Update review · November 20, 2026
draft-sweeney-wimse-credential-delegation-02 Update
What Changed
This update covers two revisions, -01 and -02. The draft roughly doubled, from 22 pages to 43.
The core architecture is the same. A Delegation Server (DS) holds the user's credentials in a vault, issues DPoP-bound Delegation Tokens to agents, and proxies every API call so the agent never sees the raw credential.
Most of the new text is in the delegation chain. The draft now leans heavily on draft-asor-wimse-agent-delegation-chain, and positions itself as the server-side half of that design.
Was the Previous Feedback Addressed?
The grant type contradiction: fixed. The IANA section no longer registers urn:ietf:params:oauth:grant-type:cred-delegation. The draft says no part of the protocol used it.
The abstract still claims the document defines no new token formats. That is harder to defend now. The draft registers six JWT claims and an HTTP header, and it adds a new "delegation grant" artifact whose token type identifier is listed as an open issue.
The agent_checksum claim: fixed. It is removed, along with the IPR footnote.
The Exercise Proxy: not addressed. Section 6.2 is unchanged. It still says the DS "maps these to the upstream API call" with no request format. This remains the biggest interoperability gap in the draft.
The new Implementation Status section makes the problem concrete. The author's own SDK does not use /exercise with an operation name. It exposes POST /api/v1/use and takes an upstream URL, method and body. The one existing implementation chose a different interface from the one the draft describes.
The revocation SLAs: partly addressed. The 1 second and 5 second bounds are still MUST-level. The draft now scopes them to a single DS and says cross-domain enforcement is out of scope. That answers the federation concern by excluding it.
The new text is honest about the limits. revoked_count "reports invalidation, not closure." A call already dispatched upstream is not recalled. A token used away from the DS can outlive a revocation by up to 1 hour. The implementation tests the 5 second cascade up to 1,111 descendants, and does not measure the 1 second bound at all.
Protocol version negotiation: not addressed. It is still there, and the draft now asks IANA to register Cred-Protocol-Version as a provisional HTTP field.
Agent lifecycle and 90-day consent retention: not addressed. Both remain.
The over-specified vault and DS internals: not addressed. The architecture section is unchanged. A new Section 10.9 does state the cost plainly: keeping credentials away from agents "concentrates" the exposure at the DS.
The February 26 disclosure reference: not addressed. [OAUTH-SPLICING] still points at the OAuth mail archive index, not a specific message.
The DID dependency: not addressed. did:key is still a normative reference to a W3C CCG draft. The draft adds optional ML-DSA keys "where a did:key encoding for it is available."
New Mechanisms
Constraint ceilings (Section 5.3). A new delegation_constraints claim replaces the old constraints member of the capability object. It carries eight typed constraints: max, min, one_of, not_one_of, prefix, rank, max_lifetime and max_subtree. The vocabulary and the subsumption rules are taken from the Asor draft. Parsing is fail-closed, and an unknown constraint type denies the action.
The cumulative types need a single writer to hold the running total. The DS is that writer. This is the best argument in the draft for keeping a server in the path.
Two-step sub-delegation (Section 5.4.1). The delegating agent asks the DS to authorize a hop and gets back a single-use delegation grant bound to the receiving agent's DID. The receiving agent then redeems it with a Token Exchange request under its own key. This closes a real gap. In -00 the child token could be requested by a party other than the agent it named.
client_id in the Delegation Token. It names the acting agent and must match act.sub. This follows the convention in draft-ietf-wimse-aims.
Chain-verifiable profile (Section 5.5). Tokens meant to be verified away from the DS use the Asor token format unchanged. The DS can issue them. How a chain crosses between DS-issued and offline-minted tokens is left to a later revision.
Workload Authorization Grant composition (Section 4.4). A WAG identity can be recorded in the delegation receipt as evidence. It must not appear in act.
Per-hop re-verification (Section 5.4). The DS must re-check the parent's signatures, expiry, DPoP binding and revocation state before issuing any child.
Implementation Status (Section 9). The author's TypeScript SDK passes all 36 Asor test vectors. The section also states that the SDK "is not wire-compatible with this document."
New security text. Section 10.3 now admits that DPoP covers the method and URI only. Every exercise goes to one endpoint with the operation in the body, so the proof does not bind the operation. Section 10.9 states that a DS-signed receipt does not prove the user approved anything.
Mechanisms Removed
- The
agent_checksumclaim - The grant type URI registration
- The
constraintsmember inside the capability object, includingmax_callsandmax_response_size
New Problems
A Standards Track draft now depends normatively on an individual draft. The Asor draft is at -03 and not adopted by any working group. Sections 5.3 and 5.5 cannot be implemented without it. Related Work even tracks the differences between Asor -02 and -03. If that draft changes or stalls, this one breaks.
There are now two token formats. A DS issues cred_delegation tokens for its own use and agent_delegation tokens for offline verifiers. The draft says they share a vocabulary but not a wire format. Implementers have to build both.
"Strict subset" is inconsistent. Section 5.2 now says a child may be "equal or narrower." The Terminology section still says an attenuated capability MUST be a strict subset.
The scope is growing, not shrinking. The previous review suggested cutting mechanisms to focus on the DS interface. This revision adds five and removes one claim. The Related Work section now reads like a changelog of other people's drafts.
Bottom Line
The author is responsive. The changelog credits six reviewers by name, and the fixes to the chain logic are real improvements. The draft is also more honest about its limits than most.
So is it closer to something a working group could adopt? Not yet. The agent-to-DS exercise interface is the part two independent implementations must agree on, and it is still one paragraph. The author's own code does not follow it. Until Section 6.2 defines a request format, the rest of the draft describes one product.
The useful next step is to split the work. The constraint and chain rules belong with the Asor draft. The DS exercise interface and revocation endpoint could stand alone as a much shorter document.