Repository navigation
fix(spec): add eth_sendRawTransaction to HYPERLIQUID as a write (MAG-4426) - #195
nimrod-teich wants to merge 1 commit into
Conversation
…4426) HyperEVM serves eth_sendRawTransaction, but HYPERLIQUID did not list it. The smart-router relays an unlisted method under a default entry with stateful 0, so every HyperEVM transaction went to one node, was retried on failure, and was never broadcast. The new entry matches the catalogue's other definitions: 10 compute units, deterministic false, stateful 1, hanging_api, DEFAULT latest and timeout_ms 10000. HYPERLIQUIDT imports HYPERLIQUID, so the testnet gets it too.
|
🤖 Spec pipeline started on PR open — running Phase 8 → 11. Results post here as comments. View run |
|
Spec pipeline failed (started at Phase 8). See the run log. Retry by commenting: To also re-run every later phase on top of the fix, use |
Manual smart-router run: eth_sendRawTransaction on HYPERLIQUID and HYPERLIQUIDT (2026-10-19)Result: PASS. On the branch spec, The router ran the static spec catalog ( The binary is smart-router PR #543 at Upstreams usedTwo upstreams per net, so the stateful fan-out is visible.
Second upstreams were vetted with Boot verificationsBranch spec, both nets (all four chain-id checks pass, both trackers Main spec, mainnet only: the same two HYPERLIQUID lines ( Results
eth_sendRawTransaction before and afterPayload:
Samples: main's spec was booted twice (once with text logs, once with JSON logs) and gave the same result both times. Branch mainnet gave the same shape in 2 of 2 boots. Branch testnet gave the row above in 2 of 3 probes. The first testnet probe hit a transient failure, described in the run notes. The router metrics agree. On the main-spec boot the series carries Read controls (branch spec, boot A)
relay_end evidence (trimmed)Fields kept: GUID, method, stateful, chain, selection, outcome, stop_reason, served_by, status, error_code, latency_ms, and per attempt the upstream, outcome, error_code, batch and returned_to_client. The message is kept only on transport errors. Run notes
|
|
🤖 Spec pipeline started on PR open — running Phase 8 → 11. Results post here as comments. View run |
|
Phase 8 — smart-router boot + method probe
PR change under test — Method Probe Report — hyperliquidGenerated: 2026-10-20T15:07:27Z Only one upstream URL per network was provided, so one Transport availabilityThe spec declares zero Per this run's transport-probing requirement, both node URLs were additionally probed over ws:
Since the spec has no subscription methods, both refusals are expected and NOT a defect or a boot blocker. Parse-directive & verification runtime check (Step 3.5) — mainnet (HYPERLIQUID)Note: the
Operational note (not a spec defect): on a separate boot made solely to probe the Addon & extension coverage (Step 1c + probes)The spec declares ONE extension, A known structural gap carried over from the orchestrator's pre-resolved facts: there is no
Evidence detail: a first archive probe attempt (before adding
Method probe table (27 stateless methods, through router at localhost:3360)All 27 non-stateful methods were probed unpaced; zero HTTP 429s were encountered against
All parameterized methods used live values resolved from the upstream immediately before probing (latest block number/hash, a tx hash + sender/receiver address drawn from that block, tx index 0) — no hardcoded placeholders were used, so the re-probe-once gate was never triggered (no method needed it). Log-scan findings (probe window, Step 4.5)
No other non-benign warn/error/fatal/panic lines appeared during either the mainnet or testnet probe windows. No PR change under test — eth_sendRawTransaction binding (PR #195)Per Step 4, Mainnet (HYPERLIQUID, localhost:3360): client response Testnet (HYPERLIQUIDT, localhost:3460, during Step 7): same payload, same result on the first attempt (no re-probe needed — no transient 500/no_answer was observed this run): Both results exactly match the expected outcome from the prior manual run on this branch. The Verdict: PASS-existence / PASS-binding on both mainnet and testnet — Testnet verification pass (Step 7)TESTNET_VERIFY: OK
PARSE_BLOCKNUM: OK ( Empirical block time (Step 8)
Both measurements are well within the 20% threshold; no TeardownBoth |
|
Phase 9 — parallel reviewers (3 fresh contexts) Three distinct reports (distinct md5s), every cited evidence line within Verdict on the PR's changed surface —
|
| # | Severity | Finding | On main? |
Agreement |
|---|---|---|---|---|
| 1 | MEDIUM | archive extension on the mainnet collection has no pruning verification, so GET_EARLIEST_BLOCK is never executed (Phase 8 confirmed it NOT_EXERCISED). Reviewers 2 and 3 add that the stock expected_value convention would not hold here anyway — mainnet's earliest call returns -32603 invalid block height: 0. |
pre-existing | 3/3 |
| 2 | MEDIUM / MINOR | allowed_block_lag_for_qos_sync: 5 is tighter than the formula value and than peer-chain convention for a 1000 ms block time. |
pre-existing | 3/3 (R1 MEDIUM, R2+R3 MINOR) |
| 3 | MEDIUM | eth_getSystemTxsByBlockHash / eth_getSystemTxsByBlockNumber use block_parsing EMPTY despite taking a block hash/number — inconsistent with this same spec's convention for the other hash/number-keyed queries. |
pre-existing | 2/3 (R1, R3) |
| 4 | MEDIUM | Standard EVM filter-polling methods (eth_newFilter family) are absent from the spec although the chain exposes them on the same /evm path. |
pre-existing | 1/3 (R2) |
| 5 | MINOR | eth_gasPrice carries an unusually tight timeout_ms: 1000. |
pre-existing | 1/3 (R2) |
| 6 | MINOR | No API-docs path was supplied to the reviewers, so a full method-coverage diff was not possible — a review-coverage limitation, not a spec defect. | n/a | 2/3 (R1, R3) |
Every finding is pre-existing on main — none was introduced by this PR. Rows 1–5 go to the Phase 10 consolidation; row 6 is dropped as a coverage note.
|
Phase 10 — synthesize gaps + fix pass Consolidation: 0 critical, 4 medium; 0 disable-suggestions stripped; 0 in-scope, 4 pre-existing. MINOR gaps dropped per the phase rules. Full list on the runner at Auto-decision at the fix gate: no spec edit applied (conservative default, recorded)This is an unattended run, so the fix gate was auto-decided with the conservative default. The consolidation verified mechanically — Applying them anyway would push four unrelated edits to a production mainnet+testnet spec under a PR whose stated scope is one write method — the exact silent-drift failure the skill guards against — and three of the four cannot be applied safely on the evidence this run has:
Rows 1–4 are carried to the watch-list below rather than silently left: they are real, they are on Post-fix static gates (Phase 10a) — all PASS
Removed-field governance guard: the spec carries none of the 15 fields smart-router dropped in #218 — no spec-level governance fields, no Watch-list (needs a follow-up PR, not this one)
Working tree: |
|
Phase 10b — smoke regression re-probe
Sampled reads = the first five non-stateful APIs alphabetically.
|
|
Phase 11 — final reviewer verdict
|
| # | Severity | Finding | Fix risk |
|---|---|---|---|
| 1 | MEDIUM | archive extension has no pruning verification; GET_EARLIEST_BLOCK is dead code. Mainnet's earliest call errors -32603 invalid block height: 0, so the stock expected_value: "0x0" template would exclude every provider at boot if copied blind. |
HIGH |
| 2 | MEDIUM | eth_getSystemTxsByBlockHash / …ByBlockNumber use block_parsing: EMPTY despite block-keyed names. PLAUSIBLE, not CONFIRMED — no API docs available to settle the parameter list. |
MEDIUM |
| 3 | MEDIUM | EVM filter-polling family absent from the spec with no rationale. | LOW (pure addition) |
| 4 | MEDIUM | allowed_block_lag_for_qos_sync: 5 on both entries vs the formula's 10 (10000 / 1000 ms); peers at the same block time (ink, katana, megaeth) use 10. |
LOW |
| 5 | MINOR | No real API-docs path was available, so a doc-driven completeness diff could not be run — findings 2 and 3 are inference-based. Review-coverage limitation, not a spec defect. | n/a |
Findings explicitly checked and dismissed this pass
block_distance_for_finalized_data: 0— standard for BFT/instant-finality chains (~40 specs in the catalog); consistent with HyperBFT. Not a defect.blocks_in_finalization_proof: 1— correct per the finality-typed rule (fast/instant finality →1).eth_gasPricetimeout_ms: 1000— raised as MINOR in Phase 9, dropped here: ETH1's owneth_gasPricecarries the identical value, so it matches upstream convention.- Disabled APIs/collections — zero
enabled: falseapis or collections anywhere in the file, matching the empty Disabled-API Justifications ledger exactly. No CRITICAL finding on the disable gate. - Removed-field guard —
RESULT: PASS (no removed fields).
Governance-field compliance (checked explicitly)
The spec carries none of the 15 fields smart-router removed in #218: no spec-level governance fields (min_stake_provider, providers_types, contributor, contributor_percentage, shares, identity, block_last_updated, reliability_threshold, data_reliability_enabled), no proposal.title / proposal.description, no top-level deposit, and no api-level extra_compute_units / category.local / category.subscription. Envelope is the canonical { "proposal": { "specs": [ … ] } }. compute_units is the only compute-unit input, and the spec declares zero FUNCTION_TAG_SUBSCRIBE directives, so no method is a subscription.
Recommended disposition
Merge PR #195 on its own merits, and open separate follow-ups for gaps 1–4. Gap 1 in particular must not be fixed by copying the standard pruning template — it needs this chain's real earliest-serviceable block determined first.
|
Phase 12 — closing summary checklist (tail of Phase 11; this pipeline run covered Phases 8–11) Run outcome
Auto-decisions taken at STOP-and-ask gates (unattended run)
Run stats — this pipeline run only (Phases 8–11)These figures cover Phases 8–11 only. This PR's spec edit was authored outside the create-spec pipeline (a focused MAG-4426 fix commit), so there is no Phases 1–7 create_spec run to add to them — unlike a fresh-chain PR, these numbers are the whole pipeline cost for this PR, not one half of it. To re-run any part: |
99b0cfe to
8c57f75
Compare
#Closes MAG-4426
Update spec: Hyperliquid (HYPERLIQUID)
Why
eth_sendRawTransaction, but HYPERLIQUID did not list it.Default-eth_sendRawTransaction. That entry has 20 compute units, stateful 0 and no timeout (defaultApiContainerin smart-routerprotocol/chainlib/base_chain_parser.go).Added
HYPERLIQUIDjsonrpc POSTeth_sendRawTransactionHYPERLIQUIDT imports HYPERLIQUID, so the testnet inherits the method. The relay window becomes 12 s: timeout_ms 10000 plus twice the 1 s block time. As an unlisted method it was 2 s.
The definition matches the catalogue. All 11 own definitions are stateful 1 at 10 CU. Ten are hanging_api with DEFAULT latest. The most common explicit timeout_ms is 10000, used by CONFLUX, GNOSIS, LIT and QUAI.
Evidence. Both public RPCs serve the method. A malformed transaction gets -32602, and an unknown method gets -32601. Probed 2026-10-19:
https://rpc.hyperliquid-testnet.xyz/evmanswers the same. The HyperEVM page says wallets add the chain by RPC URL and chain ID, and that "Pre-EIP-155 transactions are accepted as with many EVM chains." The JSON-RPC page does not list the method.Corrected
None.
Reported, not changed
eth_usingBigBlocksis on the JSON-RPC page's custom-endpoint list but not in the spec. It is out of scope here.Guard output
Test plan
eth_sendRawTransactionthrough it. Both nets broadcast it as a write to both upstreams; main's spec relays it asDefault-eth_sendRawTransaction, a read, to one. Results.The auto-started Spec Pipeline run is cancelled, because this PR adds one method to a chain that is already onboarded. A
/rerun-<phase>comment restarts it.