Cannot bundle directory inside `embed.FS`

[pete-woods]

There seems to be some confusion with it sometimes using relative paths, and otherwise not. I see an attempt to fix this previously, but I'm not sure it was a complete fix.

Edit: I created a PR #498 with some failing tests. I don't know how to fix it, though!

To reproduce, clone the DigitalOcean spec:

git clone https://github.com/digitalocean/openapi

Then in the specification directory, create the following test:

package openapi_definitions

import (
	"embed"
	"fmt"
	"io/fs"
	"testing"

	"github.com/pb33f/libopenapi"
	"github.com/pb33f/libopenapi/bundler"
	"github.com/pb33f/libopenapi/datamodel"
	"gotest.tools/v3/assert"
)

//go:embed DigitalOcean-public.v2.yaml
var v2 []byte

//go:embed description.yml resources shared
var files embed.FS

func Doc() (libopenapi.Document, error) {
	return libopenapi.NewDocumentWithConfiguration(v2, &datamodel.DocumentConfiguration{
		LocalFS:             &wrapper{files},
		AllowFileReferences: true,
	})
}

func TestDigitalOcean(t *testing.T) {
	doc, err := Doc()
	assert.NilError(t, err)

	model, err := doc.BuildV3Model()
	assert.NilError(t, err)

	_, err = bundler.BundleDocument(&model.Model)
	assert.NilError(t, err)
}

type wrapper struct {
	delegate fs.FS
}

func (w *wrapper) Open(name string) (fs.File, error) {
	open, err := w.delegate.Open(name)
	fmt.Println("!!!!", name, err)
	return open, err
}

For me it gets upset about loading all sorts of files. I've tried various tricks, like setting the base path to /, wrapping the embed.FS, etc, but nothing seems to do the trick.

!!!! resources/addons/responses/resources/gen-ai/genai_delete_workspace.yml open resources/addons/responses/resources/gen-ai/genai_delete_workspace.yml: file does not exist

You can see it doubling up on the path names: resources/addons/responses/ + resources/gen-ai/genai_delete_workspace.yml

[pete-woods]

I created a PR #498 with some failing tests. I don't know how to fix it, though!

[pete-woods]

Narrowed it down to somewhere inside rolodex.CheckForCircularReferences()

[pete-woods]

PR updated with an attempt to fix the issue, though I see now you've got an existing test that fails when you provide a non-rolodex FS implementation.

I'm now unsure if this is a "you're holding it wrong issue" now I know that you've intentionally guarded against passing a non-rolodex FS into the bundler. It seems rather unfortunate if you can't use a plain fs.FS like the interface implies, though.