fix(failover): classify transient Undici error codes (#123042)

* fix(failover): classify transient Undici error codes

* fix(failover): share transient network code policy

Signed-off-by: sallyom <somalley@redhat.com>

---------

Signed-off-by: sallyom <somalley@redhat.com>
Co-authored-by: sallyom <somalley@redhat.com>

src/agents/embedded-agent-helpers/assistant-message-failures.test.ts

@@ -25,6 +25,25 @@ describe("classifyAssistantFailoverReason", () => {
     expect(classifyAssistantFailoverReason(opencodeGoStalledStreamError)).toBe("timeout");
   });
 
+  it.each([
+    "ENOTFOUND",
+    "UND_ERR_CONNECT_TIMEOUT",
+    "UND_ERR_DNS_RESOLVE_FAILED",
+    "UND_ERR_CONNECT",
+    "UND_ERR_SOCKET",
+    "UND_ERR_HEADERS_TIMEOUT",
+    "UND_ERR_BODY_TIMEOUT",
+  ])("classifies structured %s assistant errors as timeouts", (errorCode) => {
+    expect(
+      classifyAssistantFailoverReason({
+        ...opencodeGoStalledStreamError,
+        provider: "demo-provider",
+        errorCode,
+        errorMessage: "provider connection closed",
+      }),
+    ).toBe("timeout");
+  });
+
   it("does not classify caller-aborted assistant messages as provider failover", () => {
     expect(
       classifyAssistantFailoverReason({

src/agents/failover/classification-rules.ts

@@ -1,4 +1,5 @@
 import { normalizeOptionalLowercaseString } from "@openclaw/normalization-core/string-coerce";
+import { isTransientNetworkError } from "../../infra/retryable-network-errors.js";
 import {
   extractLeadingHttpStatus,
   parseApiErrorInfo,
@@ -9,18 +10,9 @@ import {
   isRateLimitErrorMessage,
 } from "./message-patterns.js";
 import type { FailoverClassification, FailoverReason, FailoverSignal } from "./signal.js";
-const TIMEOUT_ERROR_CODES = new Set([
-  "ETIMEDOUT",
-  "ESOCKETTIMEDOUT",
-  "ECONNRESET",
-  "ECONNABORTED",
-  "ECONNREFUSED",
-  "ENETUNREACH",
-  "EHOSTUNREACH",
+const FAILOVER_TIMEOUT_ERROR_CODES = new Set([
   "EHOSTDOWN",
   "ENETRESET",
-  "EPIPE",
-  "EAI_AGAIN",
   "ERR_STREAM_PREMATURE_CLOSE",
 ]);
 const NO_BODY_HTTP_WRAPPER_RE =
@@ -342,7 +334,10 @@ export function classifyFailoverReasonFromCode(raw: string | undefined): Failove
     case "OVERLOADED_ERROR":
       return "overloaded";
     default:
-      return TIMEOUT_ERROR_CODES.has(normalized) ? "timeout" : null;
+      return FAILOVER_TIMEOUT_ERROR_CODES.has(normalized) ||
+        isTransientNetworkError({ code: normalized })
+        ? "timeout"
+        : null;
   }
 }
 export function classifyCoreFailoverReasonFromErrorType(

src/agents/failover/retry-evidence.ts

@@ -1,5 +1,6 @@
 import { parseRetryAfterHttpDateMs } from "@openclaw/ai/internal/retry-after";
 import milliseconds from "ms";
+import { isTransientNetworkError } from "../../infra/retryable-network-errors.js";
 import {
   extractErrorHttpStatus,
   extractLeadingHttpStatus,
@@ -48,12 +49,13 @@ function resolveRetrySignalStatus(signal: Pick<FailoverSignal, "message" | "stat
 
 /** Narrow evidence that replaying the same assistant request may succeed within this session. */
 export function hasTransientRetryEvidence(
-  signal: Pick<FailoverSignal, "message" | "status">,
+  signal: Pick<FailoverSignal, "code" | "message" | "status">,
 ): boolean {
   const status = resolveRetrySignalStatus(signal);
   return (
     (status !== undefined && RETRYABLE_HTTP_STATUS_CODES.has(status)) ||
-    TRANSIENT_RETRY_EVIDENCE_RE.test(signal.message ?? "")
+    TRANSIENT_RETRY_EVIDENCE_RE.test(signal.message ?? "") ||
+    isTransientNetworkError({ code: signal.code })
   );
 }
 

src/llm/utils/retry.test.ts

@@ -55,6 +55,15 @@ describe("isRetryableAssistantError", () => {
     ).toBe(false);
   });
 
+  it("retries a structured transient Undici error", () => {
+    expect(
+      isRetryableAssistantError({
+        ...errorMessage("provider connection closed"),
+        errorCode: "UND_ERR_HEADERS_TIMEOUT",
+      }),
+    ).toBe(true);
+  });
+
   it.each([
     "An error occurred while processing your request. You can retry your request.",
     "The system encountered an unexpected error. Try your request again.",

src/llm/utils/retry.ts

@@ -36,6 +36,8 @@ export function isRetryableAssistantError(message: AssistantMessage): boolean {
   const signal = {
     message: errorMessage,
     provider: message.provider,
+    code: message.errorCode,
+    errorType: message.errorType,
     ...(status === undefined ? {} : { status }),
   };
   const classification = classifyFailoverSignal(signal);