Skip to content

Cross-product ring VRF: hosts diverge (Android prompts, Rust core refuses) and the allowlist gate has no source #373

Closed

Description

@peetzweg

RFC-0024 gates foreign-key create_account_proof and ring_vrf_sign on the owning product having allowlisted the caller "in its manifest, with no user-prompt fallback", an explicitly interim position. Building on it, we found the interim state is two different behaviors across hosts, and on core-based hosts the allowlist has no source at all, so it is permanently empty. We also verified that giving the existing gate a source is a trivial change — the logic is already complete.

What we ran

truapi-host CLI from feature/people-rfc (8933422), two runs sharing one wallet session on paseo-next-v2:

  1. --product-id peopl.dot: register_ring_vrf_key(Left(1), people-lite ring) succeeds; owner ring_vrf_sign returns a 64-byte signature.
  2. --product-id dim2.dot, same session, using peopl.dot's key handle:
call stock feature branch
list_ring_vrf_keys (Anonymized / PublicKey) OK, behind account-access consent
get_account_alias OK (returned the correct product context 0x80eb3c17…)
create_account_proof NotAllowlisted, no prompt
ring_vrf_sign NotAllowlisted, no prompt

Where the refusal actually is

The effective gate is at the runtime frontend, not the authority layer: create_proof (runtime.rs ~1048) and ring_vrf_sign (~1188) both do key_handle.dot_ns_identifier != self.product_id() → NotAllowlisted before delegating. require_owned_ring_vrf_key in signing_host.rs / pairing_host.rs is a downstream duplicate of the same comparison. Nothing on either path reads a manifest, config, or flag, so no allowlist entry can ever exist.

The gate logic is complete — it only needs a source

We patched the two frontend checks to also consult an env var TRUAPI_RING_VRF_ALLOWLIST="owner:caller". With peopl.dot:dim2.dot set, the same run 2 becomes:

call with a source
ring_vrf_sign OK — 64-byte member-key signature, byte-identical to peopl.dot's own
create_account_proof NotMember — advanced past the gate; the fresh key is simply not in the ring yet

So the refusal is purely the absence of a source, not missing enforcement.

Host divergence

Android (polkadot-app-android-v2 master, RealAccountsProtocol.kt) implements the same wire calls with a one-time cross-product user prompt instead of a hard refusal, with a code TODO: keep the prompt "until the manifest RFC lands". So today the same request succeeds on Android and is refused by every core-based host (CLI now, iOS/desktop once the Rust core lands). Products cannot write against that.

Separately, the manifest RFC (#206, rfc/product-manifest branch) contains no ring-VRF allowlist field yet, so merging it as-is would still leave the gate sourceless.

Use case

Personhood as a product, as RCC-0024 describes: peopl.dot owns the member keys; dim2.dot / game.dot drive sign-up and registration flows needing a member-key ownership signature. These are mutually trusting, privileged first-party products — the allowlist entry is uncontroversial; only the mechanism is missing.

Asks, in priority order

  1. Pick one interim behavior for the core and name it in the RFC so hosts stop diverging — either adopt Android's one-time prompt (the core already routes foreign alias requests through AccountAccessReview and could reuse it), or keep the hard refusal everywhere including Android.
  2. Give the gate a deterministic source for tests/dev now, e.g. a CLI --allow-ring-vrf <owner>:<caller> flag feeding the existing check (our env-var patch shows the whole change is a few lines and needs no wire/type work).
  3. Spec the ring-VRF allowlist field in the manifest RFC (RFC: Product Manifest Format #206) so the interim state has an end date.

Reproduction (two ~80-line CLI scripts + the env-var patch) available on request.

Activity

  1. peetzweg commented on Aug 26, 2026

    @peetzweg
    MemberAuthor

    Reproduction is on a branch now, rebased onto current main (6ca0b38, post RFC-0024 landing and the Index/Raw rename):

    https://github.com/paritytech/host-rust-core/tree/spike/rfc24-cross-product-repro

    Two commits:

    • 05db993: the env-var source for the gate (TRUAPI_RING_VRF_ALLOWLIST="owner:caller"), consulted at the two runtime.rs frontend checks and the signing_host / pairing_host duplicates. Spike only, not for merging.
    • dc984ea: rfc24-cross-a-register.ts (run as peopl.dot) and rfc24-cross-b-consume.ts (same session, run as dim2.dot), plus RFC24-CROSS-PRODUCT-REPRO.md with the exact commands and expected output.

    Re-ran today on this base: stock, createAccountProof and ringVrfSign with the foreign handle return NotAllowlisted. With TRUAPI_RING_VRF_ALLOWLIST=peopl.dot:dim2.dot, ringVrfSign returns the member-key signature (byte-identical to peopl.dot's own) and createAccountProof advances to NotMember. Same result as in the description, the gate only lacks a source.

  2. filvecchiato commented on Sep 16, 2026

    @filvecchiato
    Collaborator

    closing in favour of #655

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions