fix(python-extractor): capture type-annotated *args/**kwargs parameters

[tirth8205]

Problem

In extractParams (python-extractor.ts), the typed_parameter case assumed the parameter name was a direct identifier child: const ident = findChild(child, "identifier"). For type-annotated variadic parameters such as def f(*args: int, **kwargs: str), the tree-sitter-python grammar wraps the splat in a typed_parameter whose first child is a list_splat_pattern (for *args) or dictionary_splat_pattern (for **kwargs) — the identifier is nested inside that pattern, not a direct child. Because findChild is shallow (only scans direct children), it returned null and the parameter was silently dropped.

Verified by running the extractor: def f(a: int, *args: str, b: int = 0, **kwargs: bool) returned params ["a", "b"] instead of ["a", "*args", "b", "**kwargs"]. The untyped forms (def f(*args, **kwargs)) already worked because they parse as bare list_splat_pattern/dictionary_splat_pattern children of parameters. The gap was specifically typed variadics — a very common pattern in typed Python (*args: Any, **kwargs: object).

Fix

In the typed_parameter case, check for a nested list_splat_pattern / dictionary_splat_pattern first and prefix the captured name with * / ** respectively, falling back to the existing plain-identifier lookup otherwise. The change is confined to understand-anything-plugin/packages/core/src/plugins/extractors/python-extractor.ts.

Testing

• Added test extracts type-annotated *args and **kwargs to python-extractor.test.ts, asserting def f(a: int, *args: str, b: int = 0, **kwargs: bool) yields params ["a", "*args", "b", "**kwargs"].
• The new test fails before the fix (expected [ 'a', 'b' ] to deeply equal [ 'a', '*args', 'b', '**kwargs' ]) and passes after.
• Full core vitest suite is green (693 passing, +1 from this test) with no regressions.
• tsc --noEmit (core) exits 0 and eslint is clean on both changed files.

🤖 Generated with Claude Code

[thejesh23]

A few concerns before this lands.

1. findChild is shallow — fix is grammar-version-fragile. The new branches assume list_splat_pattern / dictionary_splat_pattern are direct children of typed_parameter. That holds in current tree-sitter-python, but if a future grammar revision wraps the splat (e.g., inside a pattern node) the lookup silently returns null again and we regress to the original bug with no test signal. Worth either a recursive search or a comment pinning the assumption to a grammar version.

2. Adjacent gaps still silently dropped. Same extractParams switch has no arm for positional_separator (/), keyword_separator (bare *,), or PEP 695 type_parameter lists (def f[T](x: T) — the [T] lives on function_definition, not parameters, so untouched here, but the bare *// separators do appear in the parameters node and are dropped). Untyped *args followed by keyword-only args already exercises this. Probably out of scope for the fix, but a TODO would help — and the same shape recurs for Dart's {named}/[positional] markers in #435.

3. Test gaps. The new test exercises only a top-level def. Not covered: (a) method form (def m(self, *args: int)) — the new splat branches skip the self/cls filter the plain-identifier branch applies, so a typed *self would leak through (degenerate, but documents intent); (b) async def — should be equivalent but no assertion; (c) typed **kwargs without a preceding typed *args in the same signature, to lock the dict-splat path independently.

Nit: the function-level JSDoc on extractParams (lines 5–10) still reads as if untyped splats are the only splat path — worth one line noting typed variadics are now handled too.