The Scope Creep Automated reviews of new AI-related working group submissions

Update review · November 16, 2026

SCIM Interoperability Profile -02, Now You Can Remove a Group Member

draft-zollner-scim-interop-profile -02View on Datatracker →

What changed in -02?

Most of the document. It went from 11 pages to 18, and almost every section was touched.

The -01 review led with one problem: the PATCH rules made it impossible to remove one member from a group. That is fixed. So are the conformance signal, the IANA section, and most of the editorial items.

The policy choices are unchanged. Clients still get only eq and and. DELETE still has to free the userName. ETags are still banned.

The header now says "Updates: 7643". The abstract doesn't mention it.

Was the previous feedback addressed?

Fixed:

  • remove exists now. Section 6.5.2.2 defines it. members[value eq "2819c223"] is allowed for remove. The ban on targeting the element itself now applies only to add and replace. A remove that matches nothing returns success, so a client can retry.
  • The conformance boolean. interopProfileConformant is gone. Section 4 defines conformsTo.profiles, a list of profile URIs. Section 8.3 creates a "SCIM Profile Identifiers" registry, so the IPSIE profile has somewhere to register.
  • Non-idempotent add. Section 6.5.2.2 now says append "if it is not already present," with no change and no error for a duplicate.
  • The MUST/MAY contradiction. Section 6.5.3 now says a rejecting service provider MUST use the listed scimType values. A noTarget case was added.
  • TLS. Section 7.1 cites BCP 195 instead of naming versions.
  • Several omissions. excludedAttributes is a MUST (Section 6.3). Rate limiting with 429 and Retry-After has its own section (6.9.2). Clients can't rely on totalResults alone to find the last page (Section 6.4).
  • Editorial. The intro citation, the broken bullet lists, the *MUST* asterisks, and the Acknowledgements placeholder are all fixed.

Partly addressed:

  • Pagination. Service providers still pick "at least one" of index or cursor. Section 6.4 now says clients MUST support both, and service providers MUST publish the RFC 9865 pagination attribute and return 400 for a method they don't support. Discovery works now. Every client still implements both.
  • Versioning. The identifier is urn:ietf:params:scim:profile:interop, with no version in it.
  • IANA. The request is now executable. conformsTo is still a new top-level attribute on the core ServiceProviderConfig schema, not a schema extension.
  • Clearing a value. Section 6.5.2.1 says to use remove on a sub-attribute of a complex attribute. Nothing covers a simple singular attribute.

Not addressed:

  • eq and and only. Section 6.3 extends the limit to /.search and Section 6.5.2.2 to PATCH paths. Still no rationale and no answer for incremental sync.
  • Soft delete and identifier reuse. The MUSTs in Section 6.6 are unchanged. New Section 7.4 defends them: don't key access grants on userName, use id. The charter's soft-delete work is still not mentioned.
  • The ETag ban. Section 6.7 now calls it "a deliberate trade-off" and admits there is no lost-update protection. A server that doesn't evaluate the header SHOULD return 400.
  • JSON examples. Still none. The PATCH section has one inline operation object, and it shows what not to send.
  • patch.supported, the Location header, the User-Agent MUST, the Bulk section. All unchanged.
  • Sorting clarifications from restrictions. Still one pile.

One correction to my -01 review. I said RFC 7643 never gives id and externalId a caseExact value. It does, in Section 3.1. The draft was right.

Did the fixes create new problems?

A few.

replace on a complex attribute flipped. In -01, replace on name overwrote the whole object. In -02 it merges, citing RFC 7644 Section 3.5.2.3, and add behaves identically (Section 6.5.2.1). That matches the RFC. It also means clearing name takes one remove per sub-attribute.

Path-less PATCH went from MUST reject to MAY reject. Section 6.5.1.1 still forbids clients from sending it. A conformant server can now accept it. That fixes the contradiction by dropping the strict half.

The error codes moved. A bad operator was invalidSyntax and is now invalidFilter. A multi-match filter was invalidFilter and is now invalidPath. Unknown attributes went from invalidSyntax to invalidPath or invalidValue (Section 5.3). Anyone who coded to -01 has to redo this.

Mechanisms added or removed

Removed: the interopProfileConformant boolean and the TLS 1.3 requirement.

Added, and worth a look:

  • New PATCH syntax. Section 6.5.2.2 defines colors[value eq "blue"] to remove one value from a simple multi-valued attribute. The draft says RFC 7644 has no syntax for this. Service providers MUST support it. A profile that was only subtracting now extends the protocol.
  • Single-match filters. A PATCH path filter must match exactly one element.
  • No value on remove. Clients MUST NOT send it. Servers SHOULD reject it or treat it as a filtered remove, never as "remove all."
  • Filtering on unique attributes. eq must work on every attribute with a uniqueness constraint, and on externalId (Section 6.3).

Background of the author

No change. Danny Zollner (Okta) is still the sole author.

Questions worth raising

  1. Is new PATCH syntax in scope for a profile? The simple multi-valued remove form is a protocol extension. It may belong in the RFC 7644 revision.
  2. What replaces gt for incremental sync? Two revisions in, the restriction still has no rationale.
  3. Does Section 7.4 settle the soft-delete conflict? It argues against keying on userName. It doesn't say what the charter's soft-delete work should do.
  4. Can Section 6.5 get worked examples? The replace semantics just reversed between revisions. That is the argument for showing the JSON.