Feature Request: Free/lightweight mumbling mode (visible reasoning without extra token cost)

Problem

Cursor displays the model's reasoning process in real-time, allowing users to follow along and interrupt if the AI is going in the wrong direction.

In Claude Code, Ctrl+O shows extended thinking, but this consumes thinking tokens — making it impractical to keep on by default.

Desired behavior

A lightweight "mumbling" mode that streams Claude's intermediate reasoning as plain text, without triggering extended thinking tokens:

• Tool selection rationale ("I'm going to search for X because...")
• Plan outline before execution
• Approach decisions and tradeoffs

Similar to the existing action narration (Reading file...) but extended to the reasoning behind decisions.

Use case

• Catch wrong assumptions early, before execution begins
• Reduce wasted turns due to misunderstood intent
• Allow users to interrupt and redirect cheaply

Prior art

• Cursor's live "thinking" display
• Claude Code's existing action narration (already free, just needs extending)

Notes

Distinct from extended thinking (dedicated reasoning tokens). The idea is a best-effort, zero-cost narration of the decision process.