cartethyia 2.4.5 → 2.5.0

package/src/persistence/schema.ts

@@ -1,1257 +1,1360 @@
-/** Canonical Drizzle schema: every persistence table, enum, index, and shared type. */
-import { sql } from "drizzle-orm";
-import {
-  bigint, boolean, check, customType, index, integer, jsonb, numeric, pgEnum, pgTable,
-  primaryKey, text, timestamp, uniqueIndex, uuid, type AnyPgColumn,
-} from "drizzle-orm/pg-core";
-
-export const TENANT_STATUSES = ["active", "suspended", "deleted"] as const;
-export type TenantStatus = (typeof TENANT_STATUSES)[number];
-
-export const tenants = pgTable(
-  "tenants",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    name: text("name").notNull(),
-    status: text("status").$type<TenantStatus>().notNull(),
-    // Intentionally inline (not createdAtColumn()): this root table is imported
-    // by the helpers below; sharing the helper here would make a use-before-define.
-    createdAt: timestamp("created_at", { withTimezone: true }).notNull().defaultNow(),
-    updatedAt: timestamp("updated_at", { withTimezone: true }).notNull().defaultNow(),
-  },
-  (table) => [
-    check("tenants_status_check", sql`${table.status} IN ('active', 'suspended', 'deleted')`),
-  ],
-);
-
-// Canonical tenant/timestamp column definitions shared by every table.
-
-export function tenantRefNullable() {
-  return uuid("tenant_id").references(() => tenants.id, { onDelete: "cascade" });
-}
-
-export function tenantRefRequired() {
-  return uuid("tenant_id")
-    .notNull()
-    .references(() => tenants.id, { onDelete: "cascade" });
-}
-
-export function tenantRefPrimaryKey() {
-  return uuid("tenant_id")
-    .primaryKey()
-    .references(() => tenants.id, { onDelete: "cascade" });
-}
-
-export function tenantCascadeSetNull() {
-  return uuid("tenant_id").references(() => tenants.id, { onDelete: "set null" });
-}
-
-export function createdAtColumn() {
-  return timestamp("created_at", { withTimezone: true }).notNull().defaultNow();
-}
-
-export function updatedAtColumn() {
-  return timestamp("updated_at", { withTimezone: true }).notNull().defaultNow();
-}
-
-export function timestampColumns() {
-  return { createdAt: createdAtColumn(), updatedAt: updatedAtColumn() };
-}
-
-/**
- * `bytea` column for encrypted credential ciphertext (application-layer
- * envelope encryption: the ciphertext stored here is opaque to the
- * database). Drizzle's pg-core does not ship a first-class `bytea`
- * builder, so this is the documented `customType` escape hatch (the same
- * DDL-shape escape-hatch principle applied to column-type gaps in the
- * ORM's builder).
- */
-export const bytea = customType<{ data: Buffer }>({
-  dataType() {
-    return "bytea";
-  },
-});
-
-// Shared three-state health machine for provider_accounts and network_pools —
-// the account state machine is reused for network pools rather than
-// inventing a second one.
-//
-// `degraded` was retired. It sat between "working" and "parked" and every
-// consumer had to decide what it meant: routing treated it as a hard exclusion,
-// the sweep treated it as recoverable, and the console showed it as a third
-// badge — three readings of one value, and the operator reported the result as
-// ambiguous. The two facts it mixed are now each stated by a value that means
-// exactly one thing: a fault that clears on its own is `cooldown` (with a
-// deadline the sweep honors), and a fault that needs an operator is `disabled`.
-export const healthStatus = pgEnum("health_status", ["active", "cooldown", "disabled"]);
-
-/**
- * The health state machine shared by `provider_accounts` and `network_pools`.
- * One definition, because the two tables are driven by the same transitions and
- * a column added to one of them by hand is how they drift.
- */
-export function healthColumns() {
-  return {
-    status: healthStatus("status").notNull().default("active"),
-    consecutiveFailures: integer("consecutive_failures").notNull().default(0),
-    lastSuccessAt: timestamp("last_success_at", { withTimezone: true }),
-    lastError: text("last_error"),
-    lastErrorCategory: text("last_error_category"),
-    lastErrorAt: timestamp("last_error_at", { withTimezone: true }),
-    cooldownUntil: timestamp("cooldown_until", { withTimezone: true }),
-    lastRecoveredAt: timestamp("last_recovered_at", { withTimezone: true }),
-  };
-}
-
-export const credentialKind = pgEnum("credential_kind", ["api_key", "oauth", "none"]);
-
-// Shared wire-family domain used by providers.wire_family_default and
-// models.wire_family.
-//
-// `gemini` is Factory's native generateContent wire. Added by forward
-// migration `0002_add_gemini_wire_family.sql`; the enum here must stay in
-// step with what that migration applies, or a fresh database (built from the
-// baseline) and a migrated one would disagree.
-export const wireFamily = pgEnum("wire_family", ["chat", "responses", "messages", "gemini"]);
-
-// `https` is a hosted relay front door (Vercel, Cloudflare, Deno, Netlify):
-// plain HTTP CONNECT to an origin that terminates TLS itself, so the transport
-// is recorded as what it speaks rather than as a generic proxy.
-export const networkPoolKind = pgEnum("network_pool_kind", ["http", "https", "socks5"]);
-
-//  health_events.entity_kind — distinguishes which owning
-// table entity_id/account_id/network_pool_id references.
-export const healthEntityKind = pgEnum("health_entity_kind", ["account", "pool"]);
-
-export const telemetrySourceSurface = pgEnum("telemetry_source_surface", [
-  "chat",
-  "responses",
-  "messages",
-  "completion",
-]);
-
-export const telemetryStatus = pgEnum("telemetry_status", [
-  "completed",
-  "failed",
-  "cancelled",
-  "truncated",
-]);
-
-// The id matches the seven canonical builtin IDs (openai, anthropic, claude,
-// codex, opencode-zen, ollamacloud, cerebras) or a custom/BYOK slug
-// rejected at the identity boundary when it collides case-insensitively with
-// a reserved ID — that rejection is application-layer, not a DB constraint
-// here. `capability_profile` is a JSONB routing-capability map with no index:
-// it is read with the provider row the snapshot builder already loads, so a
-// separate index would only add write cost. `tenant_id`
-// null means the provider is global/built-in, matching the same convention
-// as `provider_accounts.tenant_id`; populated means tenant-owned/BYOK.
-// `base_url`/`compatibility_profile` are BYOK-only: `compatibility_profile`
-// carries HTTP dispatch config (extra headers/query params/streaming mode)
-// and is intentionally a separate column from `capability_profile`, which
-// drives routing capability flags — conflating the two would corrupt routing
-// data with unrelated HTTP config.
-export const providers = pgTable(
-  "providers",
-  {
-    id: text("id").primaryKey(),
-    tenantId: tenantRefNullable(),
-    wireFamilyDefault: wireFamily("wire_family_default"),
-    capabilityProfile: jsonb("capability_profile"),
-    baseUrl: text("base_url"),
-    compatibilityProfile: jsonb("compatibility_profile"),
-    enabled: boolean("enabled").notNull().default(true),
-    // Builtin default `true` (most providers need a real credential). Only a
-    // genuinely public/unauthenticated builtin (OpenCode Free) is seeded
-    // `false`. When `false`, the snapshot builder emits a routable candidate
-    // with zero `provider_accounts` rows instead of marking it `disabled` —
-    // see `src/transport/routing/route-catalog.ts`.
-    requiresAccount: boolean("requires_account").notNull().default(true),
-  },
-);
-
-// Upstream account credentials plus the full health state machine.
-// `tenant_id` null means the account is shared pool-wide; populated means
-// tenant-owned/BYOK. The concurrency ceiling lives in provider routing
-// (`provider_routing_settings.max_inflight`) and network pools
-// (`network_pools.max_inflight`); the per-account override no longer exists.
-export const providerAccounts = pgTable("provider_accounts", {
-  id: uuid("id").primaryKey().defaultRandom(),
-  providerId: text("provider_id")
-    .notNull()
-    .references(() => providers.id, { onDelete: "cascade" }),
-  tenantId: tenantRefNullable(),
-  label: text("label").notNull(),
-  credentialCiphertext: bytea("credential_ciphertext"),
-  /**
-   * HMAC fingerprint of the credential, never the plaintext secret. Nullable
-   * for credential-less accounts and legacy rows without an identity.
-   */
-  credentialFingerprint: text("credential_fingerprint"),
-  credentialKind: credentialKind("credential_kind").notNull(),
-  /**
-   * Per-account upstream auth configuration that is not a secret: auth method,
-   * region, profile ARN, OAuth client id, token endpoint, scopes. The adapter
-   * reads it where the credential is resolved (it selects the regional endpoint
-   * and the profile the upstream binds the request to) and token refresh reads
-   * it again to pick the right endpoint and client. Secrets never live here —
-   * see `credential_ciphertext` and `provider_oauth_states`.
-   */
-  authState: jsonb("auth_state"),
-  createdAt: timestamp("created_at", { withTimezone: true }).notNull().defaultNow(),
-
-  ...healthColumns(),
-  modelCooldowns: jsonb("model_cooldowns").notNull().default({}),
-  /**
-   * How many times each model has been reported unsupported for this account:
-   * `{ [modelId]: hits }`. Written only by the unsupported-model path, and it
-   * is the escalation level for that pair's cooldown.
-   *
-   * Deliberately a separate column from `model_cooldowns`: that value is
-   * `{ [modelId]: deadline }`, read by routing (`route-catalog.ts`) and by the
-   * console, and widening it to `{until, hits}` would have changed a shape two
-   * independent readers parse. Nullable — an account that never hit an
-   * unsupported model has nothing to count.
-   */
-  modelBlockHits: jsonb("model_block_hits").$type<Readonly<Record<string, number>>>(),
-  /**
-   * Stable list position within (tenant, provider). Accounts were ordered by
-   * `created_at`, which let two same-millisecond rows swap between loads; this
-   * is explicit and can be rewritten when the operator reorders or appends.
-   */
-  sortIndex: integer("sort_index").notNull().default(0),
-  /**
-   * The credential is a static bearer token (a JWT/access token pasted on its
-   * own, or an account an operator has pinned): it is used exactly as issued and
-   * must never be sent to a refresh endpoint. Distinct from `credentialKind`,
-   * which says what family the credential belongs to, not whether it can be
-   * re-minted — an `oauth` account can carry a static token when its refresh
-   * grant is absent or dead while the token itself is still valid. A static
-   * account is skipped by the refresh sweep and is never disabled by a refresh
-   * failure, because there is no refresh to run.
-   */
-  staticToken: boolean("static_token").notNull().default(false),
-  /**
-   * When this account's credential was last read for its own identity (email,
-   * display name). Null means never read.
-   *
-   * This is the convergence stamp for the enrichment sweep, and the reason an
-   * import can be instant: the batch path persists the credential without
-   * parsing it, and the sweep fills the identity in afterwards. An account is
-   * only listed while this is null, so a completed pass never re-decrypts a
-   * credential it already read.
-   */
-  identityCheckedAt: timestamp("identity_checked_at", { withTimezone: true }),
-  /**
-   * Email read out of the account's own JWT claims. Cosmetic: it names the
-   * account in the console and is never used to authenticate or route.
-   */
-  accountEmail: text("account_email"),
-
-  },
-  (table) => [
-    uniqueIndex("provider_accounts_identity_uidx").on(
-      table.providerId,
-      // No `::uuid` cast: the migrations create this index without one on
-      // SQLite (baseline l.440) and with one on Postgres (l.423), and the
-      // Drizzle definition is what a test's in-memory database is built
-      // from — a Postgres-only cast there made every account insert fail
-      // with `unrecognized token: ":"`.
-      sql`coalesce(${table.tenantId}, '00000000-0000-0000-0000-000000000000')`,
-      table.credentialFingerprint,
-    ),
-    index("provider_accounts_status_cooldown_idx").on(table.status, table.cooldownUntil),
-    index("provider_accounts_provider_tenant_idx").on(table.providerId, table.tenantId),
-  ],
-);
-
-
-//  OAuth refresh token + cross-process lease, normalized out of
-// `provider_accounts` into its own 1:0/1 table. The vast majority of
-// provider accounts are `credentialKind: "api_key"`, for which these four
-// columns stayed null forever on the shared row; splitting them here keeps
-// the hot health-machine row narrow and isolates the fenced compare-and-swap
-// lease traffic (`providers/auth/refresh-service.ts`) from unrelated
-// account reads/writes.
-export const providerOauthStates = pgTable("provider_oauth_states", {
-  providerAccountId: uuid("provider_account_id")
-    .primaryKey()
-    .references(() => providerAccounts.id, { onDelete: "cascade" }),
-  /**
-   * Encrypted refresh token. Nullable: an account can be pasted with only an
-   * access token (or a credential whose refresh token never arrived), and the
-   * account must still be tracked so the refresh sweep can flag it for re-auth
-   * rather than leaving it invisible. A null value means "cannot refresh".
-   */
-  refreshCiphertext: bytea("refresh_ciphertext"),
-  /**
-   * Encrypted companion secret for the minority of flows whose refresh is not
-   * authorized by the refresh token alone: a device-flow client registration
-   * mints a client secret that must be replayed beside the token. Nullable —
-   * most providers refresh with the refresh token by itself.
-   */
-  clientSecretCiphertext: bytea("client_secret_ciphertext"),
-  /**
-   * Access-token expiry. Nullable: a credential pasted without an expiry (an
-   * opaque token, or a bare access token) has no known deadline, and reads as
-   * "always due" so the sweep refreshes it as soon as it can.
-   */
-  expiresAt: timestamp("expires_at", { withTimezone: true }),
-  // OAuth refresh lease: fenced compare-and-swap coordination so only one
-  // process refreshes a given account's OAuth token at a time, and a losing
-  // process reloads instead of clobbering a peer's fresher token.
-  leaseOwner: text("lease_owner"),
-  leaseExpiresAt: timestamp("lease_expires_at", { withTimezone: true }),
-});
-
-
-// One row per (provider_id, model_id, route), with `endpoint_path`
-// identifying the route — the same model through two providers, or through
-// two routes of one provider, is two rows. `enabled` is a hard routing
-// invariant read by ProviderCatalogService, not a display-only flag.
-export const models = pgTable(
-  "models",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    providerId: text("provider_id")
-      .notNull()
-      .references(() => providers.id, { onDelete: "cascade" }),
-    modelId: text("model_id").notNull(),
-    wireFamily: wireFamily("wire_family").notNull(),
-    // Which protocol shape at all (`llm` = the canonical chat pipeline). A
-    // non-`llm` row (System One) is served by its native passthrough route, not
-    // by the surface codecs. Kept a plain text column with a default rather than
-    // a pgEnum: unlike `wire_family`, the set is expected to grow (embeddings,
-    // tts, …) and a widened enum needs a migration per addition.
-    serviceKind: text("service_kind").notNull().default("llm"),
-    endpointPath: text("endpoint_path").notNull(),
-    contextLimit: integer("context_limit"),
-    outputLimit: integer("output_limit"),
-    modalities: jsonb("modalities"),
-    reasoning: boolean("reasoning").notNull().default(false),
-    toolCall: boolean("tool_call").notNull().default(false),
-    cost: jsonb("cost"),
-    source: text("source"),
-    sourceUpdatedAt: timestamp("source_updated_at", { withTimezone: true }),
-    enabled: boolean("enabled").notNull().default(true),
-  },
-  (table) => [
-    uniqueIndex("models_provider_model_route_uidx").on(
-      table.providerId,
-      table.modelId,
-      table.endpointPath,
-    ),
-    index("models_provider_enabled_idx").on(table.providerId, table.enabled),
-    // Bare-id lookups, with no provider named: see migration
-    // `0003_add_models_model_id_index`. Without it those seq-scan this table,
-    // which carries one row per provider x model x route.
-    index("models_model_id_idx").on(table.modelId),
-  ],
-);
-
-// Tenant-scoped suppression of a GLOBAL/built-in model. Tenants cannot mutate
-// the shared `models` row (see `DrizzleProviderCatalogStore.setModelEnabled`),
-// so a disable is recorded here and filtered out of that tenant's routing
-// snapshot. Tenant-owned/BYOK models are still toggled directly on `models`.
-export const tenantDisabledModels = pgTable(
-  "tenant_disabled_models",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    tenantId: tenantRefRequired(),
-    providerId: text("provider_id")
-      .notNull()
-      .references(() => providers.id, { onDelete: "cascade" }),
-    modelId: text("model_id").notNull(),
-    endpointPath: text("endpoint_path").notNull(),
-    createdAt: createdAtColumn(),
-  },
-  (t) => [
-    uniqueIndex("tenant_disabled_models_tenant_provider_model_uidx").on(
-      t.tenantId,
-      t.providerId,
-      t.modelId,
-      t.endpointPath,
-    ),
-  ],
-);
-
-
-// Tenant-owned HTTP/HTTPS/SOCKS5 pools share the provider-account health state machine.
-export const networkPools = pgTable("network_pools", {
-  id: uuid("id").primaryKey().defaultRandom(),
-  createdAt: timestamp("created_at", { withTimezone: true }).notNull().defaultNow(),
-  kind: networkPoolKind("kind").notNull(),
-  /**
-   * Operator-set family/tag for this pool, e.g. `residential-id` or
-   * `datacenter-sg`. Purely organisational: it names which family the node
-   * belongs to and nothing in routing or dispatch reads it. Null when the
-   * operator has not tagged the pool.
-   */
-  family: text("family"),
-  endpointConfig: jsonb("endpoint_config").notNull(),
-  credentialCiphertext: bytea("credential_ciphertext"),
-  maxInflight: integer("max_inflight"),
-  /** Routing weight for weighted selection across the tenant pool group. */
-  weight: integer("weight"),
-  tenantId: tenantRefRequired(),
-
-  ...healthColumns(),
-  lastLatencyMs: integer("last_latency_ms"),
-  lastHealthCheckAt: timestamp("last_health_check_at", { withTimezone: true }),
-  /** Public egress address reported by the last successful probe; null until probed. */
-  egressIp: text("egress_ip"),
-  /** Operator-set egress allowance in bytes; null means unmetered. */
-  quotaBytes: bigint("quota_bytes", { mode: "number" }),
-  /** Last speed-test payload actually transferred, in bytes; null until measured. */
-  lastSpeedtestBytes: integer("last_speedtest_bytes"),
-  /** Wall-clock duration of the last speed-test transfer, in milliseconds. */
-  lastSpeedtestDurationMs: integer("last_speedtest_duration_ms"),
-  /** "ok" | "failed" for the last speed-test; null until measured. */
-  lastSpeedtestStatus: text("last_speedtest_status"),
-  /** Failure message of the last speed-test; null on success or unmeasured. */
-  lastSpeedtestError: text("last_speedtest_error"),
-  /** When the last speed-test ran. */
-  lastSpeedtestAt: timestamp("last_speedtest_at", { withTimezone: true }),
-  /**
-   * Provider-scoped probe verdict, stored beside the other last-known facts
-   * (`egress_ip`, `last_latency_ms`, the speed-test columns) rather than in the
-   * browser.
-   *
-   * It lived in `localStorage` first, and that made it decorative: the verdict
-   * was visible to the operator who ran it and to nobody else — not to another
-   * machine, not to the gateway, and not to admission. A pool could read
-   * "OpenCode blocked" on screen while dispatch kept sending traffic through
-   * it, because nothing outside that one browser had ever heard the verdict.
-   *
-   * `null` means never probed. These describe the *address* recorded in
-   * `egress_ip` at the time, so a changed endpoint clears them.
-   */
-  providerProbeProviderId: text("provider_probe_provider_id"),
-  /** `allowed` | `blocked` | `throttled` | `unreachable`; null until probed. */
-  providerProbeStatus: text("provider_probe_status"),
-  /** Upstream HTTP status when a response arrived. */
-  providerProbeHttpStatus: integer("provider_probe_http_status"),
-  /** Operator-facing reason: the upstream's own words, or why the probe failed. */
-  providerProbeReason: text("provider_probe_reason"),
-  providerProbeLatencyMs: integer("provider_probe_latency_ms"),
-  providerProbeAt: timestamp("provider_probe_at", { withTimezone: true }),
-  /** The endpoint the verdict was measured against. Compared on read, so an
-   * edited endpoint reports "never probed" instead of a stale verdict. */
-  providerProbeEndpoint: text("provider_probe_endpoint"),
-  /** Metered egress totals, flushed from the in-memory socket counters on
-   * every pool-touching write so the quota bar survives a restart. */
-  bytesSentTotal: bigint("bytes_sent_total", { mode: "number" }).notNull().default(0),
-  /** Metered ingress totals; see `bytesSentTotal`. */
-  bytesReceivedTotal: bigint("bytes_received_total", { mode: "number" }).notNull().default(0),
-  },
-  (table) => [
-    index("network_pools_tenant_id_idx").on(table.tenantId),
-    index("network_pools_status_cooldown_idx").on(table.status, table.cooldownUntil),
-    index("network_pools_family_idx").on(table.family),
-  ],
-);
-
-
-// Account and network-pool health events share one table. The entity kind
-// selects exactly one owner FK, and the CHECK constraint below preserves
-// referential integrity for both variants.
-export const healthEvents = pgTable(
-  "health_events",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    entityKind: healthEntityKind("entity_kind").notNull(),
-    accountId: uuid("account_id").references(() => providerAccounts.id, { onDelete: "cascade" }),
-    networkPoolId: uuid("network_pool_id").references(() => networkPools.id, {
-      onDelete: "cascade",
-    }),
-    fromStatus: healthStatus("from_status"),
-    toStatus: healthStatus("to_status").notNull(),
-    reason: text("reason"),
-    errorCategory: text("error_category"),
-    modelId: text("model_id"),
-    createdAt: createdAtColumn(),
-  },
-  (table) => [
-    index("health_events_account_id_created_at_idx").on(table.accountId, table.createdAt),
-    index("health_events_account_model_created_idx").on(table.accountId, table.modelId, table.createdAt),
-    index("health_events_network_pool_id_created_at_idx").on(table.networkPoolId, table.createdAt),
-    check(
-      "health_events_entity_reference_check",
-      sql`(${table.entityKind} = 'account' AND ${table.accountId} IS NOT NULL AND ${table.networkPoolId} IS NULL) OR (${table.entityKind} = 'pool' AND ${table.networkPoolId} IS NOT NULL AND ${table.accountId} IS NULL)`,
-    ),
-  ],
-);
-
-
-// Model aliasing & combos: replaces the dormant, disconnected
-// `routing_policies` scaffolding. Aliases resolve a client-facing name to a
-// real model; combos back a client-facing name with >=1 real models,
-// auto-selected via `fallback` or `round_robin`.
-export const modelComboStrategy = pgEnum("model_combo_strategy", [
-  "fallback",
-  "round_robin",
-]);
-
-/** Canonical `ComboStrategy` union, derived from the enum above. Console
- * domain and routing layers import this instead of hand-typing their own mirror.
- * `ProviderRoutingStrategy` covers the same two values but persists on
- * `provider_routing_settings` beside `rotateCount`; sticky routing is gone. */
-export type ComboStrategy = (typeof modelComboStrategy.enumValues)[number];
-
-export const modelAliases = pgTable(
-  "model_aliases",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    tenantId: tenantRefRequired(),
-    alias: text("alias").notNull(),
-    targetModel: text("target_model").notNull(),
-    /** Explicit list position. */
-    sortIndex: integer("sort_index").notNull().default(0),
-    ...timestampColumns(),
-  },
-  (t) => [uniqueIndex("model_aliases_tenant_alias_uidx").on(t.tenantId, t.alias)],
-);
-export type ModelAlias = typeof modelAliases.$inferSelect;
-
-export const modelCombos = pgTable(
-  "model_combos",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    tenantId: tenantRefRequired(),
-    name: text("name").notNull(),
-    members: jsonb("members").notNull().$type<string[]>(),
-    strategy: modelComboStrategy("strategy").notNull().default("fallback"),
-    /** Explicit list position. */
-    sortIndex: integer("sort_index").notNull().default(0),
-    ...timestampColumns(),
-  },
-  (t) => [uniqueIndex("model_combos_tenant_name_uidx").on(t.tenantId, t.name)],
-);
-export type ModelCombo = typeof modelCombos.$inferSelect;
-
-export const providerRoutingStrategy = pgEnum("provider_routing_strategy", [
-  "fallback",
-  "round_robin",
-]);
-
-/**
- * Canonical provider-routing strategy union, derived from the enum above so
- * the persisted values and the type cannot drift. The console catalog and the
- * routing engine both import this instead of restating the two literals.
- */
-export type ProviderRoutingStrategy = (typeof providerRoutingStrategy.enumValues)[number];
-
-/**
- * Per-account inflight ceiling applied when a provider has no routing row yet.
- * One literal: the console default and the route catalog's fallback both read
- * it. `null` is the "unlimited" choice, and it is also this default — an
- * operator who has never opened the routing card asked for no ceiling, and
- * inventing one silently capped their concurrency. A ceiling they *did* set is
- * still honoured, so this only decides the untouched case.
- */
-export const DEFAULT_PROVIDER_MAX_INFLIGHT: null = null;
-
-export const providerRoutingSettings = pgTable(
-  "provider_routing_settings",
-  {
-    providerId: text("provider_id")
-      .notNull()
-      .references(() => providers.id, { onDelete: "cascade" }),
-    tenantId: tenantRefNullable(),
-    strategy: providerRoutingStrategy("strategy").notNull().default("fallback"),
-    /** Requests served by one account before round robin advances. */
-    rotateCount: integer("rotate_count").notNull().default(1),
-    /** Per-account inflight ceiling applied to every account of this provider
-     * (the field sits next to the failover/round-robin strategy in the UI).
-     * `null` = UNLIMITED concurrency per account. */
-    maxInflight: integer("max_inflight"),
-    /**
-     * Per-pool inflight ceiling for this provider: how many concurrent
-     * requests one proxy pool may carry *for this provider*.
-     *
-     * A second, independent axis from `maxInflight`. That one bounds a single
-     * account; `network_pools.max_inflight` bounds a pool across every
-     * provider; this one bounds one provider's share of one pool. All three
-     * are enforced together — none replaces another.
-     *
-     * `null` = unlimited, and that is the default: a provider nobody has
-     * configured behaves exactly as it did before this column existed, with no
-     * extra Redis key and no extra round trip.
-     */
-    maxInflightPerProxy: integer("max_inflight_per_proxy"),
-    /**
-     * Minimum remaining allowance an account must hold to stay routable, for
-     * this provider. `null` means no floor. Compared against the account's own
-     * unit: absolute remaining credit when the provider reports credits, the
-     * aggregate credit percentage and lowest independent percentage window
-     * otherwise. An account below it is skipped and the request fails over to
-     * the next account instead of erroring.
-     */
-    usageFloor: integer("usage_floor"),
-    /** `credit` compares the summed remaining credit; `percent` compares the
-     * aggregate credit percentage and lowest independent percentage-only
-     * window. Ignored while `usageFloor` is null. */
-    usageFloorUnit: text("usage_floor_unit").$type<"credit" | "percent">().notNull().default("percent"),
-    enabled: boolean("enabled").notNull().default(false),
-    // Route-selected User-Agent for built-in API-key providers; OAuth and BYOK identities stay native.
-    // The column default is the Claude Code CLI identity; the live default for a
-    // row that never set one is `buildClaudeCliUserAgent()` in
-    // providers/operations/custom-cli-headers.ts, which this mirrors.
-    userAgent: text("user_agent").notNull().default("claude-cli/2.1.293 (external, cli)"),
-    // When true, this provider's requests always dial direct. When false,
-    // dispatch automatically selects among every active network pool the
-    // tenant owns (least-loaded/non-cooldown) — never an admin-pinned pool.
-    skipProxy: boolean("skip_proxy").notNull().default(false),
-  },
-  (table) => ({
-    tenantProviderUnique: uniqueIndex("provider_routing_settings_tenant_provider_idx").on(
-      table.tenantId,
-      table.providerId,
-    ),
-    globalProviderUnique: uniqueIndex("provider_routing_settings_global_provider_idx")
-      .on(table.providerId)
-      .where(sql`tenant_id IS NULL`),
-  }),
-);
-
-export const poolRoutingStrategy = pgEnum("pool_routing_strategy", [
-  "least_loaded",
-  "round_robin",
-]);
-export type PoolRoutingStrategy = (typeof poolRoutingStrategy.enumValues)[number];
-
-/**
- * Per-tenant selection strategy across the tenant's active network pools —
- * the pool-group counterpart to `provider_routing_settings`. `least_loaded`
- * (default) keeps the weighted least-loaded scan with its rotating tie-break;
- * `round_robin` serves `rotate_count` requests per pool before advancing,
- * mirroring the account strategy, and skips capacity/cooldown pools the way
- * account failover does. An absent row reads as the default, so tenants that
- * never touch the toggle own no rows.
- */
-export const poolRoutingSettings = pgTable("pool_routing_settings", {
-  tenantId: tenantRefPrimaryKey(),
-  strategy: poolRoutingStrategy("strategy").notNull().default("least_loaded"),
-  /** Requests served by one pool before round robin advances. */
-  rotateCount: integer("rotate_count").notNull().default(1),
-});
-
-
-export const API_KEY_MODES = ["personal", "share"] as const;
-export type ApiKeyMode = (typeof API_KEY_MODES)[number];
-
-/**
- * How a key's `modelList` is interpreted. Exactly two modes, never a
- * combination: `whitelist` allows only the listed names (an empty list allows
- * everything), `blacklist` refuses the listed names (an empty list refuses
- * nothing). Every model-authorization surface — the request path, admission,
- * `/v1/models`, and the share page — reads this one pair.
- */
-export const API_KEY_MODEL_ACCESS_MODES = ["whitelist", "blacklist"] as const;
-export type ApiKeyModelAccessMode = (typeof API_KEY_MODEL_ACCESS_MODES)[number];
-
-/**
- * Inbound `/v1/*` keys never store plaintext secrets. Personal keys carry an
- * authentication hash; share templates carry policy and child keys point back
- * to the template that issued them.
- */
-export const apiKeys = pgTable(
-  "api_keys",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    tenantId: tenantRefRequired(),
-    keyHash: text("key_hash"),
-    keyMode: text("key_mode").$type<ApiKeyMode>().notNull().default("personal"),
-    parentKeyId: uuid("parent_key_id").references(
-      (): AnyPgColumn => apiKeys.id,
-      { onDelete: "cascade" },
-    ),
-    issuedClientIp: text("issued_client_ip"),
-    issuedClientIpKey: text("issued_client_ip_key"),
-    label: text("label").notNull(),
-    scopes: jsonb("scopes").notNull().$type<readonly string[]>(),
-    // Only the non-secret public key prefix is returned to the owner.
-    keyPrefix: text("key_prefix"),
-    // Personal keys retain an encrypted copy for the CLI handoff. Shared
-    // children are revealed once and persist only their authentication hash.
-    keyEncrypted: bytea("key_encrypted"),
-    notesTitle: text("notes_title"),
-    notesSubtitle: text("notes_subtitle"),
-    notesBody: text("notes_body"),
-    /** Explicit list position. */
-    sortIndex: integer("sort_index").notNull().default(0),
-    requestsPerMinute: integer("requests_per_minute"),
-    dailyTokenLimit: bigint("daily_token_limit", { mode: "number" }),
-    monthlyTokenLimit: bigint("monthly_token_limit", { mode: "number" }),
-    lifetimeTokenBudget: bigint("lifetime_token_budget", { mode: "number" }),
-    lifetimeTokensConsumed: bigint("lifetime_tokens_consumed", { mode: "number" })
-      .notNull()
-      .default(0),
-    maxConcurrentRequests: integer("max_concurrent_requests"),
-    /**
-     * `whitelist` = only the names in `modelList` may be used (empty = all);
-     * `blacklist` = the names in `modelList` are refused (empty = none). One
-     * list, one mode — the old separate allow/deny columns are retired.
-     */
-    modelAccessMode: text("model_access_mode")
-      .$type<ApiKeyModelAccessMode>()
-      .notNull()
-      .default("whitelist"),
-    modelList: jsonb("model_list").$type<readonly string[]>(),
-    /**
-     * Client-router ids this key refuses. A request whose fingerprint names one
-     * of these is rejected before routing, so an operator can stop a downstream
-     * router from spending the key. Empty/NULL means "no router is refused".
-     *
-     * A label, not a security control: a client that sends no fingerprint is
-     * simply not matched, which is why this is a per-key policy rather than a
-     * boundary anything else in the system trusts.
-     */
-    clientRouterDenylist: jsonb("client_router_denylist").$type<readonly string[]>(),
-    modelPrefix: text("model_prefix"),
-    createdAt: createdAtColumn(),
-    enabled: boolean("enabled").notNull().default(true),
-    revokedAt: timestamp("revoked_at", { withTimezone: true }),
-    /**
-     * The tenant's mandatory gateway key. Seeded at first boot whether or not
-     * `CARTETHYIA_API_KEY` was configured, and the one key the console refuses
-     * to revoke — it may only be rotated. A column rather than a label match,
-     * so renaming the key cannot break the CLI lookup or let a second
-     * default be created.
-     */
-    isDefault: boolean("is_default").notNull().default(false),
-  },
-  (table) => [
-    uniqueIndex("api_keys_key_hash_idx").on(table.keyHash),
-    index("api_keys_tenant_id_idx").on(table.tenantId),
-    index("api_keys_enabled_tenant_idx").on(table.tenantId).where(sql`${table.enabled} = true`),
-    index("api_keys_parent_key_id_idx").on(table.parentKeyId),
-    check(
-      "api_keys_mode_shape_check",
-      sql`(
-        (${table.keyMode} = 'personal' AND ${table.keyHash} IS NOT NULL
-          AND ${table.parentKeyId} IS NULL AND ${table.issuedClientIp} IS NULL
-          AND ${table.issuedClientIpKey} IS NULL)
-        OR
-        (${table.keyMode} = 'share' AND (
-          (${table.parentKeyId} IS NULL AND ${table.keyHash} IS NULL
-            AND ${table.keyEncrypted} IS NULL
-            AND ${table.issuedClientIp} IS NULL AND ${table.issuedClientIpKey} IS NULL)
-          OR
-          (${table.parentKeyId} IS NOT NULL AND ${table.keyHash} IS NOT NULL
-            AND ${table.keyEncrypted} IS NULL
-            AND ${table.issuedClientIp} IS NOT NULL AND ${table.issuedClientIpKey} IS NOT NULL)
-        ))
-      )`,
-    ),
-  ],
-);
-
-/**
- * The optional popup a share link shows its recipient. Kept off `api_keys`
- * because the auth path reads that row on every request and the popup is
- * absent for nearly all of them. One row per key, present only when the owner
- * configured one.
- */
-export const sharePopups = pgTable(
-  "share_popups",
-  {
-    apiKeyId: uuid("api_key_id")
-      .primaryKey()
-      .references(() => apiKeys.id, { onDelete: "cascade" }),
-    enabled: boolean("enabled").notNull().default(false),
-    image: bytea("image"),
-    imageMime: text("image_mime"),
-    title: text("title"),
-    body: text("body"),
-  },
-  (table) => [
-    check(
-      "share_popups_image_mime_check",
-      sql`${table.imageMime} IS NULL OR ${table.imageMime} IN ('image/png', 'image/jpeg', 'image/webp', 'image/gif')`,
-    ),
-    check(
-      "share_popups_image_size_check",
-      sql`${table.image} IS NULL OR octet_length(${table.image}) <= 2097152`,
-      // SQLite has no octet_length. The generator rewrites this check to
-      // `length(`; both count bytes for a blob column.
-    ),
-  ],
-);
-
-
-// Public bearer links for enrolling one shared key per resolved client IP, and
-// for handing a personal key to its owner. The hash is the lookup key; the
-// token is additionally retained encrypted so the owner's console can show the
-// link again instead of losing it after the one response that minted it.
-/**
- * Share-link kinds, as a runtime tuple. Route schemas and DTOs project this
- * list rather than restating it.
- */
-export const SHARE_LINK_KINDS = ["enroll", "handoff"] as const;
-
-/** One share-link kind. */
-export type ShareLinkKind = (typeof SHARE_LINK_KINDS)[number];
-
-
-export const shareLinks = pgTable(
-  "share_links",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    apiKeyId: uuid("api_key_id")
-      .notNull()
-      .references(() => apiKeys.id, { onDelete: "cascade" }),
-    tokenHash: text("token_hash").notNull(),
-    // Encrypted bearer token. Present so the console can re-display a stable
-    // link; historical rows written before this column carry NULL and simply
-    // cannot be shown again.
-    tokenEncrypted: bytea("token_encrypted"),
-    kind: text("kind").$type<ShareLinkKind>().notNull().default("enroll"),
-    active: boolean("active").notNull().default(true),
-    createdAt: createdAtColumn(),
-    expiresAt: timestamp("expires_at", { withTimezone: true }),
-    lastViewedAt: timestamp("last_viewed_at", { withTimezone: true }),
-  },
-  (table) => [
-    uniqueIndex("share_links_token_hash_idx").on(table.tokenHash),
-    index("share_links_api_key_idx").on(table.apiKeyId),
-    index("share_links_active_kind_expires_idx").on(table.active, table.kind, table.expiresAt),
-    check("share_links_kind_check", sql`${table.kind} IN ('enroll', 'handoff')`),
-  ],
-);
-
-
-/**
- * Console dashboard users — distinct from API-key tenants.
- */
-export const consoleUsers = pgTable(
-  "console_users",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    tenantId: tenantRefRequired(),
-    username: text("username").notNull(),
-    email: text("email"),
-    passwordHash: text("password_hash").notNull(),
-    displayName: text("display_name"),
-    isActive: boolean("is_active").notNull().default(true),
-    isPlatformAdmin: boolean("is_platform_admin").notNull().default(false),
-    isFirstBoot: boolean("is_first_boot").notNull().default(true),
-    ...timestampColumns(),
-  },
-  (table) => [
-    index("console_users_tenant_id_idx").on(table.tenantId),
-    uniqueIndex("console_users_tenant_username_uidx").on(table.tenantId, table.username),
-  ],
-);
-
-export type ConsoleUser = typeof consoleUsers.$inferSelect;
-
-/**
- * Console sessions — HTTP-only, signed cookies.
- * Session metadata only; no secrets in JSON.
- */
-export const consoleSessions = pgTable(
-  "console_sessions",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    userId: uuid("user_id")
-      .notNull()
-      .references(() => consoleUsers.id, { onDelete: "cascade" }),
-    sessionToken: text("session_token").notNull().unique(),
-    ipAddress: text("ip_address"),
-    userAgent: text("user_agent"),
-    expiresAt: timestamp("expires_at", { withTimezone: true }).notNull(),
-    createdAt: createdAtColumn(),
-  },
-  (table) => [
-    index("console_sessions_user_id_idx").on(table.userId),
-  ],
-);
-
-export type ConsoleSession = typeof consoleSessions.$inferSelect;
-
-/**
- * Console login lockout tracking — fail-closed rate limiting keyed by
- * client IP (login attempts happen before any user identity is known, so
- * this cannot be keyed by `console_users.id`). Persisted so a ban survives
- * a process restart and is shared across every app instance, instead of a
- * static in-process map that a second instance or a redeploy would forget.
- */
-export const consoleLockouts = pgTable(
-  "console_lockouts",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    ip: text("ip").notNull(),
-    failureCount: integer("failure_count").notNull().default(0),
-    windowUntil: timestamp("window_until", { withTimezone: true }),
-    lockedUntil: timestamp("locked_until", { withTimezone: true }),
-    reason: text("reason"),
-    ...timestampColumns(),
-  },
-  (table) => [
-    uniqueIndex("idx_console_lockouts_ip").on(table.ip),
-  ],
-);
-
-
-// Configuration-change audit has its own retention lifecycle; it is separate
-// from metadata telemetry and short-lived optional payload capture.
-export const adminAuditLog = pgTable(
-  "admin_audit_log",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    createdAt: createdAtColumn(),
-    actor: text("actor").notNull(),
-    tenantId: tenantCascadeSetNull(),
-    action: text("action").notNull(),
-    target: text("target").notNull(),
-    detail: jsonb("detail"),
-  },
-  (table) => [
-    index("admin_audit_log_tenant_id_created_at_id_idx").on(
-      table.tenantId,
-      table.createdAt,
-      table.id,
-    ),
-  ],
-);
-
-
-// Runtime preferences are stored in one validated JSONB bag. The filter
-// master toggle remains a typed column because it is read on the proxy path.
-export interface ConsoleSettingsPreferences {
-  tenantConcurrencyLimit?: number | null;
-  /** Normalizes provider-native thinking config for non-user final turns. */
-  thinkingNormalizationEnabled?: boolean;
-  responsesReasoningSummary?: "auto" | "concise" | "detailed";
-  /**
-   * Request compression: compact bulky tool-result text in the message history
-   * before dispatch (RTK prune). Default off.
-   */
-  rtkPruneEnabled?: boolean;
-  /** RTK prune strength when enabled (`lite`/`full`/`ultra`); `full` when unset. */
-  rtkPruneLevel?: "lite" | "full" | "ultra";
-  /**
-   * Request compression: append a minimal-code directive to the system content.
-   * Default off; `ponyTailLevel` selects the intensity when on.
-   */
-  ponyTailEnabled?: boolean;
-  /** PonyTail directive intensity; `full` when unset. Only applied when enabled. */
-  ponyTailLevel?: "lite" | "full" | "ultra" | null;
-  /**
-   * Payload capture mode (default `none` when unset):
-   * - `metadata` — Proxy→Provider method + allowlisted headers only
-   * - `full` — full redacted bodies up to the configured capture limit
-   * - `none` — no drawer capture (request metadata events still retained)
-   */
-  telemetryPayloads?: "full" | "metadata" | "none";
-  /**
-   * Capture depth for `full` mode (default `minimum` when unset). Fixed size
-   * tiers so every debug sees all four drawer panels — only the cap changes:
-   * - `minimum` — 1 MiB combined cap, light enough to leave on
-   * - `moderate` — 8 MiB combined cap, expect CPU/memory spikes while active
-   * - `maximum` — 16 MiB combined cap, expect CPU/memory spikes while active
-   */
-  telemetryPayloadDepth?: "minimum" | "moderate" | "maximum";
-  privacyMode?: "masked" | "full";
-}
-
-export const consoleSettings = pgTable("console_settings", {
-  tenantId: tenantRefPrimaryKey(),
-  preferences: jsonb("preferences").notNull().default({}).$type<ConsoleSettingsPreferences>(),
-  updatedAt: updatedAtColumn(),
-});
-
-
-// CLI Tools (CLI tool configuration): per-tenant mapping table that lets a coding-agent
-// slot (e.g. Claude Code's "opus" slot) target a specific Cartethyia model. Rows
-// are opaque strings — validation happens at the API boundary against the
-// tool registry, not in Postgres.
-export const cliToolMappings = pgTable(
-  "cli_tool_mappings",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    tenantId: tenantRefRequired(),
-    apiKeyId: uuid("api_key_id")
-      .notNull()
-      .references((): AnyPgColumn => apiKeys.id, { onDelete: "cascade" }),
-    toolId: text("tool_id").notNull(),
-    slotKey: text("slot_key").notNull(),
-    sourceModel: text("source_model").notNull(),
-    targetModel: text("target_model").notNull(),
-    enabled: boolean("enabled").notNull().default(true),
-    updatedAt: updatedAtColumn(),
-  },
-  (table) => [
-    uniqueIndex("cli_tool_mappings_key").on(
-      table.tenantId,
-      table.toolId,
-      table.apiKeyId,
-      table.slotKey,
-    ),
-  ],
-);
-
-
-// Per-(tenant, tool, api_key) settings: whether mappings are active and which
-// mode the tool is configured for. Primary key is the composite of all three.
-export const cliToolSettings = pgTable(
-  "cli_tool_settings",
-  {
-    tenantId: tenantRefRequired(),
-    apiKeyId: uuid("api_key_id")
-      .notNull()
-      .references((): AnyPgColumn => apiKeys.id, { onDelete: "cascade" }),
-    toolId: text("tool_id").notNull(),
-    mappingsEnabled: boolean("mappings_enabled").notNull().default(false),
-    mode: text("mode").notNull().default("remote"),
-    updatedAt: updatedAtColumn(),
-  },
-  (table) => [primaryKey({ columns: [table.tenantId, table.toolId, table.apiKeyId] })],
-);
-
-
-// Metadata-only request telemetry. Retention is configured independently of
-// payload capture (`CARTETHYIA_TELEMETRY_RETENTION_DAYS`, default 30 days).
-// No prompt text, raw body, API key, encrypted-reasoning payload, or generic
-// `payload` column — the schema structurally prevents those fields.
-// Correlation columns intentionally have no catalog foreign keys, so a
-// telemetry write never depends on provider-row lifetime.
-export const telemetryEvents = pgTable(
-  "telemetry_events",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    createdAt: createdAtColumn(),
-    tenantId: tenantRefRequired(),
-    requestId: uuid("request_id").notNull(),
-    sourceSurface: telemetrySourceSurface("source_surface"),
-    requestedModel: text("requested_model"),
-    /**
-     * The `model` the upstream put in its own response, verbatim.
-     *
-     * Separate from `requested_model` because they answer different questions:
-     * one is what the client asked for, the other is what the provider said it
-     * ran. A provider that answers `muse-spark` as `deepseek-v4.1-flash` is
-     * telling the operator something, and overwriting one with the other would
-     * erase it. NULL when the response carried no model.
-     */
-    reportedModel: text("reported_model"),
-    /** Canonical reasoning effort the client asked for; NULL when it asked for none. */
-    requestedEffort: text("requested_effort"),
-    providerId: text("provider_id"),
-    accountId: uuid("account_id"),
-    networkPoolId: uuid("network_pool_id"),
-    endpoint: text("endpoint"),
-    apiKeyId: uuid("api_key_id"),
-    userAgent: text("user_agent"),
-    clientIp: text("client_ip"),
-    latencyMs: integer("latency_ms"),
-    ttfbMs: integer("ttfb_ms"),
-    stream: boolean("stream"),
-    status: telemetryStatus("status"),
-    /**
-     * HTTP status the client actually received. The `status` enum is the
-     * internal terminal state (`completed`/`failed`/…), which collapses every
-     * failure to one bucket; this column keeps the wire code, so a 503
-     * admission rejection stays distinguishable from a 500 upstream failure
-     * in the Usage status breakdown. NULL on rows written before the column.
-     */
-    httpStatus: integer("http_status"),
-    /**
-     * Which layer produced the failure: the gateway itself, the upstream
-     * provider, or the network in between. Recorded beside `error_category`
-     * because the category alone cannot separate them — `invalid_request` is
-     * both "the caller sent a bad body" (ours) and "the upstream rejected the
-     * body" (theirs), and an operator triaging a failure needs to know which.
-     */
-    errorOrigin: text("error_origin"),
-    errorCategory: text("error_category"),
-    inputTokens: integer("input_tokens"),
-    cachedInputTokens: integer("cached_input_tokens"),
-    outputTokens: integer("output_tokens"),
-    reasoningTokens: integer("reasoning_tokens"),
-    estimatedCostUsd: numeric("estimated_cost_usd", { precision: 12, scale: 6 }),
-    /** Provider-billed credits (Tencent buddy-meter `credit`); NULL when unreported. */
-    creditUsed: numeric("credit_used", { precision: 12, scale: 4 }),
-    /**
-     * Gateway-side time before the upstream dispatch began: routing,
-     * credential lease, and request preparation. `ttfbMs` is measured from
-     * the dispatch start, so this is the overhead the client waited on before
-     * the provider was even called.
-     */
-    resolveMs: integer("resolve_ms"),
-    tokensPerSec: numeric("tokens_per_sec", { precision: 10, scale: 2 }),
-    // Absolute epoch-millisecond event timestamps (~1.8e12) overflow integer
-    // (int4): completed requests carrying observed stream timings failed
-    // their telemetry insert and vanished from the console. bigint "number"
-    // mode is safe here — epoch millis stay far below MAX_SAFE_INTEGER —
-    // and keeps the columns JSON-safe like the token-budget bigints above.
-    firstContentDeltaAtMs: bigint("first_content_delta_at_ms", { mode: "number" }),
-    lastEventAtMs: bigint("last_event_at_ms", { mode: "number" }),
-    /**
-     * Why this request's drawer payloads were not captured, or `captured` when
-     * they were. Remote-diagnosis column: the drawer reads empty for reasons
-     * that are invisible from the client — the tenant's capture mode, a missing
-     * tenant, a non-terminal attempt — and without this an operator could not
-     * tell "not configured" from "broken".
-     */
-    payloadCaptureState: text("payload_capture_state"),
-  },
-  (table) => [
-    index("telemetry_events_created_at_idx").on(table.createdAt),
-    index("telemetry_events_tenant_created_idx").on(table.tenantId, table.createdAt),
-    index("telemetry_events_request_id_idx").on(table.requestId),
-    index("telemetry_events_tenant_account_created_idx").on(
-      table.tenantId,
-      table.accountId,
-      table.createdAt,
-    ),
-    /**
-     * Serves the public share page's per-key aggregate (`share-usage.ts`):
-     * equality on `api_key_id`, range and aggregate on `created_at`. Without
-     * it every share render scanned the largest table in the schema.
-     */
-    index("telemetry_events_api_key_created_idx").on(table.apiKeyId, table.createdAt),
-  ],
-);
-
-/** Durable aggregates retained after metadata telemetry expires. */
-export const TELEMETRY_USAGE_IDENTITY_TYPES = ["account", "api_key"] as const;
-export type TelemetryUsageIdentityType = (typeof TELEMETRY_USAGE_IDENTITY_TYPES)[number];
-
-export const telemetryUsageTotals = pgTable(
-  "telemetry_usage_totals",
-  {
-    tenantId: tenantRefRequired(),
-    identityType: text("identity_type").$type<TelemetryUsageIdentityType>().notNull(),
-    entityId: uuid("entity_id").notNull(),
-    requests: bigint("requests", { mode: "number" }).notNull().default(0),
-    errors: bigint("errors", { mode: "number" }).notNull().default(0),
-    inputTokens: bigint("input_tokens", { mode: "number" }).notNull().default(0),
-    outputTokens: bigint("output_tokens", { mode: "number" }).notNull().default(0),
-    lastUsedAt: timestamp("last_used_at", { withTimezone: true }).notNull(),
-    updatedAt: updatedAtColumn(),
-  },
-  (table) => [
-    primaryKey({
-      name: "telemetry_usage_totals_identity_pk",
-      columns: [table.tenantId, table.identityType, table.entityId],
-    }),
-    check(
-      "telemetry_usage_totals_identity_type_check",
-      sql`${table.identityType} IN ('account', 'api_key')`,
-    ),
-  ],
-);
-
-
-export const telemetryPayloads = pgTable(
-  "telemetry_payloads",
-  {
-    id: uuid("id").primaryKey().defaultRandom(),
-    tenantId: tenantRefRequired(),
-    requestId: uuid("request_id"),
-    capturedAt: timestamp("captured_at", { withTimezone: true }).notNull().defaultNow(),
-    expiresAt: timestamp("expires_at", { withTimezone: true }).notNull(),
-    /**
-     * Typed index into the on-disk frame store. Prompt/response bodies live
-     * only in the frame file these columns name — never in Postgres. The
-     * previous `request_body jsonb` column is gone so a body cannot be
-     * written here by accident.
-     */
-    storage: text("storage").notNull(),
-    file: text("file").notNull(),
-    offset: integer("offset").notNull(),
-    length: integer("length").notNull(),
-    checksum: text("checksum").notNull(),
-    version: integer("version").notNull(),
-  },
-  (table) => [
-    index("telemetry_payloads_request_id_idx").on(table.requestId),
-    index("telemetry_payloads_tenant_request_idx").on(table.tenantId, table.requestId),
-    index("telemetry_payloads_expires_idx").on(table.expiresAt),
-    check("telemetry_payloads_storage_check", sql`${table.storage} IN ('jsonb-file', 'jsonl-file')`),
-    check("telemetry_payloads_version_check", sql`${table.version} = 1`),
-    check("telemetry_payloads_offset_check", sql`${table.offset} >= 0`),
-    check("telemetry_payloads_length_check", sql`${table.length} > 0`),
-  ],
-);
-
-
-
-// Daily check-in ledger: one row per (account, calendar day) once the sweep has
-// attempted that account's claim.
-//
-// The day-scoped composite PK is the claim guard — an insert that conflicts is
-// exactly the "already settled today" answer — so it replaces the Redis
-// `SET NX EX` the Lite backend never had. The console reads the same table to
-// answer "was this attempted today", which is what makes the Quota page show
-// the real state instead of a permanent "Check-in due" when Redis is absent.
-export const dailyCheckinLedger = pgTable(
-  "daily_checkin_ledger",
-  {
-    dayKey: text("day_key").notNull(),
-    accountId: uuid("account_id").notNull(),
-    claimedAt: timestamp("claimed_at", { withTimezone: true }).notNull().defaultNow(),
-  },
-  (table) => [
-    primaryKey({ columns: [table.dayKey, table.accountId] }),
-  ],
-);
-
-
-/**
- * Per-day, per-model credit deductions the buddy billing meter reported.
- *
- * A snapshot, not a ledger: the upstream is the source of truth and answers the
- * whole window on every call, so a row is upserted rather than appended. What it
- * buys is that the console renders a *stored* figure — without it every page
- * open would re-hit the billing endpoint for every account, and a failed read
- * would show nothing at all instead of the last number obtained.
- *
- * Keyed (account, day, model): one row per bucket, and the conflict target is
- * what makes the write idempotent across passes.
- */
-export const buddyUsageSnapshots = pgTable(
-  "buddy_usage_snapshots",
-  {
-    accountId: uuid("account_id")
-      .notNull()
-      .references(() => providerAccounts.id, { onDelete: "cascade" }),
-    /** Calendar day the requests were metered on, `YYYY-MM-DD`. */
-    dayKey: text("day_key").notNull(),
-    model: text("model").notNull(),
-    /** Credits deducted across every request in this bucket. */
-    credit: numeric("credit", { precision: 14, scale: 4 }).notNull(),
-    syncedAt: timestamp("synced_at", { withTimezone: true }).notNull().defaultNow(),
-  },
-  (table) => [
-    primaryKey({ columns: [table.accountId, table.dayKey, table.model] }),
-    index("buddy_usage_snapshots_day_idx").on(table.dayKey),
-  ],
-);
-
-/**
- * One row per (day, account, growth task) once that day's claim was attempted.
- *
- * The composite PK is the idempotency guard: an insert that conflicts is
- * exactly the "already settled today" answer, which is what stops a sweep from
- * re-POSTing a claim on every pass. Same shape as `daily_checkin_ledger`
- * because it is the same problem — one claim per account per day — widened by
- * the task name now that there is more than one thing to claim.
- */
-export const buddyTaskLedger = pgTable(
-  "buddy_task_ledger",
-  {
-    dayKey: text("day_key").notNull(),
-    accountId: uuid("account_id")
-      .notNull()
-      .references(() => providerAccounts.id, { onDelete: "cascade" }),
-    task: text("task").notNull(),
-    claimedAt: timestamp("claimed_at", { withTimezone: true }).notNull().defaultNow(),
-  },
-  (table) => [
-    primaryKey({ columns: [table.dayKey, table.accountId, table.task] }),
-    index("buddy_task_ledger_account_idx").on(table.accountId),
-  ],
-);
+/** Canonical Drizzle schema: every persistence table, enum, index, and shared type. */
+import { desc, sql } from "drizzle-orm";
+import {
+  bigint, boolean, check, customType, index, integer, jsonb, numeric, pgEnum, pgTable,
+  primaryKey, text, timestamp, uniqueIndex, uuid, type AnyPgColumn,
+} from "drizzle-orm/pg-core";
+
+export const TENANT_STATUSES = ["active", "suspended", "deleted"] as const;
+export type TenantStatus = (typeof TENANT_STATUSES)[number];
+
+export const tenants = pgTable(
+  "tenants",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    name: text("name").notNull(),
+    status: text("status").$type<TenantStatus>().notNull(),
+    // Intentionally inline (not createdAtColumn()): this root table is imported
+    // by the helpers below; sharing the helper here would make a use-before-define.
+    createdAt: timestamp("created_at", { withTimezone: true }).notNull().defaultNow(),
+    updatedAt: timestamp("updated_at", { withTimezone: true }).notNull().defaultNow(),
+  },
+  (table) => [
+    check("tenants_status_check", sql`${table.status} IN ('active', 'suspended', 'deleted')`),
+  ],
+);
+
+// Canonical tenant/timestamp column definitions shared by every table.
+
+export function tenantRefNullable() {
+  return uuid("tenant_id").references(() => tenants.id, { onDelete: "cascade" });
+}
+
+export function tenantRefRequired() {
+  return uuid("tenant_id")
+    .notNull()
+    .references(() => tenants.id, { onDelete: "cascade" });
+}
+
+export function tenantRefPrimaryKey() {
+  return uuid("tenant_id")
+    .primaryKey()
+    .references(() => tenants.id, { onDelete: "cascade" });
+}
+
+export function tenantCascadeSetNull() {
+  return uuid("tenant_id").references(() => tenants.id, { onDelete: "set null" });
+}
+
+export function createdAtColumn() {
+  return timestamp("created_at", { withTimezone: true }).notNull().defaultNow();
+}
+
+export function updatedAtColumn() {
+  return timestamp("updated_at", { withTimezone: true }).notNull().defaultNow();
+}
+
+export function timestampColumns() {
+  return { createdAt: createdAtColumn(), updatedAt: updatedAtColumn() };
+}
+
+/**
+ * `bytea` column for encrypted credential ciphertext (application-layer
+ * envelope encryption: the ciphertext stored here is opaque to the
+ * database). Drizzle's pg-core does not ship a first-class `bytea`
+ * builder, so this is the documented `customType` escape hatch (the same
+ * DDL-shape escape-hatch principle applied to column-type gaps in the
+ * ORM's builder).
+ */
+export const bytea = customType<{ data: Buffer }>({
+  dataType() {
+    return "bytea";
+  },
+});
+
+// Shared three-state health machine for provider_accounts and network_pools —
+// the account state machine is reused for network pools rather than
+// inventing a second one.
+//
+// `degraded` was retired. It sat between "working" and "parked" and every
+// consumer had to decide what it meant: routing treated it as a hard exclusion,
+// the sweep treated it as recoverable, and the console showed it as a third
+// badge — three readings of one value, and the operator reported the result as
+// ambiguous. The two facts it mixed are now each stated by a value that means
+// exactly one thing: a fault that clears on its own is `cooldown` (with a
+// deadline the sweep honors), and a fault that needs an operator is `disabled`.
+export const healthStatus = pgEnum("health_status", ["active", "cooldown", "disabled"]);
+
+/**
+ * The health state machine shared by `provider_accounts` and `network_pools`.
+ * One definition, because the two tables are driven by the same transitions and
+ * a column added to one of them by hand is how they drift.
+ */
+export function healthColumns() {
+  return {
+    status: healthStatus("status").notNull().default("active"),
+    consecutiveFailures: integer("consecutive_failures").notNull().default(0),
+    lastSuccessAt: timestamp("last_success_at", { withTimezone: true }),
+    lastError: text("last_error"),
+    lastErrorCategory: text("last_error_category"),
+    lastErrorAt: timestamp("last_error_at", { withTimezone: true }),
+    cooldownUntil: timestamp("cooldown_until", { withTimezone: true }),
+    lastRecoveredAt: timestamp("last_recovered_at", { withTimezone: true }),
+  };
+}
+
+export const credentialKind = pgEnum("credential_kind", ["api_key", "oauth", "none"]);
+
+// Shared wire-family domain used by providers.wire_family_default and
+// models.wire_family.
+//
+// `gemini` is Factory's native generateContent wire. Added by forward
+// migration `0002_add_gemini_wire_family.sql`; the enum here must stay in
+// step with what that migration applies, or a fresh database (built from the
+// baseline) and a migrated one would disagree.
+export const wireFamily = pgEnum("wire_family", ["chat", "responses", "messages", "gemini"]);
+
+// `https` is a hosted relay front door (Vercel, Cloudflare, Deno, Netlify):
+// plain HTTP CONNECT to an origin that terminates TLS itself, so the transport
+// is recorded as what it speaks rather than as a generic proxy.
+export const networkPoolKind = pgEnum("network_pool_kind", ["http", "https", "socks5"]);
+
+//  health_events.entity_kind — distinguishes which owning
+// table entity_id/account_id/network_pool_id references.
+export const healthEntityKind = pgEnum("health_entity_kind", ["account", "pool"]);
+
+export const telemetrySourceSurface = pgEnum("telemetry_source_surface", [
+  "chat",
+  "responses",
+  "messages",
+  "completion",
+]);
+
+export const telemetryStatus = pgEnum("telemetry_status", [
+  "completed",
+  "failed",
+  "cancelled",
+  "truncated",
+]);
+
+// The id matches the seven canonical builtin IDs (openai, anthropic, claude,
+// codex, opencode-zen, ollamacloud, cerebras) or a custom/BYOK slug
+// rejected at the identity boundary when it collides case-insensitively with
+// a reserved ID — that rejection is application-layer, not a DB constraint
+// here. `capability_profile` is a JSONB routing-capability map with no index:
+// it is read with the provider row the snapshot builder already loads, so a
+// separate index would only add write cost. `tenant_id`
+// null means the provider is global/built-in, matching the same convention
+// as `provider_accounts.tenant_id`; populated means tenant-owned/BYOK.
+// `base_url`/`compatibility_profile` are BYOK-only: `compatibility_profile`
+// carries HTTP dispatch config (extra headers/query params/streaming mode)
+// and is intentionally a separate column from `capability_profile`, which
+// drives routing capability flags — conflating the two would corrupt routing
+// data with unrelated HTTP config.
+export const providers = pgTable(
+  "providers",
+  {
+    id: text("id").primaryKey(),
+    tenantId: tenantRefNullable(),
+    wireFamilyDefault: wireFamily("wire_family_default"),
+    capabilityProfile: jsonb("capability_profile"),
+    baseUrl: text("base_url"),
+    compatibilityProfile: jsonb("compatibility_profile"),
+    enabled: boolean("enabled").notNull().default(true),
+    // Builtin default `true` (most providers need a real credential). Only a
+    // genuinely public/unauthenticated builtin (OpenCode Free) is seeded
+    // `false`. When `false`, the snapshot builder emits a routable candidate
+    // with zero `provider_accounts` rows instead of marking it `disabled` —
+    // see `src/transport/routing/route-catalog.ts`.
+    requiresAccount: boolean("requires_account").notNull().default(true),
+  },
+  (table) => [
+    // `globalOrOwnedBy(providers.tenantId, …)` is the predicate behind
+    // `listModelsForTenant` and `/quota/overview`, and it is a bare
+    // `tenant_id IS NULL OR tenant_id = ?` — which no other index here can
+    // serve, so both reads scanned the whole provider list. Cheap to keep:
+    // providers are few and written rarely.
+    index("providers_tenant_id_idx").on(table.tenantId),
+  ],
+);
+
+// Upstream account credentials plus the full health state machine.
+// `tenant_id` null means the account is shared pool-wide; populated means
+// tenant-owned/BYOK. The concurrency ceiling lives in provider routing
+// (`provider_routing_settings.max_inflight`) and network pools
+// (`network_pools.max_inflight`); the per-account override no longer exists.
+export const providerAccounts = pgTable("provider_accounts", {
+  id: uuid("id").primaryKey().defaultRandom(),
+  providerId: text("provider_id")
+    .notNull()
+    .references(() => providers.id, { onDelete: "cascade" }),
+  tenantId: tenantRefNullable(),
+  label: text("label").notNull(),
+  credentialCiphertext: bytea("credential_ciphertext"),
+  /**
+   * HMAC fingerprint of the credential, never the plaintext secret. Nullable
+   * for credential-less accounts and legacy rows without an identity.
+   */
+  credentialFingerprint: text("credential_fingerprint"),
+  credentialKind: credentialKind("credential_kind").notNull(),
+  /**
+   * Per-account upstream auth configuration that is not a secret: auth method,
+   * region, profile ARN, OAuth client id, token endpoint, scopes. The adapter
+   * reads it where the credential is resolved (it selects the regional endpoint
+   * and the profile the upstream binds the request to) and token refresh reads
+   * it again to pick the right endpoint and client. Secrets never live here —
+   * see `credential_ciphertext` and `provider_oauth_states`.
+   */
+  authState: jsonb("auth_state"),
+  createdAt: timestamp("created_at", { withTimezone: true }).notNull().defaultNow(),
+
+  ...healthColumns(),
+  modelCooldowns: jsonb("model_cooldowns").notNull().default({}),
+  /**
+   * How many times each model has been reported unsupported for this account:
+   * `{ [modelId]: hits }`. Written only by the unsupported-model path, and it
+   * is the escalation level for that pair's cooldown.
+   *
+   * Deliberately a separate column from `model_cooldowns`: that value is
+   * `{ [modelId]: deadline }`, read by routing (`route-catalog.ts`) and by the
+   * console, and widening it to `{until, hits}` would have changed a shape two
+   * independent readers parse. Nullable — an account that never hit an
+   * unsupported model has nothing to count.
+   */
+  modelBlockHits: jsonb("model_block_hits").$type<Readonly<Record<string, number>>>(),
+  /**
+   * Stable list position within (tenant, provider). Accounts were ordered by
+   * `created_at`, which let two same-millisecond rows swap between loads; this
+   * is explicit and can be rewritten when the operator reorders or appends.
+   */
+  sortIndex: integer("sort_index").notNull().default(0),
+  /**
+   * The credential is a static bearer token (a JWT/access token pasted on its
+   * own, or an account an operator has pinned): it is used exactly as issued and
+   * must never be sent to a refresh endpoint. Distinct from `credentialKind`,
+   * which says what family the credential belongs to, not whether it can be
+   * re-minted — an `oauth` account can carry a static token when its refresh
+   * grant is absent or dead while the token itself is still valid. A static
+   * account is skipped by the refresh sweep and is never disabled by a refresh
+   * failure, because there is no refresh to run.
+   */
+  staticToken: boolean("static_token").notNull().default(false),
+  /**
+   * When this account's credential was last read for its own identity (email,
+   * display name). Null means never read.
+   *
+   * This is the convergence stamp for the enrichment sweep, and the reason an
+   * import can be instant: the batch path persists the credential without
+   * parsing it, and the sweep fills the identity in afterwards. An account is
+   * only listed while this is null, so a completed pass never re-decrypts a
+   * credential it already read.
+   */
+  identityCheckedAt: timestamp("identity_checked_at", { withTimezone: true }),
+  /**
+   * Email read out of the account's own JWT claims. Cosmetic: it names the
+   * account in the console and is never used to authenticate or route.
+   */
+  accountEmail: text("account_email"),
+
+  /**
+   * The plan the provider reports for this account, last seen by a quota read.
+   *
+   * Persisted rather than read from the quota cache: that cache is a short
+   * in-process TTL and empty on a cold start, so the plan would be unknown
+   * exactly when a freshly booted process makes its first routing decisions.
+   *
+   * Nullable: most providers report no plan at all, and an account never read
+   * must say so rather than have a plan invented for it. Routing reads absence
+   * as "unknown", never as "free" — guessing free would send pro-only traffic
+   * to an account that may well be refused.
+   */
+  quotaPlan: text("quota_plan"),
+
+  },
+  (table) => [
+    uniqueIndex("provider_accounts_identity_uidx").on(
+      table.providerId,
+      // No `::uuid` cast: the migrations create this index without one on
+      // SQLite (baseline l.440) and with one on Postgres (l.423), and the
+      // Drizzle definition is what a test's in-memory database is built
+      // from — a Postgres-only cast there made every account insert fail
+      // with `unrecognized token: ":"`.
+      sql`coalesce(${table.tenantId}, '00000000-0000-0000-0000-000000000000')`,
+      table.credentialFingerprint,
+    ),
+    index("provider_accounts_status_cooldown_idx").on(table.status, table.cooldownUntil),
+    index("provider_accounts_provider_tenant_idx").on(table.providerId, table.tenantId),
+    // The bare `tenant_id IS NULL OR tenant_id = ?` that `listAllAccounts`
+    // uses is served by neither index above: `(status, cooldown_until)` leads
+    // with the wrong column, and `(provider_id, tenant_id)` cannot serve a
+    // predicate that does not constrain `provider_id`. This one matches the
+    // hot listing path.
+    index("provider_accounts_tenant_id_idx").on(table.tenantId),
+  ],
+);
+
+
+//  OAuth refresh token + cross-process lease, normalized out of
+// `provider_accounts` into its own 1:0/1 table. The vast majority of
+// provider accounts are `credentialKind: "api_key"`, for which these four
+// columns stayed null forever on the shared row; splitting them here keeps
+// the hot health-machine row narrow and isolates the fenced compare-and-swap
+// lease traffic (`providers/auth/refresh-service.ts`) from unrelated
+// account reads/writes.
+export const providerOauthStates = pgTable("provider_oauth_states", {
+  providerAccountId: uuid("provider_account_id")
+    .primaryKey()
+    .references(() => providerAccounts.id, { onDelete: "cascade" }),
+  /**
+   * Encrypted refresh token. Nullable: an account can be pasted with only an
+   * access token (or a credential whose refresh token never arrived), and the
+   * account must still be tracked so the refresh sweep can flag it for re-auth
+   * rather than leaving it invisible. A null value means "cannot refresh".
+   */
+  refreshCiphertext: bytea("refresh_ciphertext"),
+  /**
+   * Encrypted companion secret for the minority of flows whose refresh is not
+   * authorized by the refresh token alone: a device-flow client registration
+   * mints a client secret that must be replayed beside the token. Nullable —
+   * most providers refresh with the refresh token by itself.
+   */
+  clientSecretCiphertext: bytea("client_secret_ciphertext"),
+  /**
+   * Access-token expiry. Nullable: a credential pasted without an expiry (an
+   * opaque token, or a bare access token) has no known deadline, and reads as
+   * "always due" so the sweep refreshes it as soon as it can.
+   */
+  expiresAt: timestamp("expires_at", { withTimezone: true }),
+  // OAuth refresh lease: fenced compare-and-swap coordination so only one
+  // process refreshes a given account's OAuth token at a time, and a losing
+  // process reloads instead of clobbering a peer's fresher token.
+  leaseOwner: text("lease_owner"),
+  leaseExpiresAt: timestamp("lease_expires_at", { withTimezone: true }),
+});
+
+
+// One row per (provider_id, model_id, route), with `endpoint_path`
+// identifying the route — the same model through two providers, or through
+// two routes of one provider, is two rows. `enabled` is a hard routing
+// invariant read by ProviderCatalogService, not a display-only flag.
+export const models = pgTable(
+  "models",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    providerId: text("provider_id")
+      .notNull()
+      .references(() => providers.id, { onDelete: "cascade" }),
+    modelId: text("model_id").notNull(),
+    wireFamily: wireFamily("wire_family").notNull(),
+    // Which protocol shape at all (`llm` = the canonical chat pipeline). A
+    // non-`llm` row (System One) is served by its native passthrough route, not
+    // by the surface codecs. Kept a plain text column with a default rather than
+    // a pgEnum: unlike `wire_family`, the set is expected to grow (embeddings,
+    // tts, …) and a widened enum needs a migration per addition.
+    serviceKind: text("service_kind").notNull().default("llm"),
+    endpointPath: text("endpoint_path").notNull(),
+    contextLimit: integer("context_limit"),
+    outputLimit: integer("output_limit"),
+    modalities: jsonb("modalities"),
+    reasoning: boolean("reasoning").notNull().default(false),
+    toolCall: boolean("tool_call").notNull().default(false),
+    cost: jsonb("cost"),
+    source: text("source"),
+    sourceUpdatedAt: timestamp("source_updated_at", { withTimezone: true }),
+    enabled: boolean("enabled").notNull().default(true),
+    /**
+     * Whether the upstream refuses this model on a free plan.
+     *
+     * Nullable three-valued: `true` refused on free, `false` served on free,
+     * `NULL` nobody established it. `NULL` is the state of every model on every
+     * provider that publishes nothing about plan access, and collapsing it to
+     * `false` would assert a measured fact for rows nobody measured.
+     *
+     * A column rather than a derived rule because CodeBuddy publishes nothing
+     * that predicts access — `credits` and `vendor` were both measured against
+     * real dispatches and neither correlates. Only the refusal itself is a
+     * signal, so the answer has to be recorded. See
+     * `.skills/cartethyia-reverse-engineering/references/codebuddy-dispatch-headers.md`.
+     */
+    proOnly: boolean("pro_only"),
+    /**
+     * Whether the provider publishes this model as free.
+     *
+     * Mirrors `proOnly` but sourced from the provider's own directory rather
+     * than inferred: Qoder states `is_free` per row, so the tier tracks a plan
+     * change without a code change. Nullable three-valued — absence means the
+     * provider did not state it, which is not the same as billed.
+     */
+    freeTier: boolean("free_tier"),
+  },
+  (table) => [
+    uniqueIndex("models_provider_model_route_uidx").on(
+      table.providerId,
+      table.modelId,
+      table.endpointPath,
+    ),
+    index("models_provider_enabled_idx").on(table.providerId, table.enabled),
+    // Bare-id lookups, with no provider named: see migration
+    // `0003_add_models_model_id_index`. Without it those seq-scan this table,
+    // which carries one row per provider x model x route.
+    index("models_model_id_idx").on(table.modelId),
+  ],
+);
+
+// Tenant-scoped suppression of a GLOBAL/built-in model. Tenants cannot mutate
+// the shared `models` row (see `DrizzleProviderCatalogStore.setModelEnabled`),
+// so a disable is recorded here and filtered out of that tenant's routing
+// snapshot. Tenant-owned/BYOK models are still toggled directly on `models`.
+export const tenantDisabledModels = pgTable(
+  "tenant_disabled_models",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    tenantId: tenantRefRequired(),
+    providerId: text("provider_id")
+      .notNull()
+      .references(() => providers.id, { onDelete: "cascade" }),
+    modelId: text("model_id").notNull(),
+    endpointPath: text("endpoint_path").notNull(),
+    createdAt: createdAtColumn(),
+  },
+  (t) => [
+    uniqueIndex("tenant_disabled_models_tenant_provider_model_uidx").on(
+      t.tenantId,
+      t.providerId,
+      t.modelId,
+      t.endpointPath,
+    ),
+  ],
+);
+
+
+// Tenant-owned HTTP/HTTPS/SOCKS5 pools share the provider-account health state machine.
+export const networkPools = pgTable("network_pools", {
+  id: uuid("id").primaryKey().defaultRandom(),
+  createdAt: timestamp("created_at", { withTimezone: true }).notNull().defaultNow(),
+  kind: networkPoolKind("kind").notNull(),
+  /**
+   * Operator-set family/tag for this pool, e.g. `residential-id` or
+   * `datacenter-sg`. Purely organisational: it names which family the node
+   * belongs to and nothing in routing or dispatch reads it. Null when the
+   * operator has not tagged the pool.
+   */
+  family: text("family"),
+  endpointConfig: jsonb("endpoint_config").notNull(),
+  credentialCiphertext: bytea("credential_ciphertext"),
+  maxInflight: integer("max_inflight"),
+  /** Routing weight for weighted selection across the tenant pool group. */
+  weight: integer("weight"),
+  tenantId: tenantRefRequired(),
+
+  ...healthColumns(),
+  lastLatencyMs: integer("last_latency_ms"),
+  lastHealthCheckAt: timestamp("last_health_check_at", { withTimezone: true }),
+  /** Public egress address reported by the last successful probe; null until probed. */
+  egressIp: text("egress_ip"),
+  /** Operator-set egress allowance in bytes; null means unmetered. */
+  quotaBytes: bigint("quota_bytes", { mode: "number" }),
+  /** Last speed-test payload actually transferred, in bytes; null until measured. */
+  lastSpeedtestBytes: integer("last_speedtest_bytes"),
+  /** Wall-clock duration of the last speed-test transfer, in milliseconds. */
+  lastSpeedtestDurationMs: integer("last_speedtest_duration_ms"),
+  /** "ok" | "failed" for the last speed-test; null until measured. */
+  lastSpeedtestStatus: text("last_speedtest_status"),
+  /** Failure message of the last speed-test; null on success or unmeasured. */
+  lastSpeedtestError: text("last_speedtest_error"),
+  /** When the last speed-test ran. */
+  lastSpeedtestAt: timestamp("last_speedtest_at", { withTimezone: true }),
+  /**
+   * Provider-scoped probe verdict, stored beside the other last-known facts
+   * (`egress_ip`, `last_latency_ms`, the speed-test columns) rather than in the
+   * browser.
+   *
+   * It lived in `localStorage` first, and that made it decorative: the verdict
+   * was visible to the operator who ran it and to nobody else — not to another
+   * machine, not to the gateway, and not to admission. A pool could read
+   * "OpenCode blocked" on screen while dispatch kept sending traffic through
+   * it, because nothing outside that one browser had ever heard the verdict.
+   *
+   * `null` means never probed. These describe the *address* recorded in
+   * `egress_ip` at the time, so a changed endpoint clears them.
+   */
+  providerProbeProviderId: text("provider_probe_provider_id"),
+  /** `allowed` | `blocked` | `throttled` | `unreachable`; null until probed. */
+  providerProbeStatus: text("provider_probe_status"),
+  /** Upstream HTTP status when a response arrived. */
+  providerProbeHttpStatus: integer("provider_probe_http_status"),
+  /** Operator-facing reason: the upstream's own words, or why the probe failed. */
+  providerProbeReason: text("provider_probe_reason"),
+  providerProbeLatencyMs: integer("provider_probe_latency_ms"),
+  providerProbeAt: timestamp("provider_probe_at", { withTimezone: true }),
+  /** The endpoint the verdict was measured against. Compared on read, so an
+   * edited endpoint reports "never probed" instead of a stale verdict. */
+  providerProbeEndpoint: text("provider_probe_endpoint"),
+  /** Metered egress totals, flushed from the in-memory socket counters on
+   * every pool-touching write so the quota bar survives a restart. */
+  bytesSentTotal: bigint("bytes_sent_total", { mode: "number" }).notNull().default(0),
+  /** Metered ingress totals; see `bytesSentTotal`. */
+  bytesReceivedTotal: bigint("bytes_received_total", { mode: "number" }).notNull().default(0),
+  },
+  (table) => [
+    index("network_pools_tenant_id_idx").on(table.tenantId),
+    index("network_pools_status_cooldown_idx").on(table.status, table.cooldownUntil),
+    index("network_pools_family_idx").on(table.family),
+  ],
+);
+
+
+// Account and network-pool health events share one table. The entity kind
+// selects exactly one owner FK, and the CHECK constraint below preserves
+// referential integrity for both variants.
+export const healthEvents = pgTable(
+  "health_events",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    entityKind: healthEntityKind("entity_kind").notNull(),
+    accountId: uuid("account_id").references(() => providerAccounts.id, { onDelete: "cascade" }),
+    networkPoolId: uuid("network_pool_id").references(() => networkPools.id, {
+      onDelete: "cascade",
+    }),
+    fromStatus: healthStatus("from_status"),
+    toStatus: healthStatus("to_status").notNull(),
+    reason: text("reason"),
+    errorCategory: text("error_category"),
+    modelId: text("model_id"),
+    createdAt: createdAtColumn(),
+  },
+  (table) => [
+    index("health_events_account_id_created_at_idx").on(table.accountId, table.createdAt),
+    // Descending `created_at` is deliberate: health history is always read
+    // newest-first. The SQLite baseline drops the direction (see
+    // `scripts/generate/sqlite-ddl.ts` — SQLite scans an index either way),
+    // so this is a Postgres-only access-path choice, not a cross-backend
+    // disagreement.
+    index("health_events_account_model_created_idx").on(
+      table.accountId,
+      table.modelId,
+      desc(table.createdAt),
+    ),
+    index("health_events_network_pool_id_created_at_idx").on(table.networkPoolId, table.createdAt),
+    check(
+      "health_events_entity_reference_check",
+      sql`(${table.entityKind} = 'account' AND ${table.accountId} IS NOT NULL AND ${table.networkPoolId} IS NULL) OR (${table.entityKind} = 'pool' AND ${table.networkPoolId} IS NOT NULL AND ${table.accountId} IS NULL)`,
+    ),
+  ],
+);
+
+
+// Model aliasing & combos: replaces the dormant, disconnected
+// `routing_policies` scaffolding. Aliases resolve a client-facing name to a
+// real model; combos back a client-facing name with >=1 real models,
+// auto-selected via `fallback` or `round_robin`.
+export const modelComboStrategy = pgEnum("model_combo_strategy", [
+  "fallback",
+  "round_robin",
+]);
+
+/** Canonical `ComboStrategy` union, derived from the enum above. Console
+ * domain and routing layers import this instead of hand-typing their own mirror.
+ * `ProviderRoutingStrategy` covers the same two values but persists on
+ * `provider_routing_settings` beside `rotateCount`; sticky routing is gone. */
+export type ComboStrategy = (typeof modelComboStrategy.enumValues)[number];
+
+export const modelAliases = pgTable(
+  "model_aliases",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    tenantId: tenantRefRequired(),
+    alias: text("alias").notNull(),
+    targetModel: text("target_model").notNull(),
+    /** Explicit list position. */
+    sortIndex: integer("sort_index").notNull().default(0),
+    ...timestampColumns(),
+  },
+  (t) => [uniqueIndex("model_aliases_tenant_alias_uidx").on(t.tenantId, t.alias)],
+);
+export type ModelAlias = typeof modelAliases.$inferSelect;
+
+export const modelCombos = pgTable(
+  "model_combos",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    tenantId: tenantRefRequired(),
+    name: text("name").notNull(),
+    members: jsonb("members").notNull().$type<string[]>(),
+    strategy: modelComboStrategy("strategy").notNull().default("fallback"),
+    /** Explicit list position. */
+    sortIndex: integer("sort_index").notNull().default(0),
+    ...timestampColumns(),
+  },
+  (t) => [uniqueIndex("model_combos_tenant_name_uidx").on(t.tenantId, t.name)],
+);
+export type ModelCombo = typeof modelCombos.$inferSelect;
+
+export const providerRoutingStrategy = pgEnum("provider_routing_strategy", [
+  "fallback",
+  "round_robin",
+]);
+
+/**
+ * Canonical provider-routing strategy union, derived from the enum above so
+ * the persisted values and the type cannot drift. The console catalog and the
+ * routing engine both import this instead of restating the two literals.
+ */
+export type ProviderRoutingStrategy = (typeof providerRoutingStrategy.enumValues)[number];
+
+/**
+ * Per-account inflight ceiling applied when a provider has no routing row yet.
+ * One literal: the console default and the route catalog's fallback both read
+ * it. `null` is the "unlimited" choice, and it is also this default — an
+ * operator who has never opened the routing card asked for no ceiling, and
+ * inventing one silently capped their concurrency. A ceiling they *did* set is
+ * still honoured, so this only decides the untouched case.
+ */
+export const DEFAULT_PROVIDER_MAX_INFLIGHT: null = null;
+
+export const providerRoutingSettings = pgTable(
+  "provider_routing_settings",
+  {
+    providerId: text("provider_id")
+      .notNull()
+      .references(() => providers.id, { onDelete: "cascade" }),
+    tenantId: tenantRefNullable(),
+    strategy: providerRoutingStrategy("strategy").notNull().default("fallback"),
+    /** Requests served by one account before round robin advances. */
+    rotateCount: integer("rotate_count").notNull().default(1),
+    /** Per-account inflight ceiling applied to every account of this provider
+     * (the field sits next to the failover/round-robin strategy in the UI).
+     * `null` = UNLIMITED concurrency per account. */
+    maxInflight: integer("max_inflight"),
+    /**
+     * Per-pool inflight ceiling for this provider: how many concurrent
+     * requests one proxy pool may carry *for this provider*.
+     *
+     * A second, independent axis from `maxInflight`. That one bounds a single
+     * account; `network_pools.max_inflight` bounds a pool across every
+     * provider; this one bounds one provider's share of one pool. All three
+     * are enforced together — none replaces another.
+     *
+     * `null` = unlimited, and that is the default: a provider nobody has
+     * configured behaves exactly as it did before this column existed, with no
+     * extra Redis key and no extra round trip.
+     */
+    maxInflightPerProxy: integer("max_inflight_per_proxy"),
+    /**
+     * Minimum remaining allowance an account must hold to stay routable, for
+     * this provider. `null` means no floor. Compared against the account's own
+     * unit: absolute remaining credit when the provider reports credits, the
+     * aggregate credit percentage and lowest independent percentage window
+     * otherwise. An account below it is skipped and the request fails over to
+     * the next account instead of erroring.
+     */
+    usageFloor: integer("usage_floor"),
+    /** `credit` compares the summed remaining credit; `percent` compares the
+     * aggregate credit percentage and lowest independent percentage-only
+     * window. Ignored while `usageFloor` is null. */
+    usageFloorUnit: text("usage_floor_unit").$type<"credit" | "percent">().notNull().default("percent"),
+    enabled: boolean("enabled").notNull().default(false),
+    // Route-selected User-Agent for built-in API-key providers; OAuth and BYOK identities stay native.
+    // The column default is the Claude Code CLI identity; the live default for a
+    // row that never set one is `buildClaudeCliUserAgent()` in
+    // providers/operations/custom-cli-headers.ts, which this mirrors.
+    userAgent: text("user_agent").notNull().default("claude-cli/2.1.293 (external, cli)"),
+    // When true, this provider's requests always dial direct. When false,
+    // dispatch automatically selects among every active network pool the
+    // tenant owns (least-loaded/non-cooldown) — never an admin-pinned pool.
+    skipProxy: boolean("skip_proxy").notNull().default(false),
+  },
+  (table) => ({
+    tenantProviderUnique: uniqueIndex("provider_routing_settings_tenant_provider_idx").on(
+      table.tenantId,
+      table.providerId,
+    ),
+    globalProviderUnique: uniqueIndex("provider_routing_settings_global_provider_idx")
+      .on(table.providerId)
+      .where(sql`tenant_id IS NULL`),
+  }),
+);
+
+export const poolRoutingStrategy = pgEnum("pool_routing_strategy", [
+  "least_loaded",
+  "round_robin",
+]);
+export type PoolRoutingStrategy = (typeof poolRoutingStrategy.enumValues)[number];
+
+/**
+ * Per-tenant selection strategy across the tenant's active network pools —
+ * the pool-group counterpart to `provider_routing_settings`. `least_loaded`
+ * (default) keeps the weighted least-loaded scan with its rotating tie-break;
+ * `round_robin` serves `rotate_count` requests per pool before advancing,
+ * mirroring the account strategy, and skips capacity/cooldown pools the way
+ * account failover does. An absent row reads as the default, so tenants that
+ * never touch the toggle own no rows.
+ */
+export const poolRoutingSettings = pgTable("pool_routing_settings", {
+  tenantId: tenantRefPrimaryKey(),
+  strategy: poolRoutingStrategy("strategy").notNull().default("least_loaded"),
+  /** Requests served by one pool before round robin advances. */
+  rotateCount: integer("rotate_count").notNull().default(1),
+});
+
+
+export const API_KEY_MODES = ["personal", "share"] as const;
+export type ApiKeyMode = (typeof API_KEY_MODES)[number];
+
+/**
+ * How a key's `modelList` is interpreted. Exactly two modes, never a
+ * combination: `whitelist` allows only the listed names (an empty list allows
+ * everything), `blacklist` refuses the listed names (an empty list refuses
+ * nothing). Every model-authorization surface — the request path, admission,
+ * `/v1/models`, and the share page — reads this one pair.
+ */
+export const API_KEY_MODEL_ACCESS_MODES = ["whitelist", "blacklist"] as const;
+export type ApiKeyModelAccessMode = (typeof API_KEY_MODEL_ACCESS_MODES)[number];
+
+/**
+ * Inbound `/v1/*` keys never store plaintext secrets. Personal keys carry an
+ * authentication hash; share templates carry policy and child keys point back
+ * to the template that issued them.
+ */
+export const apiKeys = pgTable(
+  "api_keys",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    tenantId: tenantRefRequired(),
+    keyHash: text("key_hash"),
+    keyMode: text("key_mode").$type<ApiKeyMode>().notNull().default("personal"),
+    parentKeyId: uuid("parent_key_id").references(
+      (): AnyPgColumn => apiKeys.id,
+      { onDelete: "cascade" },
+    ),
+    issuedClientIp: text("issued_client_ip"),
+    issuedClientIpKey: text("issued_client_ip_key"),
+    label: text("label").notNull(),
+    scopes: jsonb("scopes").notNull().$type<readonly string[]>(),
+    // Only the non-secret public key prefix is returned to the owner.
+    keyPrefix: text("key_prefix"),
+    // Personal keys retain an encrypted copy for the CLI handoff. Shared
+    // children are revealed once and persist only their authentication hash.
+    keyEncrypted: bytea("key_encrypted"),
+    notesTitle: text("notes_title"),
+    notesSubtitle: text("notes_subtitle"),
+    notesBody: text("notes_body"),
+    /** Explicit list position. */
+    sortIndex: integer("sort_index").notNull().default(0),
+    requestsPerMinute: integer("requests_per_minute"),
+    dailyTokenLimit: bigint("daily_token_limit", { mode: "number" }),
+    monthlyTokenLimit: bigint("monthly_token_limit", { mode: "number" }),
+    lifetimeTokenBudget: bigint("lifetime_token_budget", { mode: "number" }),
+    lifetimeTokensConsumed: bigint("lifetime_tokens_consumed", { mode: "number" })
+      .notNull()
+      .default(0),
+    maxConcurrentRequests: integer("max_concurrent_requests"),
+    /**
+     * `whitelist` = only the names in `modelList` may be used (empty = all);
+     * `blacklist` = the names in `modelList` are refused (empty = none). One
+     * list, one mode — the old separate allow/deny columns are retired.
+     */
+    modelAccessMode: text("model_access_mode")
+      .$type<ApiKeyModelAccessMode>()
+      .notNull()
+      .default("whitelist"),
+    modelList: jsonb("model_list").$type<readonly string[]>(),
+    /**
+     * Client-router ids this key refuses. A request whose fingerprint names one
+     * of these is rejected before routing, so an operator can stop a downstream
+     * router from spending the key. Empty/NULL means "no router is refused".
+     *
+     * A label, not a security control: a client that sends no fingerprint is
+     * simply not matched, which is why this is a per-key policy rather than a
+     * boundary anything else in the system trusts.
+     */
+    clientRouterDenylist: jsonb("client_router_denylist").$type<readonly string[]>(),
+    modelPrefix: text("model_prefix"),
+    createdAt: createdAtColumn(),
+    enabled: boolean("enabled").notNull().default(true),
+    revokedAt: timestamp("revoked_at", { withTimezone: true }),
+    /**
+     * The tenant's mandatory gateway key. Seeded at first boot whether or not
+     * `CARTETHYIA_API_KEY` was configured, and the one key the console refuses
+     * to revoke — it may only be rotated. A column rather than a label match,
+     * so renaming the key cannot break the CLI lookup or let a second
+     * default be created.
+     */
+    isDefault: boolean("is_default").notNull().default(false),
+  },
+  (table) => [
+    uniqueIndex("api_keys_key_hash_idx").on(table.keyHash),
+    /**
+     * One *live* default key per tenant, enforced by the database.
+     *
+     * The baseline SQL has always created this; the schema never declared it,
+     * so the two descriptions disagreed and nothing noticed. It is the only
+     * database-level guarantee behind `isDefault`, and
+     * `test/persistence/sqlite-baseline.test.ts` pins it by name — so the
+     * schema was the incomplete side, not the migration.
+     *
+     * Partial on `revoked_at IS NULL` so a tenant can rotate its default key
+     * without the old, revoked row blocking the new one.
+     */
+    uniqueIndex("api_keys_one_default_per_tenant_uidx")
+      .on(table.tenantId)
+      .where(sql`${table.isDefault} = true AND ${table.revokedAt} IS NULL`),
+    index("api_keys_tenant_id_idx").on(table.tenantId),
+    index("api_keys_enabled_tenant_idx").on(table.tenantId).where(sql`${table.enabled} = true`),
+    index("api_keys_parent_key_id_idx").on(table.parentKeyId),
+    check(
+      "api_keys_mode_shape_check",
+      sql`(
+        (${table.keyMode} = 'personal' AND ${table.keyHash} IS NOT NULL
+          AND ${table.parentKeyId} IS NULL AND ${table.issuedClientIp} IS NULL
+          AND ${table.issuedClientIpKey} IS NULL)
+        OR
+        (${table.keyMode} = 'share' AND (
+          (${table.parentKeyId} IS NULL AND ${table.keyHash} IS NULL
+            AND ${table.keyEncrypted} IS NULL
+            AND ${table.issuedClientIp} IS NULL AND ${table.issuedClientIpKey} IS NULL)
+          OR
+          (${table.parentKeyId} IS NOT NULL AND ${table.keyHash} IS NOT NULL
+            AND ${table.keyEncrypted} IS NULL
+            AND ${table.issuedClientIp} IS NOT NULL AND ${table.issuedClientIpKey} IS NOT NULL)
+        ))
+      )`,
+    ),
+  ],
+);
+
+/**
+ * The optional popup a share link shows its recipient. Kept off `api_keys`
+ * because the auth path reads that row on every request and the popup is
+ * absent for nearly all of them. One row per key, present only when the owner
+ * configured one.
+ */
+export const sharePopups = pgTable(
+  "share_popups",
+  {
+    apiKeyId: uuid("api_key_id")
+      .primaryKey()
+      .references(() => apiKeys.id, { onDelete: "cascade" }),
+    enabled: boolean("enabled").notNull().default(false),
+    image: bytea("image"),
+    imageMime: text("image_mime"),
+    title: text("title"),
+    body: text("body"),
+  },
+  (table) => [
+    check(
+      "share_popups_image_mime_check",
+      sql`${table.imageMime} IS NULL OR ${table.imageMime} IN ('image/png', 'image/jpeg', 'image/webp', 'image/gif')`,
+    ),
+    check(
+      "share_popups_image_size_check",
+      sql`${table.image} IS NULL OR octet_length(${table.image}) <= 2097152`,
+      // SQLite has no octet_length. The generator rewrites this check to
+      // `length(`; both count bytes for a blob column.
+    ),
+  ],
+);
+
+
+// Public bearer links for enrolling one shared key per resolved client IP, and
+// for handing a personal key to its owner. The hash is the lookup key; the
+// token is additionally retained encrypted so the owner's console can show the
+// link again instead of losing it after the one response that minted it.
+/**
+ * Share-link kinds, as a runtime tuple. Route schemas and DTOs project this
+ * list rather than restating it.
+ */
+export const SHARE_LINK_KINDS = ["enroll", "handoff"] as const;
+
+/** One share-link kind. */
+export type ShareLinkKind = (typeof SHARE_LINK_KINDS)[number];
+
+
+export const shareLinks = pgTable(
+  "share_links",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    apiKeyId: uuid("api_key_id")
+      .notNull()
+      .references(() => apiKeys.id, { onDelete: "cascade" }),
+    tokenHash: text("token_hash").notNull(),
+    // Encrypted bearer token. Present so the console can re-display a stable
+    // link; historical rows written before this column carry NULL and simply
+    // cannot be shown again.
+    tokenEncrypted: bytea("token_encrypted"),
+    /**
+     * The public slug this page is reachable at, in the clear.
+     *
+     * `tokenEncrypted` is the bearer token under the deployment's encryption
+     * key, so a link cannot be re-displayed once that key changes — which is
+     * exactly what a restore onto a fresh deployment does. Holding the slug
+     * unencrypted keeps every restored share page reachable under the address
+     * it was created with.
+     *
+     * Nullable: rows written before this column have no slug, and inventing one
+     * would publish an address the page never had.
+     */
+    slug: text("slug"),
+    kind: text("kind").$type<ShareLinkKind>().notNull().default("enroll"),
+    active: boolean("active").notNull().default(true),
+    createdAt: createdAtColumn(),
+    expiresAt: timestamp("expires_at", { withTimezone: true }),
+    lastViewedAt: timestamp("last_viewed_at", { withTimezone: true }),
+  },
+  (table) => [
+    uniqueIndex("share_links_token_hash_idx").on(table.tokenHash),
+    // Partial: rows without a slug (written before the column) are not
+    // constrained, so backfilling is never blocked by a uniqueness collision.
+    uniqueIndex("share_links_slug_uidx").on(table.slug).where(sql`${table.slug} IS NOT NULL`),
+    index("share_links_api_key_idx").on(table.apiKeyId),
+    index("share_links_active_kind_expires_idx").on(table.active, table.kind, table.expiresAt),
+    check("share_links_kind_check", sql`${table.kind} IN ('enroll', 'handoff')`),
+  ],
+);
+
+
+/**
+ * Console dashboard users — distinct from API-key tenants.
+ */
+export const consoleUsers = pgTable(
+  "console_users",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    tenantId: tenantRefRequired(),
+    username: text("username").notNull(),
+    email: text("email"),
+    passwordHash: text("password_hash").notNull(),
+    displayName: text("display_name"),
+    isActive: boolean("is_active").notNull().default(true),
+    isPlatformAdmin: boolean("is_platform_admin").notNull().default(false),
+    isFirstBoot: boolean("is_first_boot").notNull().default(true),
+    ...timestampColumns(),
+  },
+  (table) => [
+    index("console_users_tenant_id_idx").on(table.tenantId),
+    uniqueIndex("console_users_tenant_username_uidx").on(table.tenantId, table.username),
+  ],
+);
+
+export type ConsoleUser = typeof consoleUsers.$inferSelect;
+
+/**
+ * Console sessions — HTTP-only, signed cookies.
+ * Session metadata only; no secrets in JSON.
+ */
+export const consoleSessions = pgTable(
+  "console_sessions",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    userId: uuid("user_id")
+      .notNull()
+      .references(() => consoleUsers.id, { onDelete: "cascade" }),
+    sessionToken: text("session_token").notNull().unique(),
+    ipAddress: text("ip_address"),
+    userAgent: text("user_agent"),
+    expiresAt: timestamp("expires_at", { withTimezone: true }).notNull(),
+    createdAt: createdAtColumn(),
+  },
+  (table) => [
+    index("console_sessions_user_id_idx").on(table.userId),
+  ],
+);
+
+export type ConsoleSession = typeof consoleSessions.$inferSelect;
+
+/**
+ * Console login lockout tracking — fail-closed rate limiting keyed by
+ * client IP (login attempts happen before any user identity is known, so
+ * this cannot be keyed by `console_users.id`). Persisted so a ban survives
+ * a process restart and is shared across every app instance, instead of a
+ * static in-process map that a second instance or a redeploy would forget.
+ */
+export const consoleLockouts = pgTable(
+  "console_lockouts",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    ip: text("ip").notNull(),
+    failureCount: integer("failure_count").notNull().default(0),
+    windowUntil: timestamp("window_until", { withTimezone: true }),
+    lockedUntil: timestamp("locked_until", { withTimezone: true }),
+    reason: text("reason"),
+    ...timestampColumns(),
+  },
+  (table) => [
+    uniqueIndex("idx_console_lockouts_ip").on(table.ip),
+  ],
+);
+
+
+// Configuration-change audit has its own retention lifecycle; it is separate
+// from metadata telemetry and short-lived optional payload capture.
+export const adminAuditLog = pgTable(
+  "admin_audit_log",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    createdAt: createdAtColumn(),
+    actor: text("actor").notNull(),
+    tenantId: tenantCascadeSetNull(),
+    action: text("action").notNull(),
+    target: text("target").notNull(),
+    detail: jsonb("detail"),
+  },
+  (table) => [
+    index("admin_audit_log_tenant_id_created_at_id_idx").on(
+      table.tenantId,
+      table.createdAt,
+      table.id,
+    ),
+  ],
+);
+
+
+// Runtime preferences are stored in one validated JSONB bag. The filter
+// master toggle remains a typed column because it is read on the proxy path.
+export interface ConsoleSettingsPreferences {
+  tenantConcurrencyLimit?: number | null;
+  /** Normalizes provider-native thinking config for non-user final turns. */
+  thinkingNormalizationEnabled?: boolean;
+  responsesReasoningSummary?: "auto" | "concise" | "detailed";
+  /**
+   * Request compression: compact bulky tool-result text in the message history
+   * before dispatch (RTK prune). Default off.
+   */
+  rtkPruneEnabled?: boolean;
+  /** RTK prune strength when enabled (`lite`/`full`/`ultra`); `full` when unset. */
+  rtkPruneLevel?: "lite" | "full" | "ultra";
+  /**
+   * Request compression: append a minimal-code directive to the system content.
+   * Default off; `ponyTailLevel` selects the intensity when on.
+   */
+  ponyTailEnabled?: boolean;
+  /** PonyTail directive intensity; `full` when unset. Only applied when enabled. */
+  ponyTailLevel?: "lite" | "full" | "ultra" | null;
+  /**
+   * Payload capture mode (default `none` when unset):
+   * - `metadata` — Proxy→Provider method + allowlisted headers only
+   * - `full` — full redacted bodies up to the configured capture limit
+   * - `none` — no drawer capture (request metadata events still retained)
+   */
+  telemetryPayloads?: "full" | "metadata" | "none";
+  /**
+   * Capture depth for `full` mode (default `minimum` when unset). Fixed size
+   * tiers so every debug sees all four drawer panels — only the cap changes:
+   * - `minimum` — 1 MiB combined cap, light enough to leave on
+   * - `moderate` — 8 MiB combined cap, expect CPU/memory spikes while active
+   * - `maximum` — 16 MiB combined cap, expect CPU/memory spikes while active
+   */
+  telemetryPayloadDepth?: "minimum" | "moderate" | "maximum";
+  privacyMode?: "masked" | "full";
+}
+
+export const consoleSettings = pgTable("console_settings", {
+  tenantId: tenantRefPrimaryKey(),
+  preferences: jsonb("preferences").notNull().default({}).$type<ConsoleSettingsPreferences>(),
+  updatedAt: updatedAtColumn(),
+});
+
+
+// CLI Tools (CLI tool configuration): per-tenant mapping table that lets a coding-agent
+// slot (e.g. Claude Code's "opus" slot) target a specific Cartethyia model. Rows
+// are opaque strings — validation happens at the API boundary against the
+// tool registry, not in Postgres.
+export const cliToolMappings = pgTable(
+  "cli_tool_mappings",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    tenantId: tenantRefRequired(),
+    apiKeyId: uuid("api_key_id")
+      .notNull()
+      .references((): AnyPgColumn => apiKeys.id, { onDelete: "cascade" }),
+    toolId: text("tool_id").notNull(),
+    slotKey: text("slot_key").notNull(),
+    sourceModel: text("source_model").notNull(),
+    targetModel: text("target_model").notNull(),
+    enabled: boolean("enabled").notNull().default(true),
+    updatedAt: updatedAtColumn(),
+  },
+  (table) => [
+    uniqueIndex("cli_tool_mappings_key").on(
+      table.tenantId,
+      table.toolId,
+      table.apiKeyId,
+      table.slotKey,
+    ),
+  ],
+);
+
+
+// Per-(tenant, tool, api_key) settings: whether mappings are active and which
+// mode the tool is configured for. Primary key is the composite of all three.
+export const cliToolSettings = pgTable(
+  "cli_tool_settings",
+  {
+    tenantId: tenantRefRequired(),
+    apiKeyId: uuid("api_key_id")
+      .notNull()
+      .references((): AnyPgColumn => apiKeys.id, { onDelete: "cascade" }),
+    toolId: text("tool_id").notNull(),
+    mappingsEnabled: boolean("mappings_enabled").notNull().default(false),
+    mode: text("mode").notNull().default("remote"),
+    updatedAt: updatedAtColumn(),
+  },
+  (table) => [primaryKey({ columns: [table.tenantId, table.toolId, table.apiKeyId] })],
+);
+
+
+// Metadata-only request telemetry. Retention is configured independently of
+// payload capture (`CARTETHYIA_TELEMETRY_RETENTION_DAYS`, default 30 days).
+// No prompt text, raw body, API key, encrypted-reasoning payload, or generic
+// `payload` column — the schema structurally prevents those fields.
+// Correlation columns intentionally have no catalog foreign keys, so a
+// telemetry write never depends on provider-row lifetime.
+export const telemetryEvents = pgTable(
+  "telemetry_events",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    createdAt: createdAtColumn(),
+    tenantId: tenantRefRequired(),
+    requestId: uuid("request_id").notNull(),
+    sourceSurface: telemetrySourceSurface("source_surface"),
+    requestedModel: text("requested_model"),
+    /**
+     * The `model` the upstream put in its own response, verbatim.
+     *
+     * Separate from `requested_model` because they answer different questions:
+     * one is what the client asked for, the other is what the provider said it
+     * ran. A provider that answers `muse-spark` as `deepseek-v4.1-flash` is
+     * telling the operator something, and overwriting one with the other would
+     * erase it. NULL when the response carried no model.
+     */
+    reportedModel: text("reported_model"),
+    /** Canonical reasoning effort the client asked for; NULL when it asked for none. */
+    requestedEffort: text("requested_effort"),
+    providerId: text("provider_id"),
+    accountId: uuid("account_id"),
+    networkPoolId: uuid("network_pool_id"),
+    endpoint: text("endpoint"),
+    apiKeyId: uuid("api_key_id"),
+    userAgent: text("user_agent"),
+    clientIp: text("client_ip"),
+    latencyMs: integer("latency_ms"),
+    ttfbMs: integer("ttfb_ms"),
+    stream: boolean("stream"),
+    status: telemetryStatus("status"),
+    /**
+     * HTTP status the client actually received. The `status` enum is the
+     * internal terminal state (`completed`/`failed`/…), which collapses every
+     * failure to one bucket; this column keeps the wire code, so a 503
+     * admission rejection stays distinguishable from a 500 upstream failure
+     * in the Usage status breakdown. NULL on rows written before the column.
+     */
+    httpStatus: integer("http_status"),
+    /**
+     * Which layer produced the failure: the gateway itself, the upstream
+     * provider, or the network in between. Recorded beside `error_category`
+     * because the category alone cannot separate them — `invalid_request` is
+     * both "the caller sent a bad body" (ours) and "the upstream rejected the
+     * body" (theirs), and an operator triaging a failure needs to know which.
+     */
+    errorOrigin: text("error_origin"),
+    errorCategory: text("error_category"),
+    inputTokens: integer("input_tokens"),
+    cachedInputTokens: integer("cached_input_tokens"),
+    outputTokens: integer("output_tokens"),
+    reasoningTokens: integer("reasoning_tokens"),
+    estimatedCostUsd: numeric("estimated_cost_usd", { precision: 12, scale: 6 }),
+    /** Provider-billed credits (Tencent buddy-meter `credit`); NULL when unreported. */
+    creditUsed: numeric("credit_used", { precision: 12, scale: 4 }),
+    /**
+     * Gateway-side time before the upstream dispatch began: routing,
+     * credential lease, and request preparation. `ttfbMs` is measured from
+     * the dispatch start, so this is the overhead the client waited on before
+     * the provider was even called.
+     */
+    resolveMs: integer("resolve_ms"),
+    tokensPerSec: numeric("tokens_per_sec", { precision: 10, scale: 2 }),
+    // Absolute epoch-millisecond event timestamps (~1.8e12) overflow integer
+    // (int4): completed requests carrying observed stream timings failed
+    // their telemetry insert and vanished from the console. bigint "number"
+    // mode is safe here — epoch millis stay far below MAX_SAFE_INTEGER —
+    // and keeps the columns JSON-safe like the token-budget bigints above.
+    firstContentDeltaAtMs: bigint("first_content_delta_at_ms", { mode: "number" }),
+    lastEventAtMs: bigint("last_event_at_ms", { mode: "number" }),
+    /**
+     * Why this request's drawer payloads were not captured, or `captured` when
+     * they were. Remote-diagnosis column: the drawer reads empty for reasons
+     * that are invisible from the client — the tenant's capture mode, a missing
+     * tenant, a non-terminal attempt — and without this an operator could not
+     * tell "not configured" from "broken".
+     */
+    payloadCaptureState: text("payload_capture_state"),
+  },
+  (table) => [
+    index("telemetry_events_created_at_idx").on(table.createdAt),
+    index("telemetry_events_tenant_created_idx").on(table.tenantId, table.createdAt),
+    index("telemetry_events_request_id_idx").on(table.requestId),
+    index("telemetry_events_tenant_account_created_idx").on(
+      table.tenantId,
+      table.accountId,
+      table.createdAt,
+    ),
+    /**
+     * Serves the public share page's per-key aggregate (`share-usage.ts`):
+     * equality on `api_key_id`, range and aggregate on `created_at`. Without
+     * it every share render scanned the largest table in the schema.
+     */
+    index("telemetry_events_api_key_created_idx").on(table.apiKeyId, table.createdAt),
+  ],
+);
+
+/** Durable aggregates retained after metadata telemetry expires. */
+export const TELEMETRY_USAGE_IDENTITY_TYPES = ["account", "api_key"] as const;
+export type TelemetryUsageIdentityType = (typeof TELEMETRY_USAGE_IDENTITY_TYPES)[number];
+
+export const telemetryUsageTotals = pgTable(
+  "telemetry_usage_totals",
+  {
+    tenantId: tenantRefRequired(),
+    identityType: text("identity_type").$type<TelemetryUsageIdentityType>().notNull(),
+    entityId: uuid("entity_id").notNull(),
+    requests: bigint("requests", { mode: "number" }).notNull().default(0),
+    errors: bigint("errors", { mode: "number" }).notNull().default(0),
+    inputTokens: bigint("input_tokens", { mode: "number" }).notNull().default(0),
+    outputTokens: bigint("output_tokens", { mode: "number" }).notNull().default(0),
+    lastUsedAt: timestamp("last_used_at", { withTimezone: true }).notNull(),
+    updatedAt: updatedAtColumn(),
+  },
+  (table) => [
+    primaryKey({
+      name: "telemetry_usage_totals_identity_pk",
+      columns: [table.tenantId, table.identityType, table.entityId],
+    }),
+    check(
+      "telemetry_usage_totals_identity_type_check",
+      sql`${table.identityType} IN ('account', 'api_key')`,
+    ),
+  ],
+);
+
+
+export const telemetryPayloads = pgTable(
+  "telemetry_payloads",
+  {
+    id: uuid("id").primaryKey().defaultRandom(),
+    tenantId: tenantRefRequired(),
+    requestId: uuid("request_id"),
+    capturedAt: timestamp("captured_at", { withTimezone: true }).notNull().defaultNow(),
+    expiresAt: timestamp("expires_at", { withTimezone: true }).notNull(),
+    /**
+     * Typed index into the on-disk frame store. Prompt/response bodies live
+     * only in the frame file these columns name — never in Postgres. The
+     * previous `request_body jsonb` column is gone so a body cannot be
+     * written here by accident.
+     */
+    storage: text("storage").notNull(),
+    file: text("file").notNull(),
+    offset: integer("offset").notNull(),
+    length: integer("length").notNull(),
+    checksum: text("checksum").notNull(),
+    version: integer("version").notNull(),
+    /**
+     * Which of the four payload stages were captured, as a bitmask. The Usage
+     * list shows one icon per stage; learning presence from the frame would
+     * cost one inflate per row, so the writer records it here and the list
+     * reads it off the indexed row.
+     *
+     * Bits: 1=request, 2=provider_request, 4=provider_response,
+     * 8=client_response. Nullable: NULL means the row predates this column,
+     * which renders as all icons greyed rather than inventing presence.
+     */
+    stages: integer("stages"),
+  },
+  (table) => [
+    index("telemetry_payloads_request_id_idx").on(table.requestId),
+    index("telemetry_payloads_tenant_request_idx").on(table.tenantId, table.requestId),
+    index("telemetry_payloads_expires_idx").on(table.expiresAt),
+    check("telemetry_payloads_storage_check", sql`${table.storage} IN ('jsonb-file', 'jsonl-file')`),
+    check("telemetry_payloads_version_check", sql`${table.version} = 1`),
+    check("telemetry_payloads_offset_check", sql`${table.offset} >= 0`),
+    check("telemetry_payloads_length_check", sql`${table.length} > 0`),
+  ],
+);
+
+
+
+// Daily check-in ledger: one row per (account, calendar day) once the sweep has
+// attempted that account's claim.
+//
+// The day-scoped composite PK is the claim guard — an insert that conflicts is
+// exactly the "already settled today" answer — so it replaces the Redis
+// `SET NX EX` the Lite backend never had. The console reads the same table to
+// answer "was this attempted today", which is what makes the Quota page show
+// the real state instead of a permanent "Check-in due" when Redis is absent.
+export const dailyCheckinLedger = pgTable(
+  "daily_checkin_ledger",
+  {
+    dayKey: text("day_key").notNull(),
+    accountId: uuid("account_id").notNull(),
+    claimedAt: timestamp("claimed_at", { withTimezone: true }).notNull().defaultNow(),
+  },
+  (table) => [
+    primaryKey({ columns: [table.dayKey, table.accountId] }),
+  ],
+);
+
+
+/**
+ * Per-day, per-model credit deductions the buddy billing meter reported.
+ *
+ * A snapshot, not a ledger: the upstream is the source of truth and answers the
+ * whole window on every call, so a row is upserted rather than appended. What it
+ * buys is that the console renders a *stored* figure — without it every page
+ * open would re-hit the billing endpoint for every account, and a failed read
+ * would show nothing at all instead of the last number obtained.
+ *
+ * Keyed (account, day, model): one row per bucket, and the conflict target is
+ * what makes the write idempotent across passes.
+ */
+export const buddyUsageSnapshots = pgTable(
+  "buddy_usage_snapshots",
+  {
+    accountId: uuid("account_id")
+      .notNull()
+      .references(() => providerAccounts.id, { onDelete: "cascade" }),
+    /** Calendar day the requests were metered on, `YYYY-MM-DD`. */
+    dayKey: text("day_key").notNull(),
+    model: text("model").notNull(),
+    /** Credits deducted across every request in this bucket. */
+    credit: numeric("credit", { precision: 14, scale: 4 }).notNull(),
+    syncedAt: timestamp("synced_at", { withTimezone: true }).notNull().defaultNow(),
+  },
+  (table) => [
+    primaryKey({ columns: [table.accountId, table.dayKey, table.model] }),
+    index("buddy_usage_snapshots_day_idx").on(table.dayKey),
+  ],
+);
+
+/**
+ * One row per (day, account, growth task) once that day's claim was attempted.
+ *
+ * The composite PK is the idempotency guard: an insert that conflicts is
+ * exactly the "already settled today" answer, which is what stops a sweep from
+ * re-POSTing a claim on every pass. Same shape as `daily_checkin_ledger`
+ * because it is the same problem — one claim per account per day — widened by
+ * the task name now that there is more than one thing to claim.
+ */
+export const buddyTaskLedger = pgTable(
+  "buddy_task_ledger",
+  {
+    dayKey: text("day_key").notNull(),
+    accountId: uuid("account_id")
+      .notNull()
+      .references(() => providerAccounts.id, { onDelete: "cascade" }),
+    task: text("task").notNull(),
+    claimedAt: timestamp("claimed_at", { withTimezone: true }).notNull().defaultNow(),
+  },
+  (table) => [
+    primaryKey({ columns: [table.dayKey, table.accountId, table.task] }),
+    index("buddy_task_ledger_account_idx").on(table.accountId),
+  ],
+);