fix(daemon): retry Anthropic stream aborted by transport (LUM-1537)

assistant/src/__tests__/provider-error-scenarios.test.ts

@@ -753,6 +753,117 @@ describe("RetryProvider — streaming response handling", () => {
     expect(callCount).toBe(2);
   });
 
+  test("retries transport-aborted stream (Anthropic 'Request was aborted' with no abortReason)", async () => {
+    let callCount = 0;
+    const inner: Provider = {
+      name: "retry-transport-abort",
+      async sendMessage() {
+        callCount++;
+        if (callCount <= 1) {
+          // Mirrors the ProviderError shape produced by the catch-site in
+          // providers/anthropic/client.ts when the SDK reports
+          // ``Anthropic.APIError(status === undefined, message: "Request
+          // was aborted.")`` and the daemon's AbortController was NOT the
+          // cause (i.e. abortReason is undefined). Empirically the #1
+          // daemon error by a factor of 5x — 1,344 events in 4d on the
+          // SSE chat path, all of which used to surface as a 45s blank
+          // screen on the web client via LUM-1431.
+          throw new ProviderError(
+            "Anthropic API error: Request was aborted.",
+            "anthropic",
+            undefined,
+          );
+        }
+        return successResponse();
+      },
+    };
+    const provider = new RetryProvider(inner);
+    await provider.sendMessage(MESSAGES);
+    expect(callCount).toBe(2);
+  });
+
+  test("does NOT retry caller-aborted stream (abortReason set short-circuits retry)", async () => {
+    let callCount = 0;
+    const abortError = new ProviderError(
+      "Anthropic API error: Request was aborted.",
+      "anthropic",
+      undefined,
+      // Tagging abortReason exactly matches what the catch-site does when
+      // signal.aborted was true at the moment of failure — i.e. the
+      // daemon (or the user) cancelled the request, not the transport.
+      // The retry layer must respect this and surface the error
+      // immediately without consuming retry budget.
+      { abortReason: "user-cancelled" },
+    );
+    const inner: Provider = {
+      name: "caller-abort",
+      async sendMessage() {
+        callCount++;
+        throw abortError;
+      },
+    };
+    const provider = new RetryProvider(inner);
+    await expect(provider.sendMessage(MESSAGES)).rejects.toBe(abortError);
+    expect(callCount).toBe(1);
+  });
+
+  test("does NOT retry inner-timeout stream (deterministic 30min deadline failure)", async () => {
+    let callCount = 0;
+    // When the inner streamTimeoutMs fires, the catch-site rewrites the
+    // message to "Anthropic stream timed out after Xs (inner
+    // streamTimeoutMs)" instead of "Request was aborted." That rewrite is
+    // what allows this branch to bypass the transport-abort retry —
+    // retrying a 30min-deadline failure would almost certainly hit the
+    // same deadline on the next attempt and waste retry budget.
+    const innerTimeoutError = new ProviderError(
+      "Anthropic API error: Anthropic stream timed out after 1800s (inner streamTimeoutMs)",
+      "anthropic",
+      undefined,
+    );
+    const inner: Provider = {
+      name: "inner-timeout",
+      async sendMessage() {
+        callCount++;
+        throw innerTimeoutError;
+      },
+    };
+    const provider = new RetryProvider(inner);
+    await expect(provider.sendMessage(MESSAGES)).rejects.toBe(
+      innerTimeoutError,
+    );
+    expect(callCount).toBe(1);
+  });
+
+  test("does NOT retry OpenAI/Gemini-shaped 'Request was aborted' (no inner-timeout rewrite at those catch-sites)", async () => {
+    // The OpenAI chat-completions, OpenAI responses, and Gemini catch-sites
+    // format their errors as `"<Provider> API error (undefined): Request
+    // was aborted."` (note the `(undefined)` parenthetical that the
+    // Anthropic catch-site intentionally omits) and — unlike the Anthropic
+    // catch-site — they do NOT rewrite their inner-streamTimeoutMs
+    // deadline failures. A provider-agnostic transport-abort predicate
+    // would burn three retries on what is by construction a deterministic
+    // 30-minute deadline failure that will fire again on every attempt.
+    // Scoping the predicate to the Anthropic message prefix avoids that
+    // wasted retry budget for non-Anthropic providers until their
+    // catch-sites grow the same `innerTimeoutFired` distinction.
+    const openaiAbortError = new ProviderError(
+      "OpenAI API error (undefined): Request was aborted.",
+      "openai",
+      undefined,
+    );
+    let callCount = 0;
+    const inner: Provider = {
+      name: "openai-aborted-stream",
+      async sendMessage() {
+        callCount++;
+        throw openaiAbortError;
+      },
+    };
+    const provider = new RetryProvider(inner);
+    await expect(provider.sendMessage(MESSAGES)).rejects.toBe(openaiAbortError);
+    expect(callCount).toBe(1);
+  });
+
   test("events accumulate across retries (each attempt delivers events independently)", async () => {
     let callCount = 0;
     const inner: Provider = {

assistant/src/providers/retry.ts

@@ -72,6 +72,40 @@ const RETRYABLE_STREAM_PATTERNS = [
  */
 const RETRYABLE_PROVIDER_MESSAGE_PATTERNS = [/overloaded/i];
 
+/**
+ * Patterns that indicate the Anthropic provider SDK reported a transport-level
+ * abort (TCP close mid-stream, edge LB idle cutoff, Bun fetch deadline) rather
+ * than a caller-initiated cancellation or inner-timeout deadline. The SDK
+ * surfaces all three cases as ``Request was aborted`` with ``error.status ===
+ * undefined``; the catch-site in ``providers/anthropic/client.ts`` separates
+ * them by:
+ *   - tagging caller cancellations with ``abortReason`` (short-circuits in
+ *     {@link isRetryableError} before reaching this predicate)
+ *   - rewriting the inner-timeout message to ``"Anthropic stream timed out
+ *     after Xs (inner streamTimeoutMs)"`` (doesn't start with ``Anthropic API
+ *     error:`` so it falls through to network-error classification)
+ *   - leaving the transport-abort message verbatim as
+ *     ``"Anthropic API error: Request was aborted."``
+ *
+ * Pattern is intentionally anchored to the Anthropic-specific message prefix.
+ * The OpenAI / Gemini / OpenRouter catch-sites format their errors as
+ * ``"<Provider> API error (undefined): Request was aborted."`` (note the
+ * ``(undefined)`` parenthetical) and — crucially — do **not** rewrite
+ * inner-timeout failures, so a provider-agnostic ``/request was aborted/i``
+ * predicate would erroneously retry their 30-minute deadline failures three
+ * additional times. Once those catch-sites grow the same
+ * ``innerTimeoutFired`` distinction the Anthropic one has, the pattern set
+ * here can be expanded to cover them too.
+ *
+ * This is the daemon-side counterpart to the vembda graceful-close behavior
+ * for upstream disconnects (LUM-1536) — together they collapse the 45 s
+ * silent-stall window the web client used to observe whenever Anthropic's
+ * stream was cut mid-token.
+ */
+const RETRYABLE_TRANSPORT_ABORT_PATTERNS = [
+  /^anthropic api error:\s*request was aborted/i,
+];
+
 function isRetryableStreamError(error: unknown): boolean {
   if (!(error instanceof ProviderError)) return false;
   if (error.statusCode !== undefined) return false; // has a real HTTP status — not a stream error
@@ -84,6 +118,15 @@ function isRetryableProviderMessage(error: unknown): boolean {
   return RETRYABLE_PROVIDER_MESSAGE_PATTERNS.some((p) => p.test(error.message));
 }
 
+function isRetryableTransportAbort(error: unknown): boolean {
+  if (!(error instanceof ProviderError)) return false;
+  // Transport aborts surface with ``status === undefined`` (the SDK never
+  // saw an HTTP response). A real HTTP status here means a server error,
+  // which is handled by the status check.
+  if (error.statusCode !== undefined) return false;
+  return RETRYABLE_TRANSPORT_ABORT_PATTERNS.some((p) => p.test(error.message));
+}
+
 function isRetryableError(error: unknown): boolean {
   // Context overflow is deterministic — retrying the same oversized prompt
   // will never succeed. Short-circuit before the generic 429/5xx check so
@@ -103,6 +146,7 @@ function isRetryableError(error: unknown): boolean {
   }
   if (isRetryableProviderMessage(error)) return true;
   if (isRetryableStreamError(error)) return true;
+  if (isRetryableTransportAbort(error)) return true;
   return isRetryableNetworkError(error);
 }
 
@@ -482,7 +526,9 @@ export class RetryProvider implements Provider {
                   ? "provider_overloaded"
                   : isRetryableStreamError(error)
                     ? "stream_corruption"
-                    : "network_error";
+                    : isRetryableTransportAbort(error)
+                      ? "transport_abort"
+                      : "network_error";
           log.warn(
             {
               attempt: attempt + 1,