Add configurable output limit for ! shell commands

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 running shell commands via ! in a conversation, the full output is consumed as tokens in the context window. If a command accidentally produces large output (e.g., verbose build logs, large data dumps), it wastes context and can push out useful conversation history.

Proposed Solution

A configurable setting (e.g., in settings.json or .claude/settings.json) to cap the number of lines or bytes included from ! command output:

{
"shell_output_max_lines": 200,
"shell_output_max_bytes": 10000
}

When output exceeds the limit, it would be truncated with a message like:
[Output truncated: 1,432 lines total, showing first 200. Full output saved to /tmp/claude-cmd-output-<hash>.log]

This way the full output is still accessible (via Read on the temp file) without bloating the conversation context.

Alternative Solutions

Current workarounds:

• Piping through head/tail (requires anticipating large output)
• Redirecting to a file and reading selectively

These work but require users to predict which commands will be noisy.

Priority

Medium - Would be very helpful

Feature Category

CLI commands and flags

Use Case Example

1. Running a test suite where a single failure produces a long stack trace or diff, flooding the context with output you didn't need
2. Accidentally running a command without filters (e.g., !mix deps.tree or !git log) that dumps hundreds of lines
3. Build commands where the user doesn't anticipate them to emit verbose compilation or bundling output where only the final status or errors matter

Additional Context

_No response_