feat(go)!: publish the wrapper as moq.dev/moq - #2957
Conversation
The Go modules were named after the mirror repos they happen to be published from, so the ergonomic wrapper imported as `github.com/moq-dev/moq-go/moq` and the raw bindings as `github.com/moq-dev/moq-go-ffi/moq`. Both now use vanity paths that moq.dev owns: moq.dev/moq -> github.com/moq-dev/moq-go moq.dev/moq-ffi -> github.com/moq-dev/moq-go-ffi moq.dev serves the `go-import` meta tag the go command reads to find the repo, so the import path is no longer hostage to a repo name. Renaming or relocating a mirror costs a line in a table on the site instead of a breaking change for every consumer. The wrapper package moves from go/wrapper/moq/ to go/wrapper/, the module root. That is what makes the import path end in the package name, so `import "moq.dev/moq"` binds `moq` with no alias. package-wrapper.sh stages the sources flat to match. The bindings keep their moq/ subdirectory, since the import path already ends in `moq` and every caller aliases it to `ffi` anyway. BREAKING CHANGE: consumers must change their imports. `go get github.com/moq-dev/moq-go@latest` will fail with a module-path mismatch rather than a friendly deprecation, because both paths resolve to the same mirror repo and its newest tag now declares the new path. Versions already pinned to the old path keep resolving; only @latest and fresh installs break. VERSION goes 0.5 -> 0.6, so the next release is v0.6.0. Verified against the live resolver: with moq.dev already serving the tags, `go get moq.dev/moq@latest` today follows the meta tag, clones the mirror, and rejects v0.5.14 for declaring the old path. That is precisely what this commit fixes. `just go check` builds moq-ffi, regenerates the bindings, and reports `ok moq.dev/moq` from the staged module. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: da406c8476
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| cd "$PKG_STAGE" | ||
| go mod edit -require="github.com/moq-dev/moq-go-ffi@v${FFI_VERSION}" | ||
| go mod edit -dropreplace="github.com/moq-dev/moq-go-ffi" | ||
| go mod edit -require="moq.dev/moq-ffi@v${FFI_VERSION}" |
There was a problem hiding this comment.
When this reaches main, release-go.yml resolves the current Rust FFI version (0.3.11) and accepts its already-existing mirror tag, but that tag's go.mod still declares github.com/moq-dev/moq-go-ffi. Consequently the go mod tidy below fails with a module-path mismatch when this new requirement fetches moq.dev/moq-ffi@v0.3.11, so no v0.6 wrapper can be packaged. The inspected .github/workflows/release-go-ffi.yml only republishes on a new moq-ffi-v* tag, and this change does not cause such a tag, so the renamed FFI module must be released first or the release orchestration must explicitly bootstrap it.
Useful? React with 👍 / 👎.
Review finding from Codex. The gate checked that the ffi's mirror tag exists, which used to be the same question as "can the wrapper require it". With the rename it isn't: every tag published so far declares github.com/moq-dev/moq-go-ffi, so a main push would sail through the gate, pin moq.dev/moq-ffi@v0.3.11, and then die in package-wrapper.sh's `go mod tidy` on the path mismatch. Local `just go check` can't catch that, since it stages with --skip-tidy. So read the tagged go.mod and require the new path before packaging. A tag that predates the rename now defers exactly like a missing one: the next moq-ffi release republishes the mirror under the new path and the workflow_run chain re-cuts the wrapper. Network failures still fail loudly rather than masquerading as a deferral, since nothing retries a wrapper-only push. Verified against the live mirror: v0.3.11 reads back as the old path and defers. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5566a13fb4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| @@ -1,9 +1,9 @@ | |||
| module github.com/moq-dev/moq-go | |||
| module moq.dev/moq | |||
There was a problem hiding this comment.
This commit is based directly on the repository's main commit, but changing the published module from github.com/moq-dev/moq-go to moq.dev/moq invalidates every existing consumer import and is explicitly accompanied by a breaking 0.5 to 0.6 version-line bump. Retarget this change to dev; merging it directly into main bypasses the repository's required staging path for breaking changes to language wrappers. (Written by GPT-5.6 Sol)
AGENTS.md reference: AGENTS.md:L205-L207
Useful? React with 👍 / 👎.
1 participant
The Go modules were named after the mirror repos they happen to be published from, so the ergonomic wrapper imported as
github.com/moq-dev/moq-go/moqand the raw bindings asgithub.com/moq-dev/moq-go-ffi/moq. Both now use vanity paths that moq.dev owns:moq.dev/moqmoq.dev/moq-ffimoq.dev serves the
go-importmeta tag the go command reads to find the repo (moq.dev#126, already deployed), so the import path is no longer hostage to a repo name. Renaming or relocating a mirror now costs a line in a table on the site instead of a breaking change for every consumer.The package moved to its module root
go/wrapper/moq/*.gois nowgo/wrapper/*.go. That is what makes the import path end in the package name: Go binds the identifier from thepackageclause, but tooling adds an explicit alias whenever it disagrees with the last path element, whichmoq-godid. At the module root,import "moq.dev/moq"bindsmoqwith nothing added.package-wrapper.shstages the sources flat to match.The bindings keep their
moq/subdirectory (moq.dev/moq-ffi/moq). That path already ends inmoq, the generator emits into that layout, and every caller aliases it toffiregardless.Breaking
Consumers must change their imports.
go get github.com/moq-dev/moq-go@latestwill fail with a module-path mismatch rather than a friendly deprecation, because both paths resolve to the same mirror repo and its newest tag now declares the new path. Versions already pinned to the old path keep resolving; only@latestand fresh installs break. There is no way to soften this while the two paths share one repo's tags, and per CLAUDE.md's deprecation rule the dead name stays out of/docrather than being advertised there.VERSIONgoes 0.5 -> 0.6, so the next release is v0.6.0 (publish-wrapper.shderives the patch from existingv0.6.*tags, of which there are none).Testing
just fix,just check, andjust testall pass.just go checkbuildsmoq-ffi, regenerates the bindings withuniffi-bindgen-go, stages both modules, and reportsok moq.dev/moqfrom the staged tree, which confirms the package really is at the module root.The interesting one is the live resolver. moq.dev is already serving the tags, so the real go command can be pointed at the vanity path today:
That is the entire chain working: resolve
moq.dev/moq, read the meta tag, clone the mirror, read itsgo.mod, and reject the tag for declaring the old path. This PR is what makes the next tag declare the right one.Base
Targets
devper CLAUDE.md: a changed module path is a semver break in a published API.(written by Opus 5)