fix(combo): keep passthrough quota fallback scoped (#4194)

Integrated into release/v3.8.29 — reconciled onto post-split combo.ts (#4186); RED->GREEN validated. Thanks @Svetznaniy33.

open-sse/services/combo.ts

@@ -16,6 +16,7 @@ import {
   recordModelLockoutFailure,
   recordProviderFailure,
   isProviderExhaustedReason,
+  hasPerModelQuota,
   type ProviderProfile,
 } from "./accountFallback.ts";
 import { FETCH_TIMEOUT_MS, RateLimitReason } from "../config/constants.ts";
@@ -2521,8 +2522,12 @@ export async function handleComboChat({
           // same-provider targets are skipped immediately. API-key 429s still use
           // the short resilience cooldown, but explicit quota text should stop the
           // combo from trying another target for the same provider in this request.
+          // Passthrough/per-model-quota providers multiplex independent upstream
+          // models behind one provider connection; a quota 429 for one model must
+          // not skip fallback targets for another model on the same provider.
           const providerExhausted =
             Boolean(provider && provider !== "unknown") &&
+            !hasPerModelQuota(provider, rawModel) &&
             (isProviderExhaustedReason(fallbackResult) ||
               classifyErrorText(errorText) === RateLimitReason.QUOTA_EXHAUSTED);
           if (providerExhausted) {
@@ -3269,8 +3274,12 @@ async function handleRoundRobinCombo({
         // same-provider targets are skipped immediately. API-key 429s still use
         // the short resilience cooldown, but explicit quota text should stop the
         // combo from trying another target for the same provider in this request.
+        // Passthrough/per-model-quota providers multiplex independent upstream
+        // models behind one provider connection; a quota 429 for one model must
+        // not skip fallback targets for another model on the same provider.
         const providerExhausted =
           Boolean(provider && provider !== "unknown") &&
+          !hasPerModelQuota(provider, parseModel(modelStr).model || modelStr) &&
           (isProviderExhaustedReason(fallbackResult) ||
             classifyErrorText(errorText) === RateLimitReason.QUOTA_EXHAUSTED ||
             isAllAccountsRateLimited);

tests/unit/combo-prescreen.test.ts

@@ -203,3 +203,41 @@ test("pre-screen: backward compatible with all targets available", async () => {
   assert.equal(calls.length, 1);
   assert.equal(calls[0], "p1/m1");
 });
+
+test("priority combo: quota 429 on passthrough provider does not skip another model on same provider", async () => {
+  const calls: string[] = [];
+
+  const combo = await combosDb.createCombo({
+    name: "passthrough-quota-scope",
+    strategy: "priority",
+    models: ["antigravity/claude-opus-4-6-thinking", "antigravity/gemini-3-flash-agent"],
+  });
+
+  const response = await handleComboChat({
+    body: { ...reqBody, model: combo.name },
+    combo,
+    allCombos: [combo],
+    isModelAvailable: async () => true,
+    relayOptions: undefined,
+    signal: undefined,
+    settings: {},
+    log: makeLog(),
+    handleSingleModel: async (_body: unknown, modelStr: string) => {
+      calls.push(modelStr);
+      if (modelStr.includes("claude-opus")) {
+        return Response.json(
+          { error: { message: "quota exhausted for claude-opus-4-6-thinking" } },
+          { status: 429 }
+        );
+      }
+      return okResponse(modelStr);
+    },
+  });
+
+  assert.equal(response.status, 200);
+  assert.equal(calls.at(-1), "antigravity/gemini-3-flash-agent");
+  assert.ok(
+    calls.includes("antigravity/claude-opus-4-6-thinking"),
+    "first passthrough model should be attempted before fallback"
+  );
+});