Skip to content

platform-wallet: multi-account send persists per-account net_amount (+ null fee, stale pending_inputs, block-gated spend reconciliation) #4387

Closed

Description

@bfoss765

platform-wallet: multi-account send persists per-account net_amount (+ null fee, stale pending_inputs, INCOMING-until-confirm) — store-consistency, currently masked on Android

Summary

When a Core (L1) transaction spends inputs owned by more than one account (e.g. a plain BIP44 receive + a DashPay-friendship receive, or pooled BIP44/BIP32 funding), the persisted transactions row does not describe the wallet, only one contributing account. Concretely, on a send:

  • net_amount is stored as one account's change - spent, not the wallet total (understated by the other accounts' contribution).
  • fee is NULL.
  • The just-broadcast send is first persisted as direction = INCOMING, net_amount = +change (the change output read as a receive), and only corrected to Outgoing when the tx confirms and txos.isSpent/spendingTxid are written.
  • pending_inputs reservation rows are not cleaned up after confirmation.

None of this is user-visible on Android today — it is fully masked app-side (see below) — but the stored records are wrong until confirmation, and iOS/tooling reading these fields directly will surface the errors.

Root cause (code)

packages/rs-platform-wallet/src/changeset/core_bridge.rs:981-996 documents the mechanism precisely: upstream check_core_transaction emits one TransactionRecord per matched account, so a payment touching a contact produces two records sharing one txid —

record's account direction net_amount
funding (BIP44/BIP32/…) Outgoing change - spent
DashpayExternalAccount Incoming +paid

The persisted transactions row is keyed by txid alone, with no per-account dimension (transaction_account_involvements is only written for provider-key accounts), so last-write-wins. The watch-only exclusion (is_contact_watch_only, dashpay/rust-dashcore#926/#952) correctly drops the contact record, but the surviving funding record still reflects only that account's spent/change — not a wallet-total aggregation across every owned account that contributed inputs. fee is not computed in this projection. pending_inputs (Room table; SDK-side reservation set) is written at broadcast but has no confirmation-time cleanup.

Reproduction (Dash testnet, mainnet-equivalent code path)

Wallet holds two owned UTXOs in different accounts: 0.2 DASH in BIP44 (plain receive) + 0.01 DASH in a DashPay-friendship receiving account (received from a contact). Send-max to an address:

  • On-chain: correct — one tx spends both, delivers 0.209993, fee ~700 duffs. Recipient credited correctly.
  • transactions row: direction = Outgoing, net_amount = -20000000 (-0.2, missing the 0.01 input), fee = NULL.
  • txos: both inputs isSpent = 0 until the tx confirmed (~minutes), then isSpent = 1 + spendingTxid linked.
  • pending_inputs: 2 reservation rows remained after confirmation (isSpent already 1).

(Single-account sends do not exhibit the net_amount error — the funding record then is the wallet total.)

User-visibility & how Android masks it (for the iOS team / to size the SDK fix)

Zero user-facing impact on Android — two app-side layers absorb it:

  1. net_amount / fee correction — SdkTxStoreWalker recomputes the true wallet net (and fee) from the transaction's own inputs/outputs and overrides the stored value for display, persisting the correction once confirmed spent-evidence exists (t.spendingTxid = tx.txid). So the UI showed the correct -0.21, not the stored -0.2.

  2. INCOMING-until-confirm ("lingering receive") correction — the app treats a received output as spent-for-display iff txos.isSpent = 0 AND txos.spendingTxid IS NULL AND its outpoint ∈ pending_inputs (excluding foreign/watch-only rows). This is display-only (never persisted from a reservation; persisted corrections still require confirmed spent-evidence), and it yields automatically to any SDK authority — the moment the SDK writes isSpent/spendingTxid (at broadcast, confirmation, or via a future API), the app path contributes 0. Stale post-confirm pending_inputs rows are therefore harmless (isSpent=1 wins).

iOS has neither layer today, so it will show the wrong net_amount/fee and a phantom INCOMING row for a just-broadcast multi-account send until confirmation. iOS can either mirror the two Android workarounds or wait on the SDK fix below.

Additional finding: spend reconciliation is block-gated, not islock-gated

On Dash, InstantSend lock (~2s) is finality. The receive side honors this (txos.isConfirmed = 1 on islock). The spend side does not: after an InstantSend-locked send, the input txos.isSpent stayed 0 and the durable net_amount/fee write-back did not fire for 30+ minutes — both wait for block confirmation. Evidence: an islocked send (InstantLocked event logged) whose 5 input UTXOs remained isSpent = 0, isInstantLocked=1-then-cleared, isConfirmed = 1 well past any block interval. This is non-user-facing on Android today (the display uses the islock-independent pending_inputs reservation, so balance/amount are correct immediately), but the durable store state — and therefore any consumer reading it, plus the net_amount/fee correction — lags for no reason on a chain where the spend is already final. Spend reconciliation should trigger on InstantSend lock, consistent with the receive side.

Proposed SDK direction (additive, preserves current invariants)

Priority ordered; each is independently useful:

  1. Aggregate at persist time: compute net_amount and fee for the transactions row across all owned accounts that contributed inputs/outputs (wallet-total), rather than storing one account's view. This removes the need for the app-side recompute entirely.
  2. Clean up pending_inputs when the spending tx confirms (or is dropped/re-orged).
  3. Optionally, expose an additive pendingSpend read (reserved outpoints for in-flight local spends) so clients can render the correct pre-confirm spent state without changing txos.isSpent's "spent-and-confirmed" invariant and without reading the raw reservation table. This is the clean contract that lets both Android and iOS drop their workarounds.

Note: we deliberately did not propose flipping txos.isSpent at broadcast — that would break its on-chain-truth invariant and is unsafe in the drop/re-org case. (3) is the safe equivalent.

Not blocking

Android ships the masks above; this is store-hardening, not a user-facing defect on Android. Filed for correctness + to unblock iOS parity.

Activity

  1. bfoss765 commented on Aug 17, 2026

    @bfoss765
    CollaboratorAuthor

    Fix implemented (local branch, held until device-verified per our QA flow): fold_same_txid_records() merges the per-account records into one wallet-level row at the two seams where siblings co-occur — the BlockProcessed projection and CoreChangeSet::merge. Net = Σ of the per-account slices (disjoint detail sets ⇒ the wallet's Σreceived − Σspent), details unioned, fee from the funding record, direction recomputed from the merged net. 67/67 changeset tests green including the pre-existing contact-watch-only projections and new coverage for this issue's exact sweep shape.

    Scope notes against the issue's full list:

    • per-account net_amount — fixed at batch scope (detection emits the sibling records together). Cross-batch stragglers keep last-write semantics; the Android mirror-heal covers rows persisted before the fix and goes inert on folded rows.
    • fee NULL — the fold carries the funding record's fee onto the persisted row when any sibling knew it.
    • INCOMING-until-confirm and stale pending_inputs — not addressed here; those are display/confirm-time semantics with their own app-side mitigations (the pending-reservation display correction), and belong to a separate persistence change.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    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