fix(combos): default reasoning token buffer off

open-sse/services/combo.ts

@@ -2431,7 +2431,7 @@ async function handleRoundRobinCombo({
   const maxRetries = config.maxRetries ?? 1;
   const retryDelayMs = resolveDelayMs(config.retryDelayMs, 2000);
   const fallbackDelayMs = resolveDelayMs(config.fallbackDelayMs, 0);
-  const reasoningTokenBufferEnabled = config.reasoningTokenBufferEnabled !== false;
+  const reasoningTokenBufferEnabled = config.reasoningTokenBufferEnabled === true;
 
   const resilienceSettings: ResilienceSettings = settings
     ? resolveResilienceSettings(settings)
@@ -3032,7 +3032,10 @@ async function handleRoundRobinCombo({
           resilienceSettings.providerCooldown.enabled &&
           provider &&
           provider !== "unknown" &&
-          !(result.status === 500 && hasPerModelQuota(provider, parseModel(modelStr).model || modelStr))
+          !(
+            result.status === 500 &&
+            hasPerModelQuota(provider, parseModel(modelStr).model || modelStr)
+          )
         ) {
           recordProviderCooldown(
             provider,

open-sse/services/combo/comboSetup.ts

@@ -90,9 +90,7 @@ export function phaseComboSetup(ctx: ComboContext): ComboSetup {
 
   const universalHandoffConfig = resolveUniversalHandoffConfig(
     (combo.universal_handoff || combo.universalHandoff) as
-      | Record<string, unknown>
-      | null
-      | undefined,
+      Record<string, unknown> | null | undefined,
     relayOptions?.universalHandoffConfig as Record<string, unknown> | null | undefined
   );
 
@@ -118,7 +116,7 @@ export function phaseComboSetup(ctx: ComboContext): ComboSetup {
     FETCH_TIMEOUT_MS,
     DEFAULT_COMBO_TARGET_TIMEOUT_MS
   );
-  const reasoningTokenBufferEnabled = config.reasoningTokenBufferEnabled !== false;
+  const reasoningTokenBufferEnabled = config.reasoningTokenBufferEnabled === true;
 
   return {
     strategy,

open-sse/services/comboConfig.ts

@@ -53,7 +53,7 @@ const DEFAULT_COMBO_CONFIG = {
   maxComboDepth: 3,
   nestedComboMode: "flatten",
   trackMetrics: true,
-  reasoningTokenBufferEnabled: true,
+  reasoningTokenBufferEnabled: false,
   manifestRouting: false,
   // Complexity-aware auto routing (2026): when on, the auto router scores
   // candidates by how well their tier matches the request's classified

src/app/(dashboard)/dashboard/settings/components/ComboDefaultsTab.tsx

@@ -93,7 +93,7 @@ export default function ComboDefaultsTab() {
     retryDelayMs: 2000,
     maxComboDepth: 3,
     trackMetrics: true,
-    reasoningTokenBufferEnabled: true,
+    reasoningTokenBufferEnabled: false,
     handoffThreshold: 0.85,
     handoffModel: "",
     maxMessagesForSummary: 30,
@@ -292,8 +292,7 @@ export default function ComboDefaultsTab() {
 
   // Filtered provider list — excludes already-added ones, filtered by search query
   const filteredProviders = availableProviders.filter(
-    (p) =>
-      !providerOverrides[p.provider] && matchesSearch(p.provider, searchQuery)
+    (p) => !providerOverrides[p.provider] && matchesSearch(p.provider, searchQuery)
   );
 
   const handleDropdownKeyDown = (e: React.KeyboardEvent) => {
@@ -690,11 +689,11 @@ export default function ComboDefaultsTab() {
               </p>
             </div>
             <Toggle
-              checked={comboDefaults.reasoningTokenBufferEnabled !== false}
+              checked={comboDefaults.reasoningTokenBufferEnabled === true}
               onChange={() =>
                 setComboDefaults((prev) => ({
                   ...prev,
-                  reasoningTokenBufferEnabled: prev.reasoningTokenBufferEnabled === false,
+                  reasoningTokenBufferEnabled: prev.reasoningTokenBufferEnabled !== true,
                 }))
               }
             />

src/app/api/settings/combo-defaults/route.ts

@@ -55,7 +55,7 @@ export async function GET(request: Request) {
               maxMessagesForSummary: 30,
               maxComboDepth: 3,
               trackMetrics: true,
-              reasoningTokenBufferEnabled: true,
+              reasoningTokenBufferEnabled: false,
               zeroLatencyOptimizationsEnabled: false,
             },
       providerOverrides,

tests/unit/combo-config.test.ts

@@ -30,7 +30,7 @@ test("getDefaultComboConfig returns a fresh copy of the defaults", () => {
   assert.equal(first.failoverBeforeRetry, true);
   assert.equal(first.maxSetRetries, 0);
   assert.equal(first.setRetryDelayMs, 2000);
-  assert.equal(first.reasoningTokenBufferEnabled, true);
+  assert.equal(first.reasoningTokenBufferEnabled, false);
   assert.equal(first.zeroLatencyOptimizationsEnabled, false);
   assert.equal(first.hedging, false);
   assert.equal(first.fallbackCompressionMode, "lite");

tests/unit/combo-routing-engine.test.ts

@@ -2869,7 +2869,7 @@ test("handleComboChat aborts combo when 503 response does NOT contain the unavai
   );
 });
 
-test("#3587 reasoning model gets max_tokens buffer applied", async () => {
+test("#3587 reasoning model gets max_tokens buffer applied when explicitly enabled", async () => {
   saveModelsDevCapabilities({
     openai: {
       "gpt-4o-reasoning": capabilityEntry(12000, { reasoning: true, limit_output: 12000 }),
@@ -2882,6 +2882,7 @@ test("#3587 reasoning model gets max_tokens buffer applied", async () => {
     combo: {
       name: "reasoning-buffer",
       models: ["openai/gpt-4o-reasoning"],
+      config: { reasoningTokenBufferEnabled: true },
     },
     handleSingleModel: async (body: Record<string, unknown>) => {
       bodies.push(body);
@@ -2934,6 +2935,7 @@ test("#3587 reasoning buffer preserves max_tokens when the full buffer exceeds m
     combo: {
       name: "reasoning-buffer-near-cap",
       models: ["openai/gemini-high-cap"],
+      config: { reasoningTokenBufferEnabled: true },
     },
     handleSingleModel: async (body: Record<string, unknown>) => {
       bodies.push(body);
@@ -2983,7 +2985,7 @@ test("#3587 reasoning buffer is disabled without explicit model capability data"
   );
 });
 
-test("#3588 reasoning token buffer feature flag preserves client max_tokens", async () => {
+test("#3588 reasoning token buffer is disabled by default and preserves client max_tokens", async () => {
   saveModelsDevCapabilities({
     openai: {
       "flagged-reasoning": capabilityEntry(12000, { reasoning: true, limit_output: 12000 }),
@@ -3009,11 +3011,7 @@ test("#3588 reasoning token buffer feature flag preserves client max_tokens", as
     },
     isModelAvailable: async () => true,
     log: createLog(),
-    settings: {
-      comboDefaults: {
-        reasoningTokenBufferEnabled: false,
-      },
-    },
+    settings: null,
     relayOptions: null,
     allCombos: null,
   });
@@ -3078,6 +3076,7 @@ test("#3587 round-robin buffer does NOT compound across reasoning models", async
       name: "rr-reasoning-no-compound",
       strategy: "round-robin",
       models: ["openai/rr-reasoning-a", "openai/rr-reasoning-b"],
+      config: { reasoningTokenBufferEnabled: true },
     },
     handleSingleModel: async (body: Record<string, unknown>, modelStr: string) => {
       seen.push({ model: modelStr, maxTokens: body.max_tokens });

tests/unit/combo/combo-context.test.ts

@@ -15,7 +15,11 @@ test("createComboContext carries inputs and the body BY REFERENCE", () => {
   const body = { model: "auto", messages: [], stream: true };
   const combo = { name: "c1", models: ["a", "b"] };
   const ctx = createComboContext({ body, combo, log });
-  assert.equal(ctx.body, body, "body must be the same reference (not copied) for byte-identical pinning");
+  assert.equal(
+    ctx.body,
+    body,
+    "body must be the same reference (not copied) for byte-identical pinning"
+  );
   assert.equal(ctx.combo, combo);
   assert.equal(ctx.settings, null);
   assert.equal(ctx.relayOptions, null);
@@ -34,7 +38,11 @@ test("phaseComboSetup resolves strategy/config/stream with pinning OFF (pure, no
   assert.equal(setup.effectiveSessionId, null);
   assert.equal(setup.clientRequestedStream, true, "body.stream === true");
   assert.equal(typeof setup.comboTargetTimeoutMs, "number");
-  assert.equal(typeof setup.reasoningTokenBufferEnabled, "boolean");
+  assert.equal(
+    setup.reasoningTokenBufferEnabled,
+    false,
+    "reasoning token buffering is opt-in by default"
+  );
   assert.ok(setup.config && typeof setup.config === "object", "config cascade resolved");
   assert.ok(
     setup.resilienceSettings && typeof setup.resilienceSettings === "object",