feat(resilience): expose providerQuotaOverrides via /api/resilience (#9871)

Co-authored-by: herjarsa <herjarsa@users.noreply.github.com>

src/app/api/resilience/route.ts

@@ -111,9 +111,15 @@ function normalizeLegacyPatch(body: JsonRecord): ResilienceSettingsPatch {
 }
 
 async function syncRuntimeSettings(resilienceSettings: ResilienceSettings) {
-  const { applyRequestQueueSettings } =
-    await import("@omniroute/open-sse/services/rateLimitManager");
+  const [{ applyRequestQueueSettings }, { setProviderQuotaOverrides }] = await Promise.all([
+    import("@omniroute/open-sse/services/rateLimitManager"),
+    import("@omniroute/open-sse/services/providerDefaultRateLimit"),
+  ]);
   await applyRequestQueueSettings(resilienceSettings.requestQueue);
+  // #6846 Phase 2: re-apply per-provider RPM/concurrency overrides on the hot
+  // path so a PATCH takes effect without a process restart. Mirrors the call in
+  // rateLimitManager.ts::initializeRateLimits() (startup).
+  setProviderQuotaOverrides(resilienceSettings.providerQuotaOverrides);
 }
 
 /**
@@ -136,6 +142,7 @@ export async function GET() {
       comboCooldownWait: resilience.comboCooldownWait,
       quotaShareConcurrencyLimit: resilience.quotaShareConcurrencyLimit,
       providerCooldown: resilience.providerCooldown,
+      providerQuotaOverrides: resilience.providerQuotaOverrides,
       legacy: buildLegacyResilienceCompat(resilience),
     });
   } catch (err: unknown) {
@@ -208,6 +215,12 @@ export async function PATCH(request) {
             providerCooldown: body.providerCooldown as ResilienceSettingsPatch["providerCooldown"],
           }
         : {}),
+      ...(body.providerQuotaOverrides
+        ? {
+            providerQuotaOverrides:
+              body.providerQuotaOverrides as ResilienceSettingsPatch["providerQuotaOverrides"],
+          }
+        : {}),
       ...normalizeLegacyPatch(body),
     });
 
@@ -245,6 +258,7 @@ export async function PATCH(request) {
       comboCooldownWait: nextResilience.comboCooldownWait,
       quotaShareConcurrencyLimit: nextResilience.quotaShareConcurrencyLimit,
       providerCooldown: nextResilience.providerCooldown,
+      providerQuotaOverrides: nextResilience.providerQuotaOverrides,
       legacy: buildLegacyResilienceCompat(nextResilience),
     });
   } catch (err: unknown) {

src/shared/validation/schemas/settings.ts

@@ -158,6 +158,21 @@ export const updateResilienceSchema = z
       .strict()
       .optional(),
     defaults: legacyResilienceDefaultsSchema.optional(),
+    // #6846 Phase 2: per-provider operator overrides for the header-less
+    // "provider default" static budget (open-sse/services/providerDefaultRateLimit.ts)
+    // and its companion per-connection concurrency cap. Mirrors
+    // ProviderQuotaOverrideSettings in src/lib/resilience/settings/types.ts.
+    providerQuotaOverrides: z
+      .record(
+        z.string().min(1),
+        z
+          .object({
+            rpm: z.number().int().min(1).optional(),
+            concurrency: z.number().int().min(1).optional(),
+          })
+          .strict()
+      )
+      .optional(),
   })
   .strict()
   .superRefine((value, ctx) => {
@@ -170,7 +185,8 @@ export const updateResilienceSchema = z
       !value.quotaShareConcurrencyLimit &&
       !value.providerCooldown &&
       !value.profiles &&
-      !value.defaults
+      !value.defaults &&
+      !value.providerQuotaOverrides
     ) {
       ctx.addIssue({
         code: z.ZodIssueCode.custom,

tests/unit/resilience-settings-provider-quota-overrides.test.ts

@@ -0,0 +1,142 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import path from "node:path";
+import {
+  DEFAULT_RESILIENCE_SETTINGS,
+  mergeResilienceSettings,
+  resolveResilienceSettings,
+  type ResilienceSettings,
+} from "../../src/lib/resilience/settings.ts";
+import { updateResilienceSchema } from "../../src/shared/validation/schemas.ts";
+
+function cloneDefaults(): ResilienceSettings {
+  return structuredClone(DEFAULT_RESILIENCE_SETTINGS);
+}
+
+// --- Schema (PATCH /api/resilience body) ---
+
+test("updateResilienceSchema accepts providerQuotaOverrides entries", () => {
+  const parsed = updateResilienceSchema.safeParse({
+    providerQuotaOverrides: {
+      minimax: { rpm: 30, concurrency: 4 },
+      nvidia: { rpm: 60 },
+    },
+  });
+  assert.equal(parsed.success, true, "valid override map should parse");
+});
+
+test("updateResilienceSchema allows a body containing only providerQuotaOverrides", () => {
+  // The superRefine requires at least one field; a lone override map is a
+  // legitimate update and must not trigger "Must provide resilience settings".
+  const parsed = updateResilienceSchema.safeParse({
+    providerQuotaOverrides: { minimax: { rpm: 30 } },
+  });
+  assert.equal(parsed.success, true);
+});
+
+test("updateResilienceSchema rejects non-positive rpm/concurrency", () => {
+  assert.equal(
+    updateResilienceSchema.safeParse({
+      providerQuotaOverrides: { minimax: { rpm: 0 } },
+    }).success,
+    false,
+    "rpm must be >= 1"
+  );
+  assert.equal(
+    updateResilienceSchema.safeParse({
+      providerQuotaOverrides: { minimax: { concurrency: -1 } },
+    }).success,
+    false,
+    "concurrency must be >= 1"
+  );
+});
+
+test("updateResilienceSchema rejects unknown keys inside an override entry", () => {
+  assert.equal(
+    updateResilienceSchema.safeParse({
+      providerQuotaOverrides: { minimax: { rpm: 30, windowMs: 60_000 } },
+    }).success,
+    false,
+    "override entries are .strict() — unknown keys rejected"
+  );
+});
+
+// --- mergeResilienceSettings ---
+
+test("mergeResilienceSettings stores provider quota overrides", () => {
+  const next = mergeResilienceSettings(cloneDefaults(), {
+    providerQuotaOverrides: {
+      minimax: { rpm: 30, concurrency: 4 },
+    },
+  });
+  assert.deepEqual(next.providerQuotaOverrides, { minimax: { rpm: 30, concurrency: 4 } });
+});
+
+test("mergeResilienceSettings preserves existing overrides when patch omits the map", () => {
+  const current = mergeResilienceSettings(cloneDefaults(), {
+    providerQuotaOverrides: { nvidia: { rpm: 40 } },
+  });
+  const next = mergeResilienceSettings(current, {
+    requestQueue: { maxWaitMs: 60_000 },
+  });
+  assert.deepEqual(next.providerQuotaOverrides, { nvidia: { rpm: 40 } });
+});
+
+test("mergeResilienceSettings drops invalid override entries via normalization", () => {
+  const next = mergeResilienceSettings(cloneDefaults(), {
+    providerQuotaOverrides: {
+      minimax: { rpm: 0 }, // non-positive → dropped by normalizeProviderQuotaOverrideEntry
+    } as never,
+  });
+  assert.deepEqual(next.providerQuotaOverrides, {});
+});
+
+// --- resolveResilienceSettings round-trip ---
+
+test("resolveResilienceSettings round-trips stored provider quota overrides", () => {
+  const resolved = resolveResilienceSettings({
+    resilienceSettings: {
+      providerQuotaOverrides: {
+        minimax: { rpm: 30, concurrency: 4 },
+      },
+    },
+  });
+  assert.deepEqual(resolved.providerQuotaOverrides, { minimax: { rpm: 30, concurrency: 4 } });
+});
+
+// --- Route wiring (GET/PATCH response + hot reload) ---
+
+const RESILIENCE_ROUTE_PATH = path.resolve(process.cwd(), "src/app/api/resilience/route.ts");
+
+test("GET /api/resilience returns providerQuotaOverrides", () => {
+  const source = fs.readFileSync(RESILIENCE_ROUTE_PATH, "utf8");
+  assert.match(
+    source,
+    /providerQuotaOverrides:\s*resilience\.providerQuotaOverrides\b/,
+    "GET response should expose the resolved override map"
+  );
+});
+
+test("PATCH /api/resilience returns providerQuotaOverrides and propagates it into the merge", () => {
+  const source = fs.readFileSync(RESILIENCE_ROUTE_PATH, "utf8");
+  assert.match(
+    source,
+    /providerQuotaOverrides:\s*nextResilience\.providerQuotaOverrides\b/,
+    "PATCH response should expose the merged override map"
+  );
+  assert.match(
+    source,
+    /body\.providerQuotaOverrides/,
+    "PATCH should read providerQuotaOverrides from the validated body"
+  );
+});
+
+test("syncRuntimeSettings re-applies provider quota overrides on the hot path", () => {
+  const source = fs.readFileSync(RESILIENCE_ROUTE_PATH, "utf8");
+  assert.match(
+    source,
+    /setProviderQuotaOverrides\(resilienceSettings\.providerQuotaOverrides\)/,
+    "PATCH should hot-reload overrides without a process restart"
+  );
+});