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

Review · November 16, 2026

Transaction Tokens BCP Has Real Deployment Lessons and Several Conflicts With the Base Spec

draft-araut-oauth-transactiontokens-bcp -00View on Datatracker →

The problem it's solving

Transaction Tokens define a token format and a request to the Transaction Token Service (TTS). They don't say how to roll that out across hundreds of services.

This draft tries to fill that gap. Section 1.2 says the guidance is "derived from production deployments of Txn-Token systems." It covers token size, lifetime, propagation libraries, validation rollout, telemetry, and async workloads.

The deployment problems are real. Most of the answers are internal engineering practice, not protocol.

How it fits in the working group

It's a companion to an adopted working group draft, so the topic is in charter. It presents itself as deployment guidance that falls "outside the scope of the protocol specification," and it has no IANA actions.

That framing doesn't hold. The draft introduces three things with wire impact and no wire definition:

  • A Batch Token with a new exchange at the TTS (Section 3.9).
  • A placeholder token that travels where a Txn-Token would (Section 3.5.6).
  • A context relocation endpoint and a reference inside the Txn-Token (Section 3.2.1).

It cites Txn-Tokens -06. The current version is -11.

So does it line up with the current Txn-Tokens draft? No, in five places.

  • Validation is optional here. Section 3.6.4 says deployments MUST support TOKEN_PRESENCE_MODE, where "no cryptographic verification is performed," and SHADOW_MODE, where requests always proceed. Section 3.6.3 says services SHOULD NOT automatically fail requests when Txn-Tokens are "missing or invalid." Txn-Tokens Section 12.2 says a workload MUST validate the JWS signature.
  • Propagation is optional here. The kill switch (3.5.10) and the denylist (3.5.1) both stop a workload from forwarding the token. Txn-Tokens Section 12.2 says outbound calls MUST include the Txn-Token as received.
  • Placeholder tokens are unsigned. Section 3.5.6.3 says they "MUST NOT be cryptographically signed." Txn-Tokens defines a Txn-Token as a signed JWT and says the Txn-Token header value MUST be exactly one Txn-Token. The draft never says which header carries a placeholder.
  • Tokens are encrypted here. Section 3.5 says "tokens contain encrypted sensitive data" and Section 3.6 has libraries doing decryption. Txn-Tokens defines a signed JWT and no encryption.
  • Trust boundaries inside one organization. Section 3.5.9 has workloads request new Txn-Tokens "appropriate for the target trust domain." Txn-Tokens Section 14 points to Identity Chaining for crossing trust domains. Section 13.15 says a replacement MUST NOT change aud.

These read like descriptions of one existing system that predates the spec. Section 3.6.1 talks about cached "token schemas" for decoding, which also has no counterpart in a JWT-based Txn-Token.

What the draft actually recommends

The core advice is sensible:

  • Keep tokens small. Only add optional context when a requester asks for it (3.1).
  • Keep lifetimes under 5 minutes and don't re-request on expiry (3.3). This matches Txn-Tokens Section 6.
  • Govern what goes in the token. Once a claim is visible to every service, someone will depend on it (3.4).
  • Use request-scoped storage in propagation libraries. A token attached to the wrong request returns the wrong user's data with no error (3.5.7).
  • Add Txn-Token context to cache keys (3.5.11).
  • Roll out enforcement progressively and measure before enforcing (3.6.4).

The Batch Token is the biggest addition. A workload trades a valid Txn-Token for a long-lived "voucher," with a lifetime of "24 hours to 7 days." A worker later exchanges it at the TTS for a fresh Txn-Token with the original claims.

Txn-Tokens Section 6 already addresses batch and offline tasks. It says they still result in a short-lived Txn-Token. It defines no way to carry the original context across the pause. That gap is real, and this draft at least attempts to fill it.

Significant interoperability concerns

The Batch Token is a protocol with no specification. There's no token format, no token type URN, no request parameters, and no claim for the "use case ID." Section 3.9.3 requires an "ECDSA-signed preservation context" without saying what it contains. Section 4.7.1 says Batch Tokens MUST be "sender-constrained or scoped to specific namespaces." Those are not equivalent protections. A week-long credential that mints Txn-Tokens needs a real security analysis.

Placeholder tokens put unsigned data in the token path. Four reason codes and four camelCase fields are listed, with no format. The draft says they MUST NOT be used for access decisions. Combined with TOKEN_PRESENCE_MODE, a service checking only for presence could accept one.

Hard numbers lack justification. "MUST NOT issue tokens larger than 4KB" is derived from Apache's 8KB default header limit. That's a deployment choice, not a MUST.

MUST is used for organizational process. "Organizations MUST document and communicate enforcement promotion timelines" is one example. There are many.

Mechanisms that could be removed

Most of Section 3.5 and 3.6 is library implementation detail.

  • Language tables (3.5.3). Java ThreadLocal, Python ContextVar, and a priority of "Java first."
  • Cross-thread capture and restore (3.5.4). Good engineering. Not IETF material.
  • Kill switch details (3.5.10). JVM system properties and a 180 second polling interval.
  • Key and schema caching (3.6.1). Daemon threads, jitter percentages, Lambda cold starts.
  • Telemetry and audit fields (3.7). Includes "requester line of business."
  • Message transport (3.9.4). SQS attribute limits.

What's left is two things. One is a short list of deployment considerations that could be an appendix or security considerations text in Txn-Tokens itself. The other is the async problem, which needs its own protocol draft.

Writing quality

This is a first draft that wasn't proofread after conversion.

Raw markdown headings appear inline in the text ("### Size Limits Transaction Token Services MUST NOT..."). Bullet lists and one ASCII diagram are collapsed into paragraphs. The header says "WG Working Group." The "latest revision" link is https://example.com/LATEST. There are two reference sections, a duplicated RFC 2119 paragraph, and "TODO acknowledge."

Terminology drifts. Section 3.9 uses "TraT," the rest uses "Txn-Token." Section 4.8 mentions a "handle" token type that's never defined. Em-dashes are everywhere.

The security considerations mostly restate Section 3 with MUST swapped in for SHOULD.

Background of the author

Ashay Raut (Amazon) is the sole author. The Datatracker profile shows no RFCs and two active drafts: this one and Transaction Tokens for Agents. Both build on Txn-Tokens.

Raut is listed in the acknowledgments of the Txn-Tokens working group draft. This is a new IETF author with hands-on deployment experience in this specific area. The same pattern shows up in both drafts: a useful core idea, plus assumptions that don't match the current base spec.

What the WG should ask at IETF 127

This is on the joint OAuth/WIMSE session agenda. Questions worth putting to the author and the room:

  1. Is this describing Txn-Tokens or a predecessor system? Encryption, schemas, and unsigned placeholders suggest the latter. The draft should be rebased on -11.
  2. Should a BCP permit skipping signature validation and propagation? If rollout modes are needed, that's feedback for the Txn-Tokens editors before the base spec is finished, not a companion document that contradicts it.
  3. Does the WG want to solve async and batch? If yes, split the Batch Token into a separate, fully specified draft. Section 6 of Txn-Tokens leaves the gap open.
  4. Where does the rest belong? A short deployment considerations section in Txn-Tokens would capture the size, governance, mix-up, and caching advice in about a page.

The WG should not adopt this as is. The operational experience behind it is worth getting into the base spec while it's still open.