The Git Times

LangChain has shipped langchain-core 1.3.0, focusing on tighter integration between model usage and observability tooling. The most significant change adds chat model and LLM invocation parameters directly to traceable metadata. Developers using LangSmith can now see exact model settings, temperature, and token limits in every trace without additional instrumentation.

Additional updates reduce streaming metadata overhead, delivering measurable performance gains in high-throughput applications. Memory management received attention through reference-counted run trees that support Python’s garbage collector during long-running agent sessions. Security hardening includes refined SSRF utilities that correctly handle cloud metadata IPs and link-local ranges while blocking unauthorized outbound requests.

OpenAI response parsing was updated to gracefully handle content blocks missing explicit type keys, eliminating intermittent errors when using the Responses API. The release also bumps test dependencies and includes several internal cleanups.

These changes arrive as agent workflows grow more complex. The library’s init_chat_model("openai:gpt-5.4") pattern remains the standard entry point, while LangGraph handles stateful orchestration and Deep Agents add planning and file-system capabilities. By exposing more runtime data by default, version 1.3.0 lowers the cost of debugging multi-agent systems that combine external tools, vector stores and multiple model providers.

Why it matters now: Production LLM applications demand reproducible traces and predictable resource use. The new metadata capture and memory improvements address exactly those operational requirements without changing existing application code.

AutoGPT version 0.6.56 introduces a MemoryEnvelope metadata model that enables scoped retrieval and memory hardening for its autonomous agents. The backend change allows agents to query specific memory contexts instead of scanning undifferentiated stores, improving both precision and security during extended executions.

A related fix resolves copilot messaging errors by pre-creating assistant messages before the first yield, eliminating last_role=tool conflicts that previously disrupted conversational flows. These updates target practical weaknesses in production agent deployments where persistent state management determines reliability.

The platform lets developers create, deploy, and run continuous AI agents that automate complex workflows across OpenAI, Claude, and Llama models. Self-hosters meet modest requirements—4 CPU cores, 8 GB RAM, Docker 20.10+ and Node.js 16+—then launch via a one-line installation script. Updated documentation guides configuration of ports and outbound HTTPS connections.

For builders already operating AutoGPT instances, the release reduces memory-related failure modes that surface after days or weeks of autonomous operation. The improvements reflect a maturing focus on infrastructure stability rather than new surface features, addressing demands from teams running agents at scale.

MemoryEnvelope and its hardened retrieval now form a more robust foundation for agents that must maintain knowledge integrity without constant human oversight.

OpenClaw's v2026.4.15 release introduces targeted improvements to its self-hosted personal AI assistant. The update sets Anthropic models as default, adding opus aliases, Claude CLI defaults and bundled image understanding powered by Claude Opus 4.7.

Google plugin support now includes Gemini text-to-speech, with voice selection, WAV reply output, PCM telephony and setup guidance. A new Control UI card displays Model Auth status, showing OAuth token health and rate-limit pressure at a glance. It is backed by a models.authStatus gateway method that strips credentials and caches responses for 60 seconds.

Memory infrastructure advanced with LanceDB cloud storage support, allowing durable indexes to run on remote object storage rather than local disk. A GitHub Copilot embedding provider was added for memory search, enabling plugins to reuse transport while honoring token refresh and payload validation.

For local-model users, an experimental agents.defaults.experimental.localModelLean flag drops heavyweight tools such as browser, cron and message, reducing prompt size without affecting standard deployments.

The assistant connects to more than 20 messaging platforms including WhatsApp, Telegram, Slack, Discord, Signal, Matrix and iMessage. Configuration remains through the openclaw onboard command, which sets up the gateway daemon as a launchd or systemd user service. The stack runs on Node 24 or 22.16+ using npm, pnpm or bun.

Inria's Willow team has released Pinocchio 4.0, marking a significant update to the rigid body dynamics library. The new version focuses on constrained systems, addressing a key challenge in robotics and simulation.

Central to the release is the lcaba algorithm, which efficiently computes forward dynamics for mechanisms with closed kinematic loops. This is complemented by a new constraint API featuring models like PointContactConstraintModelTpl, FrameAnchorConstraintModelTpl and JointFrictionConstraintModelTpl. Each model includes corresponding data structures for better organization.

The update also introduces Delassus operators that calculate J M^{-1} J^T using dense, sparse and Cholesky-based methods. These feed into two new constraint solvers: ADMMConstraintSolverTpl and PGSConstraintSolverTpl. Examples in the repository show how to combine these components for realistic simulations, such as the G1 robot.

Pinocchio has long provided analytical derivatives of core algorithms including Recursive Newton-Euler and the Articulated-Body Algorithm. Version 4.0 extends this capability to constrained scenarios, facilitating gradient-based methods in optimization and control.

With its Python interface available via Conda, the library continues to bridge high-performance C++ implementations with rapid prototyping needs. It remains integral to tools like Crocoddyl for differential dynamic programming and the Humanoid Path Planner for motion planning.

This release ensures Pinocchio stays at the forefront of physics-based robotics research and industrial applications.

ArduPilot shipped Plane-4.6.3 in November 2025 as the new stable branch for fixed-wing and VTOL platforms. The release tightens transition logic between hover and forward flight, reducing altitude loss during mode changes by refining pitch and throttle scheduling.

Developers updated the extended Kalman filter tuning parameters, delivering tighter position estimates in gusty conditions. Additional support for newer barometers and magnetometers expands compatible flight-controller boards without requiring custom forks. These changes were validated through structured beta testing that logged thousands of autonomous landings.

The C++ codebase, maintained under GNU GPL v3, continues to serve as the common foundation for ArduCopter, ArduRover, and ArduSub. MAVLink telemetry remains unchanged, preserving compatibility with existing ground stations and companion computers running ROS.

Andrew Tridgell and the maintainer team merged more than 180 pull requests since the prior stable cut, focusing on edge-case robustness rather than headline features. Community forums report smoother Qautotune results on quadplanes and reduced vibration-induced drift on tailsitters.

For operators flying beyond visual line of sight, the incremental improvements translate into measurable gains in flight time and mission reliability. As commercial VTOL use expands in inspection and logistics, the release supplies a battle-tested open source option that evolves with hardware rather than locking users to proprietary stacks.

RobotCode’s v2.5.1 release tightens core analysis components that many Robot Framework developers rely on daily. The analyzer no longer reports embedded arguments as VariableNotFound when keywords are invoked through [Template] or Test Template. Previously these placeholders triggered noisy diagnostics; the fix skips variable resolution for such tokens, aligning diagnostics with Robot Framework’s own parser behavior.

A second change corrects BDD prefix recognition. French phrases such as “Étant donné que” and “Et que” now match correctly because the tool sorts prefixes by length and applies a cached regex, matching the framework’s longest-first strategy. The same logic updates keyword finding, model helpers, and semantic token generation.

Variable handling also received attention. A new helper function standardizes ${CURDIR} replacement inside variable values, removing earlier inconsistencies.

These fixes sit on RobotCode’s established architecture: native Robot Framework parsing for validation and errors, Language Server Protocol services for navigation, IntelliSense, and refactoring, plus Debug Adapter Protocol support. The VS Code extension, IntelliJ plugin, and CLI tools share identical behavior, while robot.toml configuration and an interactive REPL extend standard robot command-line workflows.

For teams maintaining large test suites or RPA projects, the release reduces editor noise and improves compatibility with non-English BDD vocabularies without altering existing workflows.

Key retained capabilities include project-wide rename operations, syntax highlighting for embedded arguments and Python expressions, and code snippets that accelerate test authoring.

SafeLine v9.3.4 introduces the ability to adjust execution order of custom rules within Allow and Deny lists. Administrators can now sequence policies with precision, reducing conflicts when layering defenses against multiple threat types.

The release corrects abnormal IP information display and resolves a user interface glitch on slave nodes during master-slave synchronization. These fixes tighten operational visibility in distributed setups. On the infrastructure side, generated Nginx configurations now include backlog=65536 and reuseport on non-default listen directives, improving connection handling at scale. License validation logic has been tightened to prevent false positives.

As a self-hosted reverse proxy and WAF implemented in Go, SafeLine sits between clients and web applications, inspecting and filtering HTTP/S traffic according to defined policies. It blocks SQL injection, XSS, command injection, path traversal, SSRF, brute-force attempts, and HTTP floods. Additional capabilities include proactive bot defense, HTML and JavaScript encryption, IP-based rate limiting, and web access control lists.

For teams operating sensitive infrastructure outside public clouds, these incremental improvements matter. They sharpen policy control and system reliability without requiring architectural changes. The project continues to evolve as a practical blue-team tool for organizations that demand full data custody and transparent protection mechanisms.

(178 words)

Caddy 2.11.2 focuses on hardening the server’s production readiness with targeted security fixes and operational upgrades. The release corrects a flaw in the forward_auth directive that could permit identity injection and privilege escalation. A second issue in vars_regexp allowed double expansion of placeholders, potentially exposing secrets in nonstandard configurations. Both have been closed.

Reverse-proxy behavior received substantial attention. Edge cases involving PROXY protocol headers, health-check port selection, and request-body closure during retries are now handled correctly. Dynamic upstreams are tracked for the first time, activating passive health checking without additional configuration.

A new global tls_resolvers option lets administrators specify DNS resolvers used for ACME DNS challenges across every site. Log rolling gains native zstd compression; the older roll_gzip option is deprecated in favor of the more general roll_compression. Metrics collection is faster, and several error messages have been clarified.

The binary is built on Go 1.26.1, inheriting its CVE patches. Updated documentation notes that file-system case sensitivity can affect the file_server handler’s hide option. These changes reinforce Caddy’s ability to serve hundreds of thousands of sites while coordinating clusters and recovering gracefully from TLS-related failures.

The maintainers of the community-scripts/ProxmoxVE repository have shipped an update that adds a one-command installer for step-ca, Smallstep’s modern certificate authority. The new script lets administrators deploy private PKI infrastructure inside an LXC container or VM on Proxmox VE 8.4–9.1 without manual certificate chaining or configuration archaeology.

Beyond the addition, the release focuses on stability. Core logic now sanitizes mount_fs input by stripping spaces and trailing commas, eliminating a class of silent failures during storage configuration. The pct create path received targeted refactoring to resolve telemetry conflicts and clean up command-line handling. Intel GPU passthrough scripts pin the IGC version to a compute-runtime-compatible tag, preventing driver mismatches that had broken accelerated workloads.

Several service scripts were also hardened. Update procedures for Umami and Bambuddy were corrected so that apt and internal migration steps no longer fail on newer Debian base images. These fixes matter because the collection is used daily by homelab operators running dozens of containers; even minor update bugs can cascade across production self-hosted stacks.

The project continues the original helper-script model: users paste a one-line command from community-scripts.org, choose Default or Advanced mode, and receive a container pre-populated with sensible defaults plus a post-install helper for routine maintenance. With hundreds of supported services—from home automation to monitoring—the incremental improvements keep the toolkit reliable as Proxmox itself evolves.

Browser monoculture poses risks to web innovation and security. Ladybird addresses this through its truly independent browser engine based on web standards.

Recent work has enhanced the multi-process design. A central UI process manages several WebContent processes, each handling an individual tab in its own sandbox. Image decoding and HTTP requests occur in isolated processes, limiting exposure to harmful content.

Many supporting libraries originate from SerenityOS, including LibWeb for the rendering engine, LibJS for JavaScript, LibWasm for WebAssembly, LibTLS for encryption, and LibGfx for graphics. These components enable steady progress toward full modern web compatibility.

The browser supports Linux, macOS and Windows via WSL2. Developers can follow the provided build instructions to compile from source.

Interest from the developer community has increased, with participation coordinated via Discord. Clear contribution guidelines help maintain focus as the project matures from its pre-alpha stage.

Licensed under the 2-clause BSD license, Ladybird offers a transparent platform for browser research and experimentation.

Fuel Labs has released fuel-core v0.48.0, updating the Rust full node implementation of the Fuel v2 protocol. The changes target production reliability, storage flexibility, and API completeness for node operators and developers already running Fuel networks.

The most practical addition is FailoverTransport, which retries GraphQL queries across multiple endpoints. This reduces the impact of individual service outages. A new adapter enables direct block storage in AWS S3 buckets, while the backup utility is now built in as an archive subcommand, simplifying long-term data management without external tools.

API improvements include a protobuf interface for the block aggregator, a quorum provider, and full coverage of proto block types. Integration tests confirm correct transaction indexing inside pre-confirmations for both single-transaction and multi-transaction blocks.

Three breaking changes require attention. The relayer server now uses only the first RPC URL; remaining entries are ignored. Transaction indexing inside native block production pre-confirmations was corrected. The minimum Rust version has advanced to 1.93.0.

Ignition and Testnet currently run v0.47.1. This release prepares infrastructure for the next upgrade cycle, giving operators clearer paths to compile from source with make build or deploy updated binaries. The focus remains on stable, observable node operation rather than headline features.

**

Fuel Labs has released Sway v0.71.0, delivering concrete gains in compilation speed and runtime efficiency for its Rust-inspired smart contract language.

The update introduces a len method to std::string::String, implements init_aggr in the intermediate representation for faster aggregate initialization, and adds encode_allow_alias to reduce logging overhead. Compiler teams tightened the size threshold inside SROA profitability checks, introduced leaf-function optimisations, and corrected the runtime memory layout of string arrays. Constant evaluation now forbids outer variables, closing a source of non-determinism.

Toolchain housekeeping continues: forc-node has migrated into the main forc monorepo, predicate root output now includes package names, and several CI jobs run faster. These changes reflect ongoing refinement of the Fuel Virtual Machine’s language stack rather than radical redesign.

Five years after its initial commit, Sway remains focused on bringing systems-language ergonomics and predictable performance to blockchain development. Developers compile contracts with forc build, test with the integrated toolchain, and rely on the language’s strict type system to limit gas surprises. The project accepts contributions through its documented guidelines, with emphasis on benchmark-driven improvements and backward compatibility.

Sway continues to evolve as the native language for the Fuel blockchain, prioritising reliability and execution speed over marketing milestones.

Delft Open Hardware's awesome-open-hardware list received its latest update in April, expanding coverage of specialized tools for laboratories and makers. New entries highlight OpenFlexure, a 3D-printed microscope with precise stage, and openUC2, a modular microscopy platform.

These additions build on foundational projects such as RepRap, the self-replicating manufacturing machine, Arduino electronics platform and Prusa3D printers. The curated collection organizes resources across 10 categories, from talks and papers to books and training programs.

The update matters as open hardware gains traction in biotechnology and environmental science. Researchers use the list to identify designs like Open Gamma Detector for spectroscopy or Biohack Academy equipment including incubators and centrifuges.

By centralizing references, the project prevents redundant development efforts. Recent inclusions such as PiKVM, based on Raspberry Pi for IP-KVM, and Mekanika tools reflect practical needs in remote diagnostics and fabrication. The repository, created in 2021, continues to serve as an essential starting point for new open hardware initiatives, from SafeCast radiation monitoring to WikiHouse for digital open housing.

Further readings and related awesome lists extend its utility. Active curation keeps the resource current, supporting the growing community of open hardware practitioners worldwide.

Five years after its debut, stack-chan/stack-chan remains a preferred open platform for builders seeking an expressive, hackable robot. The v0.2.1 release, now powering kits that first appeared at Maker Faire Tokyo 2022, has seen continued refinements addressing earlier stability issues in the embedded software.

The robot pairs an ESP32-based M5Stack with a display that renders its signature animated face. Written in TypeScript and built on the Moddable runtime, the firmware lets developers control facial expressions, gaze direction, speech output, and servo-driven head movement using simple high-level APIs. Both Serial (TTL) and PWM servos are supported, along with plug-in M5Units for sensors and expansion.

All components live in one repository: firmware sources, STL case files, KiCad schematics, and board layouts. Builders can assemble from scratch or start with pre-assembled modules. Recent community work has stabilized ChatGPT API integration, allowing the robot to hold coherent conversations while its eyes follow speakers and its face reacts to sentiment.

The project’s Apache 2.0 licensing and complete hardware documentation continue to lower the barrier for experimentation. In an era of growing interest in embodied AI, Stack-chan gives developers a concrete, programmable character rather than abstract code.

FanCtrl has shipped version 1.7.9, refreshing its dependencies on the LibreHardwareMonitor library and liquidctl. The changes, pulled from recent upstream commits, tighten sensor reporting for AMD and NVIDIA GPUs, DIMM temperatures, motherboard controllers, and supported liquid-cooling devices.

The C# application reads real-time temperatures and automatically adjusts fan and pump PWM according to user-defined curves. Target sensors can include CPU, GPU, or liquid-cooler readings; fans are added to specific graphs where operators set temperature-to-PWM mappings. Hysteresis prevents rapid oscillations, while presets save and restore complete profiles. Step granularity for temperature and PWM can be switched between 1, 5 or 10 units.

Hardware support covers standard motherboards plus NZXT Kraken, EVGA CLC, NZXT RGB & Fan Controller, and any device managed by liquidctl. Optional HWiNFO bridging and NvAPIWrapper extend the data pool. The on-screen display works with Rivatuner Statistics Server to overlay metrics without leaving full-screen applications.

Additional controls include Fahrenheit switching, tray-icon animation during automatic operation, delayed Windows startup, and one-click reset of all libraries. For users maintaining high-performance or quiet builds, the updated libraries restore compatibility with newly released components that older library versions could misread or ignore. The project remains a lightweight, open-source option for precise cooling control on Windows desktops.

Fyrox has released version 1.0.0, confirming its status as a mature, production-ready game engine written in Rust. The update caps more than seven years of development on the former rg3d codebase, delivering stable APIs that teams can now trust for commercial projects.

The engine combines a visual scene editor with full Rust control over rendering, physics, animation and GUI systems. Developers compose levels, materials and entity hierarchies in the editor, then extend behavior through native code that compiles via Cargo. Recent changes focus on reliability: breaking API adjustments have been completed, web assembly output improved, and browser-based demos now run without installation.

Documentation remains a strength. The official Fyrox book walks through setup, asset pipelines, lighting models and common gameplay systems. An active Discord server and clearly documented contribution process continue to draw developers who submit fixes and new features.

Sponsors, including JetBrains with its open-source license, have helped sustain the project. For Rust programmers who need both performance and tooling, version 1.0 removes earlier uncertainty around long-term support while preserving the engine’s lean, dependency-light design.

Release notes and downloads are available at fyrox.rs.

Ebitengine v2.9.9 focuses on incremental refinements rather than headline features. The update improves custom shader compilation times, tightens audio synchronization for Ogg/Vorbis streams, and fixes edge cases in gamepad input and offscreen rendering reported by the community.

The engine's enduring appeal is its dead-simple API. Developers write ordinary Go code that compiles to native binaries or WebAssembly with few dependencies. Automatic batching, texture atlas generation and matrix-based geometry transformations happen without manual configuration, letting programmers concentrate on gameplay loops instead of render queues.

Platform coverage remains broad. Windows and macOS builds need no Cgo. The same codebase targets Linux, FreeBSD, Android, iOS, browsers via WebAssembly, Nintendo Switch and, on a limited basis, Xbox. The ebiten, audio, vector and text/v2 packages provide the complete toolkit: mouse, keyboard, touch and gamepad input; PCM, MP3 and WAV playback; and colorm utilities for advanced blending.

For teams choosing Go over heavier engines, v2.9.9 reduces friction in continuous integration and cross-platform testing. As more studios experiment with Go for tools and lightweight titles, the release keeps the project current without altering its core philosophy of simplicity.

Community channels on Discord, Gophers Slack and Reddit continue to supply rapid feedback that shapes each point release.

Following the GPC 2026 presentation and newly released video, Electronic Arts has expanded the ShaderToHuman (S2H) samples in Gigi, demonstrating concrete gains in shader iteration speed for complex rendering tasks.

The HLSL library, with GLSL support via preprocessor, lets programmers call familiar functions such as PrintF directly inside shaders. It draws text, watch windows, and both 2D and 3D geometric primitives straight into the viewport. No C++ modifications, buffer allocations, or content setup are required. A single include file and a few lines of code complete the integration.

The design favors rapid, targeted debugging over production UI or zero overhead. Recent examples illustrate its value: one renders Gaussian splatting through a stochastic rasterizer on a procedurally generated .ply file; another adds interactive panning, zooming, and per-pixel color inspection entirely within the shader.

These capabilities matter now as teams ship more sophisticated real-time effects. By removing the usual scaffolding between shader logic and visible feedback, ShaderToHuman shortens the debug loop from minutes to seconds. EA’s SEED team continues development under a permissive license that allows commercial use, with CUDA exploration listed on the roadmap.

Interactive documentation and the Gigi browser’s “Human” collection give developers immediate hands-on access to working prototypes.