Settings file corruption: 'Always allow' stores entire command content instead of extracting pattern

Description

When clicking "Always allow" on complex bash commands (especially multi-line commands with heredocs), Claude Code stores the entire command content verbatim in the settings file instead of extracting a sensible pattern.

Example of corrupted entries

The settings.local.json file ended up containing entries like:

"Bash(git -C /path/to/repo commit -m \"$(cat <<'EOF'\nfeat(#123): Add new feature\n\nThis is a detailed commit message\nwith multiple lines\nand descriptions...\nCo-Authored-By: Claude <noreply@anthropic.com>\nEOF\n)\")"

And even content that appears to be from echoed output or documentation:

"Bash(| Some table content       |\n| with ASCII borders      |\n..."

Expected behavior

When "Always allow" is clicked on a complex command, it should extract a general pattern like:

• Bash(git:*) or
• Bash(git -C:*) or
• Bash(git commit:*)

Not store the entire command content as part of the pattern.

Impact

• Settings file grows from ~60 lines to 100+ lines
• Contains potentially sensitive content (commit messages, paths, project details)
• Patterns don't actually match future commands (too specific)
• User gets re-prompted for the same command types repeatedly
• File becomes corrupted and needs manual cleanup

Workaround

Manually clean up the settings file to use general patterns and avoid clicking "Always allow" on complex multi-line commands.

Environment

• Claude Code CLI
• macOS