Context

Implement all 5 loop prevention mechanisms from spec section 5.5. Delegation loops are a critical failure mode in multi-agent systems where agents can endlessly delegate tasks back and forth, consuming resources and never completing work.

Acceptance Criteria

Five Prevention Mechanisms

• Max delegation depth — default limit of 5 levels; configurable
• Rate limiting — max 10 delegations per agent pair per minute; configurable
• Identical request dedup — reject duplicate requests within a 60-second window
• Circuit breaker — after 3 bounces between same agents, trigger 5-minute cooldown
• Task ancestry tracking — prevent delegation to any agent in the current task's ancestor chain

On Loop Detection

• Block the delegation attempt
• Notify the sending agent with a clear error message
• Escalate to manager or human-in-the-loop
• Log the loop detection event with full context

Configuration & Testing

• All thresholds configurable via `LoopPreventionConfig` model
• Clear, actionable error messages on detection
• Escalation pathway works end-to-end
• Unit tests for each of the 5 mechanisms individually
• Integration test that creates a would-be loop and verifies prevention

Dependencies

• Depends on agent-to-agent messaging (Implement agent-to-agent messaging with channels and topics #10)

Design Spec Reference

• Section 5.5 — Loop Prevention