中文版

🎓 Student
🔭 Currently learning AI Agent development

A local Agent runtime system for desktop environments.
A desktop-oriented Agent runtime system that enables local Agents to safely access system capabilities within a controlled sandbox through strict layered contracts and defense-in-depth security.

• 📐 Agent Capabilities
  • Context engineering: Prompt orchestration, status bar support for environment information injection, and context compression.
  • User memory and knowledge base: Memory hierarchy design, including working memory and long-term memory; memory formats such as Simple Notes and Advanced JSON Cards; and long-term memory categories such as episodic memory and semantic memory.
  • Tools: Web search, code interpreter, persistent command-line tools, directory listing, grep/glob, and an old string -> new string file-editing tool with linter and syntax checks after editing; parallel tool support and more.
  • Coding capabilities: Using code tools for precise calculations and strict logical reasoning; guiding Agents to provide correct tool parameters with code; and generating UIs using an A2UI-style protocol.
  • Testing: Running scripts for automated testing.
• 🛡️ Five-layer security policy chain
  • Tool calls pass through the following layers: Scope (task authorization) ➔ Retriever (capability retrieval) ➔ Binder (parameter validation and path resolution) ➔ Executor (policy evaluation) ➔ Path-Guard (normalization and protection).
• 📊 Real-world evaluation
  • The project includes an end-to-end evaluation framework covering real-world daily office tasks, complex multi-step reasoning with GAIA, and human-Agent collaboration with τ-bench, across 38 core evaluation sets running on real hardware.
• 🏗️ Engineering
  • Built as a Monorepo with a complete pnpm verify pipeline integrating TypeScript type checking, ESLint, Python Ruff, Vitest, and Pytest as cross-language quality gates.

• SakiFlow: A frontend design project based on Vue 3.
• SakiVault: A Vue 3 frontend project using the Bangumi API.