fix(failover): classify upstream_error errorType as server_error for model fallback

secooond · secooond · commit c83f1eafb7be · 2026-06-23T21:03:26.000+08:00
When a provider returns an error with type:upstream_error (e.g. OpenAI responses API), classifyFailoverSignal returns null because the errorType is not recognized, so model fallback never triggers and the turn ends with 'LLM request failed.' Add upstream_error → server_error classification in classifyFailoverSignal so configured fallback models are attempted on transient provider-side failures. Closes #95519.
diff --git a/src/agents/embedded-agent-helpers/errors.test.ts b/src/agents/embedded-agent-helpers/errors.test.ts
@@ -98,6 +98,8 @@ describe("formatAssistantErrorText streaming JSON parse classification", () => {
   });
 });
 
+import { classifyFailoverSignal } from "./errors.js";
+
 describe("isLikelyContextOverflowError", () => {
   it("detects Codex promptError wording for a full context window", () => {
     expect(
@@ -107,3 +109,28 @@ describe("isLikelyContextOverflowError", () => {
     ).toBe(true);
   });
 });
+
+describe("classifyFailoverSignal upstream_error", () => {
+  it("classifies upstream_error errorType as server_error for model fallback (regression for #95519)", () => {
+    const classification = classifyFailoverSignal({
+      message: '{"error":{"type":"upstream_error","message":"Upstream request failed"}}',
+      errorType: "upstream_error",
+      code: null as unknown as undefined,
+      provider: "openai",
+    });
+    expect(classification).toEqual({
+      kind: "reason",
+      reason: "server_error",
+    });
+  });
+
+  it("does not classify unhandled error messages as server_error", () => {
+    const classification = classifyFailoverSignal({
+      message: "Some other error",
+      provider: "openai",
+    });
+    // Without errorType, upstream_error, or other recognized patterns,
+    // classification falls through and returns null.
+    expect(classification).toBeNull();
+  });
+});
diff --git a/src/agents/embedded-agent-helpers/errors.ts b/src/agents/embedded-agent-helpers/errors.ts
@@ -1176,6 +1176,17 @@ export function classifyFailoverSignal(signal: FailoverSignal): FailoverClassifi
           errorType: signal.errorType,
         })
       : null;
+  // upstream_error is a provider-side transient failure (e.g. OpenAI
+  // responses API returns {"error":{"type":"upstream_error"}}).  Classify
+  // it as server_error so model fallback triggers instead of ending the
+  // turn with "LLM request failed." (issue #95519).
+  if (
+    !providerPluginReason &&
+    signal.errorType != null &&
+    normalizeLowercaseStringOrEmpty(signal.errorType) === "upstream_error"
+  ) {
+    return toReasonClassification("server_error");
+  }
   const effectiveMessageClassification = providerPluginReason
     ? toReasonClassification(providerPluginReason)
     : mergeMessageAndDetailClassification(messageClassification, detailClassification);
