Context

Track every LLM API call for cost management and observability per spec 10.2. This is essential for budget enforcement, usage analytics, and understanding agent economics.

Acceptance Criteria

• Every LLM call is tracked with: agent_id, task_id, provider, model, input_tokens, output_tokens, cost_usd, timestamp
• Cost calculated from pricing configuration (per-model token pricing)
• Aggregation queries by agent, department, and time period
• Async logging that does not block LLM call responses
• In-memory storage for now (SQLite persistence planned for M5)
• Unit tests for cost calculation accuracy
• Unit tests for aggregation queries
• Thread-safe logging for concurrent agent execution

Dependencies

• Depends on Implement provider adapter layer (direct API or via LiteLLM) #5 (provider adapter layer implementation)

Design Spec Reference

Section 10.2 — Cost Tracking and Budget Management