Report llama context limit errors clearly

endpoint.py

@@ -19,6 +19,24 @@ class APIError(ValueError):
     def __init__(self,message,status=400,code='invalid_request_error'):
         super().__init__(message);self.status=status;self.code=code
 
+def chat_upstream_error(status, raw):
+    """Expose known, safe llama.cpp limits without reflecting request content."""
+    try:
+        detail=json.loads(raw)
+        error=detail.get('error',{}) if isinstance(detail,dict) else {}
+        if error.get('type')=='exceed_context_size_error':
+            needed=error.get('n_prompt_tokens');available=error.get('n_ctx')
+            if all(type(value) is int and 0<value<=2097152 for value in (needed,available)):
+                needed_text=f'{needed:,}'.replace(',','.');available_text=f'{available:,}'.replace(',','.')
+                return APIError(
+                    f'Die Anfrage benötigt {needed_text} Kontext-Tokens, das aktive Profil bietet {available_text}. '
+                    'Kontext im Profil erhöhen oder die Anfrage verkürzen.',
+                    400,'context_length_exceeded')
+    except (ValueError,TypeError,AttributeError):
+        pass
+    return APIError('llama.cpp hat die Anfrage abgelehnt. Kontext und Anfrageparameter prüfen.',
+                    status if 400<=status<600 else 502,'upstream_error')
+
 class Endpoint:
     def __init__(self,root,profiles,worker,scheduler,images,credentials,management_port):
         self.root=Path(root);self.profiles=profiles;self.worker=worker;self.scheduler=scheduler;self.images=images;self.tts=None;self.stt=None;self.video=None;self.credentials=credentials
@@ -287,7 +305,7 @@ class APIHandler(BaseHTTPRequestHandler):
                 conn.request('POST','/v1/chat/completions',body=json.dumps(body).encode(),headers={'Content-Type':'application/json','Authorization':'Bearer '+key})
                 response=conn.getresponse()
                 if response.status!=200:
-                    response.read(65536);raise APIError('llama.cpp hat die Anfrage abgelehnt. Kontext und Anfrageparameter prüfen.',response.status if 400<=response.status<600 else 502,'upstream_error')
+                    raise chat_upstream_error(response.status,response.read(65536))
                 if not data.get('stream'):
                     raw=response.read(16*1024*1024+1)
                     if len(raw)>16*1024*1024:raise InferenceError('Modellantwort überschreitet 16 MiB.')

test_endpoint.py

@@ -11,7 +11,7 @@ from pathlib import Path
 from types import SimpleNamespace
 from http.server import BaseHTTPRequestHandler,ThreadingHTTPServer
 from unittest.mock import Mock
-from endpoint import Endpoint
+from endpoint import Endpoint,chat_upstream_error
 from inference import Scheduler,InferenceError
 
 class FakeWorker:
@@ -32,6 +32,19 @@ class Upstream(BaseHTTPRequestHandler):
   else:
    body=json.dumps(dict(id='test',model=data['model'],choices=[{'message':{'role':'assistant','content':'synthetic'}}])).encode();self.send_response(200);self.send_header('Content-Length',str(len(body)));self.end_headers();self.wfile.write(body)
 
+class UpstreamErrorTests(unittest.TestCase):
+ def test_context_limit_is_actionable_without_echoing_upstream_text(self):
+  raw=json.dumps({'error':{'type':'exceed_context_size_error','message':'private request text',
+                           'n_prompt_tokens':13523,'n_ctx':8192}}).encode()
+  error=chat_upstream_error(400,raw)
+  self.assertEqual(error.code,'context_length_exceeded')
+  self.assertIn('13.523',str(error));self.assertIn('8.192',str(error))
+  self.assertNotIn('private request text',str(error))
+ def test_unknown_upstream_errors_stay_generic(self):
+  error=chat_upstream_error(400,b'{"error":{"message":"private request text"}}')
+  self.assertEqual(error.code,'upstream_error')
+  self.assertNotIn('private request text',str(error))
+
 class EndpointTests(unittest.TestCase):
  def setUp(self):
   self.tmp=tempfile.TemporaryDirectory();self.worker=FakeWorker();self.worker.http=ThreadingHTTPServer(('127.0.0.1',0),Upstream);self.worker.http.worker=self.worker;threading.Thread(target=self.worker.http.serve_forever,daemon=True).start()