Knowledge Share
Technical articles, tutorials, and insights
Code Agent Anatomy (05): How Does the Model Know Which Tools Are Available? How Is Function Calling Implemented?
Follow a tool_call from start to finish: how tools register into the Registry, how the Registry generates function schemas to inform the model, how the Orchestrator batches execution after the model triggers tool_calls, the ToolExecutor's permission/optimistic-lock/circuit-breaker pipeline, why the ToolResult protocol separates internal and external representations, and how observation results are truncated before being written back to history.
Code Agent Anatomy (06): How Do Skills Dynamically Extend an Agent's Capabilities?
A deep dive into the Skills dynamic extension mechanism in MyCodeAgent: how a single Markdown file gives an agent new capabilities without writing any Python code, covering the full lifecycle from definition, scanning, and prompt injection to invocation and execution.
Enterprise Knowledge Base (02): Classic Vector RAG Benchmark — QAnything vs LightRAG
Same 89 questions, two frameworks: QAnything v2 and LightRAG 1.5.6. This article covers the full journey from deployment to results — including Docker GPU mounting, Milvus crash recovery, user_id isolation pitfalls, and the actual numbers across single-hop, multi-hop, and boundary refusal.
Code Agent Anatomy (02): How Does an Agent Think and Act Round by Round?
A deep dive into MyCodeAgent's ReAct main loop: immutable state machine, dual-layer loop structure, three-verdict completion gate, and model error recovery. Understand why an agent isn't just 'call the model once and get a result' — it's a control system with feedback, guarantees, and well-defined termination conditions.
Code Agent Dissected (03): Different LLM Formats — How Does an Agent Unify the Interface?
A deep dive into MyCodeAgent's LLM interface layer: zero-dependency HTTP transport, provider routing table, and five-tuple response normalization. Understand why an agent avoids binding to a single SDK — and how splitting 'sending requests' from 'reading responses' into stable boundaries lets the ReAct loop see only a unified structure.
Code Agent Anatomy (04): How System Prompts Are Assembled, and Where the Agent's 'Personality' Comes From
A deep dive into MyCodeAgent's prompt assembly layer: the four-layer system message architecture — Constitution, Tool Contracts, Project Rules, and Runtime Signals — and how they combine with history projection to form the Model View sent to the LLM. Understanding that agent behavior constraints aren't 'one giant prompt', but a cacheable, fingerprintable, hot-swappable layered assembly.
Open Source Project #190: claude-video — Give Claude Eyes to Watch Videos, One Command to Analyze YouTube/Loom/Local Files, Subtitle-First Zero-Download, Scene-Aware Frame Dedup, Auto Token Budget
Brad Bonanno's Claude Code skill plugin that gives Claude the ability to analyze any video via the /watch command. Seven-step pipeline: yt-dlp download → ffmpeg frame extraction → Whisper transcription → Claude context injection. Subtitle-first optimization (use existing subtitles, skip download entirely), 16×16 thumbnail MAD algorithm for frame deduplication, automatic frame budget scaling by video duration. Four detail modes: transcript/efficient/balanced/token-burner. Works with Claude Code, Codex, Cursor, Copilot, Gemini CLI, and 50+ other agent hosts. 15.2k Stars, MIT.
Open Source Project #191: gstack — YC CEO Garry Tan's Open-Source AI Engineering Team, 23 Expert Role Commands Covering the Full Sprint from Idea to Production
Y Combinator CEO Garry Tan's Claude Code skill set that turns Claude Code into a staffed engineering team. 23 professional role commands covering planning, design, review, testing, and release: /office-hours (YC-style product challenge), /plan-ceo-review, /design-shotgun (4-6 design variants for comparison), /qa (real Playwright browser testing), /cso (OWASP + STRIDE security audit), /ship (sync→test→audit→push→PR). Skills pass context to each other, supports 10-15 concurrent sprints, works across Claude Code/Codex/Cursor and 10 other agents. 128k Stars, MIT.
Code Agent Anatomy (01): What Happens Inside the Agent When a User Types a Message?
A complete trace of a user input flowing through MyCodeAgent, from python main.py to the final response. Covers three phases — CLI entry point, dependency assembly, and the ReAct main loop — to build a mental map for diving deeper into individual modules.
Open Source Project #187: Pi — Philosophy-Driven Minimal AI Coding Agent, 86k Stars, 30+ LLM Providers, Unlimited Extensibility
Minimal philosophy-driven AI coding agent harness. 5 npm packages: unified LLM API (30+ providers), agent runtime, TUI library, coding agent CLI, telemetry contracts. Four modes (interactive/print/JSON/RPC/SDK), session branching tree, TypeScript extension system, Skills/Prompt Templates/Pi Packages ecosystem. Deliberately excludes MCP, sub-agents, permission popups, plan mode. TypeScript, MIT, 86k Stars.
Open Source Project #188: AI Agents in Depth — Li Bojie's Complete Open-Source AI Agent Book, 10 Chapters, 95 Experiments, 35k Stars
Li Bojie's fully open-source book 'AI Agents in Depth: Design Principles and Engineering Practice.' Core formula: Agent = LLM + Context + Tools. 10 chapters building progressively: context engineering, user memory and knowledge bases, MCP tool protocol, coding agents, evaluation frameworks, model post-training (SFT/RL), continuous evolution, multimodal interaction, multi-agent collaboration. 95 hands-on experiments, 13 language versions, free PDF/EPUB download. Python, Apache-2.0, 35k Stars.
Open Source Project #189: DeepTutor — Agent-Native Lifelong Learning Workspace, 3-Layer Memory + Multi-Engine RAG + Partners, 33k Stars
Agent-native AI learning workspace from HKUDS (HKU Data Intelligence Lab). Core: a single agent loop drives all modes (Chat/Quiz/Research/Visualize/Solve/Mastery Path). Three-layer memory (L1 event traces / L2 surface summaries / L3 cross-surface synthesis) with a visual Memory Graph tracing every claim to its source. Multi-engine knowledge bases: LlamaIndex/PageIndex/GraphRAG/LightRAG/Obsidian. Partners: persistent AI companions with IM channels (Slack/Discord/Telegram/WeChat/Feishu, etc.). My Agents: drive local Claude Code/Codex as subagents. Python + Next.js, Apache-2.0, 33k Stars.