Before creating a new issue, please check the FAQ to see if your question is answered there.
Environment data
- debugpy version: 1.8.22+3.g751502d (running from source at
751502d, current main)
- OS and version: macOS 26.6.2 (25G83)
- Python version (& distribution if applicable, e.g. Anaconda): CPython 3.10.6 from python.org, and 3.14.3 from Homebrew
- Using VS Code or Visual Studio: neither. I used the CLI, and also a
"code" launch sent straight to the adapter.
Actual behavior
Code passed with -c, or with "code" in a launch configuration, doesn't run as __main__. __name__ is "builtins", so an if __name__ == "__main__": block is skipped, and a class the code defines can't be pickled:
__name__ = builtins
pickle failed: Can't pickle <class 'Point'>: attribute lookup Point on builtins failed
That is from 3.10.6. 3.14.3 gives the same result, with the pickle error worded as it's not found as builtins.Point.
run_code() in src/debugpy/server/cli.py ends with eval(code, {}). That globals dict has no __name__, so the name is looked up in builtins and comes back as "builtins". A class the code defines gets that as its __module__, which is where pickle then looks for it. A file target doesn't have this problem, because run_file() uses runpy.run_path(target, run_name="__main__"), which puts a temporary __main__ module in place.
A "code" launch goes through the same function, since the adapter turns it into a -c argument for the debugpy CLI in the debuggee. With "code": ["print('__name__ is', __name__)", "if __name__ == '__main__':", " print('main block ran')"] and "console": "internalConsole", the debuggee prints __name__ is builtins and nothing else.
Expected behavior
The same as python -c:
__name__ = __main__
main block ran
pickle ok
Steps to reproduce:
CODE='import pickle
class Point: pass
print("__name__ =", __name__)
if __name__ == "__main__":
print("main block ran")
try:
pickle.dumps(Point()); print("pickle ok")
except Exception as e:
print("pickle failed:", e)'
python3 -c "$CODE"
python3 -m debugpy --listen 127.0.0.1:0 -c "$CODE"
The first command prints the expected output above and the second prints the actual output. Nothing needs to attach; the code runs right away because --wait-for-client isn't given.
I have a fix and can open a PR. It runs the code in a temporary __main__ module, the same way run_path() does for a file, and adds a test to tests/debugpy/server/test_cli.py.
Before creating a new issue, please check the FAQ to see if your question is answered there.
Environment data
751502d, currentmain)"code"launch sent straight to the adapter.Actual behavior
Code passed with
-c, or with"code"in a launch configuration, doesn't run as__main__.__name__is"builtins", so anif __name__ == "__main__":block is skipped, and a class the code defines can't be pickled:That is from 3.10.6. 3.14.3 gives the same result, with the pickle error worded as
it's not found as builtins.Point.run_code()insrc/debugpy/server/cli.pyends witheval(code, {}). That globals dict has no__name__, so the name is looked up in builtins and comes back as"builtins". A class the code defines gets that as its__module__, which is where pickle then looks for it. A file target doesn't have this problem, becauserun_file()usesrunpy.run_path(target, run_name="__main__"), which puts a temporary__main__module in place.A
"code"launch goes through the same function, since the adapter turns it into a-cargument for the debugpy CLI in the debuggee. With"code": ["print('__name__ is', __name__)", "if __name__ == '__main__':", " print('main block ran')"]and"console": "internalConsole", the debuggee prints__name__ is builtinsand nothing else.Expected behavior
The same as
python -c:Steps to reproduce:
The first command prints the expected output above and the second prints the actual output. Nothing needs to attach; the code runs right away because
--wait-for-clientisn't given.I have a fix and can open a PR. It runs the code in a temporary
__main__module, the same wayrun_path()does for a file, and adds a test totests/debugpy/server/test_cli.py.