This repository is the official code release for the Memento research project by Microsoft AI Frontiers. Memento introduces a method for large language models to learn context management by segmenting chain-of-thought reasoning into blocks, compressing each block into a concise 'memento' (summary), and then evicting the original block from the KV cache during generation. This achieves significant peak KV cache reduction (2-3x) and throughput improvements (1.75x) with minimal accuracy loss, which can be recovered via reinforcement learning (CISPO). The repository is a hybrid of a data research pipeline and a software application overlay. The data pipeline (data/) ingests raw chain-of-thought traces from the OpenThoughts dataset, filters them by domain (math, code, science), and transforms them into the Memento format. This involves LLM-based boundary scoring, dynamic programming segmentation, and iterative LLM-based summarization with judge feedback, resulting in the OpenMementos dataset (228K traces). The application component is a surgical overlay on the vLLM inference engine (vllm/). It adds a BlockMaskingProcessor that acts as a token-by-token state machine, detecting block boundaries and orchestrating physical KV cache compaction. It also includes a BlockLengthCapLogitsProcessor to enforce block size limits. This overlay is designed to be patched onto a standard vLLM 0.13.x installation, modifying only the necessary Python files while restoring compiled .so interfaces for compatibility. Overall, this is a high-quality research codebase demonstrating strong engineering craftsmanship. The data pipeline is robust and well-structured, and the vLLM overlay is surgically precise. However, it has a high technical drift risk due to its tight coupling to a specific vLLM version and its reliance on proprietary LLM APIs (GPT-5.x) for the data generation pipeline, which limits full reproducibility without access to those specific models.
XLB is a differentiable, massively parallel Lattice Boltzmann Method (LBM) library written in Python, designed for physics-based machine learning and computational fluid dynamics (CFD). It provides a modular operator-based architecture with abstract base classes and backend-specific implementations for JAX (CPU/GPU/TPU), NVIDIA Warp (single-GPU CUDA), and Autodesk Neon (multi-GPU, multi-resolution). The library supports 2D and 3D simulations with multiple velocity sets (D2Q9, D3Q19, D3Q27), collision models (BGK, KBC, Smagorinsky LES), boundary conditions (equilibrium, bounce-back, regularized, Zou-He, hybrid, extrapolation outflow), and immersed boundary method (IBM) for fluid-structure interaction. A key differentiator is its native differentiability via JAX and Warp autodiff, enabling gradient-based optimization of initial conditions, geometry, or control parameters. The codebase also includes experimental out-of-core support for simulations exceeding GPU memory, multi-resolution grid refinement, and distributed computation across multiple devices. While primarily a software library for CFD practitioners and ML researchers, it is accompanied by research validation data and academic publications, giving it a hybrid application/research character.
This repository is a large, structured dataset of maimai rhythm-game charts, not an application. It contains over 1,800 maidata.txt files organized by in-game genre such as POPSアニメ, niconicoボーカロイド, 東方Project, ゲームバラエティ, オンゲキCHUNITHM and 宴会場, plus collection manifests and an index.json song catalog. Every chart file is stamped as produced by MaichartConverter/Neskol v1.0.9.0, and uses the same key/value metadata header followed by per-difficulty &inote_N blocks in a compact note notation that encodes tap positions, holds, slides, break notes, chords, BPM changes and cabinet/version information. The data spans maimai SD through maimai DX PRiSM and beyond, making it a comprehensive cross-version chart corpus. It can serve as ground truth for simulator ingestion, chart format analysis, difficulty prediction, archival, or generation tooling. The chart notation is consistent enough that a parser/lexer could be derived by example from these files, and the stable shortid/artistid/genreid fields allow relational joins with the included index.json and collection manifests. The main weaknesses are the absence of the converter source, a formal schema, validation tooling, or license; a few files contain placeholder values such as artistid=999 or empty difficulty levels. Overall this is a well-curated data asset with high reuse potential for anything that needs maimai chart content, but consumers should treat the format as reverse-engineered rather than officially documented.
The PiSight project transforms a Raspberry Pi Zero with a camera module into a plug-and-play USB webcam by implementing the USB Video Class (UVC) standard via the Linux Gadget API. The hardware setup involves mounting the Raspberry Pi and camera inside an Apple iSight enclosure. Software-wise, the project provides scripts to configure the USB controller in OTG mode, build and install the UVC gadget framework, and activate the gadget as a systemd service. When connected to a host computer, the device appears as a standard webcam requiring no additional drivers. The core functionality resides in `rpi-uvc-gadget.sh`, which uses configfs to define a UVC device with multiple resolutions (640x480, 1280x720, 1920x1080) and formats (uncompressed and MJPEG) at various frame rates. The `setup.sh` automates dependency installation, compilation of the UVC gadget tools from freedesktop.org, and deployment of the service. A systemd unit ensures the gadget persists across reboots. This approach leverages kernel USB gadget infrastructure to create a seamless webcam experience.
pytest-flask is a pytest plugin designed to simplify testing of Flask applications. It provides fixtures for Flask test clients, live servers (for end-to-end testing), configuration access, and content negotiation headers. The plugin automatically manages request contexts and patches response classes to enable easy JSON response testing. Architecture-wise, it follows pytest's plugin system with entry points defined in setup.cfg, separating concerns into fixtures, live server management, and plugin registration. The live_server fixture spawns a subprocess running the Flask app and handles health checks via socket connections, while autouse fixtures ensure proper test isolation by pushing request contexts and applying configuration markers. External dependencies include pytest, Flask, and Werkzeug.
The Filecoin Community repository serves as a central hub for community documentation, project showcases, and governance. It contains markdown files detailing community processes, calls, polls, and working group activities. The projects directory showcases applications and tools built by the community on Filecoin. The repository provides standardized templates for project submissions, meeting agendas, and notes, facilitating consistent community engagement. While it lacks executable code, it effectively coordinates the Filecoin ecosystem through structured documentation and resource sharing.
This repository provides a command-line tool for splitting video files into smaller segments using FFmpeg. The tool supports three splitting modes: equal time-based chunks, equal file-size-based chunks, and arbitrary segments defined by a manifest file (in JSON or CSV format). The script is written in Python and relies only on the standard library and the presence of FFmpeg in the system's PATH. The architecture is a single monolithic script that parses command-line options, delegates to appropriate splitting functions, and invokes FFmpeg via subprocess calls. Key functions include get_video_length (for retrieving video duration via ffprobe), split_by_manifest (for processing manifest-defined segments), and split_by_seconds (for time- or size-based splitting). The script includes a test suite that validates the core functionality using mocks to avoid actual FFmpeg calls. While the tool is functional and well-tested, it shows signs of technical drift due to the use of the deprecated optparse module (instead of argparse) for command-line parsing. Despite this, the code remains maintainable and demonstrates solid engineering practices in error handling, modular function design, and testability.
modV is a desktop application for real-time audio-visual synthesis and effects processing. Built with Electron, Vue.js, and WebGL, it provides a modular environment where users can create chains of visual modules (2D, 3D, shader-based) that react to audio, MIDI, or user-defined parameters. The application features a dual-store Vuex architecture (main and worker processes) to keep the UI responsive while handling intensive graphics and audio processing in a background thread. Key capabilities include audio feature extraction via Meyda, MIDI input mapping, a plugin system for extensibility, and a preset system with seamless state switching. The project emphasizes real-time performance and flexibility for VJs, artists, and live visual performers.
This repository is a fork and local enhancement of Anthropic's Claude Code CLI, branded as 'claude-code-tudou' (土豆版). It is a production-grade terminal AI coding assistant that provides an interactive AI agent within the terminal, capable of reading/writing files, executing shell commands, searching code, managing git operations, and orchestrating multi-agent teams. The system features a dual UI: a rich terminal interface built with React and Ink (custom terminal rendering framework) and an optional Electron desktop application with Vue 3 frontend. It supports third-party API providers (OpenRouter, MiniMax, Gemini, Bedrock, Vertex) via environment variables, making it a versatile alternative to the official Claude Code. The architecture is modular with clear separation: tool definitions (Bash, FileEdit, Glob, Grep, LSP, MCP, etc.), a plugin marketplace system, MCP (Model Context Protocol) integration, permission/security engine, session memory, and a swarm multi-agent subsystem. The codebase includes a custom Ink-based terminal UI framework with yoga layout, ANSI parsing, and event handling. It also provides a comprehensive settings system with multiple sources (user, project, managed), keybinding configuration, and analytics/telemetry infrastructure. While the code is well-structured and functional, it carries significant upstream Anthropic markers ('ANT-ONLY' imports, conditional internal features) and ad-hoc modifications for local execution stability. The fork adds value through configurable API endpoints, desktop UI, and stability fixes, but inherits complexity from the upstream codebase. The project is actively maintained with recent commits and demonstrates professional engineering standards.
XXL-MQ is a lightweight, high-throughput distributed message queue framework implemented in Java with Spring Boot. It provides a broker (admin console) that handles message routing, persistence, and management, and a client library (xxl-mq-core) for producers and consumers. The system supports multiple message types including parallel, serial, sharded, broadcast, delayed, and failed retry messages. It uses a compute-storage separation architecture: the broker stores messages in MySQL or TiDB, partitions them using configurable strategies (hash, cycle, random, first, last, broadcast), and exposes an OpenAPI (HTTP) for client operations. Background threads handle message production/consumption, archiving, alarm, and fail-retry processing. The admin console provides a web UI for managing topics, messages, applications, access tokens, and users, with authentication via XXL-SSO. External dependencies include MySQL/TiDB, MyBatis, Spring Boot, XXL-Tool, and various frontend libraries (AdminLTE, CKEditor, CodeMirror). The framework is designed for high throughput (10W/s on MySQL, 50W/s on TiDB per node) and supports massive message accumulation with configurable archive strategies.
This repository is the Motoko Base Library, a standard library for the Motoko programming language targeting the Internet Computer platform. It provides fundamental types, data structures, and utilities for building smart contracts (canisters). The library includes purely functional collections (List, Trie, Deque), mutable containers (Buffer, HashMap, TrieMap, Stack), red-black tree based ordered sets and maps (OrderedSet, OrderedMap), fixed-width integer types (Int8-64, Nat8-64), Float, Text, Principal, and low-level IC APIs (ExperimentalCycles, StableMemory, CertifiedData, Timer, Random). The code is well-documented with examples and complexity annotations, and includes a comprehensive test suite using the mo:matchers framework. However, the README explicitly states this package is superseded by the official dfinity/motoko-core package, indicating migration is recommended. Some duplication exists between OrderedSet and OrderedMap (shared RB-tree logic), and several deprecated functions are present. The repository is in maintenance mode with periodic updates for compatibility, but active development has shifted to the core package.
amazoncaptcha is a pure Python library for solving Amazon's image-based text captchas. It does not rely on external OCR or machine learning models; instead, it uses a custom algorithm that converts the captcha image to monochrome, segments individual letters, and matches each extracted letter against a precomputed library of compressed patterns (stored as zlib-compressed byte strings in JSON files). The library provides a simple API: the `AmazonCaptcha` class can be instantiated with an image file path, URL, or a Selenium WebDriver screenshot, and its `solve()` method returns the predicted captcha text. The architecture is monolithic: a single package with image processing utilities (utils.py), the solving logic (solver.py), optional data collection tools (devtools.py), and embedded training data for letters A–Z. Externally, it depends on Pillow for image manipulation and requests for downloading captchas from URLs. The solver is designed for high accuracy (claimed 99.9% on various tests) and lightness, making it suitable for web automation and scraping tasks that encounter Amazon captchas. While the core algorithm is elegant and self-contained, its effectiveness is tightly coupled to the specific font, distortion, and noise patterns used by Amazon in its captchas. Any change to Amazon's captcha design would likely require retraining or updating the embedded pattern database. The repository also includes a test suite, CI configuration, and documentation for integration with Selenium.
This Neovim plugin provides a floating window for entering ex-commands, featuring tab completion, history navigation (with optional smart history), and customizable keybindings. It is built on top of the nui.nvim library and offers a clean API for configuration and extension via hooks. The plugin handles compatibility across different Neovim versions (pre and post 0.7) for keymap settings. The implementation is focused and modular, separating concerns between UI setup, state management, and user interaction.
LLEf (LLDB Enhanced) is a Python-based LLDB plugin for low-level reverse engineering, vulnerability research, and exploit development. It provides enhanced debugger context (registers, stack, code, threads, trace) via an automatic stop hook and adds specialized commands: context, hexdump, pattern (cyclic pattern generation/search), checksec (ELF security mitigations), xinfo (memory mappings), dereference (pointer chain dereferencing), and scan (heap/memory scanning). It supports multiple architectures (x86_64, ARM, AArch64, i386, PowerPC) and includes experimental Go binary analysis: parsing the PCLNTAB and ModuleData structures to recover function names, types, and stack frames from stripped Go executables (versions 1.7–1.24). The plugin persists configuration via JSON files in the user's home directory. The codebase is structured with clear base classes for commands, settings, and architecture definitions, though it lacks automated tests and relies heavily on the LLDB Python API.
This repository contains a collection of AWS Lambda extension examples and demonstrations across multiple programming languages (Go, Python, Node.js, Java, C#, Rust) and wrapper scripts (Bash, Python, Ruby). Each example illustrates how to build a Lambda extension that integrates deeply with the Lambda execution environment using the Extensions API for lifecycle event registration, and the Logs API/Telemetry API for observability data processing. The extensions perform tasks such as forwarding logs to external destinations (S3, Kinesis, Firehose, Elasticsearch, etc.), collecting and sending telemetry data, implementing custom runtime behavior via wrapper scripts, and caching data from AWS services like DynamoDB and Parameter Store. The repository serves as a practical guide for developers to understand and implement Lambda extensions for observability, security, and governance use cases.
CtYun is a .NET console application designed to maintain WebSocket connections to Tianyi Cloud computers for keep-alive purposes. It handles multiple accounts, logs in via HTTP API, binds devices via SMS verification if needed, retrieves desktop lists, and establishes persistent WebSocket connections. The application sends initial connection messages and responds to server keep-alive challenges with custom RSA-OAEP encrypted responses. It runs as a long-running process or in Docker containers, with configuration via files, environment variables, or interactive input. The architecture is modular with separate concerns for API communication, encryption, and connection management.
This repository implements a Model Context Protocol (MCP) server that exposes the QuickBooks Online API as tools for AI agents such as Claude Code. It provides 143 tools covering full CRUD and search operations across 29 entity types (Customer, Invoice, Bill, Vendor, Employee, Account, Item, JournalEntry, Deposit, Transfer, TimeActivity, etc.) and 11 financial reports (Balance Sheet, Profit & Loss, Cash Flow, Trial Balance, General Ledger, etc.). The architecture follows a clean separation between handler modules (wrapping the node-quickbooks library with OAuth authentication) and tool definitions (Zod schemas registered with the MCP server). A singleton QuickbooksClient manages OAuth 2.0 token refresh and persistence, while environment feature flags (QUICKBOOKS_DISABLE_WRITE/UPDATE/DELETE) suppress tool categories. The codebase is heavily tested with Jest (100% coverage enforced in CI) and uses modern TypeScript (ESM). It handles QBO-specific quirks such as automated sales tax (AST), sparse vs full updates, and fallback patterns (e.g., delete via void/inactive). Despite its broad coverage, the implementation is highly repetitive (87 handlers, 90 tools) with liberal use of `any` types and hand-rolled ambient declarations for missing upstream typings. The project is actively maintained with recent commits and detailed design documentation, though version numbers and test counts show minor drift between README and CHANGELOG. Overall, this is a comprehensive, production-grade integration that prioritizes completeness and correctness over elegance.
The repository provides a comprehensive framework for generative AI models, particularly focused on image and video generation. It integrates various diffusion-based architectures (Stable Diffusion, Stable Diffusion XL, Hunyuan DiT, Wan, etc.) and offers utilities for model loading, prompt engineering, and post-processing. The system supports text-to-image, text-to-video, image-to-video, and video-to-video generation, along with editing capabilities via ControlNet and LoRA. The architecture is modular, with core components for model management, prompting, and VRAM optimization, though it exhibits some duplication across model-specific pipelines.
Hyper-RAG is a hypergraph-driven Retrieval-Augmented Generation (RAG) system designed to reduce LLM hallucinations by modeling beyond-pairwise correlations in domain knowledge. The repository provides both a Python library (`hyperrag`) and a full-stack web application with a FastAPI backend and React frontend. The core innovation is the use of hypergraphs to represent complex relationships among entities, outperforming traditional graph-based RAG methods like Graph RAG and Light RAG. The system supports multiple query modes: hyper (full hypergraph), hyper-lite, graph, naive, and direct LLM, with streaming responses. Architecture: The Python library (`hyperrag/`) implements the RAG pipeline: text chunking, entity extraction via LLM, hypergraph construction, vector embedding storage (NanoVectorDB), and query execution. The web UI (`web-ui/`) consists of a FastAPI backend (`main.py`) exposing REST endpoints for file upload, database CRUD, and HyperRAG queries, plus a React frontend with Ant Design and Graphin for interactive chat, hypergraph visualization, and settings management. Docker Compose orchestrates the frontend, backend, and Nginx reverse proxy. Key entry points: `hyperrag/hyperrag.py` (HyperRAG class), `service_api.py` (standalone API), `web-ui/backend/main.py` (FastAPI app). External dependencies include OpenAI/Azure/Bedrock LLM APIs, NanoVectorDB, tenacity, aiofiles, and various Python packages. The system is configurable via `config_temp.py` and `settings.json`. Data flows: Raw text (uploaded or provided) is chunked, entities and relationships are extracted via LLM, hyperedges are constructed, and embeddings are stored. Queries are processed by selecting relevant hyperedges and entities, constructing a context, and generating a response. The system caches LLM responses and supports real-time logging via WebSocket during embedding.
This repository is a curated collection of reusable draw.io shape libraries, editable diagram examples, and a Docusaurus-based guideline website for creating consistent SAP Business Technology Platform (BTP) solution architecture diagrams. It provides standardized icons for SAP BTP services (organized by category and size), connectors, area shapes, text elements, brand logos, and pre-built diagram templates (e.g., SAP Task Center, SAP Cloud Identity Services, SAP Private Link Service). The guideline website documents design principles, atomic components, and best practices for diagram creation. The repository is not a runnable application but a design asset library with a static site for documentation. It serves SAP solution architects and technical writers who need to produce professional, on-brand architecture diagrams.