Problem

The security.redact_secrets feature replaces credential strings with asterisks in both tool output AND command execution contexts. This creates a functional breakage: when the agent needs to use a password in a terminal command (e.g., htpasswd -nb user Passw0rd!), the redaction either:

1. Masks the password with **** in the command itself, breaking actual command execution
2. Or the raw password appears in chat output, exposing credentials

Expected Behavior

Secret redaction should protect credentials in:

• Display output / chat messages (masked)
• Session logs / transcripts (masked)

But should NOT interfere with:

• Actual command execution (the real value must be used)
• File writes to credential files / .env files

Suggested Fix

1. Add a security.display_redaction_only flag — redact secrets in output/logs but pass real values to subprocesses
2. At minimum, add clear documentation about how to handle credentials in terminal commands without exposure