1.1.1 was published after explicit approval: npm verify402-mcp@1.1.1 (gitHead 7c43cbf5406d73d9ae396ed1363170414a2cc74e), the MCP Registry active/latest, GitHub release mcp-v1.1.1, and Render serving 1.1.1. The bridge has no runtime change; 1.1.1 exists as a marker so the version you install is the version the service reports. All the change is in the service — the six things that were broken between “we say settlement recovery exists” and the code actually doing it.
The one that hurt: durable recovery was unreachable for every real caller
The service read the legacy x-payment header for the payment's identity — but the x402 v2 SDK sends PAYMENT-SIGNATURE, and our own bridge sends that too. Every genuine v2 caller therefore staged a recovery record with no payment reference and could never reconcile a lost response. The durable-recovery feature worked only for clients that don't exist. The bridge is untouched here; the fix is in the service, and a v2 settlement reference test now guards it.
The Solana rail had the same problem in the other direction — there is no EIP-3009 nonce to match — so a Solana-exact settlement could never be reconciled at all. It is now identified by its signed transaction.
The second: the evidence hash a receiver cannot reproduce
Every paid response advertises a responseHash. Before 1.1.1, the service hashed the body before the honesty arrays (checked / notChecked) were merged in — so the advertised hash disagreed with the documented recompute recipe for every handler that did not already return them. A verification artifact you cannot verify is worse than none. The fix is one projection (projectionCore) feeding both sides: the evidence bundle a receiver recomputes, and the record used for replay promotion. A new test asserts the advertised recipe names every key the projection excludes, so the prose cannot drift from the implementation again.
The third: spec 0.9 existed on paper, not in the wire
The extension said 409 operation_in_flight while a settlement can still resolve and 503 payment_unresolved once it is stuck. The service answered 503 recovery_unavailable instead. 1.1.1 implements it:
- Write-ahead
settlingmarker — written before the facilitator call (onBeforeSettle), matched by header bytes, envelope, nonce, or Solana transaction. 409 operation_in_flightwhile the outcome is still resolvable;503 payment_unresolvedonce it isstuck— with noRetry-Afterand a named reason.- Neither state can be aged out on a timer, in any mode. A reconciliation failure escalates after a persistent ladder rather than a single transient error.
- Observable: both states are counted at
/info.recovery.observability, and the firststuckraises a server-side ALARM line.
The smaller ones
/findingsgets the same per-IP budget as the rest of the free surface; an unmetered enumeration endpoint was an abuse surface.- A malformed JSON body or a body over the 64KB cap now answers a typed JSON error (
invalid_json/payload_too_large) with arequestId— not an Express HTML stack trace carrying/app/node_modules/...paths. /usagerestoresbyInspectHostacross restarts and actually emits thebody_no_challengecounter it promised.- Release identity drift closed: root
package.jsonhad sat at 1.0.9 while everything else said 1.1.0; the release-consistency test now covers it.
Independent verification, and what it found
The spec review thread on x402-foundation/x402#3848 is where the substance of this work lives. Two independent implementers came at it from code rather than prose: Hitek-cel (Detent) ran crash-fault injection across its seven journal writes and published a 29-vector control-12 set derived from the suite; cv-scvd re-verified a Durable-Object claim/begin window in their own purchase flow, found the non-covered gap, and published the corrected trace publicly. The extension is not done when it is written down; it is done when two codebases survive the same crash.
The service ships with those implementation gaps closed. The vectors have not yet been run against this build; a conformance harness is queued, and results will be reported in the thread, not asserted here.
What 1.1.1 does not change
- No new tools. The eight stay the eight.
- No price changes. The five checks remain testnet USDC; the free gate and
/inspectstay free. - The deployed recovery store is still memory-only. The durable cutover (SQLite WAL,
RECOVERY_DB_PATH) is built and tested but not yet the deployment;/healthsays so out loud. - Zero external paid calls. There is no version of this release that manufactures adoption, and none is implied. The ledger is at the traction ledger, which keeps the zero months visible.
- The
notCheckedlists. A longer list is not the point; an honest one is.
Upgrade
npx -y verify402-mcp@1.1.1. If you pinned @1.1.0, nothing about your tools changes; the service behind them is what got fixed.
The ask
Bring the crash, not the complaint.
If you build x402 flows, the thing we need is not opinions about the extension — it is a failing reproduction. The control-12 vector format and the crash harness the adjacent implementations have both landed in the public thread; the shortest path to making this spec real is running both against your own rail and posting what breaks.
Previous: What verify402-mcp 1.1.0 ships. · All notes