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

Claude Code surfaces an SSE stream that dies after content started as
"API Error: Connection closed mid-response. The response above may be
incomplete." classifyAgentError did not recognize that string (nor the
transport-level "socket connection was closed" variants), so these
transient transport failures fell through to agent_error and hard-failed
the whole task run with a raw error message instead of being treated as
retriable upstream failures.

Generated-By: PostHog Code
Task-Id: a04c4d02-9070-4e5d-9748-bfceb251e3fa

Author: tatoalo

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

@@ -0,0 +1,27 @@
+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",
+    ],
+    ["API Error: 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 +23,18 @@ 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.
+  if (/API Error:.*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";
   }