Architecture decisions

Design history and implementation contracts. Read each record's status; an accepted design alone does not establish production qualification.

Start with the deployment profiles for current operating behavior.

• ADR-001: Minimal embedded admin key console
• ADR-002: Admin console grows into an operator dashboard (still toolchain-free)
• ADR-003: Governance differentiation vs LiteLLM — v0.2 priority ordering
• ADR-004: OIDC admin authorization — resource-server-only, break-glass preserved
• ADR-005: Provider visibility in the console; UI-write deferred
• ADR-006: Config hot-reload via a single atomic topology generation
• ADR-007: Chargeback report from the audit log (inferplane report)
• ADR-008: UI-write provider registration — DB-authoritative topology store with Git export
• ADR-009: Opt-in PII masking filter — the honest cache/cost trade-off
• ADR-010: Self-service key issuance — identity endpoint + console
• ADR-011: OpenTelemetry trace spans (GenAI conventions), opt-in
• ADR-012: S3 Object Lock audit anchoring — tamper-evident → tamper-resistant
• ADR-013: Multi-replica HA — shared state design (implementation deferred)
• ADR-014: Provider registration UX — LiteLLM-parity guided "Add Model", secret-ref-honest
• ADR-015: Analytics Mode B — shared Postgres store + fenced single-writer aggregator
• ADR-016: Teams as first-class keystore records (D3) + derived read-only Users
• ADR-017: Budget-alert webhooks (D5b)
• ADR-018: Opt-in request/response body logging (D4)
• ADR-019: Bedrock Guardrails data-plane application + per-team override (D6)
• ADR-020: Per-team region locking (D7)
• ADR-021: Ticket-driven UX fixes — discoverable errors, model aliases, self-service usage
• ADR-022: Bedrock legacy extended-thinking rewrite for newer models
• ADR-023: Declarative virtual keys (config-managed, restart-durable)
• ADR-024: Bedrock InvokeModel passthrough ingress
• ADR-026: Admin console SSO login — SPA-as-OAuth-public-client (ADR-004 revision)
• ADR-027: Admin console i18n — self-contained en/ko/zh/ja dictionary
• ADR-028: CLI OIDC login — short-lived, auto-renewing virtual keys
• ADR-029: model-level fallback (unrouted-model and cross-model-404)
• ADR-030: Cost settlement correctness — streaming usage, cache tiers, and unpriced routes
• ADR-031: Monorepo split into control plane (inferplaned) and node-local data plane (mayu)
• ADR-032: Policy money unit, user/team subjects, and governance cadences
• ADR-033: Local GovernancePolicy file channel (standalone mode)
• ADR-034: Control-plane policy distribution and budget-lease protocol
• ADR-035: Kubernetes policy channel — CRD manifest + ConfigMap-mounted documents
• ADR-036: Control-plane usage telemetry — cost/usage visibility on inferplaned
• ADR-037: inferplaned console SSO — OIDC/Cognito login for the usage console
• ADR-038: Control-plane policy store — Postgres-backed GovernancePolicy writes
• ADR-039: TPM true-up, cache-aware quota, and explicit unlimited rules
• ADR-040: Credential brokering — inferplaned vends short-lived Bedrock credentials to mayu
• ADR-041: Budget-tier model substitution (routing.budgetTiers)
• ADR-042: Calendar budget windows through the policy and lease channel
• ADR-043: Policy-aware routing with local inspection
• ADR-044: Stable adaptive routing for coding agents
• ADR-045: Durable global budget escrow
• ADR-046: Postgres shared gateway governance
• ADR-047: Codex provider compatibility
• ADR-048: Verified identities with stable accounting references
• ADR-049: Task-oriented admin workspaces with explicit evidence and change review
• ADR-050: User cost visibility and profile-aware budget/token policy guides
• ADR-051: Cost-first dashboard and comparative budget table
• ADR-052: Native Converse ingress and governed administrator comparison
• ADR-053: Request outcomes and cost evidence
• ADR-054: Local monetary liability per provider attempt
• ADR-055: Select structural inspection from effective privacy policy
• ADR-057: Audit console evidence