fix: fix errors during repeated pull and rm operations; add tag validation on pull to ensure idempotency.

[PikaByter]

related issue：#483

FIx:
add tag validation when pulling

Test:

make build
chmod +x ./output/modctl
cd output
./modctl pull <image>
./modctl rmi <image>
./modctl pulll <image>
./modctl rmi <image>

[gemini-code-assist]

Code Review

This pull request introduces a StatTag method to the Storage interface and its distribution implementation to verify tag existence. This new method is utilized in the pullIfNotExist function to ensure manifests are pushed when a tag is missing, even if the underlying blob is already present. The changes also include minor formatting adjustments to model file constants and updated mocks. A review comment suggested using errors.As for more robust error handling of wrapped ErrTagUnknown errors.

[gemini-code-assist]

Using switch err.(type) to check for distribution.ErrTagUnknown may fail if the error is wrapped. Since the project uses wrapped errors (e.g., %w), it is safer and more idiomatic to use errors.As to check for specific error types.

Suggested change

	if err != nil {
	switch err.(type) {
	case distribution.ErrTagUnknown:
	return false, nil
	}
	return false, err
	}
	if err != nil {
	var tagErr distribution.ErrTagUnknown
	if errors.As(err, &tagErr) {
	return false, nil
	}
	return false, err
	}

[aftersnow]

(duplicate, please see the English review below)

[aftersnow]

Thanks for the contribution! The fix direction is correct. A few suggestions:

Required

1. Missing digest validation on tag re-creation path: When manifest exists but tag doesn't, io.ReadAll(reader) reads content and pushes it, but the function returns early before validateDigest(). Please add validation or comment why it's safe to skip.

2. No unit test coverage: The core new path (StatManifest=true, StatTag=false) has no tests. Please add coverage.

Suggestions

3. The formatting changes in constants.go are unrelated to the fix — consider a separate commit.

4. The second commit message feat: update dis is unclear — please reword it.

[aftersnow]

Follow-up review on the current head (6125024). I reproduced #483 on today's main (f241cec) with a local registry, so the fix is still needed:

pull 1  -> Successfully pulled, `ls` shows the tag
rm 1    -> Deleted: v1
pull 2  -> Successfully pulled, but `ls` shows 0 rows (tag not recreated)
rm 2    -> Error: failed to delete manifest v1: filesystem: Path not found: .../_manifests/tags/v1

The direction of the fix (re-create the tag when the manifest blob already exists) is right. What still blocks it:

1. Digest validation is skipped on the re-tag path (pkg/backend/pull.go). When StatManifest is true and StatTag is false, the code reads the remote manifest with io.ReadAll(reader) and stores it under the tag, then returns before validateDigest runs. A corrupted or substituted response from the registry is stored and tagged without a check. Please validate the digest of body against desc.Digest before PushManifest, the same as the non-existing path does.

2. A stale tag is not refreshed. StatTag only reports existence. If the local tag exists but points to an older manifest (the remote tag moved), the pull still skips and leaves the stale tag in place. repository.Tags(ctx).Get(ctx, tag) already returns the descriptor; comparing its digest with desc.Digest and re-tagging on mismatch fixes this with no extra storage call. That would also make StatTag return the digest instead of a bool.

3. No unit test for the new branch (StatManifest=true, StatTag=false). The storage mock already has StatTag, so a table test on pullIfNotExist is cheap.

4. Unrelated reformatting of pkg/modelfile/constants.go conflicts with main today and is not part of the fix. Please drop it (or send it separately).

5. DCO fails. Both commits need a Signed-off-by line (git commit -s).

The errors.As change for distribution.ErrTagUnknown that Gemini asked for is in place.