llm-wiki

Imported from zhoushoujianwork/llm-wiki (AGENTS.md). Install upstream with npx skills add zhoushoujianwork/llm-wiki. Copyright stays with the author.

AGENTS.md — For AI Coding Assistants

This project follows the LLM Wiki architecture described in /Volumes/1TB/github/llm-wiki/docs/ARCHITECTURE.md.

Key Principles

1. Source Layer is read-only: Never modify source files; only read them.
2. Compilation is idempotent: Re-running compile should be safe and incremental.
3. Wiki pages are the source of truth for queries: Query layer reads wiki, not raw sources.
4. No vector database: Use LLM for retrieval via structured wiki pages.
5. Namespace isolation: Each source repo gets its own namespace in the wiki.

Directory Structure

cmd/llm-wiki/      # CLI entry point
internal/
  source/          # GitHub repo management, file discovery
  compiler/        # LLM compilation engine
  wiki/            # Wiki page storage and retrieval
  query/           # Query processing and answering
  index/           # Concept/entity to page mapping
sources/           # Cloned repos and cached sources (gitignored)
wiki/              # Generated wiki pages (committable)

Compilation Flow

1. Discover documents in source layer
2. For each document, check if it needs compilation (mtime, existing wiki page)
3. Send document to LLM with compilation prompt
4. LLM returns: summary page, entity pages, cross-reference updates
5. Write wiki pages to wiki/ directory
6. Update index

Query Flow

1. User asks question
2. Identify relevant namespaces/pages from index
3. Read top 3-5 relevant wiki pages
4. Send pages + question to LLM
5. Return answer with source citations

Adding New Sources

llm-wiki source add <repo-url>     # GitHub repo
llm-wiki source add <local-path>   # Local directory
llm-wiki source add <url>          # URL to scrape

ClawFlow — Issue → PR Automation

This repo participates in ClawFlow, the OpenClaw automated issue → PR pipeline.

How it works

1. Owner creates a GitHub issue and applies the ready-for-agent label
2. OpenClaw cron agent picks it up, evaluates confidence
3. If confidence ≥ 7/10, spawns a sub-agent to fix the issue
4. Sub-agent creates a branch, implements the fix, opens a PR
5. PR link is sent back to the owner

Labels

Constraint

• Do NOT apply ready-for-agent yourself — only the repo owner sets this tag
• You may evaluate issues and report confidence, but the actual tagging is owner-only

Skills

• Skill definition: ~/.openclaw/workspace/skills/llm-wiki/SKILL.md
• Full ClawFlow docs: ~/.openclaw/workspace/ClawFlow.md

Compilation Modes

• full: Re-compile everything, overwrite existing pages
• incremental: Only compile new/changed sources (default)
• single <source>: Compile specific source only