ReleaseKit/docs/consumers/check.md

Local candidate check

Use releasekit.Check to verify a candidate against the current local repository state:

result, err := releasekit.Check(ctx, releasekit.ReleaseRequest{
	RepositoryDir: "/work/project",
	Product:       "Example",
	Version:       "v1.2.3",
	Validation: releasekit.ValidationSpec{
		Go: releasekit.GoValidationSpec{Race: true, FullVet: true},
	},
	AdditionalChecks: func(ctx context.Context, candidate releasekit.Candidate) error {
		return runProjectChecks(ctx, candidate.RepositoryDir)
	},
})
if err != nil {
	return err
}
fmt.Println(result.Version, result.Commit)

The repository must have a clean index and worktree on main, and the selected version must be stable vMAJOR.MINOR.PATCH. The committed docs/releases/<version>.md note must have a first Markdown H1 containing the product name and exact version as complete title elements. No local tag with that version may already exist.

ReleaseKit runs one go test -mod=readonly -count=1 ./... suite with the available compatible local Go toolchain. Set Validation.Go.Race to run that suite with -race; set Validation.Go.FullVet to run go vet -mod=readonly ./... after successful tests, with test-time vet disabled for the suite. Set Validation.Go.CompileTargets to compile packages for each complete target pair after tests and vet. Package compilation uses CGO_ENABLED=0, writes no artifact, and does not execute a target binary; its failures use the package-compile identity and include the target. The zero values retain the ordinary native test behavior and select no compilation. Set Validation.GoFormat to check committed regular Go files using the Go formatter built into ReleaseKit. It includes vendor and nested-module source, ignores symlink blobs and worktree changes, and reports unformatted paths as go-format without rewriting files. Its zero value disables the check; it runs after tests, selected vet, and package compilation, before AdditionalChecks.

Set Validation.ModuleDistribution.WorkspaceFiles to reject committed go.work and go.work.sum files in the selected module inputs, including nested and symlink entries. It examines the captured candidate even when GOWORK=off, ignores uncommitted files, and does not change Go test workspace selection; vendored contents remain in scope. The zero value disables the check; failures use workspace-files. Set Validation.ModuleDistribution.LocalReplacements independently to parse the selected module's committed go.mod declarations. Vendored contents remain in scope, while independently released nested modules are excluded. It reports filesystem replace directives with their source path, line, and target as local-replacements; versioned replacements are permitted. Committed module symlinks are resolved within the candidate. Parse and inspection errors retain their causes, and the rule applies even when GOWORK=off. Its zero value disables the check. Set Validation.FileSize to check each committed blob against a byte limit. Nil disables the check; an empty object enables the default of 1,000,000 decimal bytes. LimitBytes is strict greater-than, accepts zero, and rejects negative values. Exclusions contains literal files or directories relative to the selected module; excluded directories include descendants. Symlink blobs are measured by link-text bytes, and gitlinks are omitted. The check applies to the captured candidate and reports one oversized file with its measurement as file-size; it does not measure historical or aggregate size. Set Validation.Markdown to check committed local Markdown destinations and supported heading fragments. Nil disables the check; an empty object enables it with outside-checkout destinations reported as out of scope. Set EnforceCheckoutBoundary to fail outside-checkout links. Submodule contents remain out of scope under either setting. The checker runs after file-size inspection and before AdditionalChecks; its nonfailure observations use the request's output writer. Markdown failures use the markdown-links identity and include the source location and destination.

Use these selections for ReleaseKit-owned validation rather than repeating the standard suite, selected vet/format/distribution checks, or Markdown navigation checks in AdditionalChecks or profile checks. Keep release-note rules and consumer-specific assertions—such as tests of maintained examples—in those product-owned checks. ReleaseKit calls AdditionalChecks once after shared validation when supplied. It does not download a different toolchain automatically. If race prerequisites are unavailable, the go-test check fails rather than being skipped; a selected full-vet failure is reported as go-vet. The suite uses the Go host target, disables workspace and module network access, and ignores ambient GOFLAGS, experiment, and CPU feature overrides. Effective cache and native cgo/C-tool settings are preserved, including values from Go's persistent environment file. The check revalidates the branch, commit, and cleanliness after validation.

For checksum failures during dependency preparation, use the module checksum investigation guide.

A successful check is evidence about the local candidate only. It does not check whether a remote branch or tag is available and does not create a tag. The check command reads the product and optional additional commands from releasekit.json.

Publishing the source tag

releasekit.Publish uses the same request and validation order: the standard suite runs first, followed by AdditionalChecks when supplied. It then performs authoritative remote checks, annotated-tag creation, and an exact tag push. The selected remote must publish main at the captured commit and must not already contain the selected version tag. A successful result includes the endpoint and tag object ID. Publication states distinguish an unattempted push, a tag that was created locally but not pushed, a successful push, a guard conflict, and an unresolved push error. See the publication guide for the complete workflow.

When a push reports an error, ReleaseKit makes one bounded query for the authoritative tag. An exact tag object match confirms success. An absent tag is reported as not_published, a different object as conflict, and an unreadable query as unknown. The local annotated tag is retained in every case; an unknown result is neither success nor proof that the remote lacks the tag.