Merge branch 'fix/request-scoped-errors-no-cooldown'

aisbf/handlers.py

@@ -39,6 +39,7 @@ from fastapi import HTTPException, Request
 from fastapi.responses import JSONResponse, StreamingResponse, Response
 from .models import ChatCompletionRequest, ChatCompletionResponse
 from .providers import get_provider_handler, RateLimitError
+from .providers.base import is_request_scoped_error
 from .config import config
 from .studio import infer_model_capabilities
 from .studio_adapters import effective_studio_adapter, infer_studio_adapter_profile, adapt_studio_payload_with_profile
@@ -97,8 +98,18 @@ def _is_upstream_rate_limit(error) -> bool:
 
 
 def _should_record_failure(error) -> bool:
-    """Whether an exception should count against the provider's failure budget."""
-    return not _is_coderai_warmup_error(error) and not _is_upstream_rate_limit(error)
+    """Whether an exception should count against the provider's failure budget.
+
+    Excludes coderai warm-up errors, upstream rate limits, and request-scoped
+    errors (bad input / unsupported capability such as a 404 "No endpoints found
+    that support image input") — none of which indicate the provider is unhealthy,
+    so they must not trip the cooldown that disables it for all traffic.
+    """
+    return (
+        not _is_coderai_warmup_error(error)
+        and not _is_upstream_rate_limit(error)
+        and not is_request_scoped_error(error)
+    )
 
 
 def _raise_if_provider_unavailable(handler, provider_id: str) -> None:

aisbf/providers/base.py

@@ -193,6 +193,53 @@ class RateLimitError(Exception):
         self.status_code = status_code
 
 
+# HTTP statuses that describe a problem with THIS request — bad input, an
+# unsupported capability for the chosen model, or a rate limit — rather than the
+# provider being unhealthy. They must not count toward the consecutive-failure
+# budget that disables the whole provider: otherwise a handful of request-scoped
+# errors (e.g. a model that lacks an image endpoint returning 404 "No endpoints
+# found that support image input", or a bad image returning 400) take the whole
+# provider offline for every request for the cooldown window.
+_REQUEST_SCOPED_STATUSES = frozenset({400, 404, 413, 415, 422, 429})
+
+
+def is_request_scoped_status(status) -> bool:
+    """True when an HTTP status code is about the request, not provider health."""
+    try:
+        return int(status) in _REQUEST_SCOPED_STATUSES
+    except (TypeError, ValueError):
+        return False
+
+
+def extract_http_status(error) -> Optional[int]:
+    """Best-effort HTTP status code from a provider exception."""
+    status = getattr(error, 'status_code', None)
+    if status is None:
+        status = getattr(getattr(error, 'response', None), 'status_code', None)
+    try:
+        return int(status) if status is not None else None
+    except (TypeError, ValueError):
+        return None
+
+
+def is_request_scoped_error(error) -> bool:
+    """True when the error is about the specific request (bad input / unsupported
+    capability / rate limit) rather than the provider's health, so it must not
+    count toward the failure cooldown that disables the whole provider.
+
+    Falls back to message sniffing for handlers that wrap the upstream error in a
+    plain Exception without preserving the status code (e.g. the streaming path).
+    """
+    if is_request_scoped_status(extract_http_status(error)):
+        return True
+    text = str(error or '')
+    return (
+        'No endpoints found that support image input' in text
+        or 'Cannot decode visions' in text
+        or '429 Too Many Requests' in text
+    )
+
+
 class AnthropicFormatConverter:
     """
     Shared utility class for converting between OpenAI and Anthropic message formats.

aisbf/providers/kilo.py

@@ -411,8 +411,19 @@ class KiloProviderHandler(BaseProviderHandler):
             return response
         except Exception as e:
             import logging
+            from .base import is_request_scoped_error, extract_http_status
             logging.error(f"KiloProviderHandler: Error: {str(e)}", exc_info=True)
-            self.record_failure()
+            # A per-request error (bad input, or the chosen model lacking an image
+            # endpoint → 404 "No endpoints found that support image input", or a
+            # 429 rate limit) says nothing about provider health. Counting it toward
+            # the three-strike cooldown would take kilo offline for every request.
+            if is_request_scoped_error(e):
+                logging.warning(
+                    f"KiloProviderHandler: request-scoped error (status={extract_http_status(e)}) "
+                    f"not counted toward the provider failure budget"
+                )
+            else:
+                self.record_failure()
             raise e
 
     async def _handle_streaming_request(self, request_params: Dict, token: str, model: str):
@@ -462,8 +473,13 @@ class KiloProviderHandler(BaseProviderHandler):
                         error_json if 'error_json' in locals() else error_message,
                         dict(response.headers)
                     )
-                
-                self.record_failure()
+
+                # Request-scoped statuses (bad input, unsupported capability, rate
+                # limit) must not trip the provider failure cooldown — see
+                # is_request_scoped_status. Only genuine provider faults count.
+                from .base import is_request_scoped_status
+                if not is_request_scoped_status(response.status_code):
+                    self.record_failure()
                 raise Exception(f"Kilo API streaming error ({response.status_code}): {error_message}")
         except Exception:
             await streaming_client.aclose()