Repository traps

Use this only when the current repository matches one below. Do not browse these notes speculatively.

python-docx / python-pptx

Public API methods commonly delegate to XML CT_* element classes. If the public method looks correct, inspect one level below before widening the search. Preserve exact XML behavior and do not touch fixture files.

astroid

Inference code is generator-sensitive. yield, return, and the Uninferable sentinel are not interchangeable. Warnings may fail tests, and strict expected-failure tests can fail if an unrelated behavior is accidentally repaired. Change only the behavior named by the issue.

pygments

Many checks compare exact or snapshot-like output. Never regenerate expected output. For delegating lexers, verify root/language lexer ordering against sibling implementations rather than inventing a new arrangement.

cantools

CLI text is exact: whitespace and line breaks matter. Keep pytest quiet and never modify files under test-data directories merely to make output agree.

oauthlib

Protocol strings are contracts: parameter names, headers, error slugs, and status values must be exact. Compare a suspect grant or endpoint with its sibling and shared base implementation.

sqlglot

Dialect implementations are highly parallel. A sibling dialect is often the strongest reference for a mutated method. Leave fixtures and expected SQL files untouched.

gpxpy

None, 0, and empty values have distinct meanings. Compare the same method across point/route/track siblings. For speed or distance logic, guard invalid or zero time intervals rather than inventing a replacement formula.

Anything else

Do not extrapolate these repository-specific rules. Return to the issue, surviving tests, docstrings, callers, and sibling code.