support 8k output with 3.5 sonnet

aider/coders/base_coder.py

@@ -866,7 +866,7 @@ class Coder:
                     self.io.tool_error(f"BadRequestError: {br_err}")
                     return
                 except FinishReasonLength:
-                    # We hit the 4k output limit!
+                    # We hit the output limit!
                     if not self.main_model.can_prefill:
                         exhausted = True
                         break
@@ -1108,7 +1108,7 @@ class Coder:
 
     def send(self, messages, model=None, functions=None):
         if not model:
-            model = self.main_model.name
+            model = self.main_model
 
         self.partial_response_content = ""
         self.partial_response_function_call = dict()
@@ -1118,7 +1118,13 @@ class Coder:
         interrupted = False
         try:
             hash_object, completion = send_with_retries(
-                model, messages, functions, self.stream, self.temperature
+                model.name,
+                messages,
+                functions,
+                self.stream,
+                self.temperature,
+                extra_headers=model.extra_headers,
+                max_tokens=model.max_tokens,
             )
             self.chat_completion_call_hashes.append(hash_object.hexdigest())
 

aider/models.py

@@ -62,7 +62,7 @@ ANTHROPIC_MODELS = [ln.strip() for ln in ANTHROPIC_MODELS.splitlines() if ln.str
 class ModelSettings:
     # Model class needs to have each of these as well
     name: str
-    edit_format: str
+    edit_format: str = "whole"
     weak_model_name: Optional[str] = None
     use_repo_map: bool = False
     send_undo_reply: bool = False
@@ -71,6 +71,8 @@ class ModelSettings:
     reminder_as_sys_msg: bool = False
     examples_as_sys_msg: bool = False
     can_prefill: bool = False
+    extra_headers: Optional[dict] = None
+    max_tokens: Optional[int] = None
 
 
 # https://platform.openai.com/docs/models/gpt-4-and-gpt-4-turbo
@@ -250,6 +252,8 @@ MODEL_SETTINGS = [
         examples_as_sys_msg=True,
         can_prefill=True,
         accepts_images=True,
+        max_tokens=8192,
+        extra_headers={"anthropic-beta": "max-tokens-3-5-sonnet-2024-07-15"},
     ),
     ModelSettings(
         "anthropic/claude-3-5-sonnet-20240620",
@@ -258,6 +262,8 @@ MODEL_SETTINGS = [
         use_repo_map=True,
         examples_as_sys_msg=True,
         can_prefill=True,
+        max_tokens=8192,
+        extra_headers={"anthropic-beta": "max-tokens-3-5-sonnet-2024-07-15"},
     ),
     ModelSettings(
         "openrouter/anthropic/claude-3.5-sonnet",
@@ -267,6 +273,8 @@ MODEL_SETTINGS = [
         examples_as_sys_msg=True,
         can_prefill=True,
         accepts_images=True,
+        max_tokens=8192,
+        extra_headers={"anthropic-beta": "max-tokens-3-5-sonnet-2024-07-15"},
     ),
     # Vertex AI Claude models
     ModelSettings(
@@ -277,6 +285,8 @@ MODEL_SETTINGS = [
         examples_as_sys_msg=True,
         can_prefill=True,
         accepts_images=True,
+        max_tokens=8192,
+        extra_headers={"anthropic-beta": "max-tokens-3-5-sonnet-2024-07-15"},
     ),
     ModelSettings(
         "vertex_ai/claude-3-opus@20240229",

aider/sendchat.py

@@ -57,7 +57,9 @@ def lazy_litellm_retry_decorator(func):
 
 
 @lazy_litellm_retry_decorator
-def send_with_retries(model_name, messages, functions, stream, temperature=0, extra_headers=None):
+def send_with_retries(
+    model_name, messages, functions, stream, temperature=0, extra_headers=None, max_tokens=None
+):
     from aider.llm import litellm
 
     kwargs = dict(
@@ -70,6 +72,8 @@ def send_with_retries(model_name, messages, functions, stream, temperature=0, ex
         kwargs["functions"] = functions
     if extra_headers is not None:
         kwargs["extra_headers"] = extra_headers
+    if max_tokens is not None:
+        kwargs["max_tokens"] = max_tokens
 
     key = json.dumps(kwargs, sort_keys=True).encode()