Pluggable safety backends

[danielbentes]

Context

The current safety model enforces all constraints strictly. Different deployment contexts need different enforcement strategies — strict for production, advisory for development, graduated for mixed environments. (Roadmap Section 4.7)

Objective

Support different safety enforcement strategies for different deployment contexts through pluggable backends and runtime policy negotiation.

Tasks

• Implement strict mode (current behavior — hard enforcement of all safety constraints)
• Implement advisory mode (log violations but allow execution — for development/testing)
• Implement graduated mode (enforcement level increases with risk level)
• Implement custom backends (organizations define their own enforcement policies)
• Implement runtime safety policy negotiation for multi-agent systems with different safety requirements

Acceptance Criteria

• Strict mode blocks all safety violations (current behavior preserved)
• Advisory mode logs violations without blocking execution
• Graduated mode enforces low-risk actions permissively and high-risk actions strictly
• Custom backends can implement organization-specific safety policies
• Multi-agent systems can negotiate a common safety policy at runtime

Related

• Safety model: schemas/capability_ontology.yaml (risk levels, mutation flags)
• Checkpoint tracker: grounded_agency/state/checkpoint_tracker.py
• Hook enforcement: hooks/hooks.json
• Grounded agent adapter: grounded_agency/adapter.py