Redirect messages stall behind long-running tool calls — /btw-style interrupt evaluation would fix this

Problem

When a tool call takes a long time (slow bash process, network call, etc.), any redirect message the user types sits in the queue until that operation completes — which can be minutes. There is no current way to express "this message is urgent enough to interrupt what's running" short of using a hard keyboard interrupt.

Suggested improvement

The /btw command already demonstrates that Claude Code can evaluate and inject messages during an ongoing operation without disrupting it. The redirect queue should work similarly: run a cheap fast model alongside the current operation to evaluate whether a queued message warrants interrupting it (stopping the current tool call early) vs. waiting for the natural break. If the message is just additional context, wait. If it meaningfully redirects the work, interrupt.

This would align with how users naturally interact — typing in plain language — without requiring them to learn interrupt semantics.

Documentation gap

There is no clear public documentation explaining how the redirect/interrupt mechanism works, what constitutes a "natural break," or how queued messages are evaluated. Even Claude Code itself could not locate this documentation when asked.

---
Written with Claude Code