feat: extend CACHE_TTL_FAILURE_SECONDS to 60s and adjust retry logic for Anthropic subscription endpoint

backend/app/services/provider_usage.py

@@ -55,7 +55,12 @@ from app.core.time import utcnow
 logger = get_logger(__name__)
 
 CACHE_TTL_SECONDS = 60
-CACHE_TTL_FAILURE_SECONDS = 5  # short TTL for results with no subscription windows
+# Anthropic's subscription endpoint rate-limit window is ~30-60s.  Retrying
+# sooner than that (old 5s value) creates a tight retry loop that keeps the
+# rate limit permanently throttled.  Use the same 60s TTL as successes so the
+# next auto-poll only fires after the window has cleared.  Manual refreshes
+# (force_refresh=True) always bypass this TTL.
+CACHE_TTL_FAILURE_SECONDS = 60
 REQUEST_TIMEOUT = 8.0  # seconds
 
 
@@ -766,6 +771,9 @@ async def _fetch_anthropic_subscription(session_key: str) -> list[SubscriptionWi
     }
     # Retry once on 429 — the subscription endpoint can be rate-limited when
     # called right after another api.anthropic.com request (e.g. /v1/models).
+    # Sleep 3s before retrying; the rate-limit window is typically >1s but
+    # rarely longer than 3s in isolation (persistent throttling is handled by
+    # the 60s CACHE_TTL_FAILURE_SECONDS at the call site).
     for attempt in range(2):
         async with httpx.AsyncClient(timeout=REQUEST_TIMEOUT) as client:
             try:
@@ -774,12 +782,12 @@ async def _fetch_anthropic_subscription(session_key: str) -> list[SubscriptionWi
                 logger.warning("provider_usage.subscription.anthropic.fetch_failed error=%s", exc)
                 return []
         if resp.status_code == 429 and attempt == 0:
-            await asyncio.sleep(1.5)
+            await asyncio.sleep(3.0)
             continue
         break
 
     if not resp.status_code == 200:
-        logger.debug(
+        logger.warning(
             "provider_usage.subscription.anthropic.http_error status=%s body=%s",
             resp.status_code,
             resp.text[:200],