[FEATURE] Shell escape that doesn't feed output to LLM context

Problem

! command runs a shell command and sends the output into the conversation as a turn, consuming context tokens. Sometimes you want to run a quick command — check a file, test something, reference output — without it eating into your context budget.

This is especially painful in long-running sessions where context is scarce, or when the command produces large output (e.g. cat on a big file, grep across a codebase) that you only need to glance at, not have the LLM reason about.

Proposal

A "quiet" shell escape — something like !! command or /shell command — that:

1. Executes the command in the user's shell
2. Displays the output in the TUI
3. Does not add the output (or the command) to the LLM context

The command and its output would be visible to the user but invisible to the model.

Use cases

• Referencing large files without burning tokens
• Running git log, git diff, ls, etc. for orientation
• Testing a command before deciding whether to ask the LLM to run it
• Scrollback reference during long sessions