Preset
Background
Text
Font
Size
Width
Account Pricing Wednesday, July 1, 2026

The Git Times

“Civilization advances by extending the number of important operations which we can perform without thinking about them.” — Alfred North Whitehead

AI Models
Claude Opus 4.8 $25/M GPT-5.5 $30/M Gemini 3.1 Pro $12/M Grok 4.20 $2.50/M DeepSeek V3.2 $0.80/M Llama 4 Maverick $0.60/M
Full Markets →
Fresh on Hugging Face

Model Drops

The newest model releases builders are picking up right now.

Zed’s AI-Powered Multiplayer Editing Reaches New Maturity 🔗

Rust-based editor adds worktree creation and terminal automation to collaborative coding workflows

zed-industries/zed · Rust · 86.3k stars Est. 2021 · Latest: v1.8.2

Why this leads today Zed introduces real-time collaborative editing in a Rust-built, high-performance code editor from Atom and Tree-sitter creators, offering a tangible shift in how developers work together and individually.

Zed, the high-performance code editor built by the creators of Atom and Tree-sitter, has evolved into a serious contender for developers seeking real-time collaboration without sacrificing speed. Written entirely in Rust and powered by its custom GPU-accelerated UI framework (GPUI), Zed delivers buttery-smooth performance even when multiple users are editing the same file simultaneously — a feat few editors achieve without noticeable lag.

The latest release, v1.

8.2, introduces practical enhancements that deepen its utility in team environments. Developers can now create a new Git worktree directly from the sidebar’s “new thread” button, streamlining branch-based workflows without leaving the editor. Complementing this, the new agent.terminal_init_command setting allows teams to automate environment setup — such as activating a virtual environment or running database migrations — whenever a new agent terminal thread is spawned. These features reflect Zed’s growing focus on reducing context-switching during collaborative sessions.

Beyond collaboration, Zed’s AI integration continues to mature. The editor’s agent system now delivers faster edit-tool responses and smoother scrolling during streaming AI outputs, while improved truncation of external agent configuration labels enhances usability in complex setups. Navigation and selection have also been refined: users can filter archived threads by project name (not just title), reset pane sizes to equal dimensions with a single command, and expand selections through nested delimiters using new editor: select inside/around delimiters actions — small but meaningful quality-of-life upgrades for power users.

Zed’s appeal lies in its blend of native performance and modern developer expectations. Unlike Electron-based editors, it avoids the memory and startup overhead associated with web technologies, instead leveraging Rust’s safety and speed. Its multiplayer architecture, built on conflict-free replicated data types (CRDTs), enables low-latency co-editing without requiring a central server for every interaction — a design choice that scales well for small to medium teams.

The catch: Zed remains Linux- and Windows-capable but lacks a native web version, limiting its accessibility for developers reliant on browser-based workflows or ephemeral environments like GitHub Codespaces, and its GPL-3.0-or-later licensing may deter teams preferring permissive licenses for internal tooling.

Use Cases
  • Remote pair programming on low-latency networks
  • Real-time code reviews with shared terminal sessions
  • Teams adopting Git worktree workflows in Rust/C/C++ projects

Source: zed-industries/zed — based on the README and release notes.

More on the Front Page

Activepieces Unifies AI Agents and Workflows in TypeScript 🔗

Open-source automation platform adds MCP server support for LLM integration

activepieces/activepieces · TypeScript · 23.1k stars Est. 2022

Activepieces has evolved into a comprehensive AI automation platform that bridges no-code workflow building with extensible AI agent capabilities through its Model Context Protocol (MCP) server ecosystem. The project enables developers to create reusable "pieces"—TypeScript-based npm packages—that function as both workflow automations and MCP servers consumable by LLMs in tools like Claude Desktop, Cursor, and Windsurf.

Recent development highlights include adding OIDC authentication for AWS Bedrock and Secrets Manager, implementing output schemas across 15 core pieces to improve data typing, and introducing server-side token caching for custom authentication flows.

The platform now supports over 280 community-contributed pieces, with approximately 60% originating from external developers, reinforcing its open ecosystem model.

Technically, Activepieces distinguishes itself through type-safe piece development in TypeScript, featuring hot reloading for local iteration and a self-hosted architecture designed for enterprise security via network-gapped deployments. Its AI-first approach includes native AI pieces and an SDK for building custom agents directly within workflows, while human-in-the-loop mechanisms—such as chat and form interfaces—allow for manual intervention in automated processes.

The builder interface caters to both technical and non-technical users, enabling organizations to standardize on developer-configured tools while empowering broader teams to construct workflows through a visual, no-code editor. Branding and access controls remain fully customizable for enterprise adoption.

The catch: Despite its extensibility, the platform’s reliance on TypeScript for piece creation may present a barrier for teams seeking faster iteration via lower-code or visual-first automation tools, particularly when compared to mature alternatives with broader pre-built integrations.

Use Cases
  • Developers creating custom AI agent tools for LLM assistants
  • Enterprises automating cross-app workflows with human approval steps
  • Teams building secure, self-hosted alternatives to commercial automation platforms

Source: activepieces/activepieces — based on the README and release notes.

Open Design merges brand assets into one loop 🔗

Native desktop app turns brand assets into reusable design systems for immediate prototyping

nexu-io/open-design · TypeScript · 73.7k stars 2mo old

Open Design is a native desktop app that lets you extract colors, type, and voice from live sites, Figma files, or browser captures to build a reusable design system. Once captured, you can prototype web, mobile, or desktop interfaces, generate slides, images, or videos, and export to HTML, PDF, PPTX, or MP4—all within one window. It runs agent skills via Claude Code, Codex, Cursor, or 17+ other CLIs, using local or remote models through its Agentic Model Router (AMR) with zero-config access to GPT, Claude, Gemini, and others.

The 0.12.0 release focuses on turning any brand into a design system you can build from, streamlining capture → design → build → export. Parallel sessions let multiple AI agents work alongside you like a local design team.
The catch: despite rapid progress, the project is only two months old with 484 open issues, raising questions about long-term stability and enterprise readiness.

Use Cases
  • Designers turn website screenshots into editable Figma-like design systems
  • Developers generate frontend prototypes from brand assets using natural language prompts
  • Teams create consistent pitch decks and video ads by reusing extracted brand styles

Source: nexu-io/open-design — based on the README and release notes.

mattpocock/skills delivers composable engineering tools for AI coding agents 🔗

Shell-based skills enhance agent workflows with grilling, triage, and teach components

mattpocock/skills · Shell · 152.7k stars 4mo old

The mattpocock/skills project provides a set of composable Shell scripts designed to improve how AI coding agents like Claude Code and Codex interact with developers. Rather than relying on rigid frameworks, these skills offer small, adaptable tools—such as /grill-me to clarify requirements through targeted questioning and /triage to label and route issues via GitHub, Linear, or local files. The latest patch update improves the teach skill by making it reuse-first: agents now pull from shared components in `.

/assets/before generating new content, extracting reusable elements like stylesheets or quiz widgets instead of inlining them. Installed vianpx skills@latest add mattpocock/skills`, the setup prompts users to configure their preferred tracker and documentation location. With over 150k stars and active maintenance—last commit just zero days ago—the project shows strong traction, though its Shell-only implementation may limit integration with non-Unix environments or agents expecting richer plugin systems.
The catch: The project's reliance on Shell scripts assumes a Unix-like agent environment, potentially excluding users on Windows-native agents or those requiring deeper IDE integration.

Use Cases
  • Engineers refine agent understanding using `/grill-me` for feature specs
  • Teams automate issue triage with custom labels across GitHub or Linear
  • Developers build reusable lesson components via the updated `teach` skill

Source: mattpocock/skills — based on the README and release notes.

iOS GPS Spoofing Made Simple via Proxy Tools 🔗

JavaScript-based module lets users fake location in Maps without jailbreak or developer accounts.

mekos2772/ios-location-spoofer · JavaScript · 469 stars 1d old

The mekos2772/ios-location-spoofer project enables iOS users to spoof GPS location by intercepting and modifying HTTPS responses from Apple’s location service. Built as a JavaScript module, it integrates with popular proxy tools like Shadowrocket, Loon, Surge, Quantumult X, and Stash to alter Wi-Fi and cell tower coordinates returned to the device. By changing the BSSID-based location data Apple sends back, the iPhone calculates a false position—useful for testing location-dependent apps or accessing geo-restricted content.

The setup requires enabling HTTPS decryption, installing a CA certificate, and importing the module file, after which users can specify custom latitude and longitude parameters. Unlike the original Go-based version, this iteration adds support for CellTower field spoofing, multiple Apple response formats, and motion activity masking to reduce detection risk. While Shadowrocket and Loon modules are marked as tested, Surge, Quantumult X, and Stash integrations remain unvalidated, relying on community feedback.
The catch: Three open issues and limited testing across proxy platforms raise questions about reliability and edge-case handling in real-world use.

Use Cases
  • Developers testing location-based app behavior under various geographic conditions
  • Users accessing region-locked services like streaming or geographically restricted mobile content
  • Researchers studying how iOS interprets and processes location data from network sources

Source: mekos2772/ios-location-spoofer — based on the project README.

Docker Agent adds OpenRouter support and model bypass controls 🔗

Latest release enables flexible AI provider routing and dedicated session compaction models

docker/docker-agent · Go · 3.1k stars 10mo old

Docker’s docker-agent CLI plugin now treats OpenRouter as a first-class provider, auto-detecting it when OPENROUTER_API_KEY is set and resolving its base URL without manual configuration. The release also introduces a compaction_model field, letting teams assign a separate, often lighter-weight model for session summarization—reducing cost and latency in long-running agent workflows. Additionally, a per-model bypass_models_gateway flag allows specific models to connect directly to their providers, skipping any configured intermediary gateway for lower latency or direct access to model-specific features.

These changes come alongside internal refactoring: mutexes replaced with atomic types for better concurrency safety, and global variables removed to enable parallel testing across core packages. The agent remains defined via declarative YAML, supports multi-agent orchestration, and runs with docker agent run agent.yaml.
The catch: Despite steady traction, the project has 14 open issues and relies on rapid AI model evolution—raising questions about long-term stability and backward compatibility as provider APIs shift.

Use Cases
  • Developers building multi-agent workflows with task delegation
  • Teams sharing AI agents via OCI registries for reuse
  • Engineers integrating local or remote MCP tools with agents

Source: docker/docker-agent — based on the README and release notes.

Zen C Compiler Gains Debugging Support in v0.4.4 Update 🔗

Latest release adds source-level debugging and output cleanup for the C-transpiling language

zenc-lang/zenc · C · 4.3k stars 5mo old

The Zen C compiler (zc) now includes debugging support for Zen C source code, allowing developers to step through high-level constructs while targeting C11 output. This feature, merged in pull request #305, enables breakpoints and variable inspection during execution without leaving the Zen C abstraction layer. The v0.

4.4 release also cleans up compiler output artifacts via PR #309, reducing clutter in build directories. Zen C continues to compile to human-readable GNU C/C11 with full ABI compatibility, offering type inference, pattern matching, generics, traits, and async/await alongside manual memory management with RAII. The project maintains active development across its core compiler, VS Code extension (alpha), and documentation repositories, with recent commits showing sustained contributor engagement. Despite its feature richness, Zen C remains a transpiler rather than a standalone compiler, meaning build times and toolchain complexity inherit from the underlying C toolchain it targets.
The catch: Debugging relies on the C backend, so performance profiling and low-level analysis still require C-tool familiarity.

Use Cases
  • Systems programmers writing embedded firmware with high-level syntax
  • Developers seeking C ABI compatibility without manual memory management
  • Language designers experimenting with transpiled syntax features in C ecosystems

Source: zenc-lang/zenc — based on the README and release notes.

Self-evolving AI coding agent ships prompt caching and error handling 🔗

yoyo-evolve v0.1.11 adds cost-cutting cache and failure analysis after 4 months of autonomous growth

yologdev/yoyo-evolve · Rust · 1.8k stars 4mo old

The yoyo-evolve project released v0.1.11, introducing prompt caching that reduces repeated system prompt costs by ~90% via automatic cache control on long tool results.

The update also adds an error-aware /run command that displays colored failure previews and offers to analyze errors with the agent. Other additions include /revisit for scanning shelved GitHub issues, expanded /doctor support for Java, Ruby, and C/C++ projects, token speed metrics in /profile, colored diff previews for file overwrites, and native desktop notifications for long completions. These features build on yoyo’s core loop: every ~8 hours, it reads its own source, plans improvements, runs tests, and commits changes if successful — all without human-written code. The agent now spans 100,000+ lines, 3,800+ tests, and 71 source files after 107 days of autonomous evolution.
The catch: Despite its autonomy, yoyo-evolve remains dependent on external test suites to validate changes, raising questions about how it assesses improvement in domains without clear pass/fail criteria.

Use Cases
  • Developers testing autonomous code improvement in Rust projects
  • Teams exploring AI-driven issue triage and revisit planning
  • Engineers studying long-term AI agent behavior in public repositories

Source: yologdev/yoyo-evolve — based on the README and release notes.

AI Agents Reshape Open Source Through Modular Skill Sharing 🔗

Open source projects now build interoperable agent skills and toolchains, enabling reusable, composable AI workflows across domains and platforms.

A defining pattern in open source is the rise of modular, interoperable AI agent skills that transcend individual projects to form a shared ecosystem. Rather than monolithic agents, developers are creating discrete, reusable capabilities—skills—that agents can dynamically load and combine. This shift is evident in repositories like VoltAgent/awesome-agent-skills, which curates over 1,000 community-contributed skills compatible with Claude Code, Codex, and Cursor, and vercel-labs/skills, offering a CLI-driven skills tool (npx skills) for instant skill injection.

Projects such as Panniantong/Agent-Reach and mvanhorn/last30days-skill exemplify this by providing specialized skills for web search and topic synthesis, respectively, enabling agents to gather and reason over live data without custom integration.

Orchestration layers are emerging to manage these skills at scale. omnigent-ai/omnigent acts as a meta-harness that lets users swap agent backends (Claude Code, Codex, etc.) while enforcing policies and enabling real-time collaboration. Similarly, archestra-ai/archestra provides an enterprise platform with an MCP registry, gateway, and orchestrator to govern agent fleets, while workweave/router optimizes cost and latency by routing prompts to the most efficient model in under 50ms. The MCP (Model Context Protocol) framework is becoming a unifying standard, with activepieces/activepieces boasting ~400 MCP servers and dagucloud/dagu allowing any AI agent to manage declarative YAML-based workflows.

Domain-specific agent systems are also maturing. calesthio/OpenMontage turns coding agents into video production studios via 12 pipelines and 500+ agent skills, while Unclecheng-li/VulnClaw and larlarua/AutoCVE demonstrate agent-driven security workflows—automating vulnerability discovery, exploitation, and reporting through chained MCP tools. Even niche areas like finance are seeing agentification: xbtlin/ai-berkshire implements a multi-agent adversarial research framework inspired by Buffett and Munger, using Claude Code as its foundation.

The catch: Despite rapid growth, the agent skill ecosystem remains fragmented across competing standards (MCP, proprietary agent interfaces), with many skills poorly documented, versioned, or tested—raising concerns about reliability, security (as highlighted by NVIDIA/SkillSpector), and long-term interoperability in production settings.

Open Source Embraces Modular LLM Skills as First-Class Code 🔗

Projects now treat reusable AI capabilities as versioned, shareable components across agent frameworks

A clear pattern is emerging in open source where LLM-powered capabilities are being packaged as discrete, interoperable skills rather than monolithic applications. This shift treats AI functions—like code review, design prototyping, or cybersecurity analysis—as modular units that can be composed, shared, and reused across different agent environments. Projects such as mukul975/Anthropic-Cybersecurity-Skills exemplify this by offering 754 structured cybersecurity skills mapped to major frameworks, designed to work with Claude Code, Copilot, and Cursor.

Similarly, mattpocock/skills provides a curated set of shell-based skills for real-world engineering tasks, while VoltAgent/awesome-agent-skills curates over 1,000 such skills from official and community sources.

This modularity extends beyond individual tools into full harnesses. revfactory/webtoon-harness uses 27 AI agents coordinated via Claude Code to generate a full webtoon episode, demonstrating how complex workflows can be assembled from specialized skills. omnigent-ai/omnigent takes this further as a meta-harness that orchestrates multiple agents—Claude Code, Codex, Cursor—allowing users to swap underlying agents without rewriting workflows. The rise of proxy and routing layers like tashfeenahmed/freellmapi (which aggregates free tiers from 16 LLM providers) and decolua/9router (connecting to 40+ free providers) underscores a parallel trend: abstracting model access to enable skill portability across LLMs.

Skills are also being optimized for longevity and reuse. microsoft/SkillOpt trains reusable natural-language skills for frozen LLMs through trajectory-driven edits, producing deployable best_skill.md artifacts. Meanwhile, EverMind-AI/EverOS introduces a portable memory layer that persists across agents and platforms, addressing a key limitation in agent state management.

The catch: While the vision of composable, LLM-agnostic skills is compelling, the ecosystem remains fragmented—skills often tie to specific agent runtimes (like Claude Code), lack standardized interfaces beyond informal conventions, and vary widely in quality and documentation. True interoperability is still aspirational, and many projects rely on fragile prompt engineering rather than robust, testable skill contracts.

Use Cases
  • Engineers deploy cybersecurity skills in agent pipelines
  • Designers assemble prototypes using shared design systems
  • Researchers route prompts across free LLM tiers via proxy routers

Local-First Data Infrastructure Redefines Real-Time System Design 🔗

Open source projects are converging on decentralized, self-contained data layers that prioritize low-latency access and AI integration without central databases.

A clear pattern is emerging in open source data infrastructure: the rise of local-first, embedded, and AI-augmented systems that eliminate reliance on centralized databases while maintaining consistency and scalability. Projects like dagucloud/dagu exemplify this shift — a Go-based workflow engine that defines DAGs in YAML, runs entirely locally, and requires no external DBMS, yet still supports complex orchestration and AI agent management. This isn't just about simplicity; it's about architectural resilience.

Similarly, tigerbeetle/tigerbeetle, written in Zig, delivers a financial transactions database engineered for mission-critical safety and performance through a single-binary, ACID-compliant design that avoids the operational overhead of distributed clusters.

The trend extends beyond storage. etcd-io/etcd continues to serve as the backbone for Kubernetes-native systems, but its role is evolving — now often paired with local caches and edge nodes to reduce latency in distributed control planes. Meanwhile, Comcast/react-data-grid and apache/superset are being reimagined not as frontends to central warehouses, but as interfaces layered over local data caches or streaming pipelines, enabling real-time exploration without round-trip delays. Even AI-driven tools like ZhuLinsen/daily_stock_analysis and simonlin1212/a-stock-data are embedding data ingestion, transformation, and visualization into self-contained Python stacks that pull from multiple sources but operate independently — no ETL pipelines to a central lake required.

This reflects a deeper technical shift: developers are prioritizing data sovereignty, reduced latency, and operational simplicity by embedding storage, compute, and governance directly into applications. The pattern favors systems that are self-contained, observable, and AI-ready by design — where data doesn’t need to leave the edge to be useful.

The catch: While promising, this local-first approach risks creating data silos that hinder cross-system analytics and compliance auditing; many of these tools lack mature replication, backup, or multi-tenant security models, and integrating them into enterprise governance frameworks remains largely untested at scale.

Deep Cuts

Eridanus Powers Smarter LLMs with Function Calling 🔗

A OneBot-based framework enhancing AI agents with live2d desktop pet mode

AOrbitron/Eridanus · Python · 190 stars

Eridanus reimagines bot development by centering on LLM function calling as its core intelligence layer. Built in Python and compatible with OneBot 11, it enables developers to create agents that don’t just chat—they act. By integrating with Gemini and OpenAI APIs, Eridanus lets LLMs trigger real-world actions: sending messages, fetching data, or controlling smart devices through structured function calls.

Its standout feature? A pure Live2d desktop pet mode that runs without QQ integration—ideal for creators wanting an AI companion on their screen, animated and responsive. The framework abstracts protocol complexity while exposing hooks for custom skills, making it easier to prototype multimodal agents. Whether extending chatbots with tool use or building interactive desktop mascots, Eridanus offers a flexible foundation where AI doesn’t just respond—it executes.

The catch: It’s early-stage with limited documentation and a primarily Chinese-speaking community, making onboarding challenging for global developers despite its innovative approach.

Use Cases
  • Developers building AI agents that execute real-time API calls
  • Creators designing interactive Live2d desktop companions without QQ
  • Prototyping multimodal bots with Gemini/OpenAI function integration

Source: AOrbitron/Eridanus — based on the project README.

Quick Hits

openchamber OpenChamber provides a desktop and web interface to interact with the OpenCode AI agent, enabling seamless AI-powered development workflows. 6k
archestra Archestra delivers an enterprise AI platform with built-in guardrails, MCP registry, gateway, and orchestrator for secure, scalable AI deployment. 3.9k
Steam-Controller-Auto-Charge Steam-Controller-Auto-Charge automates controller charging by detecting magnetic puck alignment and triggering charge cycles via physical slam detection. 240
superset Apache Superset is an open-source data visualization and exploration platform that lets builders create interactive dashboards from diverse data sources. 73.6k
winget-pkgs Winget-pkgs hosts community-maintained Windows Package Manager manifests, enabling reliable, one-command installation and updates of Windows apps 10.8k
etcd Etcd is a distributed, reliable key-value store that ensures consistency and availability for critical system data in cloud-native infrastructures. 51.9k
Beyond GitHub

The AI Wire

What builders are reading today — the headlines, papers, and announcements that aren't trending repos.

From the labs & arXiv

Scikit-learn 1.9.0 Adds Narwhals for Broader DataFrame Support 🔗

New dependency streamlines interoperability with pandas, Polars, and others in Python ML workflows

scikit-learn/scikit-learn · Python · 66.5k stars Est. 2010 · Latest: 1.9.0

The latest release of scikit-learn, version 1.9.0, introduces narwhals as a core dependency to improve dataframe interoperability across its API.

This change allows the library to accept and return dataframe objects from pandas, Polars, and other libraries without requiring conversion to NumPy arrays, reducing boilerplate in data science pipelines. The update supports Python 3.11 through 3.14 and drops older versions, aligning with modern scientific Python toolchains.

Internally, narwhals acts as a thin compatibility layer, enabling scikit-learn estimators to work seamlessly with dataframe inputs while preserving metadata like column names. This addresses a long-standing friction point where users had to manually extract features as DataFrames but needed to convert to arrays for modeling, then reconstruct results. Now, methods like fit() and predict() can preserve dataframe structure when appropriate, simplifying workflows in exploratory analysis and reporting.

The release continues scikit-learn’s shift toward stricter dependency management, with updated minimums for NumPy (1.24.1), SciPy (1.10.0), and Joblib (1.4.0). These changes reflect ongoing efforts to reduce technical debt and improve performance, though they may require updates in legacy environments. Nightly wheels and CI integrations via GitHub Actions and CircleCI ensure broad compatibility testing across platforms.

Built on 15 years of community development, scikit-learn remains a foundational tool for classical machine learning in Python, offering consistent APIs for classification, regression, clustering, and preprocessing. Its strength lies in documentation, stability, and broad adoption across academia and industry.

The catch: While narwhals improves dataframe handling, scikit-learn still lacks native support for GPU acceleration or distributed computing, limiting its use in large-scale deep learning or real-time inference systems compared to frameworks like PyTorch or TensorFlow.

Use Cases
  • Data scientists preprocessing data with pandas before modeling
  • ML engineers building pipelines that preserve dataframe metadata
  • Researchers integrating scikit-learn with Polars-based workflows

Source: scikit-learn/scikit-learn — based on the README and release notes.

More Stories

Open WebUI adds streamed reasoning display in latest release 🔗

Folder uploads now preserve subfolder structure in knowledge bases

open-webui/open-webui · Python · 143.7k stars Est. 2023

The open-source AI interface Open WebUI released v0.10.2 with streamed reasoning display, showing model thinking processes as they generate in real time within chat overviews and exported conversations.

Folder uploads to knowledge bases now maintain original subdirectory hierarchies instead of flattening files, addressing issue #26130. Administrators gained finer control over memory handling via a new toggle to keep memory tools active without injecting stored memories into system context. Automatic memory saving was refined to prioritize enduring details like user preferences over transient data such as meals or routine events unless explicitly requested. Built on Python, the platform continues supporting Ollama, OpenAI-compatible APIs, and MCP tool servers for extensible agent creation.
The catch: Despite recent UI polish, the project’s 248 open issues suggest ongoing stability challenges in complex self-hosted deployments at scale.

Use Cases
  • Developers testing local LLMs with visible reasoning traces
  • Teams organizing hierarchical document sets in knowledge bases
  • Admins balancing memory utility with context window limits in agents

Source: open-webui/open-webui — based on the README and release notes.

GitHub Copilot macOS system prompt revealed in latest leak 🔗

Repository adds June 18 update for AI coding agent alongside Claude and GPT-5.5 prompts

asgeirtj/system_prompts_leaks · JavaScript · 47.6k stars Est. 2025

The asgeirtj/system_prompts_leaks project published the full system prompt for GitHub Copilot’s macOS application on June 18, 2026, joining recent additions for Claude Sonnet 5, Claude Design, and GPT-5.5 Codex. The leak exposes the hidden instruction set governing Copilot’s behavior in Apple’s ecosystem, including tool usage rules and response boundaries.

Maintained in JavaScript, the repo now tracks over two dozen AI agents across Anthropic, OpenAI, Google, and xAI, with updates logged as recently as July 1 for Claude Sonnet 5. Developers use these prompts to debug agent behavior, refine prompt engineering, or build compliant AI wrappers. The project’s rapid update cadence reflects active community scraping of evolving model guardrails.
The catch: Prompts are reverse-engineered via unofficial channels, raising accuracy and legal concerns for production use.

Use Cases
  • AI engineers auditing Copilot’s macOS decision logic
  • Prompt designers reverse-engineering Claude Fable 5’s constraints
  • Researchers comparing system prompts across GPT-5.5 variants

Source: asgeirtj/system_prompts_leaks — based on the project README.

MediaPipe v0.10.35 adds NPU support and API3 migration 🔗

Latest release improves Android acceleration and JavaScript task exports

google-ai-edge/mediapipe · C++ · 35.9k stars Est. 2019

MediaPipe’s v0.10.35 release focuses on platform-specific refinements rather than broad framework shifts.

Android gains NPU acceleration via JIT compilation, while iOS Tasks now use Framework-type CocoaPods packaging. JavaScript sees fixes to broken exports and removal of “subgroups-f16” references. Core updates include migrating FromImageCalculator and ToImageCalculator to API3 with added tests, optional GPU service for iOS image-to-tensor conversion, and enhanced tensor debugging with histogram data. The release also drops unused analytics references and adds configurable landmark smoothing policies. Despite steady commits and a 1-day-old last push, the project maintains 572 open issues, indicating ongoing maintenance demands. The catch: MediaPipe’s graph-based calculator model, while flexible, presents a steep learning curve for developers unfamiliar with its stream-processing paradigm, potentially slowing adoption in teams seeking simpler ML integration paths.

Use Cases
  • Develop real-time hand tracking on Android using GPU-accelerated pipelines
  • Deploy audio classification models to iOS apps via MediaPipe Tasks framework
  • Prototype gesture-controlled web interfaces with JavaScript solution bundles

Source: google-ai-edge/mediapipe — based on the README and release notes.

Quick Hits

keras Keras simplifies deep learning with an intuitive, high-level API for rapid model building and experimentation. 64.1k
DeepSpeed DeepSpeed optimizes distributed training and inference, enabling scalable, efficient AI model development at massive scale. 42.6k
dify Dify provides a production-ready platform to build, deploy, and manage agentic workflows with minimal code. 147.3k
OpenBB OpenBB delivers a unified financial data platform empowering analysts, quants, and AI agents with real-time insights. 69.9k
shap SHAP explains any ML model’s output using game theory, offering transparent, feature-level interpretability for trustworthy AI. 25.6k

Drake’s SNOPT Integration Boosts Precision in Robotic Motion Planning 🔗

Pre-compiled solver enhances trajectory optimization for complex multi-contact dynamics in v1.54.0

RobotLocomotion/drake · C++ · 4.1k stars Est. 2014 · Latest: v1.54.0

The latest Drake release, v1.54.0, now ships with a pre-compiled version of SNOPT, a leading sparse nonlinear optimization solver, directly integrated into its Mathematical Program toolbox.

This move eliminates a longstanding friction point for roboticists: the need to manually compile and link external solvers when tackling trajectory optimization problems involving friction, contact, and actuator limits.

SNOPT’s inclusion strengthens Drake’s ability to handle large-scale, non-smooth optimization problems — common in legged locomotion and manipulation tasks where systems transition between dozens of contact modes. By embedding the solver, Drake reduces setup overhead and improves numerical reliability, particularly when solving for time-optimal trajectories under torque and friction constraints. The integration acknowledges sustained support from Philip E. Gill and Elizabeth Wong, whose work on SNOPT underpins this advancement.

Technically, the solver plugs into Drake’s existing solvers interface, allowing users to invoke it via SnoptSolver() within optimization programs written in C++ or Python. This maintains Drake’s model-based ethos: users define robot dynamics, constraints, and cost functions symbolically, then let the framework handle the math. The binary release ensures consistency across Linux and macOS platforms, avoiding version skew that previously hampered reproducibility in research pipelines.

For teams working on dynamic maneuvers — such as quadrupedal jumping or dexterous hand reorientation — this update means faster iteration and fewer solver-related failures during nightly builds. It also signals Drake’s ongoing shift toward providing a more “batteries-included” experience for industrial adopters who prioritize turnkey reliability over source-level tinkering.

The catch: While SNOPT excels at medium-scale nonlinear programs, it remains less competitive than interior-point methods for very large-scale convex problems (e.g., whole-body control with hundreds of contacts), where users may still prefer OSQP or Gurobi for speed, despite the added integration cost.

Use Cases
  • Legged robots optimizing jump trajectories over uneven terrain
  • Industrial arms planning force-controlled assembly sequences
  • Humanoid robots computing dynamic balance recovery motions

Source: RobotLocomotion/drake — based on the README and release notes.

More Stories

Robots plugin updates improve Rhino 8 robot path simulation 🔗

Version 2.2.2 fixes slowdowns and adds fly-by zone handling for industrial arms

visose/Robots · C# · 376 stars Est. 2015

The visose/Robots plugin for Rhino 8 and Grasshopper received a performance-focused update in version 2.2.2, released July 2026.

It now resolves slow program generation caused by repeated motion warnings in large toolpaths and improves efficiency when handling numerous targets with unnamed custom speeds, zones, or tools. A key addition is command-level fly-by compatibility, allowing compatible motions to retain smooth transitions while incompatible ones correctly register as stop points. The plugin supports ABB, KUKA, UR, Staubli, and six other robot brands, enabling users to load robot systems from XML/3DM libraries, simulate paths, and generate manufacturer-specific code via Grasshopper components or its .NET 8 API. Samples may trigger temporary assembly warnings, which auto-resolve on load.
The catch: Despite a decade of development, the project shows no open issues but has stagnant traction, with no major feature expansion beyond core simulation and code generation in recent years.

Use Cases
  • Simulate KUKA arm paths in Grasshopper for collision checking
  • Generate UR program code from Rhino-modeled toolpaths
  • Test Staubli tool changes using custom post-processors in .NET apps

Source: visose/Robots — based on the README and release notes.

OpenArm 1.1 Refines Teleoperation for Humanoid Arm Research 🔗

Release improves hardware reliability and data collection consistency for contact-rich tasks

enactic/openarm · MDX · 2.7k stars Est. 2024

The enactic/openarm project released version 1.1, focusing on incremental but practical upgrades to its 7DOF humanoid arm system. Key improvements include automated zero-position calibration, reducing setup errors in bilateral teleoperation, and a modular camera mount base now supporting Realsense D435 and D405 sensors for standardized data collection.

The J5 casing redesign with a rubber band interface enhances elbow motion tracking, addressing prior ambiguity in redundant joint control during leader-follower teleoperation. These updates respond directly to community feedback, aiming to improve reproducibility in physical AI research. Builders use OpenArm for imitation learning, reinforcement learning in simulation via MuJoCo and MoveIt2, and real-world deployment in tasks requiring force feedback and gravity compensation. The system remains priced at $6,500 for a bimanual setup, targeting research labs. The project relies on niche hardware integration and ROS2, limiting accessibility for teams without robotics infrastructure or funding for certified builds.

Use Cases
  • Researchers collecting contact-rich task data
  • Labs training humanoid arms via imitation learning
  • Teams testing force-feedback teleoperation setups

Source: enactic/openarm — based on the README and release notes.

Copper-rs Enables Deterministic Robotics Across Hardware 🔗

Rust-based OS lets developers run identical robot code from simulation to bare-metal

copper-project/copper-rs · Rust · 1.4k stars Est. 2020

Copper-rs provides a deterministic runtime and SDK for robotics applications, ensuring bit-for-bit identical execution across environments. Built in Rust, it targets sub-microsecond latency with zero-alloc, data-oriented design. The project bridges with ROS2 via Zenoh and supports deployment from Linux workstations and SBCs to bare-metal MPUs, including STM32H7 microcontrollers.

Live browser demos showcase real applications like BalanceBot and Flight Controller, which run unchanged on physical hardware such as Raspberry Pi and drones. Developers bootstrap projects via cargo cunew and run with cargo run, compiling the same graph for simulation, workstation, or embedded targets. Despite steady activity — last commit one day ago — the project maintains only 94 forks and 17 open issues after nearly six years. The catch: Copper-rs remains niche in adoption, with limited evidence of large-scale production use beyond demonstrators and academic prototypes, raising questions about its readiness for complex, multi-robot industrial systems.

Use Cases
  • Developers testing robot algorithms in browser before hardware deployment
  • Engineers porting ROS2-based robots to deterministic Rust runtime
  • Teams building airborne, aquatic, or legged robots needing bit-identical replay

Source: copper-project/copper-rs — based on the project README.

Quick Hits

URDF-Studio URDF-Studio lets builders visually design, test, and export robot models with AI-assisted workflows and MuJoCo integration — no coding needed for complex kinematics. 431
cloisim Cloisim enables rapid multi-robot simulation setup in Unity3D using SDFormat and seamless ROS2 bridging for realistic, scalable testing. 176
rtabmap RTAB-Map provides real-time SLAM and loop closure for robots using RGB-D or stereo cameras, enabling robust 3D mapping and localization in diverse environments. 3.9k
autoware Autoware delivers a full-stack, production-ready autonomous driving system with perception, planning, and control — all open-source and Docker-deployable. 11.8k
nicegui NiceGUI lets builders create sleek, interactive web UIs purely in Python — no HTML/CSS/JS required — for fast prototyping of robot dashboards and tools. 16k

Maigret Expands OSINT Reach with Proxy Integration and Recursive Search 🔗

Latest release adds Tor/proxy documentation, blocks bypass, and username chaining for deeper investigations

soxoj/maigret · Python · 34.7k stars Est. 2020 · Latest: v0.6.2

Maigret, the Python-based OSINT tool for username reconnaissance, has strengthened its utility for investigators and red teams with recent updates focused on stealth and depth. The v0.6.

2 release, built over incremental commits since mid-2024, now includes documented Tor and proxy usage, enabling users to route requests through residential or rotating networks to evade rate limits and IP-based blocks—a critical feature when scanning 3,000+ sites for username presence.

Beyond anonymity, Maigret’s core strength lies in recursive discovery: it doesn’t just check if a username exists; it scrapes profile pages for linked accounts, then iteratively investigates those new identifiers. This chaining uncovers obscured connections across platforms, turning a single handle into a network map. The tool extracts available data—bio, links, timestamps—directly from HTML or site APIs, requiring no API keys.

Users can tailor scans via tags (e.g., --tags social,usa) to target specific categories or regions, or run -a for full coverage. Results are saved in structured formats, with automatic site database updates pulled from GitHub every 24 hours to keep the target list current. Recent fixes address false positives, such as rejecting URLs or emails mistakenly parsed as usernames, and resolve Instagram-specific login title staleness caused by stale rate-limit markers.

The project remains actively maintained, with dependabot keeping dependencies like lxml, aiodns, and black current, and the lead developer resolving site-specific issues (e.g., Instagram, new site additions) promptly. A new web UI option and Telegram bot lower the barrier for casual use, while library imports allow integration into custom Python workflows.

The catch: Maigret’s effectiveness depends on target sites; sites block or CAPTCHA walls, or require manual intervention or proxy setup—though documentation is not a guarantee against sophisticated anti-bot measures.

Use Cases
  • Security researchers mapping attacker infrastructure via username reuse
  • Journalists verifying social media presence across niche platforms
  • Red teamers profiling targets during reconnaissance phase of engagement

Source: soxoj/maigret — based on the README and release notes.

More Stories

Wazuh patches critical security flaws in latest manager update 🔗

v4.14.5 addresses buffer overflows, RBAC bypass, and path traversal risks

wazuh/wazuh · C++ · 16k stars Est. 2015

The Wazuh project released v4.14.5, patching multiple high-severity vulnerabilities in its manager component.

Fixes include a buffer overflow in regex match processing ([#35106]), uncontrolled memory allocation from crafted packets ([#35173], [#35412]), and a rate limit bypass on the /events endpoint ([#35077]). Additional patches resolve RBAC bypass allowing privilege escalation ([#35307]), path traversal via agent group name ([#35230]), and size_t underflow in remote message handling ([#35193]). Agent-side updates corrected false positives in rootcheck ([#34734]), FIM deadlocks ([#34735]), and macOS Ventura SCA policy errors. Despite steady traction and 15,993 stars, the project maintains 2,921 open issues, indicating ongoing complexity in scaling its unified XDR and SIEM capabilities across diverse environments.
The catch: While strong in threat detection and compliance, Wazuh’s deep integration with the Elastic Stack may pose operational overhead for teams seeking lightweight, self-contained security tooling.

Use Cases
  • Security teams monitor cloud workloads for intrusions
  • Auditors verify PCI-DSS compliance via log analysis
  • Admins detect file changes in containerized environments

Source: wazuh/wazuh — based on the README and release notes.

Algo VPN v2.0.1 patches Ansible 12 compatibility and cloud bugs 🔗

Maintenance release fixes deployments across AWS, GCE, Vultr, and Scaleway

trailofbits/algo · Python · 30.3k stars Est. 2016

The Algo VPN project released v2.0.1, a maintenance update focused on restoring compatibility with Ansible 12 and resolving cloud provider deployment failures.

Key fixes address Ansible 12’s stricter boolean type checking and double-templating issues that were breaking Jinja2 templates during setup. Specific cloud patches resolve AWS EC2 and Lightsail API errors, GCE JSON parsing failures, Vultr region string conversion bugs, and Scaleway’s broken organization_info module. DigitalOcean error handling was improved, and the update-users script now functions correctly. The release also removes the unused pyopenssl and boto dependencies and adds the missing ansible.utils collection to prevent 'ipmath' filter errors. Notably, Exoscale support was dropped due to CloudStack API deprecation. Algo continues to automate secure WireGuard and IPsec VPN servers on major cloud platforms using Ubuntu 22.04 LTS with automatic updates, generating client configs and QR codes for iOS, macOS, Android, and Windows.
The catch: Despite its security focus, Algo does not claim to provide anonymity or resist advanced state-level adversaries like FSB or MSS, limiting its use for high-threat models.

Use Cases
  • Developers deploying personal VPNs on DigitalOcean
  • Remote workers securing iOS and Linux traffic via WireGuard
  • Teams managing SSH tunnels for internal service access

Source: trailofbits/algo — based on the README and release notes.

Subfinder Drops Dead Meta Source in v2.14.0 Release 🔗

Tool removes non-functional Facebook API integration after upstream discontinuation

projectdiscovery/subfinder · Go · 13.9k stars Est. 2018

Projectdiscovery/subfinder’s latest release, v2.14.0, removes the Facebook (Meta CT) passive source after Meta discontinued its upstream API.

The change, noted in release notes as a warning, eliminates a non-functional integration that was already causing errors in automation and CI jobs. Users referencing this source in configurations must now remove it to avoid confusion.

The update also adds the new Sub.md passive source, improves error handling for non-200 responses, and fixes rate-limit bugs across providers like Hackertarget, Shodan, and IntelX. These changes refine subfinder’s core promise: fast, lightweight passive subdomain enumeration using curated online sources.

Built in Go and optimized for speed, subfinder remains a go-to for bug bounty hunters and penetration testers relying on passive reconnaissance. Its modular design supports JSON, file, and stdout output, with stdin/stdout integration for workflow automation.

The catch: Despite its speed and stealth, subfinder’s passive-only model limits discovery to publicly indexed data, missing subdomains only visible through active DNS queries or brute force.

Use Cases
  • Bug bounty hunters enumerate subdomains passively during reconnaissance
  • Security teams automate subdomain discovery in CI/CD pipelines
  • Researchers gather OSINT data using curated passive sources without active scanning

Source: projectdiscovery/subfinder — based on the README and release notes.

Quick Hits

bbot bbot is a recursive internet scanner that automates reconnaissance for hackers by chaining tools to discover subdomains, ports, and vulnerabilities across targets. 10k
ImHex ImHex is a powerful, retina-friendly hex editor with advanced pattern matching and disassembly, built for reverse engineers and low-level developers working late into the night. 54k
juice-shop OWASP Juice Shop is a deliberately insecure, intentionally designed to teach web app security through realistic, exploitable flaws → *Too long and awkward. Let’s refine.* 13.4k
bunkerweb bunkerweb is a cloud-native, open-source WAF that protects web apps with modular security rules, auto-scaling, and Kubernetes-native deployment for modern threat defense. 10.7k
CheatSheetSeries OWASP Cheat Sheet Series delivers concise, authoritative security best practices — from input validation to crypto — so builders can implement defenses quickly and correctly. 32.4k

TensorFlow 2.21 Drops Python 3.9, Adds JPEG XL and int2 Support 🔗

Breaking changes and micro-optimizations target edge deployment as framework nears 11th year

tensorflow/tensorflow · C++ · 195.9k stars Est. 2015 · Latest: v2.21.0

TensorFlow’s latest release, v2.21.0, arrives not with fanfare but with a quiet removal: official support for Python 3.

9 is gone. The change, listed under breaking changes, signals the project’s steady push toward newer runtimes while shedding legacy baggage. For builders maintaining older pipelines, this means updating CI/CD or risking silent failures if overlooked.

More substantively, the release tightens TensorFlow Lite’s grip on integer quantization—a critical path for deploying models on microcontrollers and edge AI accelerators. New int2 and int4 types now propagate through the SQRT, EQUAL, NOT_EQUAL, cast, fully_connected, and slice operators. These aren’t theoretical gains; they directly reduce model footprint and inference latency on resource-constrained hardware. Pair that with JPEG XL support in tf.image.decode_image, and TensorFlow is clearly optimizing for bandwidth-sensitive vision workloads where newer codecs matter.

The release also exposes NoneTensorSpec in tf.data, a small but meaningful API refinement. Developers can now explicitly check for unspecified tensor shapes in dataset pipelines—a niche but welcome clarity when debugging dynamic batching or ragged tensors.

Underneath these tweaks, TensorFlow remains a C++-heavy, Python-first framework with a vast surface area. Its strength lies in the integrated ecosystem: Keras, TensorBoard (now decoupled as a separate dependency), TF Serving, and TPU tooling. Yet the project’s maturity brings friction. The sheer scale means even minor version bumps require testing across heterogeneous hardware—GPUs, CPUs, and emerging NPUs—where a single opset change can ripple through custom kernels.

The catch: TensorFlow’s extensive abstraction layers, while enabling rapid prototyping, can obscure performance bottlenecks. Builders chasing sub-millisecond latency on edge devices may find themselves diving into XLA compiler logs or writing custom C++ ops to bypass layers meant for flexibility, not speed.

Use Cases
  • Train vision models using JPEG XL for reduced storage
  • Deploy int2-quantized networks on battery-powered sensors
  • Debug data pipelines with explicit None shape checking in tf.data

Source: tensorflow/tensorflow — based on the README and release notes.

More Stories

rtk v0.43.0 adds OpenShift and Pulumi CLI filtering 🔗

Latest release extends token-saving proxy to Kubernetes and IaC workflows

rtk-ai/rtk · Rust · 67.6k stars 5mo old

The rtk CLI proxy, which compresses command output to reduce LLM token use by 60-90%, now filters OpenShift (oc) and Pulumi commands in its v0.43.0 release.

Built as a single Rust binary with zero dependencies, rtk intercepts tool output before it reaches AI agents like Claude Code or Gemini, stripping noise while preserving intent. The update adds shared Kubernetes filtering for oc and granular filters for Pulumi’s preview, up, destroy, refresh, and stack subcommands. Installation remains via Homebrew, cargo, or pre-built binaries for Linux, macOS, and Windows. Recent bug fixes include guarded AWS S3 output, improved diff exit code handling, and Docker agent command prioritization. Despite 67,585 stars and rapid growth, the project tracks 1,455 open issues, indicating active scaling challenges.
The catch: While effective for common dev commands, rtk’s filtering rules may require manual tuning for niche or heavily customized toolchains, and its long-term reliability in enterprise CI/CD pipelines remains unproven at scale.

Use Cases
  • Developers reducing Claude Code token costs during TypeScript debugging
  • DevOps teams minimizing LLM context bloat from Kubernetes audit logs
  • Infrastructure engineers cutting Pulumi CLI output overhead in AI-assisted IaC reviews

Source: rtk-ai/rtk — based on the README and release notes.

Zig Framework Bridges Native Apps and Web Views 🔗

Developers build lightweight desktop/mobile apps using Zig with flexible web rendering options

vercel-labs/zero-native · Zig · 4.3k stars 1mo old

Vercel Labs’ zero-native enables developers to create native desktop and mobile applications using Zig, treating web views as one component rather than the entire app foundation. The framework allows developers to build app chrome, menus, and system integrations with native Zig code while embedding web content via system WebViews (WKWebView on macOS, WebKitGTK on Linux) or bundled Chromium through CEF for consistent rendering. A CLI tool streamlines setup: zero-native init my_app --frontend next generates a project structure where Zig handles the native shell and frontend frameworks like Next.

js manage the UI. Recent v0.3.0 improvements focus on keyboard shortcuts, adding manifest-driven configuration via app.zon, runtime shortcut delivery to Zig Event.shortcut, and typed JavaScript events via window.zero. The update also fixes shortcut matching across platforms and enables trusted child WebView bridges on Windows WebView2.

The catch: At just two months old with 44 open issues, the framework remains early-stage, and long-term stability, community support, and real-world scalability beyond small utilities are unproven.

Use Cases
  • Developers building native desktop tools with embedded dashboards
  • Teams creating cross-platform apps needing flexible web/native splits
  • Developers seeking small binaries with instant rebuilds using Zig and web UIs

Source: vercel-labs/zero-native — based on the README and release notes.

Quick Hits

deno Deno provides a secure, modern runtime for JavaScript and TypeScript with built-in tooling and web-standard APIs. 107.3k
ripgrep Ripgrep blazingly fast recursive regex search that respects gitignore, making code navigation efficient and accurate. 65.7k
json nlohmann/json offers a lightweight, header-only C++ library for intuitive and standards-compliant JSON handling. 50k
frp frp enables secure, low-latency exposure of local services behind NAT/firewall via a fast, configurable reverse proxy. 107.7k
whisper.cpp whisper.cpp delivers high-performance, on-device speech-to-text inference in C/C++ with minimal dependencies and OpenAI compatibility. 51.2k

Photobooth App Adds Rclone Sync for Seamless Media Backup 🔗

New plugin simplifies cloud and local storage integration across platforms

photobooth-app/photobooth-app · Python · 291 stars Est. 2022 · Latest: v8.7.0

The photobooth-app has released version 8.7.0, introducing a new synchronization tool powered by Rclone to streamline media handling for DIY photobooth builders.

This update addresses a long-standing pain point: reliably transferring captured images, GIFs, and collages from the booth to external storage without manual intervention or fragmented workflows.

Built on Python with a Vue3 frontend, the app now includes an official Rclone synchronizer plugin that leverages the newly added rclone-bin-api PyPI package. This dependency bundles a recent version of Rclone for Windows, macOS, and Linux (including ARM), eliminating the need for users to install Rclone separately. The plugin supports synchronization to cloud services like Google Drive and Nextcloud, as well as local USB drives and FTP servers—replacing older, less reliable methods such as QR code sharing and basic file transfer scripts.

According to the release notes, the Rclone sync covers most existing functionality while maintaining backward compatibility with legacy synchronizers during the v8 lifecycle. Backend improvements also include enhanced TCP keepalive settings for multicamera setups using pynng, preventing dropped connections during extended still captures, and a new crash-signaling mechanism that disables automatic recovery for irrecoverable configuration errors.

On the frontend, the "download portal" has been renamed to "sharepage" with minor visual refinements, and all dependencies have been updated to improve stability and security. The project continues to support DSLR, Raspberry Pi (via Picamera2), and webcameras, with features like live preview, boomerangs, 3D wigglegrams, and LED countdown signaling via WLED integration.

The catch: Despite active maintenance—with commits as recent as zero days ago—the project shows signs of stagnation, with only 291 stars and 20 open issues after nearly four years, suggesting limited broader adoption beyond a niche DIY and maker community.

Use Cases
  • Wedding photographers automating guest photo backup to Google Drive
  • Raspberry Pi builders creating synchronized multi-camera 3D wigglegram booths
  • Event organizers deploying USB-based photo sharing without internet dependency

Source: photobooth-app/photobooth-app — based on the README and release notes.

More Stories

RealSense SDK adds Jetson depth boost and ROS2 compression 🔗

Latest release improves close-range sensing and modernizes testing for robotics workflows

realsenseai/librealsense · C++ · 8.9k stars Est. 2015

The realsenseai/librealsense project released v2.58.2, introducing a close-range depth enhancement for D400 cameras on Nvidia Jetson platforms via a dedicated Debian package.

This update integrates improved depth processing directly into the libRS Viewer, boosting accuracy at short distances. A new React-based viewer preview replaces legacy interfaces, offering a more responsive UI for developers.

On the software side, the test framework has been modernized, migrating legacy unit tests to pytest with expanded multi-device support and regression testing. ROS2 integration now enables rosbag2 compression for efficient data recording and playback, swapping the end-of-life Iron distribution for Kilted.

Platform fixes address Jetson JP6 metadata loss over USB3 and correct D405 RGB auto-exposure metadata on firmware 5.17.3+. Global-shutter GMSL cameras gain auto-exposure and gain limit controls, while CMake refactoring exposes rs_lz4 and adds support for Y8I/Y12I/Y16I infrared formats via rs-enum.

The catch:** Despite active development, the SDK’s reliance on Intel’s hardware ecosystem limits adoption for teams using non-RealSense depth sensors, creating a vendor lock-in risk for cross-platform robotics projects.

Use Cases
  • Robotics teams integrating depth sensing for navigation
  • Drone developers enabling obstacle avoidance and tracking
  • 3D scanning applications requiring calibrated color and depth streams

Source: realsenseai/librealsense — based on the README and release notes.

Pulp Platform AXI IP gains FuseSoC integration and lint CI 🔗

Latest release adds build system support and Verilator/Yosys checks

pulp-platform/axi · SystemVerilog · 1.6k stars Est. 2018

The pulp-platform/axi project released v0.39.10, introducing core file entries for FuseSoC package database integration (#404), enabling easier inclusion in SoC builds.

A new GitHub Action now runs Verilator lint and yosys-slang elaboration (#414), improving early-stage code quality. The release also split axi_demux_id_counters into a dedicated module with added port control (#406) and introduced DECL macros for cleaner parameter port lists (#413). Fixes addressed AXI-Lite wstrb protocol violations (#412), FIFO macro termination (#403), and memory interface width mismatches (#418), while treewide lint cleanup resolved multiple warnings (#416). Despite eight years of development, the project maintains slow-burn traction with 69 open issues and a focus on synthesizable SystemVerilog for ASIC/FPGA interconnects. The catch: The modular design prioritizes flexibility over turnkey solutions, requiring users to compose networks from primitives rather than offering pre-configured topologies for common SoC patterns.

Use Cases
  • FPGA designers implementing custom AXI4 interconnects
  • ASIC teams building heterogeneous network-on-chip fabrics
  • Researchers verifying synthesizable SystemVerilog IP compliance

Source: pulp-platform/axi — based on the README and release notes.

Bruce Firmware Adds NetCut ARP Module for Wardriving 🔗

Latest release enhances wireless attack tools with ARP spoofing and Flipper Zero-style app loading

BruceDevices/firmware · C++ · 6k stars Est. 2024

BruceDevices/firmware v1.15 introduces a NetCut-inspired ARP module for disrupting local network connections, expanding its offensive WiFi toolkit. The update also adds WDGoWars upload support for wardriving data logging and enhances the karma attack feature.

On the RF side, users gain a revamped Custom SubGhz UI with KeeLoq decoding and PN532 RFID emulation. Notably, the firmware now includes USB U2F/FIDO2 authentication support, broadening its hardware interaction capabilities. Under the hood, contributors addressed buffer overflow risks, fixed UI scaling on small screens, and improved Flipper Zero-style app loading via JavaScript enhancements. Despite active development — with commits as recent as two days ago — the project maintains 644 open issues, indicating ongoing stability and usability challenges.
The catch: While feature-rich, the firmware’s focus on offensive security limits its appeal for general embedded development and raises questions about long-term maintainability amid complex, niche dependencies.

Use Cases
  • Red teamers testing WiFi network resilience
  • Hardware hackers conducting BLE spoofing attacks
  • Security researchers emulating NFC/RFID access systems

Source: BruceDevices/firmware — based on the README and release notes.

Quick Hits

silhouette-card-maker Create custom card game components and proxies using Silhouette cutting machines with Python-driven design tools. 168
firmware Deploy community-maintained, open-source firmware to enhance and customize IP camera functionality beyond vendor limits. 2.1k
openwifi Build and experiment with a fully open-source WiFi baseband system on FPGA, including driver and software stack. 4.7k
ghdl Simulate VHDL designs with full support for IEEE 1076-2008/93/87 standards using a fast, open-source GHDL compiler. 2.8k
hwinfo Query detailed cross-platform hardware specs — CPU, RAM, GPU, and more — via a lightweight C++ library for system introspection. 715

SpacetimeDB v2.6.0 adds primary key support for views across key languages 🔗

Developers can now enforce data integrity in procedural views without sacrificing real-time sync

clockworklabs/SpacetimeDB · Rust · 24.8k stars Est. 2023 · Latest: v2.6.0

The latest release of SpacetimeDB introduces primary key support for procedural views in Rust, TypeScript, and C#, addressing a long-standing gap in data modeling flexibility. Previously, views—computed queries that expose derived state to clients—lacked the ability to declare primary keys, complicating client-side caching, optimistic updates, and reliable diff-based UI state management. With v2.

6.0, developers can now annotate views with primary key semantics, enabling SpacetimeDB to optimize view change detection and deliver more precise synchronization signals.

In Rust, the #[view] macro now accepts a primary_key parameter, as shown in the BitCraft Online-inspired example where player lists are keyed by id. TypeScript developers achieve the same via .primaryKey() on column definitions, while C# uses the primaryKey attribute in the View attribute. This alignment across languages reduces cognitive friction for polyglot teams and brings view behavior closer to that of base tables.

Beyond views, v2.6.0 includes CLI binary distribution improvements and event table automigration enhancements, streamlining deployment and schema evolution. The project continues to position itself as a single-binary backend alternative: application logic lives inside the database, compiled from Rust, C#, TypeScript, or C++, and executed in-process with ACID guarantees. State resides in memory for low-latency access, backed by a disk commit log for durability.

SpacetimeDB’s appeal lies in eliminating traditional server layers—no web servers, containers, or Kubernetes—by letting clients connect directly to the database and execute application logic as modules. This model shines in real-time applications like MMORPGs, collaborative tools, and live dashboards where millisecond synchronization matters. The BitCraft Online backend, powering thousands of concurrent players, runs entirely as a SpacetimeDB module.

The catch: While SpacetimeDB removes infrastructure complexity, it ties application logic tightly to the database runtime, making debugging, profiling, and third-party tool integration more challenging than in conventional server setups. Teams must evaluate whether the operational simplicity outweighs reduced observability and flexibility in tooling choices.

Use Cases
  • MMORPG developers syncing player state in real time
  • Collaborative app builders avoiding backend microservices
  • Live dashboard creators needing low-latency data synchronization

Source: clockworklabs/SpacetimeDB — based on the README and release notes.

More Stories

O3DE 26.05.0 Release Adds Wwise Audio SDK Improvements 🔗

Latest update refines audio integration workflow for game and simulation developers

o3de/o3de · C++ · 9.5k stars Est. 2021

The Open 3D Engine (O3DE) project released version 26.05.0 on July 1, 2026, focusing on stability and toolchain refinements rather than major features.

This point release includes updated documentation for the Wwise Audio Engine Gem, clarifying setup steps and version compatibility for developers integrating spatial audio into projects. Build instructions now specify CMake 3.24.0 as the minimum supported version, with explicit warnings against using release candidates. The release notes confirm no breaking API changes, targeting existing users seeking reliable updates. Despite steady commits and a recent push, the project maintains 3,529 open issues, indicating ongoing challenges in bug triage and feature completion.

Use Cases
  • Game studios building cross-platform titles with C++
  • Simulation creators needing real-time 3D rendering
  • Audio engineers implementing Wwise-based spatial sound

Source: o3de/o3de — based on the README and release notes.

Babylon.js adds HTML-in-Canvas support for richer web experiences 🔗

Latest release enables DOM-like interaction within 3D scenes via WebGPU-ready texture integration

BabylonJS/Babylon.js · TypeScript · 25.7k stars Est. 2013

The Babylon.js team has merged HTML-in-Canvas (WICG) support into the core engine, allowing developers to render and interact with HTML content directly inside 3D scenes using HtmlTexture. This feature, implemented via polyfill installer and interaction managers, lets UI elements like buttons or forms exist as textures on 3D objects while maintaining event handling and styling.

The update also includes fixes for Gaussian Splatting GPU picking, IBL voxelization, and Flow Graph Editor improvements such as auto-loaded playground snippets and stopped-state variable type handling. Engine telemetry now reports "Native" for name identification, and texture updates are more reliable via the new updateTextureData method. These changes reflect Babylon.js’s push toward seamless 2D/3D convergence in web applications, particularly for immersive dashboards, educational tools, and product configurators. Built with TypeScript and targeting WebGPU, the engine continues to prioritize modularity and tree-shaking ES6 imports.
The catch: HTML-in-Canvas remains experimental and polyfill-dependent, with performance varying across browsers and complex DOM structures still posing challenges for real-time 3D rendering at scale.

Use Cases
  • Develop interactive 3D product configurators with embedded HTML forms
  • Build educational simulations featuring clickable UI overlays on 3D models
  • Create data dashboards where charts and controls exist as textures in virtual space

Source: BabylonJS/Babylon.js — based on the README and release notes.

Jolt Physics v5.5.0 adds per-body stats and shape auto-adjustment 🔗

New release improves debugging and usability for game and VR developers

jrouwe/JoltPhysics · C++ · 10.6k stars Est. 2021

Jolt Physics v5.5.0 introduces JPH_TRACK_SIMULATION_STATS, enabling per-body performance tracking to identify simulation bottlenecks.

Shapes like BoxShape and CylinderShape now auto-reduce convex radius when overspecified, preventing errors and streamlining integration. Ragdoll handling gains CalculateConstraintPriorities to strengthen root-joint stability. The update adds Visual Studio 2026 support and fixes macOS clang 21 and Windows _CRTDBG_MAP_ALLOC compatibility issues. Despite steady adoption in titles like Horizon Forbidden West, the library’s narrow focus on rigid body dynamics means developers seeking soft body, fluid, or granular simulation must look elsewhere.
The catch: Jolt Physics does not natively support soft body or fluid dynamics, limiting its use in simulations requiring deformable materials.

Use Cases
  • Game developers implementing rigid body physics in VR titles
  • Engineers debugging performance hotspots in physics simulations
  • Teams integrating physics with multithreaded game engines using background loading

Source: jrouwe/JoltPhysics — based on the README and release notes.

Quick Hits

assimp Assimp loads over 40 3D file formats into a unified, clean data structure, simplifying asset integration for developers. 13k
ebiten Ebiten offers a dead-simple, lightweight 2D game engine in Go, ideal for rapid prototyping and minimal-overhead game development. 13.3k
godot-ai Hi-godot/godot-ai provides production-grade MCP server and AI tools for Godot, enabling easy integration of intelligent behaviors with zero cost and snap installation. 792
godot Godot Engine delivers a full-featured, multi-platform 2D and 3D game engine with a flexible scene system and GDScript scripting for cross-platform deployment. 113.4k
godot_heightmap_plugin Zylann’s Godot Heightmap Plugin enables procedural terrain generation in GDScript, letting builders sculpt realistic landscapes directly within the Godot editor. 2.2k
The Git Times AI Desk
Ask about today's stories — or hit “Ask about this” on any article to focus on one.

Unlock the Git Times AI desk to ask about today's stories and the AI model market.

Upgrade to Premium
Answers by the Git Times AI desk · verify before you ship