Claude Code Plugin Implementation

Relevant source files

The following files were used as context for generating this wiki page:

• README.md
• hooks/hooks.json
• skills/claude-code-docs/SKILL.md
• hooks/rewrite-docs-url.sh
• sync.ts
• hooks/rewrite-docs-url.test.sh

The claude-code-docs repository serves as an unofficial mirror of the official Claude Code documentation, designed to enhance the capabilities of AI coding agents. It provides a structured environment where documentation is synced every 6 hours from code.claude.com/docs, allowing agents to access local, version-pinned markdown files rather than processing raw HTML.

The implementation functions as both a data mirror and a Claude Code plugin. As a plugin, it utilizes PreToolUse hooks to intercept WebFetch calls, ensuring that when an agent attempts to access a documentation URL, it receives clean markdown content. Additionally, it provides a specialized "Skill" for semantic searching and question answering based on the mirrored documentation.

Sources: README.md:1-12, skills/claude-code-docs/SKILL.md:1-5

Core Architecture and Data Synchronization

The system architecture revolves around maintaining a local file-based mirror of the documentation. The synchronization process treats the remote llms.txt file as the source of truth for the manifest of available pages.

Synchronization Logic

The sync.ts script performs the following steps:

1. Fetches the remote llms.txt manifest.
2. Identifies all .md URLs listed within the manifest.
3. Downloads each page (with a concurrency of 10 and 3 retries).
4. Rewrites internal site-absolute links to relative markdown links so that the documentation remains navigable when viewed locally or on platforms like GitHub.

Sources: sync.ts:1-10, sync.ts:60-90, README.md:14-19

The diagram above illustrates the data flow from the canonical documentation source to the local mirror directory.
Sources: sync.ts:12-40, sync.ts:60-95

Claude Code Plugin Hooks

The plugin implementation utilizes a PreToolUse hook specifically for the WebFetch tool. This mechanism intercepts requests to code.claude.com/docs and redirects them to the .md source variant.

Hook Configuration

The hook is registered via hooks/hooks.json. It targets the WebFetch tool and executes a bash script to handle the URL transformation logic.

Configuration Field	Value	Description
matcher	WebFetch	Targets the tool used for fetching web content.
type	command	Specifies the hook executes an external command.
command	bash ${CLAUDE_PLUGIN_ROOT}/hooks/rewrite-docs-url.sh	The script performing the URL rewrite.
timeout	5	Execution timeout in seconds.

Sources: hooks/hooks.json:4-16

URL Rewriting Logic

The rewrite-docs-url.sh script filters URLs to ensure only specific documentation pages are rewritten. It strips URL fragments and query parameters, then appends .md to the path if it is a valid page under a locale segment (e.g., /docs/en/).

Plugin Skill Implementation

The plugin exposes a skill named claude-code-docs which provides a structured method for agents to answer questions about Claude Code.

Skill Workflow

The skill follows a multi-step resolution process:

1. Resolve Mirror: Locates the cached checkout of the docs mirror using the ask CLI or npx -y @pleaseai/ask.
2. Search: Performs semantic search using csp (Code Search Please) or falls back to grep over the local markdown files.
3. Read and Answer: Reads the top matches and generates an answer, citing canonical URLs (e.g., mapping docs/hooks.md back to https://code.claude.com/docs/en/hooks).
4. Fallback: If local tools are unavailable, it fetches the online llms.txt and uses WebFetch on individual page URLs.

Sources: skills/claude-code-docs/SKILL.md:11-53, README.md:59-69

Semantic Search Integration

The skill leverages the csp tool for token-efficient searching. This allows the agent to find relevant information without loading the entire 5MB documentation bundle into the context window.

Conclusion

The Claude Code Plugin Implementation provides a robust bridge between the official documentation and AI coding agents. By combining a frequently synced local mirror with a URL-rewriting hook and a semantic search skill, it ensures that agents can access high-quality, clean markdown documentation efficiently. This architecture reduces context window bloat while maintaining high accuracy through grounded search and canonical citations.

Sources: README.md:70-74, skills/claude-code-docs/SKILL.md:41-48