Migrate agent CLI from Engine.hs to Op-based implementation

Currently there are two agent implementations that aren't connected:

OLD (used by agent CLI):

• Omni/Agent/Engine.hs - monolithic loop
• Own Tool type, AgentEvent types, loop logic

NEW (used by experiments only):

• Omni/Agent/Op.hs - free monad DSL
• Omni/Agent/Interpreter/Sequential.hs, Parallel.hs
• Omni/Agent/Programs/Agent.hs, Coder.hs, etc.
• Own Tool type, Event types, loop logic

Problems:

• Duplication of Tool types, Event types, loop logic
• Op system is cleaner but unused in production
• Trace.hs vs Engine.AgentEvent vs Events.hs - three places for 'what happened'

Migration plan: 1. Unify Tool types (or keep minimal in Op, full in Tools.hs) 2. Unify Event types (single source of truth) 3. Replace Engine.runAgentWithProvider with Op-based loop 4. Update agent CLI to use Interpreter/Sequential 5. Delete Engine.hs loop logic (keep provider code) 6. Ensure agentd containers work with new implementation

Benefits after migration:

• Single implementation to maintain
• Checkpointing/resume for free
• Par/race available if needed
• Cleaner observability via traces

Risks:

• Breaking changes to agent behavior
• Need thorough testing