Adopt a consistent rule: anything coming from persistence, other platforms/accounts, terminal escape sequences, or secret-managed inputs must be validated/normalized and handled with collision-safe, non-injectable, non-leaky logic.

Apply this as standards:

Example patterns (condensed):

// 1) Shell quoting helper
fn shell_single_quote(value: &str) -> String {
    format!("'{}'", value.replace("'", "'\\''"))
}

// 2) Secret env insertion with collision safety
fn build_secret_env_vars(secrets: &HashMap<String, ManagedSecretValue>) -> HashMap<OsString, OsString> {
    let mut env_vars = HashMap::new();
    for (secret_key, secret) in secrets {
        // typed secrets: if any env var is already set non-empty, skip that secret entirely
        // (and never override existing process env)
        for (env_name, env_value) in typed_secret_entries(secret) {
            if std::env::var(env_name).is_ok_and(|v| !v.is_empty()) {
                continue; // collision: skip
            }
            env_vars.insert(OsString::from(env_name), OsString::from(env_value));
        }
    }
    env_vars
}