- 6 phases: Event Bus, Agent Runtime, AI Budget, AI Evals, Reviews, Support Cases
- 7-sprint mapping (14 weeks) with parallelization opportunities
- Cross-referenced all claims against actual codebase — 15 bugs/gaps found and fixed:
- DurableEventBus already exists in @bytelyst/events (not just in-memory)
- jobs/ has 25 tests (not 6), support-cases/ has 4 (not 3)
- ai-budgets already has verdict engine (252 LOC), alert generation, model allowlists
- BudgetScopeType only supports product+agent (not org/workspace yet)
- reviews/notifications.ts already has notifyReviewAssigned()
- Phase 2 effort overflowed (17d in 15d sprint) — rebalanced
- Test baseline corrected to 1,278 (not 1,308)
- Identified Cosmos QueueStore gap as critical path for Phase 1
- ai-diagnostics has 5,235 LOC but 0 tests — flagged as risk
- Estimated ~138 new tests bringing total to ~1,416
