code-puppy 0.0.378__py3-none-any.whl → 0.0.379__py3-none-any.whl

code_puppy/http_utils.py

@@ -102,17 +102,24 @@ class RetryingAsyncClient(httpx.AsyncClient):
 
     This replaces the Tenacity transport with a more direct subclass implementation,
     which plays nicer with proxies and custom transports (like Antigravity).
+
+    Special handling for Cerebras: Their Retry-After headers are absurdly aggressive
+    (often 60s), so we ignore them and use a 3s base backoff instead.
     """
 
     def __init__(
         self,
         retry_status_codes: tuple = (429, 502, 503, 504),
         max_retries: int = 5,
+        model_name: str = "",
         **kwargs,
     ):
         super().__init__(**kwargs)
         self.retry_status_codes = retry_status_codes
         self.max_retries = max_retries
+        self.model_name = model_name.lower() if model_name else ""
+        # Cerebras sends crazy aggressive Retry-After headers (60s), ignore them
+        self._ignore_retry_headers = "cerebras" in self.model_name
 
     async def send(self, request: httpx.Request, **kwargs: Any) -> httpx.Response:
         """Send request with automatic retries for rate limits and server errors."""
@@ -131,32 +138,39 @@ class RetryingAsyncClient(httpx.AsyncClient):
                 # Close response if we're going to retry
                 await response.aclose()
 
-                # Determine wait time
-                wait_time = 1.0 * (
-                    2**attempt
-                )  # Default exponential backoff: 1s, 2s, 4s...
-
-                # Check Retry-After header
-                retry_after = response.headers.get("Retry-After")
-                if retry_after:
-                    try:
-                        wait_time = float(retry_after)
-                    except ValueError:
-                        # Try parsing http-date
-                        from email.utils import parsedate_to_datetime
+                # Determine wait time - Cerebras gets special treatment
+                if self._ignore_retry_headers:
+                    # Cerebras: 3s base with exponential backoff (3s, 6s, 12s...)
+                    wait_time = 3.0 * (2**attempt)
+                else:
+                    # Default exponential backoff: 1s, 2s, 4s...
+                    wait_time = 1.0 * (2**attempt)
 
+                    # Check Retry-After header (only for non-Cerebras)
+                    retry_after = response.headers.get("Retry-After")
+                    if retry_after:
                         try:
-                            date = parsedate_to_datetime(retry_after)
-                            wait_time = date.timestamp() - time.time()
-                        except Exception:
-                            pass
+                            wait_time = float(retry_after)
+                        except ValueError:
+                            # Try parsing http-date
+                            from email.utils import parsedate_to_datetime
+
+                            try:
+                                date = parsedate_to_datetime(retry_after)
+                                wait_time = date.timestamp() - time.time()
+                            except Exception:
+                                pass
 
                 # Cap wait time
                 wait_time = max(0.5, min(wait_time, 60.0))
 
                 if attempt < self.max_retries:
+                    provider_note = (
+                        " (ignoring header)" if self._ignore_retry_headers else ""
+                    )
                     emit_info(
-                        f"HTTP retry: {response.status_code} received. Waiting {wait_time:.1f}s (attempt {attempt + 1}/{self.max_retries})"
+                        f"HTTP retry: {response.status_code} received{provider_note}. "
+                        f"Waiting {wait_time:.1f}s (attempt {attempt + 1}/{self.max_retries})"
                     )
                     await asyncio.sleep(wait_time)
 
@@ -219,12 +233,14 @@ def create_async_client(
     verify: Union[bool, str] = None,
     headers: Optional[Dict[str, str]] = None,
     retry_status_codes: tuple = (429, 502, 503, 504),
+    model_name: str = "",
 ) -> httpx.AsyncClient:
     config = _resolve_proxy_config(verify)
 
     if not config.disable_retry:
         return RetryingAsyncClient(
             retry_status_codes=retry_status_codes,
+            model_name=model_name,
             proxy=config.proxy_url,
             verify=config.verify,
             headers=headers or {},
@@ -290,6 +306,7 @@ def create_reopenable_async_client(
     verify: Union[bool, str] = None,
     headers: Optional[Dict[str, str]] = None,
     retry_status_codes: tuple = (429, 502, 503, 504),
+    model_name: str = "",
 ) -> Union[ReopenableAsyncClient, httpx.AsyncClient]:
     config = _resolve_proxy_config(verify)
 
@@ -309,12 +326,15 @@ def create_reopenable_async_client(
         kwargs = {**base_kwargs, "client_class": client_class}
         if not config.disable_retry:
             kwargs["retry_status_codes"] = retry_status_codes
+            kwargs["model_name"] = model_name
         return ReopenableAsyncClient(**kwargs)
     else:
         # Fallback to RetryingAsyncClient or plain AsyncClient
         if not config.disable_retry:
             return RetryingAsyncClient(
-                retry_status_codes=retry_status_codes, **base_kwargs
+                retry_status_codes=retry_status_codes,
+                model_name=model_name,
+                **base_kwargs,
             )
         else:
             return httpx.AsyncClient(**base_kwargs)

code_puppy/model_factory.py

@@ -585,7 +585,12 @@ class ModelFactory:
                 return None
             # Add Cerebras 3rd party integration header
             headers["X-Cerebras-3rd-Party-Integration"] = "code-puppy"
-            client = create_async_client(headers=headers, verify=verify)
+            # Pass "cerebras" so RetryingAsyncClient knows to ignore Cerebras's
+            # absurdly aggressive Retry-After headers (they send 60s!)
+            # Note: model_config["name"] is "zai-glm-4.7", not "cerebras"
+            client = create_async_client(
+                headers=headers, verify=verify, model_name="cerebras"
+            )
             provider_args = dict(
                 api_key=api_key,
                 http_client=client,

{code_puppy-0.0.378.dist-info → code_puppy-0.0.379.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: code-puppy
-Version: 0.0.378
+Version: 0.0.379
 Summary: Code generation agent
 Project-URL: repository, https://github.com/mpfaffenberger/code_puppy
 Project-URL: HomePage, https://github.com/mpfaffenberger/code_puppy

{code_puppy-0.0.378.dist-info → code_puppy-0.0.379.dist-info}/RECORD

@@ -8,10 +8,10 @@ code_puppy/config.py,sha256=eAyVqiu8SwzJQpaJu80rlJvd8XXY51DqafMsP5lBRI4,55827
 code_puppy/error_logging.py,sha256=a80OILCUtJhexI6a9GM-r5LqIdjvSRzggfgPp2jv1X0,3297
 code_puppy/gemini_code_assist.py,sha256=KGS7sO5OLc83nDF3xxS-QiU6vxW9vcm6hmzilu79Ef8,13867
 code_puppy/gemini_model.py,sha256=UHb5vFC9zrEdFJ-yCN3vNCdp0UxP156BL_fwbnEhaw8,27988
-code_puppy/http_utils.py,sha256=SAH6EOdbR6Cbfmi-4EtHDqRDBUV5bWtGc-5nr44F0Is,10418
+code_puppy/http_utils.py,sha256=Xel6UyQGO6X_5BlrtPYO1hVXN24aqP72tn5g1uUkluU,11491
 code_puppy/keymap.py,sha256=IvMkTlB_bIqOWpbTpmftkdyjhtD5todXuEIw1zCZ4u0,3584
 code_puppy/main.py,sha256=82r3vZy_XcyEsenLn82BnUusaoyL3Bpm_Th_jKgqecE,273
-code_puppy/model_factory.py,sha256=VFMUY0u8oqHn9f1QqpWsxvXujlK0Ku5NEwfuaTcv2rs,32060
+code_puppy/model_factory.py,sha256=5ndms7O0_bJftizpJMbsdYs3geixn3vZKL2105zEgmM,32338
 code_puppy/model_switching.py,sha256=3IsnSWKHLWzI5d2WDYNg0Xr78BeYNN1WrZuzas-lYJ4,2064
 code_puppy/model_utils.py,sha256=cG1V4fsIOEQIb0W88FyGcsMWNv8SpmRAXN3A7LBEoyE,5116
 code_puppy/models.json,sha256=SC7N2lV1Q8ikXlalRNqABkNvuuL_8fIIk638739-gGY,3319
@@ -225,10 +225,10 @@ code_puppy/tools/browser/chromium_terminal_manager.py,sha256=w1thQ_ACb6oV45L93TS
 code_puppy/tools/browser/terminal_command_tools.py,sha256=9byOZku-dwvTtCl532xt7Lumed_jTn0sLvUe_X75XCQ,19068
 code_puppy/tools/browser/terminal_screenshot_tools.py,sha256=J_21YO_495NvYgNFu9KQP6VYg2K_f8CtSdZuF94Yhnw,18448
 code_puppy/tools/browser/terminal_tools.py,sha256=F5LjVH3udSCFHmqC3O1UJLoLozZFZsEdX42jOmkqkW0,17853
-code_puppy-0.0.378.data/data/code_puppy/models.json,sha256=SC7N2lV1Q8ikXlalRNqABkNvuuL_8fIIk638739-gGY,3319
-code_puppy-0.0.378.data/data/code_puppy/models_dev_api.json,sha256=wHjkj-IM_fx1oHki6-GqtOoCrRMR0ScK0f-Iz0UEcy8,548187
-code_puppy-0.0.378.dist-info/METADATA,sha256=qZ-JMYspwKU2-m7ccFxWT7A4hm4XJh1helqEooIl2KY,27604
-code_puppy-0.0.378.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-code_puppy-0.0.378.dist-info/entry_points.txt,sha256=Tp4eQC99WY3HOKd3sdvb22vZODRq0XkZVNpXOag_KdI,91
-code_puppy-0.0.378.dist-info/licenses/LICENSE,sha256=31u8x0SPgdOq3izJX41kgFazWsM43zPEF9eskzqbJMY,1075
-code_puppy-0.0.378.dist-info/RECORD,,
+code_puppy-0.0.379.data/data/code_puppy/models.json,sha256=SC7N2lV1Q8ikXlalRNqABkNvuuL_8fIIk638739-gGY,3319
+code_puppy-0.0.379.data/data/code_puppy/models_dev_api.json,sha256=wHjkj-IM_fx1oHki6-GqtOoCrRMR0ScK0f-Iz0UEcy8,548187
+code_puppy-0.0.379.dist-info/METADATA,sha256=ldoaFKnfpIocHpmV_ZBTVEjfW0l7YWxIX7qRXJt1rQM,27604
+code_puppy-0.0.379.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+code_puppy-0.0.379.dist-info/entry_points.txt,sha256=Tp4eQC99WY3HOKd3sdvb22vZODRq0XkZVNpXOag_KdI,91
+code_puppy-0.0.379.dist-info/licenses/LICENSE,sha256=31u8x0SPgdOq3izJX41kgFazWsM43zPEF9eskzqbJMY,1075
+code_puppy-0.0.379.dist-info/RECORD,,