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

Review · November 21, 2026

Extend-Only, Threshold-Attested Authority Renewal for Partitioned Agent Meshes

draft-chandra-wimse-partition-tolerant-renewal -00View on Datatracker →

Problem Statement

Short-lived workload credentials assume the issuer is reachable. This draft covers the case where it isn't: a workload is mid-task, its credential is about to expire, and the network path to the issuer is down.

The proposed answer is to let a quorum of nearby peers vouch for extending the expired credential, within limits the issuer signed in advance.

How the Spec Fits In

The draft positions itself under the WIMSE charter's local-issuance work item, which has no adopted draft yet. That makes the topic in scope.

The mechanism is a poor match for the rest of WIMSE, though. It doesn't issue a WIMSE token. It produces a separate JSON proof chain that travels next to an expired credential.

It also doesn't build on the WIMSE credential formats. Every identity is an Ed25519 did:key, and every signature is raw Ed25519 over JCS-canonicalized JSON. The WIMSE credentials draft uses JWTs and X.509 certificates, with URI identifiers. The draft says did:key is "consistent with WIMSE Workload Identifier practice". It isn't.

All WIMSE references are informative. The normative dependencies are JCS, I-JSON, CDDL, base64 and RFC 3339.

This is an entirely new protocol with five new object types and a new role (the attesting peer). It also creates a new relationship: relying parties must opt in to accepting expired credentials when a valid chain is attached.

The title says "Agent Meshes", but nothing in the body is specific to AI agents.

High-Level Summary

There are four players: the issuer, the requester (the cut-off workload), its peers, and the relying party.

Before any outage, the issuer signs a Mesh Manifest. It names at least 5 member workloads, a ceiling on total extension, a maximum scope, a clock tolerance, and a validity window.

During an outage, the requester broadcasts a signed Renewal Request to its peers, along with its original credential. Each peer runs ten checks on its own. These include: the credential is authentic, the scope is unchanged, the ceiling isn't exceeded, the credential isn't in the peer's cached revocation list, and the peer has itself recently seen the requester active.

A passing peer signs a Peer Attestation over the hash of the full request. Once two-thirds of the peers have signed, the requester bundles them into a Renewal Proof.

Longer outages produce a Renewal Chain. Each round's request includes the hash of the previous proof.

The relying party verifies the manifest, the original credential, and the chain. If all pass, it treats the credential as valid through the extended expiry at the original scope.

After the outage, the requester must get a fresh credential from the issuer and hand over the chain for audit.

The draft is upfront about what it doesn't do. It isn't consensus. A revocation issued during the outage has no effect until the network heals. Peer "health" observations show the workload was active, not that it's uncompromised.

Significant Interoperability Concerns

No mapping to real WIMSE credentials. Peers must compare original_scope, requester_did and original_issuer_did against the credential's claims. A Workload Identity Token has no scope claim, its subject is a URI, and its issuer is not a did:key. The draft says credential authentication is out of scope. Two implementations can't interoperate on a WIT without inventing this mapping themselves.

No transport. The draft doesn't say how a request reaches peers, or how a chain is presented to a relying party next to an HTTP request. This is listed as an open issue. Without it, this is a data format, not a protocol.

Revocation state is undefined. Peers and relying parties check "cached revocation state" by credential hash. No WIMSE mechanism distributes revocation that way.

Health evidence is unverifiable. The four values are self-reported labels. One of them is other. A relying party can't distinguish a real observation from a rubber stamp.

Is the quorum needed at all? The issuer already accepts exposure up to the ceiling. In the worked example, a 4-hour token has a 12-hour ceiling. The draft admits a requester can claim the entire ceiling in a single round on one set of observations. At that point, the issuer could have issued a 12-hour credential, or a pre-signed backup credential, with no new protocol. The working group should ask what the quorum buys beyond a one-time liveness check.

Mechanisms That Could Be Removed

  • The redundant attestation fields (request_id, nonce, original_credential_hash, granted_extension_seconds). The draft itself says the request digest already covers them.
  • The health_evidence vocabulary. It doesn't change any verifier's decision.
  • Rules like "MUST NOT represent this mechanism as providing partition-proof revocation". These constrain marketing, not wire behavior, and can't be tested.
  • The Related Work material on registry leases, the NANDA Index, and the author's other draft.

Writing Quality

The text is grammatical and technically careful. The verification steps are precise, and the CDDL appendix is useful. The security considerations are honest about limits, more so than most individual drafts.

It is hard to read. The abstract is one 330-word paragraph. Object fields are listed in long run-on paragraphs instead of tables. The same points (self-attestation, digest binding, "this is not consensus") are restated in three or four places.

Much of the text reads as a rebuttal to an earlier security review, which the acknowledgements confirm. The spec would be a third shorter if it stated each rule once.

The wire version is renewal/0.3 in a -00 draft, with a reference implementation at 0.4.0 on PyPI. This was developed as a product first and brought to the IETF afterward.

Background of the Author

Sharath Chandra (Stellarminds.ai)

  • RFCs published: 0
  • Active Internet-Drafts: 2, this one and draft-chandra-agent-registry-corroboration
  • Expired drafts: 0

This is a new IETF author. Both drafts are recent individual submissions in the AI agent space.

Bottom Line

The problem is real, and the charter names it. This draft solves it with a standalone threshold-signature protocol that shares almost nothing with the WIMSE credential formats.

Before this is useful to the working group, it needs a concrete binding to the WIMSE credentials draft, and a clear answer to why a longer-lived or pre-issued backup credential doesn't solve the same problem.