fix(llm): cap codex retry delays

src/llm/providers/openai-codex-responses.test.ts

@@ -304,4 +304,40 @@ describe("streamOpenAICodexResponses transport", () => {
       expect(setTimeoutSpy).toHaveBeenCalledWith(expect.any(Function), 1000);
     },
   );
+
+  it("caps oversized Retry-After delays before sleeping", async () => {
+    const fetchMock = vi
+      .fn<typeof fetch>()
+      .mockResolvedValueOnce(
+        new Response("rate limited", {
+          status: 429,
+          headers: { "retry-after": String(Number.MAX_SAFE_INTEGER) },
+        }),
+      )
+      .mockRejectedValueOnce(new Error("usage limit: stop after retry delay"));
+    vi.stubGlobal("fetch", fetchMock);
+    const setTimeoutSpy = vi
+      .spyOn(globalThis, "setTimeout")
+      .mockImplementation((callback: TimerHandler) => {
+        if (typeof callback === "function") {
+          callback();
+        }
+        return 0 as unknown as ReturnType<typeof setTimeout>;
+      });
+
+    const stream = streamOpenAICodexResponses(model, context, {
+      apiKey: createJwt({
+        "https://api.openai.com/auth": {
+          chatgpt_account_id: "acct-1",
+        },
+      }),
+      transport: "sse",
+    });
+
+    const result = await stream.result();
+
+    expect(result.stopReason).toBe("error");
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+    expect(setTimeoutSpy).toHaveBeenCalledWith(expect.any(Function), MAX_TIMER_TIMEOUT_MS);
+  });
 });

src/llm/providers/openai-codex-responses.ts

@@ -20,7 +20,7 @@ if (typeof process !== "undefined" && (process.versions?.node || process.version
   });
 }
 
-import { resolveTimerTimeoutMs } from "../../shared/number-coercion.js";
+import { resolveTimerTimeoutMs, clampTimerTimeoutMs } from "../../shared/number-coercion.js";
 import { getEnvApiKey } from "../env-api-keys.js";
 import { clampThinkingLevel } from "../model-utils.js";
 import { registerSessionResourceCleanup } from "../session-resources.js";
@@ -343,7 +343,7 @@ export const streamOpenAICodexResponses: StreamFunction<
               const trimmedRetryAfterMs = retryAfterMs.trim();
               const millis = Number(trimmedRetryAfterMs);
               if (/^\d+(?:\.\d+)?$/.test(trimmedRetryAfterMs) && Number.isFinite(millis)) {
-                delayMs = Math.max(0, millis);
+                delayMs = clampTimerTimeoutMs(millis, 0) ?? delayMs;
               }
             } else {
               const retryAfter = response.headers.get("retry-after");
@@ -351,11 +351,11 @@ export const streamOpenAICodexResponses: StreamFunction<
                 const trimmedRetryAfter = retryAfter.trim();
                 const seconds = Number(trimmedRetryAfter);
                 if (/^\d+$/.test(trimmedRetryAfter) && Number.isFinite(seconds)) {
-                  delayMs = Math.max(0, seconds * 1000);
+                  delayMs = clampTimerTimeoutMs(seconds * 1000, 0) ?? delayMs;
                 } else if (RETRY_AFTER_HTTP_DATE_RE.test(trimmedRetryAfter)) {
                   const date = Date.parse(trimmedRetryAfter);
                   if (!Number.isNaN(date)) {
-                    delayMs = Math.max(0, date - Date.now());
+                    delayMs = clampTimerTimeoutMs(date - Date.now(), 0) ?? delayMs;
                   }
                 }
               }