Skip to content

Hold an app's spm.modules names to the same rules as a library's - #58060

Open
chrfalch wants to merge 2 commits into
spm/reserved-name-guardfrom
spm/spm-modules-guard
Open

Hold an app's spm.modules names to the same rules as a library's#58060
chrfalch wants to merge 2 commits into
spm/reserved-name-guardfrom
spm/spm-modules-guard

Conversation

@chrfalch

Copy link
Copy Markdown
Collaborator

Summary:

An app declares extra native modules through spm.modules in its react-native.config.js. Those names went into the generated package graph unvalidated, and two of the ways they can go wrong fail silently.

This PR fixes this by using the same Swift name collision detection/resolving as we introduced in #58044

NOTE: #58044 must be merged before this one so that we can change the base branch for this one to main

Changelog:

[IOS] [FIXED] - Reject colliding or invalid spm.modules names instead of silently dropping a module from the build

Test Plan:

✅ Unit tests/CI

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Aug 21, 2026
@chrfalch
chrfalch requested a review from cipolleschi August 21, 2026 13:41
@github-actions

Copy link
Copy Markdown

Caution

Invalid Base Branch

The base branch for this PR is spm/reserved-name-guard, which is not main or a -stable branch.
Are you sure you want to target this branch?

A dependency's Swift name is derived from its npm package name with the scope
dropped, which makes two collisions unavoidable: `@powersync/react-native`
derives `ReactNative`, one of React Native's own names, and `@a/foo` and
`@b/foo` both derive `Foo`. Either was emitted into the package graph as-is, and
SwiftPM then failed deep inside dependency resolution with a duplicate-name
error that named neither the library nor the react-native.config.js that caused
it.

The scope that was dropped is the fix. A scoped dep whose derived name is
reserved gets the TitleCased scope prepended (`PowersyncReactNative`), and deps
that resolved to the same name get it prepended too (`AFoo`, `BFoo`), so the
library author has nothing to do. Every disambiguation logs one line naming the
package, the name it would have taken and the name it got. Nothing can regress
on this — both collisions fail SwiftPM resolution today, so no working library
carries such a name and no consumer imports headers under it.

The two cases need different shapes. A reserved name is decidable per dep, so it
resolves in `resolveSwiftName`. A collision with another dep is not visible from
there, so it is a pass over the resolved set: group by name, and prepend the
scope to every scoped member of a group larger than one. Every member moves
rather than one arbitrary winner staying put, since there is no non-arbitrary
winner. Two members never move: a name the author set with `spm.name` (their
choice wins, and the others move around it) and an unscoped one (no scope to
borrow).

The pass runs once, and then the whole set is validated — this is the part that
has to be right. A borrowed scope can land on a name another dep already holds
(`@a/foo` → `AFoo`, next to a package `a-foo`) or on a reserved one, and two
libraries silently sharing a name is worse than the error this replaces. So both
existing checks now run over the final set, and anything a scope could not
resolve still fails with the message it did before: two unscoped deps deriving
the same name, an explicit `spm.name` that is reserved, a group whose only
scoped member's new name is taken. Retrying instead of failing would trade a
diagnosable error for a name nobody can predict.

Reserving the names React Native puts in a manifest is what makes the first case
diagnosable at all. Matching is case-insensitive throughout: a name that differs
from another only in case is not distinct enough for the build to keep the two
apart. The reserved check runs before the dep-vs-dep one, so the more specific
diagnosis wins, and it runs for a library that ships an autolinking plugin too:
`spm scaffold` knows nothing about plugins, so an exemption there would leave
the two commands disagreeing about the same library.

`SpmNameCollisionError` distinguishes a misconfiguration from a resolution
failure, so `scaffoldAll` — which degrades to the direct deps when a transitive
dep can't be found — still surfaces it instead of scaffolding manifests SPM
will reject. Its remote package config moves out of the same try for the same
reason.

Every Swift name reaching a manifest now comes from that one resolved set: the
autolinker's two `toSwiftName` fallbacks would have re-derived the
pre-disambiguation name and emitted a reference nothing matches, so they are
replaced by a required lookup that fails loudly.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@chrfalch
chrfalch force-pushed the spm/reserved-name-guard branch from c9aa6ae to b8cc821 Compare August 24, 2026 09:08
`spm.modules` names go into the generated manifest exactly as written, with no
validation at all: a name React Native reserves produced the same opaque
SwiftPM duplicate-name failure a library's name did, a name that is not a Swift
identifier produced a manifest SwiftPM refuses to parse, and two modules — or a
module and an autolinked library — could quietly claim the same target name.

Run the reserved-name and charset checks the library surface already has, and
check each name against the targets already emitted, so the app author is told
which entry to rename in their own react-native.config.js.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@chrfalch
chrfalch force-pushed the spm/spm-modules-guard branch 2 times, most recently from 8b362bc to cdaf50b Compare August 24, 2026 09:11
@chrfalch
chrfalch force-pushed the spm/reserved-name-guard branch from b8cc821 to 881411a Compare August 24, 2026 13:45
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. p: Expo Partner: Expo Partner

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant