Distribute version-aware GAIA docs to coding agents #201

Evaluate three options with the same small task suite:

1. Mintlify's hosted documentation MCP and Markdown exports. Verify the actual GAIA deployment, domain/path prefix, available tools, retrieval quality, and access requirements before documenting an endpoint. Mintlify documents a hosted search/page MCP; this is distinct from its authenticated content-editing MCP.
2. A small coding-agent plugin or skill. Supply connection/setup instructions and a concise index that points to the canonical docs. Avoid embedding a stale full documentation copy in a prompt or plugin.
3. A release-pinned offline bundle. Generate Markdown, an index, and tested examples from the same source in CI. Evaluate whether this is sufficient before introducing a custom local MCP server, indexer, or service.

Deliverables:

• A short comparison covering setup effort, client compatibility, offline support, version selection, maintenance cost, and retrieval accuracy; recommend one default and an offline path if needed.
• Working setup instructions validated in two coding-agent clients, including removal/upgrade steps.
• Each retrieved page/example identifies its source URL and release or commit. Detect an SDK/documentation version mismatch and report it explicitly.
• Search and retrieval cover: first custom agent; constructor/output options; tool validation and approval; mixin initialization; TUI launch and session lifetime.
• A fresh coding-agent session can retrieve and execute the custom-agent tutorial's deterministic tests without relying on obsolete imports, constructor keywords, or commands.
• Automated generation excludes retired plans and legacy agent tutorials from current guidance, preserves citations, and checks links/examples before publishing.
• Bound retrieval size and distinguish documentation text from executable instructions. Public documentation access should not require write access to the repo or runtime secrets.

Out of scope: changing the agent runtime, exposing local user files or memory, a general MCP marketplace, and adopting a new documentation source of truth. Do not build semantic search infrastructure until the hosted/static options have been measured.