Repository navigation
fix(testnet): get InstantSend locks from the DAPI stream instead of trpc.digitalcash.dev - #53
Conversation
trpc.digitalcash.dev is stuck on Core v23 and stops serving IS locks after the testnet V24 fork (block 1,578,000), so the bridge hung waiting for the lock. Testnet now uses the existing DAPI subscription path with an explicit snapshot of healthy evonodes, which skips dapi-client's seed/SML discovery (broken on V24 CbTx v4 coinbases). - TESTNET: drop rpcUrl, add 15 explicit https://<ip>:1443 dapiAddresses - dapi.ts: remove the implicit testnet JSON-RPC default so hasRpcUrl is false and no request goes to any digitalcash.dev host - tests: pin testnet to the subscription path; JSON-RPC deadline tests now use mainnet Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…t lock The bloom-filter IS lock stream used one dapi-client instance that picks a random node per call. subscribeToTransactionsWithProofs returns before any connection exists, so a dead node only surfaced ~12 s later as 'Response closed without headers', after the asset-lock tx was already broadcast, with no retry. A single live node can also silently drop a lock. With two or more configured addresses, open streams on up to 3 distinct nodes at once, each through its own one-address DAPIClient so nodes cannot repeat. The first matching lock wins and the other streams and clients are torn down; late locks and errors from losers are ignored. A stream that errors or ends is replaced from the untried shuffled list (9 attempts max), and when every node has failed the error names the nodes tried and the last cause. Timeout and abort messages are unchanged. The function returns after the first stream message or a 3 s readiness cap, so the broadcast is not held up by slow nodes. disconnect() now also stops in-flight subscriptions, since DAPIClient.disconnect() does not cancel streams. Fewer than two distinct addresses keep the single-client path. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
attempts already indexes the next untried node and slots.size is the number of live streams, so nextNode and running were duplicates. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Every testnet node omits the optional state_sync message from platform getStatus, and @dashevo/dapi-client 3.1.0-dev.7 dereferences it unconditionally, so the network-health indicator reported "degraded" and the chain-lock fallback never got a height. Testnet now reads chain-locked height, Platform block height and block time via sdk.system.status(); devnets and mainnet keep their existing paths. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Testnet certificates rotate in 2-6 days and nodes come and go, so the static dapiAddresses snapshot goes stale. At client init, testnet now fetches the quorum service /masternodes list (CORS-enabled for browsers; 5 s, one attempt), keeps ENABLED nodes with a successful version check and dials them on :1443. The static list is the fallback when the fetch fails and tops up the result when fewer than 3 nodes come back. The static list is refreshed to a current 20-node snapshot with regeneration notes. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Testnet no longer uses JSON-RPC, so fix the stale mainnet/testnet wording in islock.ts, main.ts and the network-status tests, document the per-network InstantSend source in CLAUDE.md, and add a mock-mode e2e test asserting that testnet discovers evonodes from the quorum service and sends no request to any digitalcash.dev host. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Review follow-ups: - initClients no longer installs a stale islock service when the user switched networks while the evonode list was loading. - evonode hosts that pass the character check but are not valid URLs (999.1.1.1) are dropped instead of making the DAPI client constructor throw for the whole list. - Platform status routes on the network config type rather than its name, so an unknown name that falls back to the testnet config still uses the SDK path. - connectPlatformSdk shares one in-flight connect between concurrent callers (status polls and write paths) instead of orphaning SDKs. - Trim duplicated comments and redundant tests. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ails The caller broadcasts the asset-lock tx as soon as subscribeForInstantSendLock returns. When every node attempt had already failed inside the readiness window or initial open loop, it used to return a handle whose wait() rejected later, so funds were committed with no listener. It now rejects itself with the same "N nodes tried, last error" message; main.ts's existing catch turns that into the retryable ISLOCK error with nothing broadcast. Failures after subscribe has returned still surface through wait(). Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… pick The one-off getBestBlockHeight before opening streams went through the shared random-pick client. When it picked an unreachable node the browser's TCP connect hung ~75 s before failover, with nothing broadcast meanwhile (measured: 75 s from subscribe to broadcast in the funded browser test). The multi-node path now asks the first three nodes of the shuffled list at once, each through the single-address client its stream will reuse, with an 8 s per-request deadline. The first valid height wins; it fails only when all of them fail or time out, with the usual "N nodes tried, last error" message and the probe clients disconnected, before anything is broadcast. Caller aborts also stop the wait. The single-node path is unchanged. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Since a057e84 the testnet health poll read Platform status through sdk.system.status(), which loads and connects the whole EvoSDK (~7 MB gz) on every testnet page load. Testnet is the default network, so that undid the load-time work from PR #28. What loads when now: - Page load (idle warmup): the JS modules and the IslockService, the evonode list fetch, and the Core height poll via Insight. The EvoSDK chunk and wasm are NOT requested. - Health polls (every 30 s) call IslockService.peekPlatformStatus(). On testnet that resolves undefined while no SDK is connected, without touching the loader. The verdict then rests on Core alone: healthy with only the Core height in the tooltip, no reasons, never "degraded"; if Insight also fails it is "unknown" ("Could not reach Insight"). - The SDK is loaded by the flows that need it (creating an identity, top-up, DPNS, ...). The first poll after connectPlatformSdk has cached a connected SDK reads Platform status through it and shows chain-lock lag and block age as before. A failure of that real read still degrades. - The chain-lock fallback keeps using the loading getPlatformStatus path, since the user is mid-flow and needs the height. - Devnets are unchanged: dapi-client needs no SDK. Mainnet is unchanged (JSON-RPC). The mock-mode e2e test now also asserts no evo-sdk request on page load; it fails if the poll uses the loading path. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Two bugs, both from the poller skipping any result whose verdict matched the previous one: - A flow reset (Try Again, Contract -> Back) builds a fresh state without networkStatus, and the poller then skipped the unchanged verdict, so the indicator stayed hidden until it changed (75+ s observed). - The tooltip numbers (heights, block age) never refreshed while the verdict was unchanged (stuck over 8 polls). Poll results are now always written to the state. A full re-render still happens only when the verdict or its reasons change, which is all the label and banner show; otherwise only the tooltip is patched in place, so focus isn't disturbed every 30 s. updateState re-attaches the last known status of the same network to any state that lacks one, and a poll that finds the state without a status renders it. Logic lives in ui/network-status-sync.ts so it can be tested. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Log "[islock-sub] IS lock for <txid> from <host>:<port> after <ms> ms since subscribe" on the winning stream (node address and elapsed time only), so a slow or failing lock can be traced to a node. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- The indicator says so when Platform was not read yet (tooltip line "Platform status: not read yet", NetworkStatus.platformChecked), so the Core-only "healthy" is not mistaken for a Platform check. - A poll result for a network the user already left is ignored, and switching networks clears the old network's status instead of showing it under the new badge. - Re-render decisions compare with the verdict actually on screen rather than with the previous poll. - The height probes use retries 0 and a gRPC deadline, so no hidden background retries outlive the 8 s deadline. - disconnect() (network switch) now also stops a pending height race, so no streams open on the network being left. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The health poll no longer downloads the SDK on page load (35eddc7), which left testnet showing a Core-only "healthy" until registration, i.e. after the asset lock was broadcast. Now, on testnet only, create / top-up / send-to-address start connecting the EvoSDK as soon as their clients are ready (before the deposit address is shown and before any funds can be committed) via warmPlatformSdkForFunding in the existing loaders, and poll the network status right after the connect settles. Registration needs the SDK anyway, so funding users download nothing extra; idle visitors still never do. Nothing blocks funding; mainnet and devnets are untouched. Indicator behaviour: - Before any funding flow, and while the first connect is in flight: Core only, tooltip "Platform status: not read yet", not degraded. - Once a connect has been attempted (connected, failed, or evicted after a connection error), the poll always reads Platform through the SDK, so a failing read is "degraded" and no longer reverts to the Core-only healthy when withConnectedPlatformSdk evicts the SDK. The next poll reconnects (module already loaded); success shows healthy again. This replaces hasConnectedPlatformSdk with isPlatformSdkSettled (attempted and not currently connecting). Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Review follow-ups to the funding-flow SDK warmup: - A reconnect in flight after an outage no longer reads as "not read yet" (which flipped the badge between degraded and healthy every poll): only the very first connect is shown that way, and only for a 60 s grace period, so a hung first connect cannot read as healthy forever. Later reads join the pending connect and are bounded by their own 20 s timeout. isPlatformSdkSettled becomes isPlatformSdkReadable. - The status poll no longer passes maxAttempts 1 to the shared connect, so a write path that joins a poll-started connect keeps its normal retries. - Overlapping polls (the extra one after the SDK connects) apply in order; a slower older poll cannot overwrite a newer result. - The funding warmup is skipped in e2e mock mode, and the extra poll only runs when a connect was actually attempted (testnet), not on mainnet or devnets. - Note: the warmup starts when a funding flow starts, which can be before the 7 MB download finishes; until it connects the indicator shows the Core-only "Platform status: not read yet". Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
|
Caution Review failedThe pull request is closed. 📝 Walkthrough
Merge Risk: 🔵 Low · up to
|
Why
The testnet V24 hard fork activated at block 1,578,000.
trpc.digitalcash.dev, which the bridge polls for InstantSend locks and ChainLocks on testnet, is still on Dash Core v23 and stopped following the chain, so every testnet funding hangs after the asset lock is broadcast.InstantSend locks do not need a central service. Any evonode delivers them over
subscribeToTransactionsWithProofs, and the bridge already had that path for devnets. This PR moves testnet to it and removes the testnet dependency ontrpc.digitalcash.dev. Mainnet is unchanged.What changed (testnet)
TESTNEThas explicitdapiAddressesand norpcUrl. Defaultdapi-clientdiscovery is broken on V24: the legacy@dashevo/dashcore-libcannot parse CbTx v4 coinbases while syncing the masternode list, so the client never connects. Explicit addresses skip discovery.subscribeForInstantSendLockthrows before the asset lock is broadcast.quorums.testnet.networks.dash.org/masternodes(ENABLED + version check, CORS-enabled), with a static 20-node fallback insrc/config.ts.platform.getStatus()through@dashevo/dapi-client3.1.0-dev.7 throws on every testnet node (state_syncis omitted andcreateFromProtoreads it unconditionally), which made the network-health indicator say "degraded" and left the ChainLock fallback without a height source. Testnet now readssdk.system.status()(src/platform/status.ts).digitalcash.devhost on testnet (asserted in a new mock-mode e2e test).Testing
tsc --noEmit,vitest run(720 tests),npm run build, and the mock-mode Playwright suite (53/53) pass.*.digitalcash.dev, healthy indicator with real Platform values at funding time.Known limitations
quorums.testnet.networks.dash.org. If it is down, registration/top-up/send fail after broadcast (recoverable through the ChainLock fallback, but the asset lock stays unspent until retried). The health indicator shows degraded in that case once the SDK has started connecting.dapiAddressesnow use the multi-node path (side effect of the subscription change).rpc.digitalcash.dev. It needs its own DAPI stream test and an evonode source before the mainnet V24 fork.sdk.system.status()read but waits for a ChainLock instead of an InstantSend lock).🤖 Generated with Claude Code
Summary by CodeRabbit