Repository navigation
Proposal: define what the resume.metadata tamper-proof signature covers — a WYSIWYS consent binding for rendered surfaces #3309
Replies: 3 comments
|
Quick addition while this matures: I have a runnable reference implementation of the binding (resolves a surface, computes the consent digest over displayed content + the action's resolved target, binds it to a signed receipt, and refuses on mismatch), plus a protocol-neutral write-up (DOI: https://doi.org/10.5281/zenodo.23933180) and a repo at https://github.com/cleitonaugusto/wysiwys-consent-binding . Happy to turn this into a concrete spec PR whenever the maintainers see it as the right direction — just say the word. I also opened a parallel discussion on the AP2 side (google-agentic-commerce/AP2 #394) about how an agent-rendered surface can meet their "Trusted Surface" requirement, since the same primitive applies there — AG-UI being the runtime that carries the UI protocols seems like the natural place to define it once. |
|
Picking this back up with something concrete enough to implement against. Since opening this, I've written the property up as a protocol-neutral draft with a normative byte layout and conformance vectors, and sharpened it in public review. The short version for AG-UI: What { "wysiwys": { "alg": "sha256", "digest": "<base64>", "nonce": 7, "sig": "<optional>", "interruptId": "int_42" } }
The core is a gate at dispatch, not a stored receipt: before executing the resumed action, the verifier recomputes the surface it is about to act on and refuses on any mismatch. That closes the render→act (TOCTOU) drift with no trusted display required — the gate runs at the party that performs the call, which in an AG-UI stack is already downstream of the renderer. Being precise about what it is, because it came up in review: this is integrity of the surface about to be acted on, not authenticity of what a human's eyes saw. It catches a surface that drifted after approval; a renderer that lied before approval is out of scope (that needs a surface the verifier owns). The same line shows up in the lineages this borrows from — PSD2 dynamic linking (bind the auth code to amount + payee; any change invalidates it) and EIP-712 / ERC-7730 clear-signing in wallets — with the difference that in an agent stack the dispatch point is often downstream of the renderer, so the binding can cover the dispatched action, not just the displayed values.
I have a reference implementation with this exact |
|
One data point from payments, where we had to decide what the digest covers. At send21 (agents prepare payment drafts; a person approves and signs each one in a wallet the agent doesn't control) we hash the canonical action rather than the rendered surface. Each draft has a Two reasons we went that way:
So I'd lean toward minimal: a digest over the resolved action fields, plus a rule that the displayed values are rendered from those same fields. |
Uh oh!
There was an error while loading. Please reload this page.
AG-UI already anticipates the right primitive: the interrupt/resume model carries an optional
resume[].metadatadescribed as "a signature proving the human decision was not tampered with", andtoolCallIdlinks a decision back to a tool call. What the spec intentionally leaves open is what that signature covers — today it's producer-defined and "never interpreted by a consumer".That open-ness is also a gap worth closing, especially for consequential / HITL approvals. Because AG-UI is the runtime that carries generative-UI surfaces (A2UI, MCP-UI, Open-JSON-UI), the content a human actually sees comes from one side, while the decision/resume carries
{approved: true}on the other — and nothing defined binds the two. For payments/commerce surfaces that means a surface can show one thing and the agent act on another, within a fully conformant run.Proposal (additive, opt-in): define a small consent binding — a deterministic digest over the resolved displayed surface and the action's resolved target, carried in
resume.metadata(and optionally signed). A verifier recomputes the surface it intends to act on and refuses on mismatch. This is "what you see is what you sign" (trusted-path / hardware-wallet), applied to agent tool-calls. Defined at the AG-UI layer, A2UI / MCP-UI / Open-JSON-UI all inherit it from one place.I've written a protocol-neutral draft of the primitive — WYSIWYS Consent Binding for Agent UIs (DOI: 10.5281/zenodo.23933180) — and have a runnable reference implementation (resolves a surface, computes the digest over displayed content + resolved target, binds it to a signed receipt, refuses on mismatch). Happy to turn this into a concrete spec PR if there's interest.
Would love the maintainers' and partners' take on scope (minimal vs full surface) and on whether an explicit signature contract belongs in-spec or in deployments.
All reactions