Handle retries at a higher level; exceptions come out of the streaming completion object

aider/coders/base_coder.py

@@ -30,7 +30,7 @@ from aider.llm import litellm
 from aider.mdstream import MarkdownStream
 from aider.repo import GitRepo
 from aider.repomap import RepoMap
-from aider.sendchat import send_with_retries
+from aider.sendchat import retry_exceptions, send_completion
 from aider.utils import format_content, format_messages, is_image_file
 
 from ..dump import dump  # noqa: F401
@@ -891,6 +891,8 @@ class Coder:
         else:
             self.mdstream = None
 
+        retry_delay = 0.125
+
         self.usage_report = None
         exhausted = False
         interrupted = False
@@ -899,6 +901,14 @@ class Coder:
                 try:
                     yield from self.send(messages, functions=self.functions)
                     break
+                except retry_exceptions() as err:
+                    self.io.tool_error(str(err))
+                    retry_delay *= 2
+                    if retry_delay > 60:
+                        break
+                    self.io.tool_output(f"Retrying in {retry_delay:.1f} seconds...")
+                    time.sleep(retry_delay)
+                    continue
                 except KeyboardInterrupt:
                     interrupted = True
                     break
@@ -1161,7 +1171,7 @@ class Coder:
 
         interrupted = False
         try:
-            hash_object, completion = send_with_retries(
+            hash_object, completion = send_completion(
                 model.name,
                 messages,
                 functions,

aider/sendchat.py

@@ -14,53 +14,28 @@ CACHE = None
 # CACHE = Cache(CACHE_PATH)
 
 
+def retry_exceptions():
+    import httpx
+
+    return (
+        httpx.ConnectError,
+        httpx.RemoteProtocolError,
+        httpx.ReadTimeout,
+        litellm.exceptions.APIConnectionError,
+        litellm.exceptions.APIError,
+        litellm.exceptions.RateLimitError,
+        litellm.exceptions.ServiceUnavailableError,
+        litellm.exceptions.Timeout,
+        litellm.exceptions.InternalServerError,
+        litellm.llms.anthropic.AnthropicError,
+    )
+
+
 def lazy_litellm_retry_decorator(func):
     def wrapper(*args, **kwargs):
-        import httpx
-
-        def should_giveup(e):
-            if not hasattr(e, "status_code"):
-                return False
-
-            if type(e) in (
-                httpx.ConnectError,
-                httpx.RemoteProtocolError,
-                httpx.ReadTimeout,
-                litellm.exceptions.APIConnectionError,
-                litellm.exceptions.APIError,
-                litellm.exceptions.RateLimitError,
-                litellm.exceptions.ServiceUnavailableError,
-                litellm.exceptions.Timeout,
-                litellm.exceptions.InternalServerError,
-                litellm.llms.anthropic.AnthropicError,
-            ):
-                return False
-
-            # These seem to return .status_code = ""
-            # litellm._should_retry() expects an int and throws a TypeError
-            #
-            # litellm.llms.anthropic.AnthropicError
-            # litellm.exceptions.APIError
-            if not e.status_code:
-                return False
-
-            return not litellm._should_retry(e.status_code)
-
         decorated_func = backoff.on_exception(
             backoff.expo,
-            (
+            retry_exceptions(),
-                httpx.ConnectError,
-                httpx.RemoteProtocolError,
-                httpx.ReadTimeout,
-                litellm.exceptions.APIConnectionError,
-                litellm.exceptions.APIError,
-                litellm.exceptions.RateLimitError,
-                litellm.exceptions.ServiceUnavailableError,
-                litellm.exceptions.Timeout,
-                litellm.exceptions.InternalServerError,
-                litellm.llms.anthropic.AnthropicError,
-            ),
-            giveup=should_giveup,
             max_time=60,
             on_backoff=lambda details: print(
                 f"{details.get('exception', 'Exception')}\nRetry in {details['wait']:.1f} seconds."
@@ -71,8 +46,7 @@ def lazy_litellm_retry_decorator(func):
     return wrapper
 
 
-@lazy_litellm_retry_decorator
+def send_completion(
-def send_with_retries(
     model_name, messages, functions, stream, temperature=0, extra_headers=None, max_tokens=None
 ):
     from aider.llm import litellm
@@ -108,9 +82,10 @@ def send_with_retries(
     return hash_object, res
 
 
+@lazy_litellm_retry_decorator
 def simple_send_with_retries(model_name, messages):
     try:
-        _hash, response = send_with_retries(
+        _hash, response = send_completion(
             model_name=model_name,
             messages=messages,
             functions=None,