Add *safe* global option

[clavery]

We want to control the safety level of the tools when used by LLMs in agentic tooling loops. Potential options:

• --safe flag that disables all destructive actions
• --safety-level flag that allows for levels (NODELETE, NOUPDATE, etc)
• env var support
• http middleware to block actions

flag should be configurable at a global level (env var, config, etc) and not easily ejected by the LLM (perhaps standard CLI flags should not be used for this reason).