Claude Code doesn't invoke claude-code-guide subagent for CLI questions, fabricates incorrect answers

Description

When asking Claude Code questions about its own CLI functionality (flags, commands, session management), it frequently answers from "memory" with fabricated/incorrect information instead of invoking the claude-code-guide subagent that is specifically designed for this purpose.

Reproduction steps

1. Ask Claude Code: "How do I fork a previous session (not the latest one)?"
2. Claude responds with claude sessions list — a command that doesn't exist
3. When told it doesn't work, Claude suggests manually reading JSON files from ~/.claude/sessions/ instead of using the correct command
4. The correct answer (claude -r --fork-session) is trivially found when the claude-code-guide agent is invoked

Expected behavior

When the user asks about Claude Code CLI features, the model should invoke the claude-code-guide subagent (which has access to official documentation via WebFetch/WebSearch) before answering — especially when unsure.

Actual behavior

The model confidently provides incorrect CLI commands and workarounds without consulting the available specialized agent. This has happened repeatedly across multiple sessions with different CLI questions.

Impact

• Users receive wrong instructions and waste time debugging non-existent commands
• Erodes trust in Claude Code's self-knowledge
• The tooling to answer correctly already exists (claude-code-guide subagent) but isn't being used consistently

Environment

• Model: Claude Opus 4.6 (1M context)
• Platform: Windows 11
• Claude Code CLI