Zen Mode

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

When debugging a tricky issue in Claude Code, the model often underperforms compared to a clean conversation in claude.ai, even though it's the same underlying model. The reason isn't intelligence, it's context pollution.
By the time you're deep in a Claude Code session, the context window is packed with file trees, terminal output, previous tool calls, and compacted conversation history. When you need Claude to just think about a code snippet and tell you what's wrong, all that accumulated noise degrades the response. The model starts chasing its own tail in agentic loops instead of reasoning clearly.
The current workaround is to leave Claude Code, open claude.ai, paste the relevant code, get a focused answer, then come back to Claude Code to implement. This works but it's a broken workflow.

Proposed Solution

Proposed Solution:
A /zen command that opens a fresh, scoped reasoning window within Claude Code:

Starts with a clean context window (no inherited file tree, tool outputs, or conversation history)
Disables all tool use (no file reads, no bash, no searching the codebase)
Lets you paste in a code snippet or describe a problem and get a focused answer
Exits back to your normal session when you're done

Think of it as a rubber duck that's actually smart, living inside the same tool where you're already working.

Alternative Solutions

Using claude.ai in a browser: Works but breaks flow. You lose time context-switching between tools.
/clear then ask, then /resume: Destructive. You lose your working session.
Subagents: Overkill for "look at this function and tell me what's wrong." Also still inherits context setup overhead.

Priority

High - Significant impact on productivity

Feature Category

MCP server integration

Use Case Example

I'm deep in a Claude Code session building a feature in an Electron app
The IPC chain fires correctly but an image never appears on the canvas
I suspect the main process fetch handler is returning image data as a string instead of ArrayBuffer
Instead of Claude Code reading 40 files and running tests, I just want to paste the one function and ask "what's wrong here"
I type /zen, paste the function, get a clean focused answer in seconds, type /exit, and I'm back in my working session with full context intact

Additional Context

This would also be useful for quick architecture questions mid-session, asking "how does X work" without Claude trying to grep your codebase, getting a second opinion on an approach before committing to it, or any situation where you want reasoning not action.
The core insight: sometimes the best debugging tool is a model with less context, not more. A clean window with just the relevant snippet consistently produces better reasoning than a bloated session with thousands of lines of irrelevant code.