README.md

Architecture Decision Records — Moira

Canonical home for Moira's product architecture decisions. These were previously authored under an ai-sdlc working copy whose git index still tracked an unrelated project's ADRs; they now live with the code they govern (ADR-005/006 P5 hygiene fix).

ADR	Decision	Status
000	ADR template	—
001	Desktop shell on Tauri	Accepted
002	Orchestration engine	Superseded for v0.2 runtime by ADR-006 (custom DAG engine; LangGraph deferred)
003	Model routing via LiteLLM	Accepted
004	Dev execution is delegated to agent backends	Accepted (amended by ADR-006)
005	Pluggable persistence (SQLite/Postgres + git sink)	Accepted (amended by ADR-006)
006	Durable runner execution model (one model, two hosting modes)	Accepted
007	Governance packs as enforceable controls (repo-only, deterministic-first)	Accepted (MVP)
008	API identity (JWT, local/oidc) + default-deny RBAC (5 roles)	Accepted (backend MVP)
009	System-generated rework feedback on gate reject (closed quality loop, part 1)	Accepted (implemented)
010	Bounded rework loop — max_loop caps system rejects, audit-derived counter (closed quality loop, part 2)	Accepted (implemented)
011	Informed backend retry — previous errors in the retry prompt + linear backoff (closed quality loop, part 3)	Accepted (implemented)
012	Backend install/login probes + fail-fast launch gating (asymmetric-TTL cache, unknown never blocks)	Accepted (implemented)
013	"Retry" as the third decision on an escalated failed node; approve-the-gap becomes explicit	Accepted (implemented)
014	Closed test-fix loop — failing check output feeds the rework prompt (opt-in per gate, audit-derived)	Accepted (implemented)
015	Fail-loud model identity — litellm has no silent default model; validation at save/launch	Accepted (implemented)
016	Escalating, verified process termination — group SIGTERM→SIGKILL, reader unblock, visible verdict	Accepted (implemented)
017	Server-enforced cost budgets (run + workspace-month) — settings store, pause-not-kill, governed continue	Accepted (implemented)
018	Prompt & metering hygiene — [UNTRUSTED DATA] framing, fail-loud skills + references delivery, cache-aware tokens	Accepted (implemented)
019	Auth on by default, end-to-end — process default local, Tauri token handshake, CORS narrowing, one-command onboarding	Proposed (TODO)
020	Run isolation via git worktrees — code_path only, fail-closed, branch as deliverable; unblocks multi-run and ×N variants	Proposed (design settled)

ADR-007-agentic-engineering-workflow was left in the ai-sdlc framework repo — it describes the engineering workflow, not the Moira product, so its ownership is intentionally triaged separately.