0009 — Product-first sequencing (supersedes ADR-0006 phase order)

Status: accepted Date: 2026-04-30 Supersedes: ADR-0006 (phase order only; tracer-bullet philosophy retained)

Context

ADR-0006 set vertical (tracer-bullet) implementation in foundation-first order: identity → billing → agents → platform. plan/00 D2 (“RBAC at foundation”) drove this — auth/RBAC before product features.

Solo dev pivoted: minimal first, defer non-blocking foundation work, demoable agent in Phase 1, retrofit auth/billing/audit on real product.

Decision

Phase 1 starts with paper-board/agents as the tracer bullet (not identity).

New phase order:

Phase	Service	Scope
1	agents	Minimal runtime: 3 tables, REST + SSE, 1 LLM provider, no auth
2	identity	Auth (API keys + JWT signing); retrofit agents middleware
3	billing	Subscriptions + usage metering; agents emit usage
4	platform	Audit log + memory + artifacts
5	(production hardening)	DNS, TLS, partitioning, hash-chain, cosign, prod cluster
6	runtime + compute	Multi-tenant data plane split from agents
7	gateway	Centralize auth + routing
8+	RBAC + marketplace + advanced

Phase 1.0 prep (sdk + proto + infra) order unchanged from ADR-0006.

Rejected alternatives

Foundation-first (ADR-0006 original) — defers demoable product 4-6 weeks; foundation over-engineered before product feedback.
Big-bang all-services — solo dev cannot maintain 4 services in flight simultaneously.
Agents-first WITHOUT multi-repo — would defeat ADR-0001; multi-repo cost paid, sequencing only changes.

Consequences

Gain:

Demoable agent in 2-3 weeks (vs 4-6 weeks foundation-first).
Real product feedback shapes auth/billing/audit design (vs upfront speculation).
Phase 1 acceptance simple: curl POST /v1/sessions/:id/prompt returns SSE.
Parallel deferral of plan/00 ops items (DNS, mailbox, Stripe, Postmark, Vanta, Statuspage, prod K8s) — minimal Phase 0.

Risk accepted:

Auth retrofit refactor on existing endpoints (Phase 2).
No multi-tenancy in Phase 1 (single hardcoded org_id placeholder).
plan/00 D2 (“RBAC at foundation”) violated — RBAC deferred to Phase 8.
Phase 1 minimal schema → Phase 5 partitioning is migration work (acceptable: data is dev-only Phase 1-4).

Mitigation:

Phase 1 endpoints minimal (~6); retrofit blast radius small.
Phase 5 schema retrofit while data volume is dev-only (truncate-and-recreate acceptable).
Catalog tables (roles, role_permissions) added in Phase 5 even if assignments deferred to Phase 8 — minimizes Phase 8 schema change.

Tracer-bullet philosophy retained (ADR-0006)

Each service still implemented end-to-end before next. Phase 1 = agents end-to-end (migrator + schema + server + LLM + Helm + smoke), then Phase 2 = identity end-to-end. The vertical-vs-horizontal slicing principle is the same; only the service order changes.

Reference

Roadmap: tasks/2026-04-30-backend-roadmap.md
Archived foundation-first plan: tasks/archive/2026-04-30-backend-phase-1-foundation-first.archive.md