From ec70bbb4b91ea661380d28a77f78a790705f6e36 Mon Sep 17 00:00:00 2001
From: Eason Goodale <easong@openai.com>
Date: Fri, 18 Apr 2025 03:50:27 -0700
Subject: [PATCH] handle more API errors

Signed-off-by: Eason Goodale <easong@openai.com>
---
 codex-cli/src/utils/agent/agent-loop.ts       |   8 ++
 .../tests/agent-server-error-stream.test.ts   | 111 ++++++++++++++++++
 2 files changed, 119 insertions(+)
 create mode 100644 codex-cli/tests/agent-server-error-stream.test.ts

diff --git a/codex-cli/src/utils/agent/agent-loop.ts b/codex-cli/src/utils/agent/agent-loop.ts
index b9105f64a7..4508553a1b 100644
--- a/codex-cli/src/utils/agent/agent-loop.ts
+++ b/codex-cli/src/utils/agent/agent-loop.ts
@@ -955,6 +955,14 @@ export class AgentLoop {
           return true;
         }
 
+        // Explicit check for OpenAI "server_error" types which are surfaced
+        // when the backend encounters an unexpected exception. The SDK often
+        // omits the HTTP status in this case (leaving it undefined) so we
+        // must inspect the structured error fields instead.
+        if (e.type === "server_error" || e.code === "server_error") {
+          return true;
+        }
+
         if (typeof e.status === "number" && e.status >= 500) {
           return true;
         }
diff --git a/codex-cli/tests/agent-server-error-stream.test.ts b/codex-cli/tests/agent-server-error-stream.test.ts
new file mode 100644
index 0000000000..9d66f8facd
--- /dev/null
+++ b/codex-cli/tests/agent-server-error-stream.test.ts
@@ -0,0 +1,111 @@
+import { describe, it, expect, vi } from "vitest";
+
+// ---------------------------------------------------------------------------
+// Utility helpers & OpenAI mock – tailored for server‑side errors that occur
+// *after* the streaming iterator was created (i.e. during iteration).
+// ---------------------------------------------------------------------------
+
+function createStreamThatErrors(err: Error) {
+  return new (class {
+    public controller = { abort: vi.fn() };
+
+    async *[Symbol.asyncIterator]() {
+      // Immediately raise the error once iteration starts – mimics OpenAI SDK
+      // behaviour which throws from the iterator when the HTTP response status
+      // indicates an internal server failure.
+      throw err;
+    }
+  })();
+}
+
+// Spy holder swapped out per test case.
+const openAiState: { createSpy?: ReturnType<typeof vi.fn> } = {};
+
+vi.mock("openai", () => {
+  class FakeOpenAI {
+    public responses = {
+      create: (...args: Array<any>) => openAiState.createSpy!(...args),
+    };
+  }
+
+  class APIConnectionTimeoutError extends Error {}
+
+  return {
+    __esModule: true,
+    default: FakeOpenAI,
+    APIConnectionTimeoutError,
+  };
+});
+
+// Approvals / formatting stubs – not part of the behaviour under test.
+vi.mock("../src/approvals.js", () => ({
+  __esModule: true,
+  alwaysApprovedCommands: new Set<string>(),
+  canAutoApprove: () => ({ type: "auto-approve", runInSandbox: false } as any),
+  isSafeCommand: () => null,
+}));
+
+vi.mock("../src/format-command.js", () => ({
+  __esModule: true,
+  formatCommandForDisplay: (c: Array<string>) => c.join(" "),
+}));
+
+// Silence debug logging so the test output stays uncluttered.
+vi.mock("../src/utils/agent/log.js", () => ({
+  __esModule: true,
+  log: () => {},
+  isLoggingEnabled: () => false,
+}));
+
+import { AgentLoop } from "../src/utils/agent/agent-loop.js";
+
+describe("AgentLoop – server_error surfaced during streaming", () => {
+  it("shows user‑friendly system message instead of crashing", async () => {
+    const apiErr: any = new Error(
+      "The server had an error while processing your request. Sorry about that!",
+    );
+    // Replicate the structure used by the OpenAI SDK for 5xx failures.
+    apiErr.type = "server_error";
+    apiErr.code = null;
+    apiErr.status = undefined; // SDK leaves status undefined in this pathway
+
+    openAiState.createSpy = vi.fn(async () => {
+      return createStreamThatErrors(apiErr);
+    });
+
+    const received: Array<any> = [];
+
+    const agent = new AgentLoop({
+      model: "any",
+      instructions: "",
+      approvalPolicy: { mode: "auto" } as any,
+      additionalWritableRoots: [],
+      onItem: (i) => received.push(i),
+      onLoading: () => {},
+      getCommandConfirmation: async () => ({ review: "yes" } as any),
+      onLastResponseId: () => {},
+    });
+
+    const userMsg = [
+      {
+        type: "message",
+        role: "user",
+        content: [{ type: "input_text", text: "ping" }],
+      },
+    ];
+
+    await expect(agent.run(userMsg as any)).resolves.not.toThrow();
+
+    // allow async onItem deliveries to flush
+    await new Promise((r) => setTimeout(r, 20));
+
+    const sysMsg = received.find(
+      (i) =>
+        i.role === "system" &&
+        typeof i.content?.[0]?.text === "string" &&
+        i.content[0].text.includes("Network error"),
+    );
+
+    expect(sysMsg).toBeTruthy();
+  });
+});