Stricter shell command checks with a new runsTextHeld option

What

Before Claude Code runs a shell command, it analyses the command to decide whether it is safe to run without asking you. Several of these checks changed.

• A new option, runsTextHeld, is added to the command analysis. It is on unless explicitly set to false, and sits next to other checks such as forNameDo and rubyAttachedPattern.
• For Python commands run with -P or -I, when runsTextHeld is set, code passed with -c that mentions exec, eval or execfile is flagged, as are programs fed in through standard input. Before, -P or -I ended the check with nothing further examined. Other interpreters keep their existing behaviour.
• The part that reads command prefixes for run, foreach, -x and --exec now looks at the correct list of words.
• The shell -s check now tests a path with a helper instead of only checking whether it starts with /, ~ or $.

Why

These changes affect whether certain commands are run automatically or need your approval. Some Python commands that previously passed may now ask for permission.