"""Configured chat-completion protocol; no credential discovery or automatic replay.""" import json import re import time import hashlib from urllib.parse import urlsplit import httpx from src.modules.llm_gateway.service import ModelReply, ProviderFailure class ChatCompletionProvider: def __init__(self, *, name, endpoint, model, token, timeout=60, transport=None, enable_thinking=None, json_mode=True): url = urlsplit(endpoint) if (url.scheme != 'https' or not url.hostname or url.username or url.password or url.query or url.fragment or not 1 <= timeout <= 300 or not token or not model): raise ValueError('INVALID_PROVIDER_CONFIGURATION') self.name, self.endpoint, self.model = name, endpoint, model self._token, self.timeout, self.transport = token, timeout, transport self.enable_thinking = enable_thinking self.json_mode = json_mode def payload(self,messages,max_tokens): payload={'model':self.model,'messages':messages,'max_tokens':max_tokens,'stream':False} if self.json_mode:payload['response_format']={'type':'json_object'} if self.enable_thinking is not None:payload['enable_thinking']=self.enable_thinking return payload def complete(self, messages, max_tokens, cancelled): deadline = time.monotonic() + self.timeout evidence={} response_received=False received_bytes=0 try: payload = self.payload(messages,max_tokens) with httpx.Client(timeout=httpx.Timeout(self.timeout, connect=10), follow_redirects=False, trust_env=False, transport=self.transport) as client: with client.stream('POST', self.endpoint, headers={'Authorization': 'Bearer ' + self._token}, json=payload) as response: response_received=True if response.status_code in {402, 429}: raise ProviderFailure('MODEL_QUOTA_REJECTED', rejected=True, retryable=True) if response.status_code in {400, 401, 403, 404, 422}: # Inspect only bounded error text for categorical diagnostics, never expose it. error_bytes = bytearray() for part in response.iter_bytes(8192): error_bytes.extend(part) if len(error_bytes) > 16384: break error_text = bytes(error_bytes).decode('utf-8', errors='replace').lower() quota = response.status_code == 403 and any(word in error_text for word in ('quota', 'arrear', 'allocation', 'billing', 'balance')) details = {'http_status': response.status_code, 'quota_rejection': quota, 'thinking_parameter': 'enable_thinking' in error_text, 'response_format_parameter': 'response_format' in error_text, 'model_parameter': 'model' in error_text} # Retain categorical provider evidence without returning the message, # which can contain account details or credential fragments. try: envelope=json.loads(error_bytes) error=envelope.get('error',envelope) code=error.get('code') if isinstance(error,dict) else None if (isinstance(code,str) and re.fullmatch(r'[A-Za-z][A-Za-z0-9_.]{0,95}',code) and code != self._token): details['provider_error_code']=code except (ValueError,TypeError,AttributeError): pass details['response_evidence']={'response_received':True, 'received_bytes':len(error_bytes),'response_sha256':hashlib.sha256(error_bytes).hexdigest(), 'raw_retained':False} raise ProviderFailure('MODEL_QUOTA_REJECTED' if quota else 'MODEL_REQUEST_REJECTED', rejected=True, retryable=quota, details=details) if response.status_code != 200: raise ProviderFailure(details={'http_status':response.status_code,'failure_category':'http_error'}) raw = bytearray() for chunk in response.iter_bytes(chunk_size=8192): if cancelled() or time.monotonic() > deadline: raise ProviderFailure(details={'failure_category':'cancelled_or_deadline'}) raw.extend(chunk) received_bytes=len(raw) if len(raw) > 1024 * 1024: raise ProviderFailure('MODEL_RESPONSE_TOO_LARGE',details={'failure_category':'output_size', 'reason_code':'PROVIDER_RESPONSE_LIMIT','response_evidence':{'response_bytes':len(raw), 'response_sha256':hashlib.sha256(raw).hexdigest(),'response_complete':False,'raw_retained':False}}) if cancelled() or time.monotonic() > deadline: raise ProviderFailure(details={'failure_category':'cancelled_or_deadline'}) evidence={'response_bytes':len(raw),'response_sha256':hashlib.sha256(raw).hexdigest(),'raw_retained':False} try:data = json.loads(raw) except json.JSONDecodeError as error: raise ProviderFailure(details={'failure_category':'output_parse','reason_code':'PROVIDER_ENVELOPE_JSON_INVALID', 'parse_layer':'provider_envelope','parse_position':error.pos,'parse_line':error.lineno,'parse_column':error.colno, 'response_evidence':evidence}) from None choice = data['choices'][0] usage = data['usage'] input_tokens, output_tokens = usage['prompt_tokens'], usage['completion_tokens'] if (type(input_tokens) is not int or type(output_tokens) is not int or input_tokens < 0 or not 0 <= output_tokens <= max_tokens): raise ProviderFailure('MODEL_USAGE_INVALID') request_id = str(data.get('id', '')) request_id = request_id if re.fullmatch(r'[A-Za-z0-9_-]{1,128}', request_id) else '' reason = choice.get('finish_reason') reason = reason if reason in {'stop','length','tool_calls','content_filter'} else 'unsupported' value=choice.get('message',{}).get('content') evidence.update(content_type=type(value).__name__) if isinstance(value,str):evidence.update(content_bytes=len(value.encode()),content_sha256=hashlib.sha256(value.encode()).hexdigest()) details = {'finish_reason':reason,'request_id':request_id,'response_evidence':evidence} if reason != 'stop' or choice['message'].get('tool_calls'): raise ProviderFailure('MODEL_OUTPUT_TRUNCATED' if reason=='length' else 'MODEL_OUTPUT_UNSUPPORTED', details={**details,'failure_category':'output_truncated' if reason=='length' else 'output_schema', 'reason_code':'OUTPUT_LENGTH_LIMIT' if reason=='length' else 'OUTPUT_UNSUPPORTED_FINISH'},input_tokens=input_tokens,output_tokens=output_tokens) try: content=json.loads(value) except json.JSONDecodeError as error: raise ProviderFailure('MODEL_OUTPUT_INVALID',details={**details,'failure_category':'output_parse','reason_code':'COMPLETION_JSON_INVALID', 'parse_layer':'completion_content','parse_position':error.pos,'parse_line':error.lineno,'parse_column':error.colno}, input_tokens=input_tokens,output_tokens=output_tokens) from None except (ValueError,TypeError): raise ProviderFailure('MODEL_OUTPUT_INVALID',details={**details,'failure_category':'output_schema','reason_code':'COMPLETION_CONTENT_TYPE_INVALID'},input_tokens=input_tokens,output_tokens=output_tokens) from None return ModelReply(content,input_tokens,output_tokens,self.model,reason,request_id,evidence) except ProviderFailure: raise except httpx.TimeoutException as exc: reason=('CONNECT_TIMEOUT' if isinstance(exc,httpx.ConnectTimeout) else 'READ_TIMEOUT' if isinstance(exc,httpx.ReadTimeout) else 'WRITE_TIMEOUT' if isinstance(exc,httpx.WriteTimeout) else 'POOL_TIMEOUT') raise ProviderFailure(details={'failure_category':'transport_timeout','reason_code':reason, 'response_evidence':{'response_received':response_received,'received_bytes':received_bytes,'raw_retained':False}}) from None except httpx.TransportError as exc: reason='REMOTE_PROTOCOL_ERROR' if isinstance(exc,httpx.RemoteProtocolError) else 'CONNECTION_INTERRUPTED' raise ProviderFailure(details={'failure_category':'transport_error','reason_code':reason, 'response_evidence':{'response_received':response_received,'received_bytes':received_bytes,'raw_retained':False}}) from None except Exception: # Do not expose HTTP headers, response bodies, tokens or raw exception messages. raise ProviderFailure(details={'failure_category':'output_unknown','reason_code':'PROVIDER_ENVELOPE_FIELDS_UNKNOWN', 'parse_layer':'provider_envelope','response_evidence':evidence}) from None