Risk Register

Risk	Why it matters	Current mitigation	Evidence milestone
Agent escalation is too rare or too noisy	Decision Inbox is a core product bet.	Structured injected tool, explicit capability, preflight verification, measure real tasks before preset tuning.	M10
Worktree lifecycle damages or loses work	Trust in Forge collapses immediately.	Stable ID paths, real Git executable, dirty-work invariant, durable cleanup operations, real-repo tests.	M3 onward
Provider protocols change	Adapters can break silently.	Raw-before-parse, adapter fixtures/contracts, explicit unknown-event handling, version/probe diagnostics.	M7 onward
Crash creates split brain between DB and external state	Tasks/runs become stuck or lie.	Durable Operations, reconciliation, idempotency, crash-boundary tests.	M2 onward
Long histories make desktop slow	Agent sessions can generate enormous streams.	Separate live/history models, paging, coalescing, bounded replay, perf fixtures.	M6/M12
UI displays stale state as live	User makes decisions on incorrect information.	Separate Connection and Sync health, snapshot fallback, authoritative refresh for sensitive actions.	M12
Too much architecture delays a usable tool	The project becomes documentation instead of product.	Vertical milestones; first usable UI before all subsystems are complete.	Every milestone
AI agents produce clever but unreadable code	User is learning and must maintain it.	AGENTS.md, lesson completion gates, simple architecture, explanation requirement, contract tests.	Every milestone
Repository setup scripts execute untrusted code	Local compromise/security risk.	Project trust state; explicit first execution; centralized ProcessRunner.	M3/M13
Full event sourcing adds compatibility/perf complexity too early	Can bury the project in infrastructure.	SQLite current-state model + durable Operations; only add event-sourcing machinery if future evidence requires it.	Revisit post-MVP