[JosefAschauer]

Problem

A Python class with a parameterized base, class Leaf(Base[str]), has no extends edge. The grammar parses Base[str] in the base list as a subscript, and the base-list reader only took bare identifiers and attributes, so the base was skipped. Everything that follows the class hierarchy misses such a class. This is the Python counterpart of #2856 (Java/TS Store<String>), and the "generic bases" item left open in #2076.

The resolver also reads a class's base list as text (pythonBases in name-matcher.ts, behind module-global and self. dispatch). It looked up Base[str] verbatim and split Pair[str, int] at its inner comma, so current = Leaf() then current.run() did not reach Base.run.

Fix

• Extraction, both arms: a subscript base is unwrapped to its value field before the existing identifier | attribute check, in src/extraction/tree-sitter.ts and in the kernel's codegraph-kernel/src/python.rs. The reference sits at the value's start, the position the plain form had.
• Resolver: pythonBases drops bracketed type arguments (nested ones too) before splitting the list.
• Not linked, as before: Generic[T] / Protocol[T] (outside the project, like their plain forms), a computed base like make()[0], a subscript of a subscript like Base[int][str], and keyword arguments like metaclass=Meta.

Validation

• New __tests__/python-generic-bases.test.ts, run on both arms: wasm (CODEGRAPH_KERNEL=0) and kernel (CODEGRAPH_KERNEL_LANGS=python; skipped when no kernel is staged). 16 tests:

  • Base[str], Base[list[int]], Pair[str, dict[str, int]], a base written over several lines, a mixed list (Mixin, Base[str]), and models.Base[int] all link like their plain forms;
  • a keyword argument, a computed base and Base[int][str] link nothing; Generic[T] links nothing;
  • a module global holding Leaf / Two resolves run to Base::run / Pair::run, with a same-named decoy class in the file.

  Each fix was mutation-checked on its own: disabling the wasm unwrap fails only wasm tests, disabling the kernel unwrap fails only kernel tests, and disabling the pythonBases change fails the module-global test on both arms.

• npx tsc --noEmit: clean. Kernel built with npm run build:kernel.

• Full suite on Linux with the kernel staged: 9281 passed, 4 failed. The failures are environmental and also fail on main:

  • bundle-launcher ×2: no zip binary.
  • cli-sync ×2: the host runs Node 26, which prints the unsupported-Node banner.

• Together with feat(resolution): link Python base methods to their overrides (python-override) #2076, Base.run gets its override link to Leaf.run on both arms.

Follow-ups

• pythonBases reads only the class line, so a base list written over several lines is invisible to it (plain bases too). Fixed in fix(python): a base list wrapped over several lines is read whole #3001, which is stacked on this PR.
• declaredSupertypes' Python branch counts type arguments (str, User) as supertypes, so Repo.save() on class Repo(Base[User]) reached User.save. Fixed in fix(python): a base's type argument is no supertype, so Repo.save() is not User.save #3000, which is independent.

Independent of #2076.

🤖 Generated with Claude Code

https://claude.ai/code/session_018hJjBUPFiLSFBssshFwHzq

[colbymchenry]

Thanks for this, and for testing both extraction paths! On current main a parameterized base like class Leaf(Base[str]) gets no extends edge, and a module global holding one reaches nothing. Applied to the newest main, your change fixes both on the wasm path. When you mark it ready, we'll run the kernel arm and an A/B on real Python repos before landing it.

[JosefAschauer]

Thanks for checking both paths. It's finished from my side; marking it ready for the kernel arm and the A/B.

[colbymchenry]

Thanks @JosefAschauer! This landed as #3101 with you as the commit author, and your fix went in unchanged. The kernel arm and the A/B on real Python repos both came out well: on polar, 491 calls now reach RepositoryBase instead of a protocol stub. We made the test's kernel arm fail rather than skip when a build expects a kernel, and moved the CHANGELOG bullet next to the other Python entries.