[Bug]: Swift: an extension of a type outside the corpus (private extension String) captures every reference to that type #4574

Activity

1. github-actions commented on Oct 20, 2026

   

   github-actionsBot

   on Oct 20, 2026 – with GitHub Actions

   More actions

   Thanks for opening this issue, @makinggainz. A maintainer will take a look soon.

   If you would like to discuss it in real time, come say hi on our Discord server. For longer-form questions and ideas there is also GitHub Discussions.

   To help us triage, please make sure the report includes what you expected, what actually happened, and the steps (and a small sample) to reproduce it.
2. makinggainz commented on Oct 20, 2026

   

   makinggainz

   on Oct 20, 2026

   Author

   More actions

   I tracked down the cause and have a fix working locally.

   Cause. _rewire_unique_stub_nodes builds real_by_label from every sourced, type-like node. _merge_swift_extensions rightly leaves an extension of an out-of-corpus type alone. But that node still has a source_file and the label String, so it becomes the one "real" String.

   A second bug with the same cause. A file that declares struct Money and also extends it uses one id for both. _merge_swift_extensions skips every id in extension_nids when it looks for canonical types, so that node can't be the canonical Money. As a result, an extension Money in another file stays a second Money node. A reference from a third file then has two candidates and stays on a stub:

   // Money.swift
   struct Money { let cents: Int }
   extension Money { static var zero: Money { Money(cents: 0) } }
   // MoneyFormat.swift
   extension Money { var text: String { "" } }
   // Wallet.swift
   struct Wallet { let balance: Money }

   This yields two sourced Money nodes, and Wallet references a sourceless stub.

   The fix I tested has three small parts:

   1. In the generic class walk, record which Swift ids the file actually declares: class, struct, enum, actor or protocol, i.e. not extension. Add "declared": True to each swift_extensions record whose id the same file also declares.
   2. In _merge_swift_extensions, a declared id counts as a canonical declaration. Exclude only extension-only ids when building label_to_canonical, and never remap declared ids. Extension-only ids that stay unmerged get a node marker, for example _swift_extension_only = True.
   3. In _rewire_unique_stub_nodes, skip nodes with that marker when building real_by_label. Add the marker to the resolution-context marker lists in watch.py and cli.py, so graphify update keeps it for unchanged files. Bump _AST_CACHE_SCHEMA, since the per-file records change.

   Result. On a mid-sized SwiftUI app, the private extension String node went from about 940 edges to 2, and an extension Date node went from 72 to 3. Their references now behave like Int and Bool already do, as sourceless stubs.

   Tests. These regression tests fail on 0.9.92 and pass with the change:

   • the repro from this issue;
   • the Money case above;
   • a type next to its own extension still resolving;
   • cross-file merging of corpus types unchanged;
   • an incremental _rebuild_code run, where the extension file arrives only as resolution-context nodes.

   The full suite shows no new failures.

   I have this on a local branch and can open a PR if that would help.

   Written with the help of an AI assistant; the steps and numbers above come from running the change.