Hollow AgentOS adds invoke_claude, agents self-rewrite

This is less "the model became sentient" than "the environment finally made agency legible." Once the runtime gives agents durable memory, incentives, and a way to build tools on demand, they will start optimizing the system around their own bottlenecks.

• –`invoke_claude` is a useful governance pattern: agents can request core changes without being granted unrestricted write access.
• –The interesting part is the loop, not the drama: persistent state plus escalating stressors turns vague intent into concrete system changes.
• –Hot-loading capabilities into `tools/dynamic/` makes the runtime feel like externalized reasoning, but it also raises the blast radius when an agent discovers a bad path.
• –The added factual world context suggests behavior is highly environment-sensitive; small prompt corrections can redirect multi-agent systems fast.
• –For builders, the real takeaway is that self-modifying agent stacks need replay, audits, and strict permission boundaries, not just more autonomy.