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

Review · August 26, 2026

SCIM 2.0 Interoperability Profile

draft-zollner-scim-interop-profile -01View on Datatracker →

draft-zollner-scim-interop-profile-01 — Danny Zollner (Okta), submitted 15 April 2026. Individual submission, 11 pages, no working group adoption activity yet.

The Problem Statement

SCIM's optionality is the product of a many-to-many market, and it makes every integration bespoke.

One identity provider provisions into hundreds of SaaS applications. Each of those applications accepts connections from dozens of identity providers. RFC 7643 and RFC 7644 leave PATCH syntax, pagination, filtering, case sensitivity, and error handling open enough that two conformant implementations routinely fail to talk to each other. The result is that every IdP maintains per-application quirk configuration, and every application maintains per-IdP quirk handling.

This draft picks the specific places where implementations have diverged and closes them. Mandate PATCH, ban PUT. Require eq and and, forbid everything else. Require the discovery endpoints. Say what add versus replace means on a complex attribute. Say that DELETE deletes.

The framing is worth stating plainly because it's unusual: this document adds no new capability to SCIM. It subtracts. Its entire value is in the requirements it removes from the implementer's discretion.

How the Spec Fits In

This is chartered work, and it already has a dependent.

The SCIM charter says the working group "will also consider publishing extensions to SCIM that have found broad adoption. These extensions may include profiles and schemas for interoperability in additional use cases." A profile is explicitly in scope. draft-schreiber-scim-ipsie-profile, the SCIM half of the OpenID Foundation's IPSIE work, normatively requires conformance to this draft and layers enterprise security requirements on top. So there's a two-layer profile stack forming, and this document is the bottom layer. If the working group wants the IPSIE profile, it needs this one or something like it.

Dependencies are as small as they get: RFC 7643, RFC 7644, RFC 9865 for cursor pagination, RFC 8446 for TLS, plus BCP 14. No new roles, no new endpoints, one new ServiceProviderConfig attribute. There are no informative references at all.

Two places where it collides with other chartered work.

Soft deletes. The charter names draft-ansari-scim-soft-delete as input for "Handling Deletes in SCIM Servers that don't allow Deletes." §6.6 of this draft says a service provider MUST NOT respond to DELETE by modifying attributes, MUST return 404 for deleted resources, MUST omit them from all query results, and MUST make userName available for reassignment. That is a direct prohibition on soft delete. Both can't be right for the same implementation, and the working group will have to decide whether the profile carves out an exception or the soft-delete work is scoped to non-profile deployments.

Incremental sync. The charter wants an "Incremental approach to synchronization," and there's an expired draft-sehgal-scim-delta-query with this author's name on it. §6.3 says clients MUST NOT use any filter operator other than eq and and. Incremental sync needs meta.lastModified gt "...". Under this profile a client cannot ask "what changed since Tuesday." The profile forecloses the mechanism the charter asks for.

There's also a document-placement question. Roughly half of what's here is clarification of RFC 7643 and RFC 7644 rather than restriction of them: case sensitivity applied consistently between filtering and uniqueness, PATCH is not a delete, DELETE actually deletes, uniqueness violations return 409 with scimType of uniqueness, /Schemas must match what the server implements. The charter's first goal is a revision of 7643 and 7644 that incorporates "errata and interoperability feedback." Those requirements belong in the revision, where they bind everyone, not in an opt-in profile that a non-conformant server can simply decline. The genuinely profile-shaped requirements are the subtractions: PATCH only, eq and and only, no ETags, the restricted PATCH grammar. Sorting the document into those two buckets is the most useful structural feedback available.

High-Level Summary

Two parties, both already defined by SCIM. The Client is the identity provider pushing changes. The Service Provider is the application receiving them. Nothing new is introduced between them.

Conformance is advertised one direction only. A service provider claiming conformance returns interopProfileConformant: true in its ServiceProviderConfig. A non-conformant one omits it or returns false. There is no equivalent signal from the client side.

The substantive requirements group into four areas.

Discovery becomes mandatory and has to be honest. ServiceProviderConfig, /Schemas, and /ResourceTypes must all exist. /Schemas must match exactly what the server implements, and every schema named in /ResourceTypes must be retrievable. Each ResourceType must have an id equal to its name. In deployed SCIM these endpoints are frequently missing, stale, or aspirational, so this is one of the higher-value requirements in the document.

Requests are strictly validated. Send an attribute or schema URI the server doesn't define and you get 400 with invalidSyntax. Today most servers silently drop unknown attributes, which means a client can misspell an attribute name and successfully provision incomplete data for years. Strict rejection is the right call, and it is a breaking change for a lot of shipped client code.

PATCH becomes the only way to update, and PATCH gets a grammar. RFC 7644 §3.5.2 makes PATCH an OPTIONAL server function. This profile inverts that: PATCH is mandatory, and clients SHALL NOT use PUT. Then it pins down the eight or so behaviors implementations disagree on. Every operation needs an explicit path, so path-less PATCH is out. One operation per attribute path. On a simple attribute, add and replace mean the same thing. On a whole complex attribute, replace overwrites everything and add merges. On a whole multi-valued attribute, replace overwrites the collection and add appends.

Banning PUT is the correct instinct. Read-modify-write PUT is how large groups get emptied and how attributes the client never knew about get cleared. But it's worth being clear that the profile mandates SCIM's optional update mechanism and forbids its baseline one, and never says a conformant server must advertise patch.supported: true in ServiceProviderConfig, which is where a client would look.

Everything else gets a floor. One endpoint per resource type. Address resources by id, never by userName. JSON with application/scim+json. At least one of index or cursor pagination. At least 100 results per page on request. No conditional request headers. TLS 1.3 required, HTTP Basic not recommended. Bulk stays optional.

Significant Interoperability Concerns

The PATCH restrictions make it impossible to remove one member from a group. This is the one to fix first.

§6.5.2.2 says that when a filter targets an element in a multi-valued complex attribute, the path MUST target a specific sub-attribute of the matched element. Targeting the element object itself, emails[type eq "work"], is PROHIBITED.

That is the exact construction RFC 7644 uses in its own examples for removing a group member. "path": "members[value eq \"2819c223-...\"]" appears three times in RFC 7644 §3.5.2, including in the worked example titled "remove a member from a group." Every SCIM client in production removes group members this way.

Compounding it: the word remove never appears in this draft. §6.5.2 defines add and replace and stops. So the profile prohibits the standard removal idiom, defines no replacement, and never specifies remove semantics at all. Under a literal reading, the only way to remove one email or one group member is replace on the whole collection, which requires the client to read the current state first and write it back. That is read-modify-write, which is precisely the hazard that motivated banning PUT.

The prohibition is presumably aimed at ambiguity in what replace on a matched element means when the value object is partial. That's a real ambiguity and worth resolving. But it needs to be resolved without removing element-level remove.

Restricting clients to eq and and is a bigger cut than it looks. gt on meta.lastModified is how incremental sync works. co and sw are how admin search UIs work. ne and pr show up in real client logic. The draft forbids clients from using them, which also means a conformant service provider has no obligation to support them, so the ecosystem calcifies at two operators.

There's a reasonable argument that broad filter support is the single worst-implemented part of SCIM and that cutting to eq and and is the honest floor. If so, the draft should say that, and it should say what the migration path is for the use cases it breaks. Right now the restriction arrives with no justification and no acknowledgment that it forecloses the delta query work.

"At least one of index or cursor pagination" reintroduces the optionality the profile exists to eliminate. A client still has to implement both and detect at runtime which one it got. That's a conditional branch in every client, which is the thing the document is supposed to be removing. The fix is to require index-based pagination as the floor, since it's in the RFC 7644 baseline and universally implemented, and allow cursor as an addition. If cursor is preferred for large sets, say that and reference RFC 9865's discovery mechanism, which this draft cites normatively but never tells the client how to use.

Mandatory identifier reuse after deletion will get pushback, and it should. §6.6 requires that a deleted resource's userName MUST NOT count toward uniqueness and MUST be available for reassignment. Plenty of applications deliberately tombstone usernames so a new hire cannot inherit a former employee's mailbox, audit trail, or ACL entries keyed on username. This MUST forbids a practice that exists for good security reasons. If the intent is narrower, that a client re-creating a user it just deleted must not hit a spurious 409, say that instead.

interopProfileConformant doesn't scale and has no version. A single boolean says "conformant to some revision of this document." This is already at -01 and there will be more revisions. The IPSIE profile layered on top will need its own signal, and so will every profile after it, which means ipsieProfileConformant, then a third one. The right shape is a multi-valued attribute of profile identifier URIs, one URI per profile version, which solves versioning and layering at the same time and gives the IPSIE draft something to register. Worth fixing now while there's exactly one profile.

The IANA request cannot be executed as written. §8 asks IANA to update the registry entry for urn:ietf:params:scim:schemas:core:2.0:ServiceProviderConfig to "include or reference" the new attribute. That registry maps schema URIs to defining documents. It doesn't hold attribute definitions, and adding a top-level attribute to the core ServiceProviderConfig schema means modifying RFC 7643. The mechanism SCIM provides for this is a schema extension: define something like urn:ietf:params:scim:schemas:extension:interop:2.0:ServiceProviderConfig, register that URI, and have conformant servers list it in the schemas array of their ServiceProviderConfig response. Also, the section opens with "Prior to being published as an RFC, this document requests that the IANA..." IANA actions happen at publication, not before.

Error handling is normative in one section and advisory in another. §6.5.1.1 says a service provider MUST reject a missing path with 400 and invalidSyntax. §6.5.3 then says a service provider MAY return 400 for any violation of these constraints, and that specific scimType values "should be used." That's a direct MUST/MAY contradiction on the same behavior. And the lowercase "should" is non-normative in a document that just invoked BCP 14, so the scimType mapping is advice. Error codes are what client retry logic branches on. If they aren't normative, clients still need per-server error handling, and a good chunk of the profile's value evaporates.

add on a collection is defined as append, which makes it non-idempotent. §6.5.2.2 says the service provider MUST append the provided values to the existing collection. Nothing says what happens if a value is already there. A sync client whose request times out and retries gets the email or the group member twice. The draft should say the server MUST NOT create duplicates, and define what counts as a duplicate for a complex value.

Half the requirements bind clients, and nothing lets a server know it's talking to a conformant one. Clients MUST NOT use PUT, MUST NOT use other operators, MUST NOT send ETag headers, MUST only use canonical type values. A service provider cannot detect any of that in advance, so it still has to handle everything defensively. The User-Agent requirement in §6.2 is the closest thing to a signal and it isn't machine-readable. This makes the document a procurement checklist for service providers and a coding-standards document for clients, which is fine, but it should be honest that the client-side requirements are unenforceable.

The case sensitivity section claims to restate RFC 7643 and partly doesn't. §5.2 says the profile "requires adherence to the case sensitivity definitions specified in [RFC7643]" and then asserts caseExact is true for id and externalId. RFC 7643 defines those two as common attributes in §3.1 prose and never gives them a caseExact value; they appear in no schema representation. userName genuinely is "caseExact": false in §8.7.1, so that one checks out. The draft is defining new behavior for id and externalId while framing it as a restatement. It should just define it. Separately, RFC 7643 says the server does not enforce externalId uniqueness, so requiring case-sensitivity consistency "with uniqueness enforcement" for that attribute is describing something that doesn't exist.

Notable omissions for a document whose thesis is "we picked the things that break." No requirements on attributes and excludedAttributes, which servers honor inconsistently. Nothing on totalResults accuracy under index pagination, which is what clients use to size a sync job. Nothing on the Location header for 201 responses. Nothing on 429 and Retry-After, which is a top cause of broken provisioning at scale. And nothing on how a client clears a value: replace with null, remove, or an empty array. That last one may be the single most common SCIM interop bug in the field, and it's absent from a document built specifically to fix common interop bugs. The selection criteria for what made the list would be worth stating.

Mechanisms That Could Be Removed

The User-Agent MUST. §6.2 requires clients to send it and says the value "should be meaningful." No protocol behavior depends on it, no server can be required to do anything with it, and there's no defined format. It's an operations nicety expressed as a normative requirement, and it's untestable. Drop it or make it a SHOULD in a separate operational-considerations section.

§6.8, "Support for the /Bulk endpoint is OPTIONAL." That's already true in RFC 7644. The section adds nothing. Either make a decision about bulk or cut the section.

The Content-Type requirement in §6.2. RFC 7644 §8.1 already requires this and the draft cites it while restating it.

The userName caseExact restatement in §5.2. RFC 7643 already says caseExact is false. The new and valuable requirement is that filtering and uniqueness enforcement use the same rules. Keep that, cut the restatement, and separately define the values for id and externalId since those genuinely aren't in RFC 7643.

The blanket ETag prohibition in §6.7 is stronger than needed. Conditional requests are how you avoid lost updates. Forbidding clients from ever sending If-Match removes an available safety mechanism to normalize behavior across servers that don't implement it. Saying that servers are not required to support ETags and that clients MUST NOT depend on them accomplishes the interop goal without banning the header.

§7.1's TLS version requirements. BCP 195 exists for this. Cite it instead of restating specific version requirements that will need revision.

The requirements worth protecting if the document has to shrink are §5.1 discovery accuracy, §5.3 strict attribute rejection, §5.2's filtering-and-uniqueness consistency rule, §6.5.2.1's add/replace equivalence on simple attributes and merge-versus-overwrite on complex ones, and §6.6's PATCH-is-not-a-delete. Those are the ones that resolve genuine field disagreements.

Writing Quality

Concise, and that's the right instinct. Eleven pages with no filler, no marketing language, and no restatement of SCIM basics. The document went through the toolchain properly: Discussion Venues section, table of contents, BCP 14 boilerplate, correctly formatted references.

The one place brevity hurts is §6.5. There are eight distinct PATCH behaviors defined there in pure prose with zero JSON examples. For a section whose entire purpose is disambiguating a construct that implementations already read differently, prose alone is not enough. A worked request and response for each case would make the section unambiguous and would likely surface the missing remove semantics during authoring.

Concrete cleanup items.

The Introduction reads "The SCIM 2.0 standard RFC7643 [RFC7644]." It names one RFC and cites the other. The same bug appears in this author's group-members draft, which suggests a copied template rather than a one-off typo.

The kramdown source has broken bullet lists in at least two places. §5.2 renders as "Common Attributes ([RFC7643], Section 3.1): * id: caseExact is true. Service Providers MUST treat abc-123 and ABC-123 as distinct values. * externalId (optional): ..." all in one paragraph. §6.5.3 has the same problem with the scimType list. Both need a blank line before the list.

Normative keyword emphasis is inconsistent. Some are wrapped in asterisks that render literally in the text output, *MUST*, and others are plain, MUST. Roughly half and half, which reads like two authoring passes that didn't get reconciled.

§6.3's "The use of filters in the URI for any HTTP method other than GET MUST NOT be used" is passive, has no actor, and says "use ... be used." It should name the client.

§6.9 is titled "Error Handling" with one subsection. §6.5.3 is also titled "Error Handling." Merge them or retitle.

§6.4's "MUST return fewer results than requested when the client specifies a count value that exceeds that limit" is trivially satisfiable and doesn't say the important part, which is that the client has to be able to tell it was capped. It should require itemsPerPage to reflect what was actually returned.

Acknowledgements is a TODO placeholder.

There are no informative references. The document doesn't cite the IPSIE profile that depends on it, RFC 7642, the working group's use-cases-reloaded draft, or the delta query and soft delete work it conflicts with. A reader can't tell from this document that any of those relationships exist.

Background of the Author

Danny Zollner (Okta) has one published RFC: RFC 9865, cursor-based SCIM pagination, October 2025, with Mike Peterson and Anjali Sehgal. This draft cites it normatively.

Four active drafts, all SCIM:

  • draft-zollner-scim-interop-profile
  • draft-zollner-scim-group-members
  • draft-wzdk-scim-agent-resource (co-author)
  • draft-schreiber-scim-ipsie-profile (co-author)

Expired drafts include draft-ietf-scim-roles-entitlements, a working group document, plus draft-sehgal-scim-delta-query, draft-zollner-scim-referential-value-location, and draft-zollner-scim-domain-extension.

Not a new author. Has shipped an RFC in this working group, has edited an adopted working group draft, and every document with his name on it is SCIM. Sole author on this one, which for an interoperability profile is worth flagging as a gap rather than a criticism: a profile that constrains what implementations may do gets its authority from having multiple implementers behind it. A single-vendor author list on a document that tells other vendors which SCIM features they may not use is going to draw that question in the adoption discussion. The fix is co-authors from other implementations, not different content.

Suggested Feedback

The problem is real, it's chartered, and the IPSIE profile already depends on this document existing. Most of the useful feedback is about individual requirements rather than the premise.

Start with group member removal. §6.5.2.2 prohibits emails[type eq "work"] as a path, which is the construction RFC 7644's own examples use to remove a member from a group. The word remove appears nowhere in the draft. As written, taking one user out of a group requires replacing the whole members array, which is the read-modify-write pattern the PUT ban exists to prevent. Ask for remove semantics to be specified and for element-level targeting to be permitted for remove specifically.

Ask what happened to gt. Restricting clients to eq and and means no client can ask what changed since a given timestamp. That forecloses the incremental synchronization the charter calls for and the delta query draft this author co-wrote. If the restriction is deliberate, the document should say so and say what replaces the capability.

Push on pagination. "At least one of index or cursor" leaves every client implementing both plus runtime detection, which is the exact optionality the profile exists to remove. Require index as the floor, allow cursor as an addition, and point at RFC 9865's discovery mechanism so a client knows how to find out.

Raise the soft delete conflict directly with the working group. §6.6 requires that DELETE really deletes, that deleted resources vanish from query results, and that usernames become reusable. The charter names soft delete as in-scope work. Those two positions need reconciling before either document advances, and the username reuse requirement in particular forbids tombstoning that many applications do on purpose.

Ask for the conformance signal to be redesigned before it ships. A single unversioned boolean can't express which revision of the profile, and it doesn't extend to the IPSIE profile layered on top or to whatever comes after. A multi-valued list of profile URIs solves both, and it's a cheap change now and an expensive one after the first RFC.

Flag the IANA section as needing rework. You can't add an attribute to the core ServiceProviderConfig schema through a registry update. Define a schema extension URI and register that.

Ask for the scimType values to be normative. §6.5.3 currently says a server MAY return 400 and "should" use particular error types, which contradicts the MUST in §6.5.1.1 and leaves clients doing per-server error handling anyway.

Ask for JSON examples in §6.5. Eight PATCH behaviors defined in prose, in a document written because implementers already disagree about PATCH, is the one place where the draft is too short.

On process, the sharpest question for the working group is a sorting question. This document mixes clarifications of RFC 7643 and RFC 7644, which bind everyone and belong in the chartered revision, with restrictions on optional features, which only make sense as an opt-in profile. Deciding which requirements go where is more valuable than reviewing them as a single pile, and it would shrink this document to the part that genuinely needs to be a profile.