7.4 billion tokens per month. 34 free LLM providers. 635 free model endpoints. All behind one /v1 endpoint, plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypted keys. Personal experimentation only.
TypeScript
feat(db): stamp request history with the concrete model row (#1212)
suantea (asuan_suan)abe4c08bff6f
diff --git a/server/src/__tests__/db/migrate/roundtrip.test.ts b/server/src/__tests__/db/migrate/roundtrip.test.tsindex fdc216bd..13ee35a1 100644--- a/server/src/__tests__/db/migrate/roundtrip.test.ts+++ b/server/src/__tests__/db/migrate/roundtrip.test.ts@@ -39,6 +39,7 @@ const ANALYTICS_LATENCY_PERCENTILE_INDEX_FILENAME = '20260902_000001_analytics_lconst MCP_ENABLED_DEFAULT_FILENAME = '20260903_000001_mcp_enabled_default.ts';const RESPONSE_CACHE_FILENAME = '20260903_000002_response_cache.ts';const KEY_MONTHLY_BUDGET_FILENAME = '20260904_000001_key_monthly_budget.ts';+const REQUEST_MODEL_ATTRIBUTION_FILENAME = '20260913_000001_request_model_attribution.ts';const KEY_MONTHLY_USAGE_FILENAME = '20260914_000001_key_monthly_usage.ts';const QUOTA_SNAPSHOT_FRESHNESS_FILENAME = '20260915_000001_quota_snapshot_freshness.ts';@@ -126,6 +127,7 @@ describe('migration round trip', () => {MCP_ENABLED_DEFAULT_FILENAME,RESPONSE_CACHE_FILENAME,KEY_MONTHLY_BUDGET_FILENAME,+ REQUEST_MODEL_ATTRIBUTION_FILENAME,KEY_MONTHLY_USAGE_FILENAME,QUOTA_SNAPSHOT_FRESHNESS_FILENAME,]);diff --git a/server/src/__tests__/routes/custom-endpoint-identity.test.ts b/server/src/__tests__/routes/custom-endpoint-identity.test.tsindex 3c755f67..22ed3514 100644--- a/server/src/__tests__/routes/custom-endpoint-identity.test.ts+++ b/server/src/__tests__/routes/custom-endpoint-identity.test.ts@@ -7,6 +7,7 @@ import { setCooldown, isOnCooldown, clearCooldownsForKey } from '../../services/import { getModelGroups, resolveRequestedIdToMembers, setUnifyOverrides } from '../../services/model-groups.js';import { noteModelRetirementSignal, resetModelRetirementObservations } from '../../services/model-retirement.js';import { endpointHandle, qualifiedModelMemberId } from '../../lib/endpoint-scope.js';+import { logRequest } from '../../lib/request-log.js';import { buildModelListing } from '../../services/model-listing.js';import { mintDashboardToken, isGatedApiPath } from '../helpers/auth.js';@@ -234,6 +235,30 @@ describe('per-endpoint identity for custom relay models (#651)', () => {expect(rowsFor(SHARED_MODEL).map(r => r.enabled)).toEqual([1, 1]);});++ it('records the concrete model row selected for each relay (#1187)', async () => {+ await registerBothRelays(app);+ const a = rowOn(RELAY_A);+ const b = rowOn(RELAY_B);+ expect(a.id).not.toBe(b.id);++ // Same platform + model id on both relays: only the key tells them apart.+ logRequest('custom', SHARED_MODEL, b.key_id!, 'success', 10, 20, 30, null);+ const logged = getDb().prepare(`+ SELECT model_db_id FROM requests+ WHERE platform = 'custom' AND model_id = ?+ ORDER BY id DESC LIMIT 1+ `).get(SHARED_MODEL) as { model_db_id: number };+ expect(logged.model_db_id).toBe(b.id);++ logRequest('custom', SHARED_MODEL, a.key_id!, 'success', 10, 20, 30, null);+ const loggedA = getDb().prepare(`+ SELECT model_db_id FROM requests+ WHERE platform = 'custom' AND model_id = ?+ ORDER BY id DESC LIMIT 1+ `).get(SHARED_MODEL) as { model_db_id: number };+ expect(loggedA.model_db_id).toBe(a.id);+ });});describe('naming', () => {diff --git a/server/src/db/migrate/defaults.ts b/server/src/db/migrate/defaults.tsindex 601f1023..82569d19 100644--- a/server/src/db/migrate/defaults.ts+++ b/server/src/db/migrate/defaults.ts@@ -36,6 +36,7 @@ import * as responseCache from '../migrations/20260903_000002_response_cache.js'import * as keyMonthlyBudget from '../migrations/20260904_000001_key_monthly_budget.js';import * as keyMonthlyUsage from '../migrations/20260914_000001_key_monthly_usage.js';import * as quotaSnapshotFreshness from '../migrations/20260915_000001_quota_snapshot_freshness.js';+import * as requestModelAttribution from '../migrations/20260913_000001_request_model_attribution.js';export interface MigrationModule {up(db: Db): void;@@ -82,6 +83,7 @@ export const ANALYTICS_LATENCY_PERCENTILE_INDEX_FILENAME = '20260902_000001_analexport const MCP_ENABLED_DEFAULT_FILENAME = '20260903_000001_mcp_enabled_default.ts';export const RESPONSE_CACHE_FILENAME = '20260903_000002_response_cache.ts';export const KEY_MONTHLY_BUDGET_FILENAME = '20260904_000001_key_monthly_budget.ts';+export const REQUEST_MODEL_ATTRIBUTION_FILENAME = '20260913_000001_request_model_attribution.ts';export const KEY_MONTHLY_USAGE_FILENAME = '20260914_000001_key_monthly_usage.ts';export const QUOTA_SNAPSHOT_FRESHNESS_FILENAME = '20260915_000001_quota_snapshot_freshness.ts';@@ -121,6 +123,7 @@ export const DEFAULT_MIGRATIONS: readonly DefaultMigration[] = [{ filename: MCP_ENABLED_DEFAULT_FILENAME, module: mcpEnabledDefault },{ filename: RESPONSE_CACHE_FILENAME, module: responseCache },{ filename: KEY_MONTHLY_BUDGET_FILENAME, module: keyMonthlyBudget },+ { filename: REQUEST_MODEL_ATTRIBUTION_FILENAME, module: requestModelAttribution },{ filename: KEY_MONTHLY_USAGE_FILENAME, module: keyMonthlyUsage },{ filename: QUOTA_SNAPSHOT_FRESHNESS_FILENAME, module: quotaSnapshotFreshness },];diff --git a/server/src/db/migrations/20260913_000001_request_model_attribution.ts b/server/src/db/migrations/20260913_000001_request_model_attribution.tsnew file mode 100644index 00000000..e6b64d21--- /dev/null+++ b/server/src/db/migrations/20260913_000001_request_model_attribution.ts@@ -0,0 +1,75 @@+// Migration: stamp request history with the concrete model row (#1187)+//+// DOWN: reversible+//+// Custom relays can expose the same upstream model id, so (platform, model_id)+// no longer uniquely identifies the `models` row whose pricing, reliability,+// and speed apply. Today request identity is derived at read time from+// api_keys.base_url: deleting a key, or editing its base URL, retroactively+// reattributes (or orphans) all of that key's historical requests.+//+// This migration freezes request identity by stamping each request with the+// concrete `models.id` that served it. Catalog requests resolve by+// (platform, model_id); custom requests resolve through the request's key+// (its normalized base_url is the row's endpoint_scope). Custom requests+// whose key never reached routing (key_id IS NULL) stay NULL — an+// unattributable request must not land on the wrong relay.+//+// No foreign key on models(id) — request history must survive the user+// deleting the model row that once served it.++import type { Db } from '../types.js';++function hasColumn(db: Db, table: string, column: string): boolean {+ const columns = db.prepare(`PRAGMA table_info(${table})`).all() as { name: string }[];+ return columns.some(c => c.name === column);+}++export function up(db: Db): void {+ if (hasColumn(db, 'requests', 'model_db_id')) return;++ // Add the column (nullable — unattributed requests keep NULL).+ db.prepare('ALTER TABLE requests ADD COLUMN model_db_id INTEGER').run();++ // Non-unique index for backfill scans and future analytics queries.+ db.prepare('CREATE INDEX IF NOT EXISTS idx_requests_model_db_id ON requests(model_db_id)').run();++ // Backfill catalog requests: (platform, model_id) → models.id+ // These are unambiguous — catalog rows are unique on (platform, model_id).+ // SQLite UPDATE has no table alias; correlate via the full table name.+ db.prepare(`+ UPDATE requests+ SET model_db_id = (+ SELECT m.id FROM models m+ WHERE m.platform = requests.platform AND m.model_id = requests.model_id+ LIMIT 1+ )+ WHERE requests.platform != 'custom'+ `).run();++ // Backfill custom requests: key_id → api_keys.base_url → models.endpoint_scope+ // Only while the saved key still resolves to a matching row.+ db.prepare(`+ UPDATE requests+ SET model_db_id = (+ SELECT m.id FROM models m+ JOIN api_keys k ON k.id = requests.key_id AND k.platform = 'custom'+ WHERE m.platform = 'custom'+ AND m.model_id = requests.model_id+ AND m.endpoint_scope = rtrim(trim(k.base_url), '/')+ LIMIT 1+ )+ WHERE requests.platform = 'custom' AND requests.key_id IS NOT NULL+ `).run();++ // Orphaned custom history stays NULL — better than being attributed to the+ // wrong relay. The two UPDATEs above are best-effort: if a custom key was+ // deleted or its base_url changed, the JOIN drops out and model_db_id stays+ // NULL (which is exactly the "unattributable" semantics we want).+}++export function down(db: Db): void {+ if (!hasColumn(db, 'requests', 'model_db_id')) return;+ db.prepare('DROP INDEX IF EXISTS idx_requests_model_db_id').run();+ db.prepare('ALTER TABLE requests DROP COLUMN model_db_id').run();+}diff --git a/server/src/lib/request-log.ts b/server/src/lib/request-log.tsindex 50288258..ca917ddd 100644--- a/server/src/lib/request-log.ts+++ b/server/src/lib/request-log.ts@@ -29,6 +29,34 @@ function setSettingIfMissing(db: LogTx, key: string, value: string): void {`).run(key, value);}+// Resolve the concrete models.id for a request (#1187). Catalog requests+// resolve by (platform, model_id); custom requests resolve through the+// request's key (its normalized base_url is the row's endpoint_scope).+// Returns NULL when the request is unattributable — a custom request whose+// key never reached routing, or whose key no longer resolves to a matching+// model row. An unattributable request must not land on the wrong relay.+function resolveModelDbId(+ db: LogTx,+ platform: string,+ modelId: string,+ keyId: number | null,+): number | null {+ if (platform !== 'custom') {+ const row = db.prepare('SELECT id FROM models WHERE platform = ? AND model_id = ? LIMIT 1')+ .get(platform, modelId) as { id: number } | undefined;+ return row?.id ?? null;+ }+ if (keyId == null) return null;+ const key = db.prepare("SELECT base_url FROM api_keys WHERE id = ? AND platform = 'custom'")+ .get(keyId) as { base_url: string | null } | undefined;+ if (!key?.base_url) return null;+ const scope = key.base_url.trim().replace(/\/+$/, '');+ const row = db.prepare(+ 'SELECT id FROM models WHERE platform = ? AND model_id = ? AND endpoint_scope = ? LIMIT 1',+ ).get('custom', modelId, scope) as { id: number } | undefined;+ return row?.id ?? null;+}+// Append a row to the request analytics table. Shared by the chat proxy, the// responses path, and the fusion panel so every served (or failed) sub-request// is logged identically. Lives in a neutral lib module to avoid an import cycle@@ -86,9 +114,9 @@ export function logRequest(const client = getClientContext();const tx = db.transaction(() => {const insert = db.prepare(`- INSERT INTO requests (platform, model_id, key_id, status, input_tokens, output_tokens, latency_ms, error, ttfb_ms, requested_model, served_model, client_ip, client_user_agent, client_agent, caller)- VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)- `).run(platform, modelId, keyId, status, inputTokens, outputTokens, latencyMs, error, ttfbMs, requestedModel, servedModel, client.ip, client.userAgent, client.agent, caller);+ INSERT INTO requests (platform, model_id, key_id, status, input_tokens, output_tokens, latency_ms, error, ttfb_ms, requested_model, served_model, client_ip, client_user_agent, client_agent, caller, model_db_id)+ VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)+ `).run(platform, modelId, keyId, status, inputTokens, outputTokens, latencyMs, error, ttfbMs, requestedModel, servedModel, client.ip, client.userAgent, client.agent, caller, resolveModelDbId(db, platform, modelId, keyId));// Report the row id back to the fallback loop's attempt trace (if one is// active): the LAST id noted during a loop run is the terminal row the