The Git Times

Hermes Agent has released a Tool Gateway that lets paid Nous Portal subscribers access external capabilities without managing additional API keys. Web search via Firecrawl, image generation with FLUX 2 Pro, OpenAI text-to-speech, and Browser Use automation are now available through the existing subscription. Users run hermes model, select Nous Portal, and toggle tools with the use_gateway setting. The runtime automatically prefers gateway implementations even when direct keys exist.

The integration tightens the agent’s closed learning loop. With easier access to live tools, the system can more fluidly create new skills after complex tasks, self-improve those skills in use, and incorporate fresh data into its persistent memory and Honcho user model. Full-text search across past sessions, periodic knowledge nudges, and autonomous subagent spawning all benefit from reduced friction.

Deployment flexibility remains unchanged. The agent runs on a $5 VPS, Modal or Daytona serverless instances that hibernate when idle, or any of six terminal backends. A single gateway process feeds Telegram, Discord, Slack, WhatsApp, Signal, and a rich local TUI with multiline editing and streaming output.

Version v2026.4.16 contains more than 180 commits addressing stability, CLI reliability, and tool orchestration. The gateway replaces earlier environment-variable hacks with clean subscription detection.

**

The OpenCV project has released version 4.13.0, updating its foundational open source computer vision library with targeted performance gains and improved support for contemporary neural networks.

The update refines the deep neural network module, delivering faster inference on both CPU and GPU targets while reducing memory overhead for quantized models. These changes address bottlenecks in real-time video pipelines and enable smoother integration with newer ONNX-exported architectures. The release also tightens several classical image-processing functions, including accelerated feature detectors and more stable camera calibration routines.

OpenCV supplies battle-tested building blocks for production systems. Its C++ core implements feature detection, optical flow, stereo matching, object tracking, and high-performance matrix operations that power everything from factory inspection to surgical navigation. The library's strength lies in combining classical algorithms with deep learning inference inside a single, optimized binary.

This matters now as teams ship vision models to edge devices in robotics, automotive, and medical imaging. Version 4.13.0 lowers the barrier to deploying accurate, low-latency pipelines without sacrificing the reliability that has made the project the default choice for over a decade. Contributors followed the project's established rules—one pull request per issue, clean history, tests, and documentation—ensuring the update maintains the library's production-grade stability.

Five years after its initial release, ultralytics/yolov5 continues evolving with version 7.0. The update introduces YOLOv5-seg models that set new records for real-time instance segmentation on the MSCOCO dataset, delivering both higher accuracy and faster inference than prior state-of-the-art systems.

The models maintain the project's defining simplicity. Training, validation and deployment use the same command-line interface established for object detection. After cloning the repository and installing dependencies in a Python≥3.8 environment with PyTorch≥1.8, developers can start segmentation experiments with minimal code changes.

Export pipelines remain central. Models convert directly to ONNX, CoreML or TFLite, enabling consistent performance from server GPUs to iOS devices and embedded hardware. This capability has made the project a standard choice for teams needing reliable cross-platform computer vision.

Documentation covers custom dataset training, hyperparameter tuning and inference optimization. A dedicated Colab notebook demonstrates end-to-end segmentation workflows. The v7.0 release focuses on practical production use rather than architectural novelty, lowering the cost of adding pixel-level masking to existing detection pipelines.

As Ultralytics advances the broader YOLO series toward YOLO11, v7.0 keeps the original repository relevant for teams requiring mature, well-tested segmentation tools today.

GTSAM 4.2 is now available, bringing hybrid factor graph inference to the Georgia Tech smoothing and mapping library. The update allows discrete and continuous variables to be optimized together, a capability long requested by teams building systems that must reason about both symbolic decisions and metric states.

The C++ library implements smoothing and mapping using factor graphs and Bayes networks instead of sparse matrices. This computing paradigm has powered perception and sensor-fusion pipelines in robotics and vision for nearly a decade. Version 4.2 adds substantially expanded Python and MATLAB wrappers; the Python package installs directly with pip install gtsam. It also incorporates Shonan averaging, which delivers more reliable rotation estimates critical for consistent SLAM results.

The develop branch has entered “Pre 4.3” mode. Maintainers plan to adopt C++17 and remove most Boost dependencies, with several deprecated APIs scheduled for deletion. Projects relying on the stable 4.2 release, including the roboticsbook.org curriculum, can continue without immediate code changes.

These improvements arrive as autonomous platforms demand tighter integration between learning-based perception and geometric estimation. Hybrid inference lowers the cost of combining object recognition with continuous trajectory optimization, while the polished wrappers let researchers prototype quickly before deploying performance-critical C++ code.

**

The Autoware Foundation has released version 0.50.2 of Autoware Universe, addressing a narrow but practical build issue in its mapping pipeline. The patch modifies the compare_map_segmentation component to ignore -Werror=array-bounds, eliminating compilation failures under strict compiler flags while leaving the underlying 3D map processing logic untouched.

This incremental change reflects the project's ongoing focus on production readiness. Autoware Universe supplies the bulk of functional packages that sit above the core Autoware foundation, delivering production-grade implementations for perception, planning, localization, control, sensing and vehicle interfaces. All are written in C++ and built on ROS 2, supporting calibration routines, trajectory generation and map fusion required for full autonomous operation.

Updated code coverage reporting now tracks component-level metrics more granularly, giving maintainers clearer visibility into test completeness for evaluator, simulator and system modules. For teams already running Autoware-based fleets or research platforms, the release reduces friction during continuous integration and platform upgrades.

As regulatory and technical requirements for autonomous systems tighten in 2026, these maintenance updates keep the stack reliable without forcing downstream rework. Documentation hosted via MKDocs details exact integration steps for each package.

Word count: 178

Newton 1.1.0 brings major upgrades to deformable simulation and rendering for robotics researchers.

The implicit MPM solver now supports new material models, additional solver options and expanded examples. A new TetMesh class enables loading of volumetric deformable meshes from USD files.

Kinematic and VBD workflows gain support for prismatic, revolute and D6 joints. Rendering additions feature Gaussian splats, tiled-camera support and PBR lighting in the viewer.

Contact sensing improvements include friction aggregation, while collision performance and SDF memory usage see notable gains. The release bolsters validation for asset import pipelines including MJCF and USD.

Initiated by Disney Research, Google DeepMind and NVIDIA, the Linux Foundation project builds upon NVIDIA Warp with MuJoCo Warp integration. It prioritizes GPU computation, differentiability and extensibility under Apache-2.0 licensing.

Examples now demonstrate these features across basic shapes, joints, conveyors and advanced robots such as the Unitree G1 and H1.

Developers report faster iteration times thanks to these enhancements in sensor support and multi-GPU execution. Bug fixes address issues in viewer behavior and contact handling.

Kubernetes Goat has received its most significant refresh in over a year with the v2.3.0 release, adding a new Kyverno Policy Engine Security Hardening scenario and mapping every exercise to the MITRE ATT&CK framework for Kubernetes.

The project remains an intentionally vulnerable cluster that lets practitioners safely attack and defend real Kubernetes components. Version 2.3 reorganizes existing labs to align with the OWASP Kubernetes Top 10, giving users clearer pathways through common cloud-native risks. It also updates the Falco runtime detection scenario with fresh visuals and improves integration with Cilium Tetragon for eBPF-based observability.

Infrastructure changes include native support for Arm-based Macs, fixing crashes in the system monitor and resource-check components. Minor fixes address broken links, typos in setup scripts, and dependency updates to the Go backend.

Setup is unchanged for users with kubectl and helm: clone the repository, run setup-kubernetes-goat.sh, verify pods, then execute access-kubernetes-goat.sh to reach the local dashboard at http://127.0.0.1:1234. The environment now spans more than 20 scenarios covering container escapes, RBAC misconfigurations, SSRF, crypto-miners, namespace bypasses, and resource exhaustion attacks, plus defensive labs using KubeAudit, Popeye, and network security policies.

As Kubernetes attack surfaces continue expanding, these concrete, hands-on updates keep the tool relevant for both offensive and defensive security teams.

The OWASP Cheat Sheet Series has updated its contribution tooling, replacing ad-hoc editing with a reproducible local build system that lets builders quickly test changes before they reach the official website.

Markdown files remain the single source of truth. Running make install-python-requirements, make generate-site and make serve spins up a preview on port 8000 in seconds. npm scripts now enforce consistent terminology and markdown style; npm run lint-markdown-fix automatically corrects many issues. An automated ZIP build supplies a complete offline copy for air-gapped environments or internal wikis.

These changes matter as application security guidance must track rapid shifts in cloud deployments, API architectures and supply-chain threats. The core team, supported by an active Slack channel, continues to invite pull requests that fix errors, expand existing sheets or add new ones. Because the rendered site—not the raw Markdown—is the authoritative reference, contributors focus on substance while the tooling handles presentation and quality.

The result is a living set of concise, practical references that developers and security teams consult daily rather than treat as static documentation.

Cilium has released version 1.19.3, focusing on stability and performance fixes for its eBPF dataplane in production Kubernetes environments.

The update corrects a slow memory leak triggered by incremental policy updates, an issue that could affect clusters under continuous change. It also resolves a performance bug in L7 policy proxy redirect handling, improving throughput for application-aware traffic control.

Cluster operators will benefit from several targeted fixes. BGP service advertisements no longer race during error retries. ClusterMesh CRD installation logic has been corrected to prevent accidental version downgrades. The agent now initializes correctly when using KVStore identity mode with etcd placed behind a Kubernetes Service.

Additional changes include guaranteed timeouts for completion WaitGroups, refined Envoy XDS server accounting for NPDS listeners, and corrected policy service selector handling. New Helm values add config drift detection, giving platform teams better visibility into configuration state.

These improvements reinforce Cilium's architecture: a flat L3 network that spans clusters in native routing or overlay mode, identity-based security decoupled from IP addresses, and kube-proxy replacement using eBPF hash tables. The project continues to integrate ingress/egress gateways, bandwidth management, and service mesh functions while maintaining the prior three minor releases as stable.

The changes are surgical yet significant for teams running at scale.

NullClaw v2026.4.9 introduces production-oriented refinements to its fully autonomous AI assistant infrastructure, written entirely in Zig. The project compiles to a static 678 KB binary with zero runtime dependencies beyond libc, consuming roughly 1 MB of RAM and starting in milliseconds on any compatible $5 single-board computer.

The latest release focuses on operational stability. It fixes immediate WebSocket disconnects on Windows, adds retry logic for transient outbound messaging failures, and preserves reasoning traces in OpenRouter streaming responses. Telegram and QQ channel integrations now feature smoother interactive reply flows and better handling of delayed messages when message IDs expire. CLI onboarding received improvements to model catalog selection and JSON configuration formatting.

NullClaw operates as complete assistant infrastructure. It manages conversation state, routes tasks across providers, and exposes a gateway API for external orchestration. Administrators run commands such as nullclaw status directly from the binary. Build reproducibility remains trivial: zig build -Doptimize=ReleaseSmall generates the optimized executable.

The architecture prioritizes minimalism over feature bloat. By avoiding interpreters and heavy runtimes common in Python-based agents, NullClaw achieves deployment footprints orders of magnitude smaller than conventional frameworks while retaining autonomous execution.

Recent changes also include Windows binary packaging as zip archives and expanded beginner documentation.

Windows Terminal 1.24 has advanced to the Release Preview ring, bringing a series of targeted fixes that improve reliability for daily command-line work.

The update tones down the “invalid media resource” warning and removes it entirely from the Stable channel. Terminal now correctly emits an empty bracketed paste sequence (\e[200~\e[201~) when users paste images into agentic coding CLIs. Restarting sessions with “Restart Connection” returns the buffer to a cleaner default state instead of leaving residual formatting.

Selection behavior during searches has been corrected so dragged selections properly restore keyboard focus for copy operations. Korean IME users no longer see characters inserted at incorrect cursor positions when pressing arrow keys. The console host launches successfully on Windows editions lacking the full text input framework.

Mark Mode indicators survive scrolling without permanent disappearance, and “Copy on select” no longer overwrites the clipboard when pasting into another terminal with an active selection. ConPTY receives a corrected function signature for ConptyShowHidePseudoConsole and eliminates MSB4019 build errors.

These changes matter because the microsoft/terminal repository supplies both the modern tabbed interface and the original conhost.exe that underpins every command prompt, PowerShell window, and WSL session on Windows. The C++ codebase remains the single source for console infrastructure, with ongoing maintenance focused on compatibility rather than flashy redesigns. Microsoft Store deployment continues as the recommended path for automatic updates.

Ultraworkers has unlocked ultraworkers/claw-code, releasing the canonical Rust implementation of the claw CLI agent harness. The project supplies a high-performance command-line tool for orchestrating agent operations, with emphasis on correct build processes, authentication, session control and compatibility checking.

Developers compile the Cargo workspace located in the rust/ directory. After building, claw doctor runs as the mandatory first health check. USAGE.md details the supported workflows while PARITY.md records progress toward feature alignment with the original implementation. Full ACP and Zed daemon support remains pending; claw acp and claw --acp currently report status only.

The repository deliberately excludes the crates.io claw-code package, which installs a deprecated stub that prints a rename notice. Users must build from source to obtain the functional binary. Companion Python code in src/ and tests/ provides reference implementations and audit tooling. Container-first deployment guidance appears in docs/container.md, reflecting the project's preference for isolated, reproducible environments.

ROADMAP.md lists active priorities including daemon entrypoints and cleanup tasks. PHILOSOPHY.md frames the architectural choices that favor Rust's safety and speed for agent harness responsibilities.

The chrisns/childrens-clock project has shipped v1.10.107-rc.10, delivering more flexible scheduling and expanded sensor capabilities that address longstanding parental pain points. The device, built on ESPHome and an ESP32, now makes it simpler to adjust weekday versus weekend rules or insert one-off holiday lie-ins without manual intervention at the hardware itself.

At its heart the clock combines a £1.91 ESP32 with a £2.59 WS2812B 8×32 LED matrix mounted in a standard 15×5-inch photo frame. Three solder joints—5 V, ground, and GPIO 13—complete the hardware. Once flashed via PlatformIO, the unit syncs time over NTP, retains settings across power cuts, and automatically observes daylight-saving changes.

Colour patterns on the matrix tell children whether it is sleep time, quiet-play time, or get-up time. The visual language works at distance, removing the need for kids to leave bed or decipher small text. Home Assistant integration lets parents tweak parameters from a phone without lighting up the bedroom.

The new release adds first-class support for temperature sensors and Bluetooth relays while maintaining the project’s strict no-camera policy. Total build cost remains under £5. After 18 months of nightly use in the maintainer’s own household, the clock has become a stable, grown-up-looking fixture that also feeds environmental data back into the smart-home dashboard.

**

After more than ten years of service, joergi/MagPiDownloader has streamlined its platform support. The shell script no longer offers working Windows or Docker options, with the maintainer noting these will likely be removed. Only Mac and Linux users retain full functionality.

The utility automates downloading every issue of MagPi, the official Raspberry Pi magazine, from the free archives at https://www.raspberrypi.org/magpi/issues/. Instead of clicking through dozens of PDFs, users execute a single script that retrieves the complete collection of over 140 issues.

Each issue contains detailed tutorials on Python, hardware projects, camera usage, and Linux system administration. Local copies enable offline reference, faster searching, and preservation against potential website changes. The project saves significant time for educators, makers, and developers who regularly consult back issues.

Recent updates reflect practical maintenance rather than expansion. The README clearly marks broken components while preserving clear setup instructions for supported operating systems. Contributors have kept the core functionality reliable across a decade of Raspberry Pi hardware evolution, from the original Model B to the current Raspberry Pi 5.

As interest in single-board computing grows in education and hobbyist communities, dependable access to historical content matters. MagPiDownloader delivers exactly that: a focused, no-frills tool that fetches what users need without unnecessary complexity.

Installation requires cloning the repository and running the provided shell commands. No heavy dependencies are needed on modern Linux distributions or macOS.

egui has shipped version 0.34.1, delivering two focused fixes that matter to teams shipping Rust interfaces to the web.

The eframe framework’s wgpu backend now automatically falls back to WebGL when WebGPU is unavailable. Maintainer emilk’s change removes a hard dependency that previously broke deployment on certain browsers and older hardware. Applications continue running without code changes, broadening the platforms reachable from a single Rust codebase.

A parallel tweak ensures cursor styles affect only the <canvas> element. Contributor mkeeter’s update prevents egui from altering cursor appearance elsewhere on the host page, eliminating a frequent source of friction when embedding egui inside larger web applications.

These updates reflect the project’s shift toward production reliability after seven years of steady evolution. Immediate-mode design keeps UI code colocated with logic, while eframe handles windowing, input, and rendering across Linux, macOS, Windows, Android, and Wasm targets. The library still requires nothing beyond the ability to draw textured triangles, preserving easy integration into custom game engines.

For builders already using egui, the release lowers deployment risk rather than adding surface area. Rerun’s sponsorship of core development keeps the focus on practical, low-overhead GUI needs in visualization and tooling.

The changes are live in the web demo and available immediately via Cargo.

O3DE has released version 25.10.2, delivering targeted stability and toolchain updates to its Apache 2.0 open-source 3D engine. Five years after its debut, the project remains a fee-free option for teams building AAA games, cinema-quality worlds, and high-fidelity simulations.

The release tightens compatibility with current development environments. Projects now clone via Git LFS to manage large binary assets efficiently. On Windows, the documented requirements specify Visual Studio 2019 16.9.2 or later with the Game Development with C++ workload, MSVC v142 tools, and CMake 3.24.0 minimum. Optional Wwise SDK integration continues to support advanced audio pipelines.

Core engine systems written in C++ handle real-time rendering, animation, and physics at production scale. The modular Gem architecture lets teams enable or disable features without rebuilding the entire stack. Roadmap tracking at o3de.org shows continued focus on graphics fidelity and multi-platform performance.

For studios already invested in the ecosystem, 25.10.2 removes friction in dependency management and build reliability. The project’s transparent contribution process and absence of commercial obligations keep it relevant against proprietary alternatives.

Godot 4.6.2 is a maintenance release that prioritizes stability and usability after months of community bug reports. The update fixes rendering glitches, physics edge cases, and editor workflow issues while preserving full compatibility with projects built on earlier 4.6 versions.

Developers shipping to Linux, macOS, Windows, Android, iOS, web, and consoles will notice fewer export failures and more consistent runtime behavior. Core systems governing scene handling, scripting, and asset pipelines received targeted repairs, reducing crashes during complex 2D tilemap work or 3D lighting passes.

The release continues Godot’s community-driven model. Contributors identified problems through the GitHub tracker, submitted patches, and verified fixes across hardware configurations. Official binaries and export templates are available immediately; source compilation instructions remain unchanged.

By keeping the engine reliable rather than chasing new features, the Godot Foundation reinforces its value for teams that treat the codebase as a long-term foundation. The curated changelog highlights the most practical improvements, allowing developers to adopt 4.6.2 with confidence and minimal disruption.

This steady refinement illustrates how an independent, MIT-licensed engine matures through disciplined iteration instead of marketing-driven resets.