Built-in claude-code-guide subagent runs on Haiku, prone to fabrication on Claude Code factual lookups

Description

The built-in claude-code-guide subagent is defined in the CLI with model:"haiku" (found via string search in the installed @anthropic-ai/claude-code package, version 2.1.266, file bin/claude.exe). Its stated purpose is to answer factual questions about Claude Code, the Claude Agent SDK, and the Claude API — exactly the domain where wrong details (flag names, subcommands, model IDs) mislead users who trust the answer at face value.

Reproduction

Asked (via the Agent tool, subagent_type: claude-code-guide) how to set a persistent default model and effort level from the terminal, the agent answered:

claude config set model claude-sonnet-4-20250514
claude config set effortLevel medium

Both are wrong:

• claude config is not a subcommand — claude --help lists no config command at all.
• claude-sonnet-4-20250514 is a stale model ID, not a currently offered model.

The actual mechanism, confirmed by inspecting the installed binary, is: --model/--effort CLI flags apply to the current session only, while the interactive /model and /effort slash commands persist the choice to the user's settings.json as the default for new sessions.

Impact

Haiku-class models are more prone to fabricating specific technical details (exact flag/subcommand names, IDs) than larger models. Using one as the default reasoning engine for a "how does Claude Code work" lookup agent seems to work against the agent's own purpose.

Suggestion

Use a stronger default model for claude-code-guide (or make it configurable), and/or have it verify claims like available subcommands against --help output rather than answering from recall.