Skip to content

Confirm a finalized root's payments in one Multicall3 read - #238

Merged
RunTimeAdmin merged 1 commit into
mainfrom
fix/sdk-evidence-tool-reads
Nov 8, 2026
Merged

RunTimeAdmin merged 1 commit into
mainfrom
fix/sdk-evidence-tool-reads

Conversation

@RunTimeAdmin

Copy link
Copy Markdown
Owner

The bug

sigvara_verify_evidence asked the reputation contract's verifyEvidence about each covered payment in a separate call, up to 100, eight at a time. Arc's public RPC limits calls a second. Measured on 7 November with every read logged: the 29th to 32nd calls within about a second came back rate limit exceeded (ethers reports it as "missing revert data"), and one refusal failed the whole check. So the tool failed 8 of 8 tries on the simulation's solo agent, on both rpc.testnet.arc.io and rpc.testnet.arc.network. It worked only on agents with few payments.

The fix

  • SigvaraVerifier.verifyEvidenceMany(did, items) asks once, through Multicall3's aggregate3 at its usual address, which Arc testnet has and the site already uses. Where Multicall3 has no code, it asks one at a time, in turn. Whether it has code is read once per verifier, and a failed read is asked again.
  • The tool calls it with the last 100 covered payments. Nothing else in its result changes.

Tests

  • New test/evidence-confirm.test.ts. It runs the tool against a local node that refuses past 20 eth_calls a second, over 60 payments, and checks all 60 are confirmed with no refusal and one aggregate3. It also checks each answer is read from the batch (a wrong leaf comes back false), and the fallback without Multicall3. Both tests fail on the old code.
  • test/mcp.test.ts stubs verifyEvidenceMany in place of verifyEvidence, and the first test now expects one read for all twenty payments.
  • 371 SDK tests, all passing locally. Type-check clean.

Live check

The built branch against the live solo agent on 8 November, at about 00:12 UTC, three times across both public RPCs: each time it reproduced the finalized root over 70 payments, including the two Base payments, and the contract confirmed all 70, in about a second.

Also

  • Bumps the SDK to 1.0.0-alpha.18, with a changelog entry, not yet published.
  • Adds a sentence about 1.0.0-alpha.18 to the three docs that say alpha.17 still fails. The SDK README says the confirmations are one read. The README and whitepaper SDK counts are updated in place.
  • After publishing, the plugin PR feat(wallet-mcp): add Sigvara plugin coinbase/agents#20 gets its pin bumped and its evidence flow back.

🤖 Generated with Claude Code

sigvara_verify_evidence asked the reputation contract about each covered payment
in a separate call, up to 100, eight at a time. Arc's public RPC refused the 29th
within a second with "rate limit exceeded", and one refusal failed the whole
check, so the tool failed on every agent with a long payment history.

SigvaraVerifier.verifyEvidenceMany asks once, through Multicall3's aggregate3,
and one at a time where Multicall3 has no code. The regression test runs the
tool over 60 payments against a node that refuses past 20 calls a second, and
fails on the old code. Against the live agent that failed 8 of 8 times, the
check now confirms all 70 covered payments in about a second.

Bumps the SDK to 1.0.0-alpha.18, not yet published.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@RunTimeAdmin
RunTimeAdmin merged commit 3908d04 into main Nov 8, 2026
7 checks passed
@RunTimeAdmin
RunTimeAdmin deleted the fix/sdk-evidence-tool-reads branch November 8, 2026 00:22
RunTimeAdmin added a commit that referenced this pull request Nov 8, 2026
Published 8 November 2026 with the one-read evidence check (#238).

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant