@scality/cloudserverclient points to a local file path (file:../cloudserverclient/scality-cloudserverclient-v1.0.9.tgz). This will break CI and other developers' builds. Must be changed to a proper registry version or git-pinned tag before merge.

— Claude Code

Arsenal is pinned to 8.3.9, but checkCrrCascadeEvent and decode are imported from versioning.VersionID. Neither checkCrrCascadeEvent nor the getMicroVersionId() method (called on source/dest entries throughout this PR) appear to exist in arsenal 8.3.9 or in backbeat's own models. The arsenal dependency likely needs a version bump for this PR to work.

— Claude Code

Any 409 from the destination is assumed to be a cascade-stale scenario and the replication is silently marked COMPLETED. If cloudserver ever returns 409 for a different reason, the object would never be replicated. Consider checking for a more specific signal (e.g. a response body field or custom error code) rather than relying solely on the HTTP status code.

— Claude Code

Can already check this pr, but should really be reviewed after all the other cascade prs, as changes in these pr would also mean changes here

I think we can functional tests instead of just these,
But waiting for Arsenal/cloudserver to be merged, as it will be easier to make these tests (functional tests in backbeat rely on an image of cloudserver)

@scality/cloudserverclient is pinned to a local file path (file:../cloudserverclient/scality-cloudserverclient-v1.0.9.tgz). This will break CI and other developers' builds. Before merging, this needs to be changed to a proper npm version (e.g., ^1.0.9) or a git tag reference consistent with the other git-based deps in this project.

— Claude Code

When cascadeLoopDetected or cascadeDataComplete is returned as an error from a part, this error flows through the retry wrapper in _getAndPutPart. The BackbeatTask.retry() method checks err.retryable === undefined and then inspects err.code, err.name, and err.message to detect network errors. Since these cascade signals are plain objects ({}), none of those properties exist, so no mutation occurs today. But if a future change accidentally adds such a property to these sentinels, retry() would mutate the shared singleton (setting retryable = true) and corrupt all subsequent cascade checks. Consider using Object.freeze() on the sentinel objects at line 33-35 to prevent accidental mutation.

— Claude Code

Bug: when err is truthy (line 434 returns early), destLocations is used in the _deleteOrphans call, but destLocations was computed on line 432-433 with the !err guard already making allPartsAlreadyAtDest false. The filter on line 433 still runs (partResults || []).filter(...) which accesses partResults — but after an error, partResults from mapLimitWaitPendingIfError may contain partial results mixed with partAlreadyAtDest sentinel values. These sentinel objects (empty {}) will pass the result && result !== partAlreadyAtDest filter and be passed to _deleteOrphans, which tries to read .key and .dataStoreName from them. Since they're empty objects, they'll produce undefined keys — likely harmless because _deleteOrphans filters out falsy .key, but worth a defensive check or comment explaining why this is safe.

— Claude Code

Arsenal dependency is pinned to a raw commit hash (2c429ab) instead of a release tag. All other git-based deps in this file use tags (e.g. breakbeat#v1.0.3, bucketclient#8.2.4). Pin to a tagged release once the arsenal PR is merged to keep the dependency auditable and consistent with the project convention.

— Claude Code

Vendoring a .tgz into the repo adds a binary blob to git history that can never be garbage-collected. Prefer publishing @scality/cloudserverclient@1.0.9 to the registry (or using a git tag reference like the other deps) and pin to it here. If vendoring is intentional for CI reasons, it should be called out in the PR description.

— Claude Code

Trailing whitespace on this line (after the opening brace {). Minor, but some CI lint checks catch this.

— Claude Code

In _processQueueEntryRetryFull, the second waterfall stage now passes allPartsAlreadyAtDest as the mdOnly argument to _putMetadata. This means a full retry where all parts happened to already exist at the destination will send ReplicationContent: 'METADATA' in the putMetadata call. Verify this is the intended behavior — it changes the semantics from "always full replication on retry" to "maybe metadata-only on retry if cascade detected all parts present." If a retry was triggered because metadata was stale but data was fine, this seems correct; if the retry was triggered for a data integrity reason, skipping the data write could mask the issue.

— Claude Code

cascadeLoopDetected is passed as the err argument via doneOnce(cascadeLoopDetected). This sentinel is an empty object {}, so upstream retry logic in BackbeatTask._retry will see it as a truthy error and may retry the operation before _handleReplicationOutcome gets to check for it. Confirm that _getAndPutPart (the retry wrapper around _getAndPutPartOnce) won't retry on this sentinel — if it does, the loop detection is bypassed and the part is retried unnecessarily.

— Claude Code

Arsenal is pinned to a raw commit hash (2c429ab...) instead of a semantic version tag. Per project conventions, git-based deps (arsenal, vaultclient, etc.) should pin to tags (e.g. #8.x.y). Commit hashes are opaque — it's unclear which features/fixes are included, and there's no semver contract. This also makes it harder for reviewers and operators to reason about what changed.

— Claude Code

A binary .tgz is committed to the repo and referenced via file: in package.json. This permanently inflates the git history (binary blobs can't be diffed or garbage-collected), prevents code review of the dependency contents, and bypasses the normal package registry workflow. Consider publishing @scality/cloudserverclient@1.0.9 to the npm registry (or a private registry) and referencing it as a versioned dependency instead.

— Claude Code