v2.0.2

[2026-04-19] — v4.2: Visual Compilation Engines

Overview

Expanded the wiki ecosystem with native visual compilation capabilities. Introduced three specialized diagramming engines (Canvas, Excalidraw, Mermaid) to generate highly precise architecture diagrams and mind maps, treating visual thinking as a first-class citizen alongside text.

New Features

• visual-canvas skill — Built-in Obsidian Canvas visualization engine with native layout algorithms for spatially organizing concepts and mind maps.
• visual-excalidraw skill — Specialized Excalidraw architecture diagram engine featuring strict JSON crash-prevention rules for reliable drawing.
• visual-mermaid skill — Dedicated Mermaid graph generation workflow with guidelines to avoid common Markdown rendering conflicts.
• Visual Mind Library — Added raw/visual/ directory to store and manage the raw source files for these three visual engines.

v2.0.1

[2026-04-19] — v4.1: Web Re-verification, Refresh Subcommand & Skill Optimization

Overview

Major upgrade to the llm-wiki skill (v3.1 → v4.1). Introduces the refresh subcommand for web-based knowledge re-verification, enhances query with proactive freshness checks, strengthens lint with semantic evaluation, and removes output deliverable lifecycle tracking to reduce complexity.

New Features

• refresh subcommand — Re-verify a specific topic via web search, compare with existing wiki content, persist updates to raw/ after user confirmation, then trigger ingest
• query freshness check — Pages with confidence < 0.6 or stale status now trigger proactive web search for latest information
• lint semantic enhancements — One-way link semantic relevance assessment; staleness probing suggests refresh for high-frequency topics in stale pages
• publish verification checklist — Added 3-item verification checklist (no wikilinks, complete frontmatter, proper citations)
• Web re-verification reinforcement — New +0.20 confidence boost when web re-verification finds no conflict; any refresh/web verification resets time decay timer

Removed

• Output deliverable lifecycle tracking — Removed ## Output Deliverable Tracking section from lifecycle-spec.md, removed calc_output_staleness() from lifecycle-check.py, removed output-related references from SKILL.md. Layered consolidation (including Procedural layer) is preserved.

Modified

• skills/llm-wiki/SKILL.md (v3.1 → v4.1) — Added refresh subcommand workflow; enhanced query step 3-4 with freshness check and incremental save; enhanced lint step 3 with one-way link evaluation and staleness probing; added publish verification checklist; added gotcha about refresh requiring user confirmation; updated description with natural language trigger phrases
• skills/llm-wiki/references/lifecycle-spec.md — Added +0.20 reinforcement event for web re-verification; added clarification paragraph about refresh resetting time decay; removed Output tracking from data distribution table; removed ## Output Deliverable Tracking section
• skills/llm-wiki/references/log-spec.md — Enriched lint log format (check results + fix actions sections); added source field to publish log; added refresh log entry type
• skills/llm-wiki/references/output-page-spec.md — Aligned frontmatter with wiki-page-spec pattern (aliases, tags, sources, description instead of type, audience, status, wiki_sources); added body structure rules (h2/h3 hierarchy, code block language annotation)
• skills/llm-wiki/scripts/lifecycle-check.py — Removed calc_output_staleness() function and Output section parsing; simplified parse_lifecycle() return value
• AGENTS.md — Fixed lifecycle reference path (lifecycle.md → lifecycle-spec.md); updated subcommand count to five; added refresh subcommand; added core principle #9 (Web Re-verification)
• CLAUDE.md — Updated subcommand count to five; added refresh subcommand; added core principle #9
• GEMINI.md — Added core principle #9 (Web Re-verification)
• README.md — Updated subcommand count to five; added Refresh capability and workflow rows; updated Query descriptions with freshness check; added "Web Re-verification & Fact Preservation" section; added "7. Re-verify Topics" Quick Start section

v2.0.0

[2026-04-15] — i18n: Full English Translation + Architecture Upgrade to Skill-based System

Overview

Translated the entire project from Chinese to English (except docs/), and restructured the schema from monolithic inline specs to a modular skill-based architecture with progressive disclosure.

Architecture Changes

The three schema files (AGENTS.md, CLAUDE.md, GEMINI.md) were significantly refactored:

• Before: Each file was a self-contained ~220-line monolith embedding all rules inline — page format specs, tag system, naming conventions, index/log format, full workflow steps, and directory structure descriptions.
• After: Each file is a ~50-70 line overview referencing the shared skills/llm-wiki/ skill. Detailed rules are loaded on demand via 7 reference spec documents.

Key structural additions:

• Added Skills layer to the architecture table (between Schema and Raw Sources)
• Added knowledge lifecycle management section (confidence scoring, state transitions, forgetting curve, supersession, demotion, layered consolidation)
• Added "Progressive Disclosure" as the 8th core principle

Modified (tracked files)

• AGENTS.md (295 → 53 lines) — Translated to English; refactored from monolithic spec to concise overview referencing skills/llm-wiki/SKILL.md
• CLAUDE.md (221 → 70 lines) — Translated to English; refactored from monolithic spec to concise overview with Claude Code tool adaptation notes (Read/Write/Edit/MultiEdit/Glob/Grep, TodoWrite for task management)
• GEMINI.md (224 → 76 lines) — Translated to English; refactored from monolithic spec to concise overview with Gemini CLI tool adaptation notes (read_file/write_file/edit_file/list_directory/shell, context window management)
• README.md — Translated to English; language switcher flipped from 简体中文 | English to English | 简体中文; added sections on lifecycle management, layered consolidation (four-tier memory model), progressive disclosure, core capabilities table; architecture tree now includes skills/, log-archive/, lifecycle.md

Deleted

• README.en.md — Removed; README.md is now the English version (was previously the Chinese version)

Added (new untracked files)

• README.zh.md — Placeholder for future Chinese version (referenced by README.md language switcher, not yet created with content)
• CHANGELOG.md — This file
• docs/Agent-Skills-Specification.md — Agent Skills specification document
• skills/llm-wiki/SKILL.md — Main skill definition with four subcommands (ingest/query/lint/publish), including detailed workflow steps, verification checklists, gotchas, and on-demand reference loading table
• skills/llm-wiki/references/wiki-page-spec.md — Wiki page format spec (directory structure, frontmatter schema, body structure requirements)
• skills/llm-wiki/references/output-page-spec.md — Output deliverable page spec (directory structure, frontmatter, type-specific format requirements)
• skills/llm-wiki/references/tags-spec.md — Tag system (3 dimensions: domain/type/maturity); tags converted from Chinese to lowercase hyphenated English (e.g. 软件工程 → software-engineering, 对比分析 → comparative-analysis)
• skills/llm-wiki/references/naming-spec.md — Page naming conventions by type
• skills/llm-wiki/references/index-spec.md — index.md table format spec (with confidence and status columns)
• skills/llm-wiki/references/log-spec.md — log.md append-only format spec with 30-day retention and archive rules
• skills/llm-wiki/references/lifecycle-spec.md — Full lifecycle management spec (confidence initial values/reinforcement/event decay/time decay with type coefficients, state transition diagram, supersession mechanism, demotion mechanism, layered consolidation promotion rules, output deliverable tracking)
• wiki/lifecycle.md — Lifecycle registry (pluggable module tracking access counts, last_accessed, supersession chains, and output deliverables)

Not Changed

• docs/ — Excluded from translation per request
• wiki/index.md, wiki/log.md — Already minimal English content, unchanged
• LICENSE, .gitignore — Not applicable

v1.0.0

LLM Wiki — 个人知识库

简体中文 | English

一个基于 Obsidian + LLM Wiki 持续维护的个人知识库。灵感来自 Karpathy 的 LLM Wiki 模式。

核心理念：知识不是每次查询时从零推导，而是被编译一次、持续积累。人类负责策展素材和提出问题，LLM 负责所有繁琐的整理工作——摘要、交叉引用、归档、一致性维护。

与传统 RAG 的区别

	传统 RAG	LLM Wiki
知识存储	原始文档 + 向量索引	结构化、互相链接的 Markdown 页面
查询方式	每次从零检索、拼凑	基于已编译的知识综合回答
知识积累	无——每次重新推导	有——每次操作都让 wiki 更丰富
交叉引用	无	自动维护的 [[双链]] 网络
矛盾处理	不感知	主动标注、注明来源

工作流

操作	触发方式	说明
Ingest	新素材放入 raw/，告诉 LLM 处理	LLM 读取素材 → 提取要点 → 创建/更新 Wiki 页面 → 维护交叉引用
Query	向 LLM 提问	LLM 基于 Wiki 已编译知识回答，有价值的回答可存回 Wiki
Lint	定期要求 LLM 检查	检查矛盾、孤立页面、缺失引用、过时内容，保持 Wiki 健康
Publish	要求 LLM 从wiki提炼成品输出	LLM 生成，人类审核定稿

核心原则

• Schema 层（AGENTS.md / CLAUDE.md / GEMINI.md）：告诉 LLM 如何维护 wiki
• Raw 层（raw/）：不可变的原始素材，是事实来源
• Wiki 层（wiki/）：LLM 的编译产物，所有知识的结构化呈现
• Output 层（output/）：从 wiki 提炼的成品，面向外部受众，独立可读
• 知识复利：每次操作都让 Wiki 更丰富
• 详细规范见 AGENTS.md