Thread lock 2

app.py

@@ -36,16 +36,32 @@ if not OPENROUTER_API_KEY:
 # outbound OpenRouter calls behind a single lock and enforce a minimum gap
 # between them. Whichever request arrives second simply waits its turn.
 # ---------------------------------------------------------------------------
-MIN_INTERVAL_SECONDS = float(os.environ.get("OPENROUTER_MIN_INTERVAL", "2.0"))
+MIN_INTERVAL_SECONDS = float(os.environ.get("OPENROUTER_MIN_INTERVAL", "6.0"))
+MAX_RETRIES = int(os.environ.get("OPENROUTER_MAX_RETRIES", "3"))
+RETRY_BACKOFF_SECONDS = float(os.environ.get("OPENROUTER_RETRY_BACKOFF", "5.0"))
 
 _or_lock = threading.Lock()
 _last_call_time = 0.0
 
+def _is_rate_limited(response: requests.Response) -> bool:
+    """Detect a 429 whether it comes as an HTTP status or embedded in a 200 body
+    (OpenRouter sometimes returns 200 with {"error": {"code": 429, ...}})."""
+    if response.status_code == 429:
+        return True
+    try:
+        body = response.json()
+    except ValueError:
+        return False
+    return isinstance(body, dict) and body.get("error", {}).get("code") == 429
+
+
 def call_openrouter(payload: dict, headers: dict, timeout: int = 20):
     """Thread-safe, rate-limited call to OpenRouter's chat completions endpoint.
 
-    Ensures at least MIN_INTERVAL_SECONDS has elapsed since the previous
-    OpenRouter call (across ALL endpoints) before firing this one.
+    - Ensures at least MIN_INTERVAL_SECONDS has elapsed since the previous
+      OpenRouter call (across ALL endpoints) before firing this one.
+    - If the provider still returns a 429 (as a status code or embedded in the
+      JSON body), retries with exponential backoff up to MAX_RETRIES times.
     """
     global _last_call_time
 
@@ -62,6 +78,16 @@ def call_openrouter(payload: dict, headers: dict, timeout: int = 20):
         response = requests.post(url, headers=headers, json=payload, timeout=timeout)
         _last_call_time = time.monotonic()
 
+        attempt = 0
+        while _is_rate_limited(response) and attempt < MAX_RETRIES:
+            backoff = RETRY_BACKOFF_SECONDS * (2 ** attempt)
+            logger.warning(f"OpenRouter rate limited (attempt {attempt + 1}/{MAX_RETRIES}); "
+                            f"backing off {backoff:.1f}s before retry")
+            time.sleep(backoff)
+            response = requests.post(url, headers=headers, json=payload, timeout=timeout)
+            _last_call_time = time.monotonic()
+            attempt += 1
+
     return response