fix(agent): classify mid-response stream deaths as transient upstream failures

packages/agent/src/adapters/error-classification.test.ts

@@ -0,0 +1,34 @@
+import { describe, expect, it } from "vitest";
+import { classifyAgentError } from "./error-classification";
+
+describe("classifyAgentError", () => {
+  it.each([
+    ["API Error: terminated", "upstream_stream_terminated"],
+    [
+      "API Error: Connection closed mid-response. The response above may be incomplete.",
+      "upstream_stream_terminated",
+    ],
+    [
+      "API Error: The socket connection was closed unexpectedly.",
+      "upstream_stream_terminated",
+    ],
+    [
+      "The socket connection was closed unexpectedly. For more information, pass `verbose: true`",
+      "upstream_stream_terminated",
+    ],
+    ["socket connection closed", "upstream_stream_terminated"],
+    ["API Error: Connection error.", "upstream_connection_error"],
+    ["API Error: Request timed out.", "upstream_timeout"],
+    ["API Error: 429 rate limited", "upstream_provider_failure"],
+    ["API Error: 529 overloaded", "upstream_provider_failure"],
+    ["API Error: 400 invalid request", "agent_error"],
+    [
+      "Connection closed mid-response without the API Error prefix",
+      "agent_error",
+    ],
+    ["some unrelated failure", "agent_error"],
+    [undefined, "agent_error"],
+  ] as const)("classifies %j as %s", (message, expected) => {
+    expect(classifyAgentError(message)).toBe(expected);
+  });
+});

packages/agent/src/adapters/error-classification.ts

@@ -23,6 +23,19 @@ export function classifyAgentError(
   if (/API Error:\s*terminated\b/i.test(text)) {
     return "upstream_stream_terminated";
   }
+  // Claude Code surfaces an SSE stream that dies after content started
+  // (no message_stop) as "Connection closed mid-response". Seen when a
+  // gateway pod is replaced mid-stream or an intermediary cuts the socket
+  // during a long silent stretch.
+  if (/API Error:.*Connection closed mid-response/i.test(text)) {
+    return "upstream_stream_terminated";
+  }
+  // Transport-level socket deaths reported by fetch implementations
+  // (Bun/undici wording varies) — same failure mode as above. These are raw
+  // transport errors, so they don't always carry the "API Error:" prefix.
+  if (/socket connection (?:was )?closed/i.test(text)) {
+    return "upstream_stream_terminated";
+  }
   if (/API Error:\s*Connection error\b/i.test(text)) {
     return "upstream_connection_error";
   }

packages/shared/src/errors.test.ts

@@ -133,6 +133,9 @@ describe("isTransientUpstreamError", () => {
     "API Error: 500 internal server error",
     "API Error: 529 overloaded_error",
     "Internal error: API Error: request timed out",
+    "Internal error: API Error: Connection closed mid-response. The response above may be incomplete.",
+    "The socket connection was closed unexpectedly.",
+    "socket connection closed",
   ])("recognises %j", (message) => {
     expect(isTransientUpstreamError(message)).toBe(true);
   });

packages/shared/src/errors.ts

@@ -94,6 +94,10 @@ const FATAL_SESSION_ERROR_PATTERNS = [
 const UPSTREAM_TRANSIENT_ERROR_REGEXES = [
   /API Error:\s*terminated\b/i,
   /API Error:\s*Connection error\b/i,
+  /API Error:.*Connection closed mid-response/i,
+  // Raw transport-level socket death — wording varies by fetch
+  // implementation and doesn't always carry the "API Error:" prefix.
+  /socket connection (?:was )?closed/i,
   /API Error:.*\b(?:timed out|timeout)\b/i,
   /API Error:\s*(?:429|5\d\d)\b/i,
 ] as const;