[dexhunter]

TL;DR

• This PR makes ParserTreeFilter.values() and ParserTreeFilter._filter faster, so name completion gets faster.
• values() now skips name keys that have no definitions and looks up the module cache once.
• Measured on one CPU core with a completion benchmark (9 stdlib modules plus Interpreter completions): 462.7 ms -> 419.4 ms, 9.4% faster.
• Tests: pytest test/ shows the same 44 failures as master (environment-dependent), 3883 passed.

What & Why

Most keys in used_names are plain usages, not definitions. Before, values() still called _get_definition_names() and _filter() for each of these keys. Each call did a weak dict lookup, a super() call, two list builds and a generator, only to return an empty list.

Now values() gets the per-module definition cache once. Then it skips keys whose definition tuple is empty. _filter() returns early for empty input and checks until_position and reachability in one list comprehension. _check_flows() sorts only when there are two or more names, and it returns a list.

The output stays identical. Every _filter() override (ClassFilter, SelfAttributeFilter, FixtureFilter, StubFilter) returns nothing for empty input. The cache keeps the same keys and values. Names keep the same order. The flow checks run in the same order and stop at the same name.

Changes

• _AbstractUsedNamesFilter.values(): get the definition cache once, fill it in the same way, and skip empty keys.
• ParserTreeFilter._filter(): return early for no names; use one list comprehension; no super() call.
• ParserTreeFilter._check_flows(): sort only for two or more names; build and return a list.

Results

Median times on one dedicated CPU core, Python 3.12, InterpreterEnvironment.

Situation	Before	After	Change
Completion workload (9 stdlib modules + Interpreter completions)	462.7 ms	419.4 ms	-9.4%
Script.complete() at end of argparse.py	19.79 ms	18.25 ms	-7.8%
Script.complete() at end of bisect.py	13.27 ms	12.80 ms	-3.5%
Interpreter completions (C, C.foo, C., os.path.)	34.96 ms	33.07 ms	-5.4%
Interpreter("str.").complete()	28.91 ms	24.98 ms	-13.6%
Script.infer() on one local name (does not use values())	2.39 ms	2.37 ms	no change

Why this is safe and easy to maintain

The change is small and stays in jedi/inference/filters.py. It keeps the same structure and only removes extra work. The completion results, their order and the cache contents do not change. test/test_api/test_interpreter.py, test_classes.py, test_signature.py, test_imports.py and the completion integration tests cover this code.

Validation

• pytest test/test_api/test_interpreter.py test/test_inference/test_signature.py test/test_api/test_classes.py test/test_inference/test_imports.py: 485 passed, 5 skipped, 4 xfailed (same as master).
• pytest test/: 44 failed, 3883 passed, 111 skipped, 5 xfailed. The 44 failures are the same test IDs on master (environment-dependent test_integration pytest/conftest cases and test_namespace_package.py::test_relative_import).
• The completion results of the workload above are identical before and after (names, complete, name_with_symbols, order).
• Short reproduction:

python -m timeit -s "import jedi, argparse; s = open(argparse.__file__).read(); n = s.count(chr(10)) + 1; jedi.Script(s, path=argparse.__file__).complete(n, 0)" "jedi.Script(s, path=argparse.__file__).complete(n, 0)"

AI assistance

Codex (gpt-6.1-sol) generated the candidate changes, and Weco Observe tracked and benchmarked them. Claude Code selected this candidate and ran the tests and benchmarks above. It also rejected a faster candidate because it crashed on empty modules.

Weco share link: https://dashboard.weco.ai/share/FvOQKGvG0Xt8f_28B1bRfiP_tq4kRO_z