I am not convincing this is the right solution. This will hide the real error and as the traceback is missing will make it very difficult to find out what the problem is. The real solution is probably to filter the candidates in the traceback module before submitting it to the_generate_suggestions function.

@ambv@iritkatriel what do you think?

I agree with@pablogsal.

Either filter intraceback, or error out inrunpy when the globals dict is weird.

Thank you both for the review, sorry for taking so long to address it.

Either filter intraceback, or error out inrunpy when the globals dict is weird.

runpy is just an easy path to aNameError with a user provided globals, but it also works with e.g.:

>>>g= {0:1,"p":1}>>>exec(compile("pa",'g',"exec"),g)[...]NameError:name'pa'isnotdefinedDuringhandlingoftheaboveexception,anotherexceptionoccurred:[...]TypeError:allelementsin'candidates'mustbestrings[AndtheREPLexits]

So, the current situation intrackeback._compute_suggestion_error() is:

• ImportError: if the module has a non-string in its__dict__, no suggestion is offered and no error happens.
• AttributeError: object returns a non-string in its__dir__ (or has one in__dict__), no suggestion is offered and no error happens.
• NameError: iff_locals,f_globals orf_builtins has a non-string key, an error is raised that exits the new REPL.

I suggest we just guard against an exception in theNameError case and offer no suggestions (by returningNone). An alternative that gets the same result is to returnNone if a check likeany(not isinstance(x, str) for x in candidates) succeeds.

Or we can filter non-strings from the candidates in theNameError case, which would provide suggestions while avoiding the error. If this is preferred, I suggest to also filter non-strings in theImportError andAttributeError cases, for the same effect.

@devdanzin If you just want to fix the crash, this one-liner does the job:

diff --git a/Lib/traceback.py b/Lib/traceback.pyindex 647c23ed782..cd5e2a88f07 100644--- a/Lib/traceback.py+++ b/Lib/traceback.py@@ -1527,6 +1527,7 @@ def _compute_suggestion_error(exc_value, tb, wrong_name):             + list(frame.f_globals)             + list(frame.f_builtins)         )+        d = [name for name in d if isinstance(name, str)]          # Check first if we are in a method and the instance         # has the wrong name as attribute

All commit authors signed the Contributor License Agreement.

Gentle ping@iritkatriel@pablogsal . Should we go with guarding against the exception and not offering suggestions onNameError case? What's your preference?

Non-string keys are legal in dicts. Hints are an excellent by optional (and CPython-specific, I presume) enhancement. The hinters should just ignore, by filtering out, non-string keys and give a hint if possible. There is no reason to not give hint. Hinters are not linters.

A decision on this (#131001 (comment)) is blocking#130997. If no different opinions are presented, I'm going with@terryjreedy's approach in a day or two.

I agree with Terry and your final suggestion in the comment, let's filter and provide whatever other suggestions still apply. This is interactive, so unlikely to be "too slow for comfort".

I've implemented filtering of the candidates, and also guarded againstdir() raising, so it's protected from exiting the REPL in more cases, while correctly displaying the remaining suggestions.

One doubt: whendir() raises, I chose to fetchobj.__dict__.keys(), but that misses attributes defined in the class, so I also set it to fetch fromobj.__class__.__dict__.keys(). What is preferred, more candidates for suggestions (current code) or simpler/more strict suggestions (only fetch from the instance dict)?

@devdanzin What you did is what users will expect 👍🏻

Thanks@devdanzin for the PR, and@ambv for merging it 🌮🎉.. I'm working now to backport this PR to: 3.13, 3.14.
🐍🍒⛏🤖

GH-135019 is a backport of this pull request to the3.14 branch.

GH-135020 is a backport of this pull request to the3.13 branch.

This PR causes failures when doing refleak hunting. I'm on it.