are available), then convert. Neural networks handle alignment, feature extraction and final compositing on both images and video. Eight years after the project made face-swapping accessible to non-academics, the v3.0 installer lowers the hardware and system-administration barriers that previously limited experimentation.

Maintainers continue to stress legitimate applications in film VFX, academic research and digital artistry, pointing to an active forum and Discord for support. The source remains fully open for audit.

**

Scikit-learn has released version 1.8.0, adding support for free-threaded CPython along with compatibility for Python 3.11 to 3.14. The update allows the machine learning library to run on experimental Python builds without the global interpreter lock, promising improved parallelism for suitable workloads.

The project provides production-ready implementations of dozens of algorithms. These range from simple linear regression to complex ensemble methods like gradient boosting and random forests. All follow a uniform interface centered on fit and predict methods.

Data pipelines benefit from built-in tools for feature scaling, imputation and encoding. Model evaluation utilities include cross-validation and hyperparameter tuning capabilities that streamline development workflows.

This release arrives as many teams migrate to newer Python versions. Free-threaded support represents a proactive step that could yield performance gains in CPU-bound tasks common in machine learning.

Installation is unchanged. The command pip install -U scikit-learn fetches the latest version, while conda users run conda install -c conda-forge scikit-learn.

Maintenance remains strong, with the release accompanied by updated benchmarks and comprehensive documentation. The changelog highlights numerous incremental improvements that enhance stability and usability across the ecosystem.

For practitioners, the library continues to bridge the gap between research prototypes and deployable solutions. Its maturity after 15 years of development provides confidence for critical applications in finance, healthcare and technology sectors.

Dify 1.14.0 introduces real-time collaboration to its visual canvas for agentic workflows. Multiple workspace members can now edit the same graph concurrently, with live synchronization, online presence indicators, and visibility into who is modifying which nodes.

Self-hosted instances disable the feature by default. Administrators must set ENABLE_COLLABORATION_MODE=true, configure SERVER_WORKER_CLASS=geventwebsocket.gunicorn.workers.GeventWebSocketWorker, and supply the correct NEXT_PUBLIC_SOCKET_URL for WebSocket connectivity.

The release adds a Service API for human-in-the-loop interactions, allowing programmatic insertion of human review steps into production flows alongside existing console controls. MCP tool improvements refresh metadata automatically after updates, fix double /v1 URL segments that broke OAuth, and correctly map checkbox and json_object types during schema publishing. Plugin handling now persists auto-upgrade strategies and tightens tenant scoping for internal API calls.

These changes respond to the practical demands of scaling agentic applications. Teams building complex RAG pipelines, multi-LLM orchestration, and tool-augmented agents benefit from reduced merge conflicts and faster iteration. The platform’s existing strengths—visual workflow editor, support for GPT-4, Mistral, Llama 3 and OpenAI-compatible models, prompt IDE, and observability hooks for Langfuse and Arize Phoenix—gain new relevance for coordinated enterprise development.

Gazebo Sim has shipped its Jetty release, gz-sim10_10.0.0, the newest iteration of the open-source robotics simulator maintained by Intrinsic. The update refines components that trace their lineage through 16 years of continuous development from Gazebo Classic.

The simulator supplies high-fidelity dynamics via Gazebo Physics, supporting multiple engines selected at runtime. Gazebo Rendering employs OGRE v2 to generate production-grade lighting, shadows and material textures. Gazebo Sensors produces synchronized data streams from laser scanners, RGB-D cameras, IMUs, force-torque units and GPS, each with tunable noise profiles.

Plugin interfaces let developers inject custom logic for robot behavior, sensor fusion or environment control without altering core binaries. The Gazebo GUI, itself plugin-driven, provides interactive scene building, real-time introspection and drag-and-drop model placement. TCP/IP message passing through Gazebo Transport allows simulation to run on remote clusters while clients on local workstations issue commands or subscribe to sensor topics.

Command-line utilities have been extended for batch testing and headless CI runs. Ready-to-use models including PR2, TurtleBot and iRobot Create are available from Gazebo Fuel; new assets are defined with SDF. For teams moving between simulation and hardware, Jetty tightens the ROS 2 bridge and reduces the iteration cost of perception and control validation.

The release demonstrates that mature open-source simulation stacks continue to evolve in lockstep with industry requirements rather than being displaced by proprietary alternatives.

XRobot has shipped version 0.3.1, updating its generated module templates to match the current LibXR STDIO literal API, switching CI templates to C++20, and dropping stdlib argparse from runtime dependencies. The changes reduce the tool's footprint and eliminate compatibility friction that had accumulated since its initial 2022 release.

The toolkit continues to automate the boilerplate that consumes engineering time on embedded robotics projects. It manages module repositories with recursive dependency resolution and version-consistency checks, parses constructor and template parameters from C++ headers, then emits editable YAML files. From those files it produces a complete XRobotMain() entry point supporting multiple modules, repeated instances, and nested argument structures.

A manifest parser extracts descriptions, dependencies, and parameter data directly from header comments. A one-command generator creates standardized module directories containing headers, READMEs, and CMakeLists files. These capabilities remain especially useful for teams targeting STM32, ESP32, Raspberry Pi, and Webots-based simulators.

The v0.3.1 refinements matter because modern C++ standards and evolving LibXR conventions now propagate automatically into new projects. Developers spend less time reconciling API drift and more time on control algorithms and hardware integration. By tightening the feedback loop between module definition and executable entry point, XRobot keeps large robotics codebases maintainable as both hardware and library requirements change.

**

CLOiSim, the Unity-based multi-robot simulator, has moved fully to Unity 6 with the 5.1.3 release. Earlier branches tied to Unity 2022.3 LTS are no longer maintained, and the main branch now tracks the 5.x series built on 6000.4.0f1.

The simulator ingests SDF files to instantiate complete 3D environments and robot descriptions, then connects to ROS2 via the companion cloisim_ros package. It implements joint models with control and physics, plus visual, collision, and sensor components directly inside Unity. This architecture was created to overcome the frame-rate collapse Gazebo experienced when loading several robots carrying multiple sensors.

Version 5.1.3 delivers targeted efficiency gains. Sonar sensing now uses RaycastCommand.ScheduleBatch() with NativeArray structures, eliminating per-frame sequential Physics.Raycast loops. The MicomSensor pools Bumper, USS, and IR sub-messages to remove repeated GC allocations in GenerateMessage(). Editor gizmos for GroundTruth and Link have been refined—Link’s center-of-mass sphere now scales with mass contribution—and object-initializer syntax has been adopted across plugins, devices, and the SDF pipeline.

New Copilot skill documents help developers extend compute shaders, noise models, and UI modules. For robotics teams running fleets with dense sensor arrays, these changes translate into sustained real-time performance where earlier simulators faltered.

**

Matomo 5.9.0 landed this week with targeted stability improvements and refined visualisation components. The PHP and MySQL application, formerly known as Piwik, remains the primary self-hosted alternative for organisations unwilling to hand visitor data to third-party processors.

Installation still follows the familiar five-minute process: upload to a server running PHP 7.2.5 or newer and MySQL 5.5+, run the setup, then embed the provided JavaScript tag. Once live, the system logs traffic, session paths, device types and campaign outcomes in real time. Administrators gain immediate access to raw logs, heatmaps and conversion funnels without external API calls.

The release reinforces built-in consent management, IP anonymisation and configurable data-retention policies. These features matter as GDPR enforcement actions and CCPA-style rules proliferate across jurisdictions. Because Matomo runs on infrastructure operators control, export and deletion requests can be satisfied directly from the database.

Development continues at steady pace more than 15 years after the initial commit. Contributors working in the Git repository can spin up a full DDEV environment for testing changes before submitting pull requests. For teams avoiding operations overhead, the hosted Matomo Cloud option provides an on-ramp with a 21-day trial.

The 5.9 update demonstrates that mature open-source infrastructure can keep pace with commercial analytics suites while preserving the original mission of ethical data stewardship.

Dirsearch has received a material update with the v0.4.3 release. The Python web path scanner, actively maintained by maurosoria and shelld3v since 2013, now removes several operational friction points that security teams encounter during extended reconnaissance.

The tool automatically selects the correct URI scheme (http or https) when none is supplied, eliminating a common misconfiguration. New options let users overwrite unwanted extensions, inspect complete redirect histories, and crawl paths extracted from server responses. Session management allows scans to be saved and resumed later, an essential capability when processing large wordlists that exceed reasonable time windows.

Reporting has been expanded with SQLite output alongside existing formats. All HTTP traffic is written to a log file whose maximum size can be capped through configuration. Client certificate authentication is now supported, and the maintainers report measurably higher accuracy in response classification.

Platform support remains comprehensive. Security engineers can run the tool with Python 3.9 or higher, deploy self-contained binaries for Linux x86_64, Windows x64, macOS Intel or Apple Silicon, or build Docker images. These changes arrive as web attack surfaces continue to fragment across hidden APIs, administrative interfaces and forgotten endpoints. The refinements keep dirsearch effective for structured enumeration without altering its core brute-force approach.

Installation is unchanged: a shallow git clone remains the recommended route, with pip and standalone binaries as alternatives.

TheSpeedX/PROXY-List pushed its latest update on 29 April 2026, bringing the total number of compiled proxies to 6,358. The repository has maintained near-daily refreshes since its creation in 2018, pulling public proxies from across the internet and publishing them in protocol-specific text files.

Direct raw links allow immediate consumption. SOCKS5, SOCKS4 and HTTP lists sit at predictable GitHub URLs, each containing IP:port pairs ready for scripts. A companion Python validation tool is recommended to test functionality before use.

The maintainer notes the lists are provided for educational purposes only and disclaims responsibility for proxy behaviour or downstream activity. Users are asked to credit the repository with stars and follows when integrating the data.

The resource matters now because public proxies typically expire within hours. Consistent updates reduce the manual effort required to maintain working pools for network experiments. Plain-text format enables quick integration into rotation libraries without additional parsing overhead.

Current lists

• SOCKS5: https://raw.githubusercontent.com/TheSpeedX/SOCKS-List/master/socks5.txt
• SOCKS4: https://raw.githubusercontent.com/TheSpeedX/SOCKS-List/master/socks4.txt
• HTTP: https://raw.githubusercontent.com/TheSpeedX/SOCKS-List/master/http.txt

Lazygit maintainers have released v0.61.1, delivering targeted fixes to the GitHub pull request integration added in recent versions. The update hides closed pull requests on main branches, normalizes repository owner casing to prevent lookup failures, and stops defaulting the base repository to "origin". These adjustments resolve real-world friction when working with forks, organizations, and inconsistently named remotes.

Since 2018 the Go-based terminal UI has simplified git's most cumbersome operations. Instead of editing TODO files for interactive rebasing, users navigate and reorder commits visually. Staging individual lines or splitting hunks requires only keystrokes rather than hand-crafted patch files. The interface further supports cherry-picking, bisect sessions, worktree management, amending older commits, and "rebase magic" for custom patch application.

A commit graph view and side-by-side commit comparison aid history navigation. This release also adds a justfile for developer convenience and includes a security fix that removes vulnerable variable interpolation from GitHub Actions workflows. Two first-time contributors supplied patches.

For developers who live in the terminal, the changes tighten an already mature tool, reducing context switches between CLI and browser while preserving the speed that made lazygit popular.

The OpenWrt project has released version 25.12.2, the second service update in the 25.12 stable series. The patch focuses on device support and stability fixes for hardware that manufacturers have long since abandoned.

Key corrections include renaming the airoha PWM kernel module from kmod-pwm-an7581 to kmod-pwm-airoha, requiring reinstallation for affected users. On apm821xx targets, U-Boot environment definitions were fixed for the NETGEAR WNDR4700, Western Digital MyBookLive, Meraki MR24 and MX60, resolving PCIe boot failures on the latter. Ath79 platforms now correctly handle initramfs boot for Huawei AP5030DN and AP6010DN devices and VLAN CPU port tagging on dual-CPU switch configurations. Mikrotik RB750r2 images no longer ship unused WiFi packages.

ipq40xx fixes restore proper ART partition naming on Linksys Velop WHW03 V1 for WiFi calibration data and improve MAC address reading from eMMC-based NVMEM on other Linksys models. Lantiq xrx200 gains corrected failsafe mode for BT Home Hub devices.

The Firmware Selector remains the fastest way to locate compatible factory images. Builders still need a case-sensitive filesystem and baseline tools (gcc-6+, make4.1+, python3.7+, rsync) to compile from source.

A decade after its creation, OpenWrt continues to matter because its writable filesystem and package manager let users and developers treat routers as programmable platforms rather than fixed appliances.

gRPC has released version 1.80.0, delivering concrete improvements to its core security layer and runtime components. The update implements a TLS private key signer for Python and introduces private key offload for TLS credentials. An InMemoryCertificateProvider now allows certificates to be refreshed independently of other configuration.

Runtime changes include enabling EventEngine by default for Python, with added fork support for both Python and Ruby. Load-balancing logic for round-robin and weighted round-robin was adjusted to permit random index starts. The xDS implementation advances with gRFC A101 compliance, while DNS resolver code received a correctness fix.

Ruby users gain official builds and native gems for version 4.0. A platform fix corrects maximum sockaddr struct sizing on OpenBSD.

More than eleven years after its initial release, gRPC continues to serve as the standard high-performance RPC layer for production distributed systems. Built on a C++ core and HTTP/2 transport, it abstracts connection management, flow control and binary serialization so engineering teams can focus on service logic rather than transport plumbing. The 1.80.0 changes strengthen its credentials handling and language runtime stability, areas under pressure as organizations move more workloads to zero-trust environments and heterogeneous language stacks.

The release contains only targeted refinements and bug fixes with no breaking API changes.

openwifi has released v1.5.0, backed by comprehensive testing that positions its FPGA design as equal to or better than commercial off-the-shelf Wi-Fi chips. The NLnet-funded maturity effort produced a 47-page benchmark report detailing head-to-head results against proprietary silicon.

Key FPGA improvements target indoor multipath environments. New DSP algorithms provide finer frequency offset estimation, more robust time-frequency tracking, and low-complexity LLR calculation in the equalizer for soft decoding. Designers removed FIFO buffers from ADC and DAC paths, yielding deterministic IQ sample timing critical for distributed MIMO and radar sensing.

The CSI fuzzer received bug fixes that now enable reliable channel response manipulation at the transmitter. Timing closure was optimized across the CSMA/CA module and PHY, allowing the design to run efficiently on the lowest-cost Xilinx Zynq 7020 while preserving 10 µs SIFS performance in hardware.

The stack remains fully compatible with Linux mac80211, supporting 802.11a/g/n at 20 MHz bandwidth across 70 MHz to 6 GHz. It implements DCF in FPGA, offers real-time CSI and IQ capture to userspace, and exposes extensive runtime tuning of CCA thresholds, interframe spaces, and queue scheduling.

Developers must still observe local spectrum rules or use cabled setups. The project maintains AGPLv3 licensing for open-source work alongside commercial options.

GPU-T has received a targeted usability update in version 0.1.4, four months after its initial release. The C# application built with Avalonia UI continues to serve Linux users seeking GPU-Z-style detail on AMD hardware without root privileges or external dependencies.

The self-contained AppImage now offers a manual theme toggle with Auto, Dark, and Light options, decoupling appearance from system-wide settings. Both the Advanced and Sensors windows permanently remember their resized heights, storing preferences in ~/.config/GPU-T/gpu_t_settings.json for consistent layouts across restarts. Build scripts have been streamlined and ICU globalization libraries added, delivering reliable text rendering even on minimal distributions such as Alpine Linux.

The embedded hardware database has grown with specifications for the AMD Radeon Graphics 448SP Mobile (Barcelo). Core functionality remains unchanged: GPU-T reads directly from sysfs, Vulkan, and its JSON lookup table to report die size, transistor count, clock speeds, hotspot and edge temperatures, fan RPM, PPT power draw, ReBAR status, and support for ROCm, OpenCL, and ray tracing. Sensor data can still be logged to file.

These incremental changes reduce daily friction for users who treat the tool as their single source of truth for AMD GPU verification and monitoring.

FireSim’s latest point release strengthens its ties to Chipyard, reflecting the project’s steady evolution as the preferred FPGA-accelerated simulation platform for RISC-V hardware teams. Version 1.20.1 now invokes Chipyard’s makefile directly from FireSim’s build system, removing an earlier layer of indirection that complicated iterative RTL development. Paper workloads previously maintained inside FireSim have migrated to Chipyard, centralizing reference benchmarks and easing cross-project maintenance.

A targeted fix resolves a missing workload definition for the enumeratefpgas command, restoring reliable operation on multi-FPGA setups. These refinements matter because FireSim routinely delivers 10s to 100s of MHz simulation speeds on real FPGA hardware, far outpacing software simulators when validating full SoCs.

Users configure everything from a single Xilinx Alveo card on a desktop to hundreds of Amazon EC2 F1 instances simulating datacenter-scale racks. The platform co-simulates ASIC RTL with cycle-accurate models of DRAM, networks, and peripherals, enabling realistic profiling of BOOM and Rocket Chip designs under production-like conditions.

More than 40 papers from 20-plus institutions demonstrate its ongoing value in architecture, systems, networking, and security research. With the 1.20.1 changes, the toolchain becomes marginally smoother for the growing community that already depends on it.

The Open Industry Project has released v4.7-beta1, updating its free warehouse and manufacturing development framework to Godot 4.7-beta1. The new version introduces live snapping while dragging, allowing parts to connect to compatible neighbors in real time rather than requiring separate alignment steps. Curved conveyors now auto-flip their reverse_belt property when snapped in opposing orientations, maintaining consistent material flow without manual correction.

The package provides an executable build of the project's custom Godot fork together with template project files. Users extract the archive, launch the included editor, and start new simulations via the "New Simulation" option. Parts from the dedicated tab are dragged into the 3D viewport where they inherit configurable properties for industrial communication.

The framework supports OPC UA through open62541, Ethernet/IP and Modbus TCP via libplctag, and Siemens S7 1200/1500 Put/Get. Simulation objects can connect directly to PLCs or SCADA systems such as Ignition, enabling realistic testing of control logic against virtual equipment.

Three years after its initial release, the project continues to lower barriers for developers and educators who need to prototype manufacturing systems using standard industrial protocols. The beta remains open for issue reports to guide final 4.7 development.

NVIDIA has released version 4.17.3 of its NRD library, delivering concrete improvements to the five-year-old spatio-temporal denoising toolkit still embedded in more than 15 AAA titles and professional visualization applications.

The headline changes target REBLUR. Performance is improved, noise and pixelation on specular lobe boundaries are reduced, and hit distances are now left spatially unprocessed to minimize bias while preserving shadow detail. Occlusion-only modes are unaffected. The NRD_REJITTER_AMPLITUDE default drops from PI to 2, and several internal calculations around virtual history confidence and motion-vector selection have been simplified to single-register operations.

Two breaking changes require integration updates. IN_BASECOLOR_METALNESS and its downstream logic have been removed, and the legacy ReblurHitDistanceParameters::D is gone; smc now manages the remapped roughness internally. These clean-ups were stress-tested in the Dark Souls 2 path-tracing integration.

NRD remains API-agnostic HLSL code that ingests per-pixel G-buffer data (normal, roughness, viewZ, motion vectors) to denoise diffuse and specular radiance, including Spherical Harmonics variants that rival DLSS-RR quality without machine learning. RELAX supports RTXDI workflows, while SIGMA handles shadow denoising for both infinite and local lights. On an RTX 4080 at 1440p native, the default settings continue to deliver real-time results suitable for production engines.

The sample continues to demonstrate a denoising-free glass path that combines SHARC, temporal reprojection and dithering with TAA or upscaling.