Subgraph mode: compatibility fixes, 30% → 83% of 860 real subgraphs (85% of 2026 repos) by moose-code · Pull Request #1759 · enviodev/hyperindex

moose-code · 2026-10-20T14:22:49Z

Every subgraph-mode compatibility fix in one PR against #1626's branch.

Round 1: the 13 earlier PRs from the hackathon, listed at the end, plus three branches that never had a PR.
Round 2 (20–21 Oct): measured on a corpus of 860 real subgraphs and checked against live graph-node deployments.

447 commits (393 non-merge). Each fix comes with a test that failed first, and cites graph-node or graph-ts behaviour where it matters.

Results: unmodified real subgraphs

Real subgraph repositories (subgraph.yaml + schema + AssemblyScript mappings), cloned at a pinned commit and run unmodified with envio start. ok means at least one batch was committed with no error for 60 s, or the run synced to the end of its block window.

set	entries	`envio@3.14.0-subgraph` (#1626 as published)	this PR when opened (`d2777248b`)	this PR now
recent: repos created or pushed since March 2026	632	233 (37%)	470 (74%)	538 (85%)
legacy: the October measurement's repos	128	17 (13%)	58 (45%)	99 (77%)
subgraphs deployed on hosted subgraph platforms	117	12 (10%)	66 (56%)	89 (76%)
all	860	257 (30%)	582 (68%)	713 (83%)

Size: 708 distinct repositories. The recent set has 537 repos, all pushed between 1 March and 20 Oct 2026; 389 of its 632 entries are repos created in 2026.
Deploy options: of the 713 ok, 601 run as-is. 112 need a deploy option a deploying user passes anyway, the same input graph build --network takes: ENVIO_SUBGRAPH_NETWORK for a network pick, or ENVIO_CONFIG for a manifest that isn't subgraph.yaml.
Zero-event syncs: 57 of the ok synced a window with 0 events.
- 40 of those were checked against HyperSync for logs their handlers match, and none had any. The other 17 are on chains HyperSync doesn't serve, so they're unverified.
- One sync that missed matching logs is counted as not ok.
Regressions: 0. Each build re-ran every entry not ok on the previous one, plus a 20% random sample of the ok ones. Any drop was re-run on the previous build with the same entry before being called a regression.

Data parity: 113 unmodified subgraphs were compared against a live graph-node deployment of the same subgraph at a pinned block, using envio subgraph diff.

93 are equal in every entity type. That's 1,194 of 1,280 types, with 41k entities compared field by field. They include:
- SSV network (53/53 types, 571k events) and BOLT (14/14, 38M events);
- GMX synthetics (38/38) and Balancer v3 vault and pools;
- harbor-app (29/29, 3.6M events) and aave-v4-omnigraph (25/25).
Every mismatch was triaged. Two were envio bugs and are fixed here: the event.logIndex byte width, and Bytes.empty(). The rest:
- The deployed build differs from the repo: deploy-time constants or older mappings. Patched copies confirm it.
- Two deployments are short of events: ours equals the on-chain log count.
- IPFS files no gateway served.

Remaining failures, by first blocker:

57 are repo setup: mostly ABIs the manifest names that are contract build output not committed to the repo.
13 manifests are written by package scripts envio deliberately won't run (they also deploy).
10 are mapping imports that don't resolve.
23 are external: chains whose only keyless RPC is pruned or too slow.
13 use network names no registry knows and have no RPC.
The rest are scattered: perf, one substreams subgraph, and so on.

What changed, by area

Mapping runtime (AssemblyScript semantics):
- === follows AssemblyScript identity semantics.
- Module-cycle initialisation.
- Top-level code re-runs per trigger, as graph-node instantiates per trigger.
- AssemblyScript Map/Set.
- Arithmetic:
  - u32/i64 bit operations and integer casts (as i64, <T>x);
  - float division and toString;
  - zero-initialised fields and locals.
- Pointer-level operations:
  - memory builtins and StaticArray;
  - String.UTF8/UTF16 and reinterpret;
  - changetype retagging onto the generated getters.
- The default sort() comparator.
- Minimal-byte BigInts from chain data, everywhere graph-node builds them.
graph-ts API:
- Bytes.empty();
- BigInt.toI32/toU32/toHexString widths;
- ethereum.encode per token, and decodeParams;
- JSON number text and key order;
- dataSource.address();
- YAML;
- reserved-word fields;
- null into a value-typed setter;
- the missing-non-null-field error on store.set, as graph-node raises it.
Generated code:
- The bundled codegen follows the project's graph-cli version.
- When a mapping imports a generated file that only existed in the author's tree, it's rebuilt from an unambiguous source (a sibling manifest, an ABI, a fragment), with a warning each time.
- Workspace packages resolve without an install, and missing libraries are fetched when the repo's install fails.
- Pre-0.18 graph-ts names are supported.
Repo shapes:
- Manifests rendered with standard tools (mustache, cp, cat, graph).
- ENVIO_SUBGRAPH_NETWORK picks the network, and --config picks a manifest.
- Case-only path mismatches.
- Clear, named errors for everything envio won't do.
Newer graph-node features:
- specVersion 1.4/1.5 and apiVersion 0.0.10.
- Aggregations and timeseries: graph-node's own aggregation store test is ported here and passes.
- once/polling handlers on templates.
- skipDuplicates.
- IPFS ContentPath, timeouts and hedged gateways.
- A clear "can't index" refusal, with the reason, for substreams, composition, Amp and non-EVM data sources.
Networks:
- A network name envio doesn't know resolves through the RPC's eth_chainId.
- New network slugs.
- Hardened RPC-only sync for chains HyperSync doesn't serve.
Contract calls and throughput:
- JSON-RPC batches;
- preload lookahead and a per-batch overlay;
- adaptive concurrency on rate limits;
- call caches that survive envio start -r;
- effect cache keys hashed above 2 kB;
- BigDecimal and hex fast paths, bit-identical against the crate's vectors.
- Uniswap-v3 forks run about 35% faster, and request counts fall by 80–90% on call-heavy subgraphs.
Templates created from block handlers: they now register with graph-node's timing. Before, they were silently missed (yieldyak strategies: 1 → 7,002 events).
envio subgraph diff: --wait, --config, aggregation intervals, and pacing for rate-limited references.
Tests and CI:
- Test servers take OS-assigned ports.
- Cherry-picked from main, with -x: Speed up envio-tests CI: skip unused dev addon build, shorten MultichainStuckAtHead #1690 (skip the unused dev build in envio-tests' global setup; it ate 3–5 min of CI), Fix "Worker exited unexpectedly" in envio-tests #1674 ("Worker exited unexpectedly") and Skip a log line still being written in the tests' log reader #1694 (log reader).

Please review closely

envio now runs a repo's own scripts, narrowly (docs §6a, §6d). It already ran the steps of the codegen script.
- Codegen steps: the codegen script's steps before graph codegen now also run when the manifest or its ABIs are missing.
- Manifest writer: when there is no manifest, envio runs the one package.json script that writes it, picked by network, and only up to its first graph-cli step.
- Guards: each command is logged, ENVIO_* variables are stripped, and stdin is closed. Deploy, publish and docker scripts, and anything running git, curl or ssh, are never run. Tests prove a deploy script and another network's script never run.
Core engine changes, gated to subgraph mode unless noted:
- Block-handler register pass: FetchState, ChainState and ChainFetching (blockRegisterThrough is None unless an onBlock registration carries a contractRegister).
- Other gated changes:
  - a once-first block comparator;
  - preload lookahead and overlay in EventProcessing, UserContext and SyncWrites;
  - the register-store frontier.
- Applies to all effects: the effect cache key becomes sha256: above 2 kB.
- Coverage: the full envio-tests suite, native included, is green on Postgres and ClickHouse.
Network resolution:
- Disagreements: when ENVIO_SUBGRAPH_RPC serves a different chain than the registry gives a manifest's only network name, the RPC's chain wins, with a warning, as on graph-node.
- Translation probe: translation asks each RPC endpoint for eth_chainId once.
Behaviour now matching graph-node. It changes outputs that used to be accepted:
- a missing non-null field fails at store.set;
- ipfs.cat returns null after 60 s;
- BigInt.toI64 overflow aborts;
- top-level code re-runs per trigger;
- a non-anonymous topic0 override drops its handler with a warning.
Heuristics, each warning when used: the generated-file fill and the re-export fill for stale generated/ trees.
Test indexer: it reads the undocumented process._getActiveHandles() to avoid probing its own RPC stubs. Without it, it falls back to waiting as before.

Known gaps

Refused (§7): a create from a template's unfiltered or polling block handler, and a data-source block handler running over more than 1,000 blocks whose create isn't certain.
Accepted but not honoured: nonFatalErrors; Timestamp's sub-millisecond digits.
Not done:
- ABIs that only exist after compiling the repo's contracts are not built. envio names the missing file and the step that writes it.
- Not emulated: AssemblyScript integer literals above 2^53, i32 arithmetic overflow, f32's shortest toString.
For envio-serve: a null list of entities should be served as [], as graph-node does.

Tests and CI

CI at c2ea844bb: green.
- envio-tests: Postgres 3.7 min, ClickHouse 5.4 min.
- cargo-test, e2e, scenarios and templates all pass.
- hypersync-health-check fails only because 51.hypersync.xyz reports unhealthy, the same as on main. It's continue-on-error.
Locally:
- cargo test --lib: 1,176 passed.
- Full envio-tests on Postgres: 430 files passed.
- test/Subgraph: 227 files.
- Replay stress: green.

What's in it

Round 1 (merged as one branch, d2777248b):

Subgraph mode: compatibility fixes from real subgraphs (wildcard data sources, register-pass scope, struct params, encode, …) #1732: compat fixes from real subgraphs.
Subgraph mode: translator and storage compatibility fixes #1735: translator and storage.
Subgraph mode: throughput #1736: throughput.
Subgraph mode: graft (ignored with warnings) and timeseries/aggregations #1740: graft and timeseries.
Subgraph mode: long-tail load failures (BigInt/Address/matchstick shims, AS compound assignment, entries field) #1741: long-tail load failures.
Subgraph mode: file data sources (file/ipfs, file/arweave) #1742: file data sources.
Subgraph mode: receipt.logs, block.hash in block handlers, call-heavy polling batches #1743: receipt.logs and block headers.
Subgraph mode: fetch mapping libraries and graph-cli when the repo's install fails #1744: fetch toolchain.
Subgraph mode: runtime blockers (null BigInt, integer casts, createWithContext, typed-array wrap) #1745: runtime blockers.
Subgraph mode: call handlers #1746: call handlers.
Subgraph mode: bundled codegen (generated/ without graph-cli) #1747: bundled codegen.
Subgraph mode: network names and small compat classes #1748: network names.
envio subgraph diff: parity report against a graph-node endpoint #1731: envio subgraph diff.
Plus sg-callbound, sg-fixer5 and sg-fixer6.

Round 2, one branch per area, merged in order:

runtime, codegen, shapes, hosted, modern;
mapping (runtime round 2), recent, calls;
hygiene, sweep, shapes2, parity, polish, ci, blockcreate.

🤖 Generated with Claude Code

coderabbitai · 2026-10-20T14:23:00Z

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI
Review profile: CHILL
Plan: Essentials
Run ID: d555d8aa-8b6b-4cd8-9cd3-37f3c6c46dd1

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

graph-cli before 0.18 (mapping apiVersion 0.0.3 and older) generates `super.call("ownerOf", [params])` and `super.tryCall(...)` with no function signature; graph-node then takes the first function of that name in the data source's ABI (chain/ethereum/src/runtime/ runtime_adapter.rs: `function_signature: None` → `contract.function(name)`). The shim only took the three-argument form of later graph-ts and failed with 'Unreadable contract call signature ""'. SmartContract.call/tryCall now accept the two-argument form; the runtime installs a lookup that reads the manifest's ABI of that name and builds graph codegen's `name(inputs):(outputs)` signature for the first function so named. A name the ABI lacks is an error, as on graph-node. Test: SubgraphLegacyContractCall_test.res Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

graph-node's handlers_for_log returns every handler of a data source whose event matches a log and runs them in manifest order. The translation gave each extra handler of an event its own envio event (`Ping_1`), which envio refuses as two events "the indexer can't tell apart", so the subgraph didn't start. envio runs every registration on an event since #1469, so a repeated event now registers its handler on the first one's event; the event selects receipt fields when any of its handlers asks for them. Groundwork for swaap-labs/swaap-subgraph-v1 (a subgraph from a hosted subgraph platform), which hangs a dozen handlers on LOG_CALL. Test: SubgraphSameEventTwice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…sked of RPC concurrently On a chain HyperSync doesn't serve, a block handler's block.timestamp (and header) comes from ENVIO_SUBGRAPH_RPC, one eth_getBlockByNumber per block, and the fallback awaited them one after another: on BlockDAG (evm-1404) a handler on every block took 277 s to commit its first 500 blocks. graph-node fetches a range's blocks concurrently; the fallback now issues them at once, bounded by the RPC client's requests in flight (ENVIO_SUBGRAPH_RPC_CONCURRENCY). Test: SubgraphRpcBlockTimestamps_test.res (mock RPC records the most eth_getBlockByNumber requests in flight). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

graph-ts before 0.22 named the transaction's sixth field gasUsed; graph-node fills that slot with the transaction's gas limit (gas_limit in every AscEthereumTransaction_0_0_*, chain/ethereum/src/runtime/abi.rs), which later graph-ts calls gasLimit. Uniswap v3's 2022 mappings (arbitrum-minimal) store event.transaction.gasUsed on every Transaction, and it read null, so the batch write failed on the non-null column. The shim's transaction now carries gasUsed from envio's gas field, and the field scan selects gas for it. Test: SubgraphLegacyTransactionGasUsed_test.res, cargo test selects_the_gas_limit_for_a_pre_0_22_gas_used. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Omni-DeFi's mappings import `../generated/staking/staking`, and graph codegen writes `generated/Staking/Staking.ts` for the data source `Staking`. asc resolves that on the macOS or Windows filesystem the subgraph builds and deploys from; Node on Linux didn't find the module, and the mapping failed to load. A relative import from a project file that Node can't find is now matched against the filesystem one path component at a time: when each missing component has exactly one entry differing from it only in case, that file is loaded, with one warning per import. Imports into node_modules and files outside the project are left alone. Test: SubgraphImportCase_test.res (failed to load the mapping before). Omni-DeFi/subgraphs (mode-testnet, hosted): failed to load src/staking.ts; now indexes (44 events to block 20,032,984 in 150 s). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… the network's graph-node checks each provider against the chain its network was configured for (genesis hash and net version) and refuses one that serves another. Here the endpoint was taken on trust: one of the wrong chain answered a mapping's contract calls with another chain's state, or on a chain HyperSync doesn't serve, indexed another chain outright. At startup the endpoint is now asked its eth_chainId and refused on a mismatch, naming both chains; an endpoint that doesn't answer is left to the runtime's retries. Test: subgraph::project_tests::refuses_an_rpc_endpoint_of_another_chain. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

graph codegen writes schema.ts, templates.ts, <data source>/<ABI>.ts and templates/<template>/<ABI>.ts for the manifest it runs on, and no index.ts. A mapping importing anything else fails `graph build` too (asc TS6054), and builds only on a machine whose tree is left from an earlier manifest: revert's uni-v3-staker imports generated/UniV3Staker/ after renaming its data sources UniV3Staker1.0.0/1.0.2; WOOFi's per-chain manifests import data sources only another chain's manifest declares; tornado imports a hand-written generated/index.ts. envio reported only Node's "Cannot find module"/"Directory import"; it now adds which data source, template or ABI the manifest lacks and what it has. Test: SubgraphGeneratedNotWritten_test.res Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

A cumulative aggregate carries each group's previous bucket forward (graph-node rollup.rs, `cumulative: true`), so a rollup needs every group's carried values. They were read one `getSync` at a time, and each miss suspends the sync bridge: the handler that crossed a bucket's end ran once per group, every replay redoing the rollup. graph-node's aggregation example (graphprotocol/graph-tooling examples/aggregations, a group per block number modulo 1000) replayed block 930's handler 934 times and committed nothing in three minutes. The rollup state rows now carry `aggregation` = `<Aggregation>|<interval>` (indexed), and a rollup reads its groups' rows with one getWhereSync: at most one round for all of them. Test: SubgraphAggregationGroups_test.res, the same 2,000 rows rolled up as one group and as a group each (4.8 s vs 0.26 s before; now comparable). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…de does graph-node validates a saved entity, merged onto the stored one with explicit nulls removed, against the schema (EntityCache::set -> Entity::validate) and fails the handler with "Entity T[id]: missing value for non-nullable field `f`". envio accepted the save and the batch write failed later, on Postgres' NOT NULL or on a null it couldn't serialize, with nothing pointing at the mapping: Opolis' ensureOpolisPayToken saves an OpolisPayToken without createdAt: BigInt!. The translated schema now records which fields are declared non-null, and store.set checks the merged row before writing it (not in the register pass). Two tests saved entities that graph-node would refuse: the receipt-logs mapping now stores a fallback name for an unparsed log, and the changetype test seeds an entity type without required fields. Test: SubgraphMissingRequiredField_test.res. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The registry added aliases since the snapshot was taken (arctest for arc-testnet, basesepolia, unisepolia, zksyncera), and the CI drift check (sync-graph-networks.mjs --check) fails on it. Refreshed with the script. Test: subgraph::manifest::tests::maps_graph_network_names_to_chain_ids (arctest, basesepolia). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Preload runs a batch's handlers with their writes dropped, so each one read the rollup clock as the previous batch left it, found its block past the end of that clock's bucket, and rolled the bucket up again: a batch of 5,000 block handlers redid one rollup 5,000 times, ten aggregations each in graph-node's aggregation example, which then committed nothing from its second batch on. The execute pass, where the clock advances handler by handler, rolls up as graph-node's Layout::rollup does; preload still reads the clock, so the execute pass finds it in memory. graphprotocol/graph-tooling examples/aggregations (mainnet, block handler on every block, five aggregations, cumulative groups): 0 blocks committed in 180 s before; 345,000 blocks in 90 s after, with this and the previous commit. Test: SubgraphAggregationPreload_test.res, a batch after a split run vs the same rows in one run (2.5 s vs 0.35 s before; now comparable). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ore translating cope-market's erc4626-vault commits subgraph.template.yaml and writes subgraph.yaml in the first step of its codegen script: `"codegen": "node scripts/build-manifest.mjs ${NETWORK:-arc-testnet} && graph codegen"`. envio runs a codegen script's other steps anyway (§6a), but the missing manifest stopped it at translation, before it got there. When the manifest is missing and no mustache script renders it, the codegen script's steps before `graph codegen` now run first, in the package directory with its bins on PATH. A variable a step reads is set to ENVIO_SUBGRAPH_NETWORK when that is set; unset, the script's default applies. The manifest is recorded as envio's, so a later start writes it again, and the runtime's codegen skips the steps already run, which would otherwise write the script's default network over the one picked. A failing step stops envio with the script's own error. The "config.yaml not found" hint no longer offers a `*.template.yaml` as a manifest to run. Tests: project_tests::runs_the_codegen_steps_that_write_the_manifest, SubgraphCodegenPrelude_test.res (graph codegen got the default network's manifest without the skip). cope-market/subgraphs erc4626-vault (arc-testnet), unmodified: "Found subgraph.template.yaml but no subgraph.yaml" before; now starts and indexes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

balancer/gauges-subgraph builds one tree from two manifests: `graph codegen subgraph.yaml --output-dir src/types/ && graph codegen subgraph.arbitrum.yaml --output-dir src/types/`, and its mappings import data sources only the arbitrum manifest declares. envio ran every `graph codegen` step of the codegen script on its own manifest, dropping the one the step names, so that half was never written; the bundled codegen wrote only envio's manifest. A step naming another manifest of the project now generates that one, into the directory the step names; envio's own manifest is generated where a step names it or no manifest at all, and last otherwise, so its files win. The bundled codegen follows the same plan. Test: SubgraphCodegenScriptManifests_test.res (graph-cli), SubgraphBundledCodegenManifests_test.res (bundled) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…s topics graph-node keys an event handler on `topic0` when the manifest gives one (MappingEventHandler::topic0) and parses an anonymous event's log with every topic an indexed param. Balancer v1's pools emit an anonymous `LOG_CALL(bytes4 indexed sig, address indexed caller, bytes data)` and their subgraphs hang one handler per called selector on it; translation refused those as "topic0 is not the event signature's hash". - manifest: a topic0 other than the signature's hash is kept on the handler (a value that isn't a 32-byte word is an unknown-error). - translation: on an event the ABI declares anonymous, each (event, topic0) is its own envio event carrying that topic0; elsewhere, or with topic1-3 filters, it is still refused. - config: an EVM event config may carry a `topic0` (not in config.yaml's schema): on an anonymous ABI event it becomes the sighash and the event is marked anonymous in the internal config. - runtime: an anonymous event's topic count is its indexed params alone, and the decoder reads a registration whose topic count equals its indexed params as anonymous (no signature topic). swaap-labs/swaap-subgraph-v1, a subgraph from a hosted subgraph platform: indexes on polygon (setSwapFee/finalize/bind decode to the values on chain) until its next blocker, a Transaction saved without its non-null gasUsed. Tests: SubgraphAnonymousTopic0, evm_hypersync_source::decode::tests::decodes_an_anonymous_event_from_all_of_its_topics, subgraph::manifest keeps_a_topic0_other_than_the_signature_hash and refuses_a_topic0_that_is_not_a_word, and the translation's topic0 test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…s created Preload runs a batch's handlers against the store as it was before the batch, so a handler reading an entity an earlier event of the same batch creates (a v3 pool's Mint after its PoolCreated) finds nothing, throws or returns early, and makes none of its contract calls. The execute pass then asks them one event at a time: a round trip per call, in series. Preload reads that find an entity absent are now remembered per item. When execute creates that entity, those items are preloaded again against the current store, as long as they are within ENVIO_SUBGRAPH_PRELOAD_LOOKAHEAD (default 256, 0 = off) items of the one executing; an item whose preloads already ran for 20 ms of CPU is not preloaded again. Subgraph mode only. Preload never writes, and every value it fetches is keyed by its full input, so what execute computes is unchanged. Test: SubgraphPreloadLookahead_test (16 handlers reading a pool created earlier in the batch: calls in flight at once 1 -> 16, results equal). sushiswap v3 (mainnet, from its start block), 180 s cap, 200 ms per RPC request: 44 -> 1,044 events committed; execute rounds waiting on eth_call 489 (100 s) -> 91 (33 s). Row-for-row equal to a lookahead-off run over the same 544 events (23 tables, 0 differing rows). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…othing else runs The codegen script's steps that run before translation (ca39bb2) now print a line naming each step, and the variables set from ENVIO_SUBGRAPH_NETWORK, before running it. A test pins the scope: with preinstall, install, postinstall, prepare, build and manifest scripts that would write the manifest, and a codegen script that is only `graph codegen`, none runs, and the error names the one that looks like a writer. §6d says so. Test: project_tests::runs_no_script_but_the_codegen_scripts_steps. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

graph-ts' ByteArray.toI32()/toI64(), which BigInt's defer to (common/collections.ts), assert only that the bytes past the width equal the sign padding of the last byte, then read the low bytes as two's complement. A uint32 of 2^32 - 1 is `ff ff ff ff 00` and reads as -1, and any value from -2^32 to 2^32 - 1 (2^64 for i64) wraps into the signed range. The shim refused everything outside i32 (and BigInt.toI64 returned values past i64 unchecked). poa-box/subgraph-pop calls `event.params.maxMembers.toI32()` on `type(uint32).max` in handleSubjectCreated: the run stopped there; now synced (1,307 events in the window). Test: SubgraphBigIntSignedWrap_test.res. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…repo's build passes iExec's PoCo subgraph names no network in subgraph.yaml and keeps addresses in networks.json; its build runs `graph build --network ${NETWORK_NAME:-arbitrum-sepolia}`. A bare `envio start` stopped at "declares no indexable data source", asking for ENVIO_SUBGRAPH_NETWORK. When ENVIO_SUBGRAPH_NETWORK is unset, a data source names no network and networks.json exists, the network is now the one the package.json scripts pass `--network` — a `${VAR:-default}` default or a literal — when they all agree on it and networks.json has it. Startup logs the network built. Scripts that pass several still leave the choice to ENVIO_SUBGRAPH_NETWORK. Test: project_tests::builds_the_manifest_for_the_network_the_repo_build_passes (also: two literal networks stay an error). iexec-poco, unmodified: now starts on arbitrum-sepolia; its next blocker is an RPC without archive state. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

graph-node instantiates the mapping's WASM module for every trigger (runtime/wasm/src/mapping.rs: instantiate_module before each handle_trigger): the start function runs again in that trigger's context and no global keeps what an earlier handler left in it. envio evaluated each file once, so module state leaked across events, replays and the preload pass: SporkDAO numbers TransferData from a module counter, morpho-blue caches its protocol entity in a module variable, fabcrowd keeps a per-block cache; and a top-level dataSource.context() read the data source the file was loaded for, not the current instance's. The post-transform pass now splits a module into stages. Classes whose definition runs no code, constant globals and globals built only from graph-ts values (BigInt.fromI32(0), Address.fromString(...)) are set once in __envio_static, as before. Every other top-level statement moves, in order, into __envio_start, a declaration becoming an assignment to its var; it runs as the module loads (a failure there is left to the first trigger, which runs it in a real context) and again through __envio_init(generation), which first calls the same function of each relative module imported or re-exported. The runtime calls the mapping's __envio_init as every handler, register or file data source round starts. Test: SubgraphModuleState_test.res, cargo test splits_a_module_into_static_and_per_trigger_stages. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

graph codegen never cleans its output directory, so the tree a project with one manifest per network was built with holds the data sources of every manifest it generated, and the mappings import them: WOOFi's bsc mappings import generated/WooRouterV1_3/ (only avax.yaml declares it), DODO's bsc ones src/types/dodoex/DVMFactory/ (mainnet), flaunch's generated/BuyBackAndBurnFlay/ (base-sepolia). `graph build` of that one manifest fails on a clean checkout (asc TS6054), but the deployed subgraph was compiled against the developer's tree. After codegen (graph-cli or bundled), the generated files the mappings still miss are written from the first other manifest of the project that declares that data source or template with that ABI (codegen::fill_from_manifests, exported as subgraphFillGenerated), with a warning naming the files, the manifest and why. Only missing files are written; schema.ts and envio's own manifest's files are never touched. ENVIO_SUBGRAPH_CODEGEN=graph-cli keeps graph-cli's behaviour. WOOFi bsc (with its documented constants copy) and DODO bsc now index. Test: SubgraphGeneratedFromSiblingManifest_test.res Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…receipt logs from RPC alone A block handler's timestamp and header, and event.receipt.logs, asked <chain>.hypersync.xyz first and fell back to ENVIO_SUBGRAPH_RPC on failure, even for a chain HyperSync doesn't serve. A refused connection costs little, but where a firewall drops it instead every batch waited out the request deadline (30 s) first. The translation knows which chains HyperSync serves; it now hands the runtime the others (rpcOnlyChains), whose headers and receipts go to the RPC directly. Test: SubgraphRpcBlockTimestamps_test.res (a HyperSync host that never answers is not asked). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

graph-node parses the link handed to `ipfs.cat`, `ipfs.map` and a file data source with `ContentPath::new` (graph/src/ipfs/content_path.rs): `<CID>[/<path>]`, `/ipfs/…`, `ipfs://…`, or an http(s) URL with `/ipfs/<CID>…` in its path or the CID in `?arg=`; an empty string, a non-CID or a path with a query fail to parse, and `ipfs.cat` returns null for them (runtime/wasm/src/module/context.rs) while `ipfs.map` fails. The shim sent anything it couldn't normalize to the gateway as is: SoulSystemProtocol/SoulSystem-Subgraph's `ipfs.cat("")` (the last segment of an empty URI) fetched the gateway's root, which answered 429, and the run stalled asking again; a gateway URL was appended to the gateway. SubgraphIpfs_test.res used `QmMissing`/`QmFlaky`, which graph-node refuses as CIDs; they are CIDv0-shaped now. Test: SubgraphIpfsContentPath_test.res. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…s, with its variables Aave's protocol-subgraphs writes subgraph.yaml from one script per deployment: `"deploy:hosted:mainnet-v3": "VERSION=v3 BLOCKCHAIN=v3 NETWORK=mainnet env-cmd npm run prepare:subgraph && …"`, where `prepare:subgraph` is `mustache ./config/${NETWORK}-${VERSION}.json ./templates/${BLOCKCHAIN}.subgraph.template.yaml > subgraph.yaml && … npm run generate:schema && npm run subgraph:codegen`, and `generate:schema` copies `schemas/${VERSION}.schema.graphql` to schema.graphql. envio saw three unbound variables and named the script instead of rendering. A step that is `npm run <script>` (`yarn <script>`, `pnpm <script>`) is now read as that script's steps, with the variables the call sets bound in them, as a shell hands them on; `cp <from> <to>` is repeated, and `env-cmd` is a wrapper. A script that only calls the one that renders (zora's `codegen` runs `yarn gen-yaml-config`) is dropped for it, and the runtime's codegen skips every script the rendering one inlined. Also fixed: a mustache view holding a variable (`mustache config/$NETWORK.json …`) was taken for an option. Test: project_tests::renders_through_the_scripts_a_deployment_script_runs (names the network per deployment: mainnet-v3, arbitrum-v3). Aave v3 mainnet, unmodified, ENVIO_SUBGRAPH_NETWORK=mainnet-v3: "no subgraph.yaml" before; now renders, copies its schema and indexes (57,751 events, blocks 16.29M -> 16.80M in 180 s). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Every contract call, balance and block lookup was its own HTTP request, so a batch's thousands of calls cost thousands of round trips through 48 connections. Requests that queue while those are busy now leave together as one JSON-RPC batch of up to ENVIO_SUBGRAPH_RPC_BATCH (default 32; 1 = off). Each request keeps the answer it would get alone: answers are matched by id (in any order), a JSON-RPC error stays that item's error, so a revert reaches only the try_ call that reverted, and an HTTP failure fails the items it carried for untilAnswered to ask again. An endpoint that refuses a batch (HTTP 4xx, an error object in place of the array, or a per-item "batch of more than N requests" as dRPC's free plan answers) gets smaller batches from then on, down to plain single requests. viem's own batching matches answers by position and can't shrink, so the transport is built on its HTTP client and error types instead. Test: SubgraphRpcBatch_test (40 calls, one reverting: one HTTP request per >= 4 calls, results equal; all answered by an endpoint allowing 3 per batch and by one that takes no batches). sushiswap v3 at 200 ms per HTTP request, 180 s: HTTP requests 1,899 -> 335 for ~1,900 calls, 544 -> 1,044 events; row-for-row equal to a run without (23 tables). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

graph-ts' `ByteArray.empty()` returns `ByteArray.fromI32(0)` and `Bytes.empty()` casts it (common/collections.ts, unchanged from 0.27 to 0.38): four zero bytes. The shim returned an empty array, so every field a mapping defaults with `Bytes.empty()` stored `0x` where graph-node stores `0x00000000`, and `Bytes.empty().length` read 0 instead of 4. 148 mapping files in modern's 2025-26 survey call it, among them KlimaDAO's polygon-digital-carbon, valory-xyz/autonolas-subgraph and Proof-Of-Humanity v2. Test: SubgraphBytesEmpty_test.res. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

A node's own failure arrives as JSON-RPC -32603 "Internal error" over a 200, and the same request answers on the next try. The host-op retry only knew rate-limit codes and stopped the run on it: Miljan9602/dragonswap-subgraph-sei's register pass failed on one symbol() call ('An internal error was received') that answered when asked again. graph-node retries every eth_call error that isn't a revert (chain/ethereum/src/call_helper.rs interpret_eth_call_error). Test: SubgraphRpcResilience_test.res (an internal error, then the value). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

AssemblyScript's std Map and Set aren't JavaScript's (asc 0.19.10 std/assembly/map.ts, set.ts): keys() and values() return arrays, map[key] is the @operator("[]") get and map[key] = v its set, set[key] is has, and get of a missing key throws "Key does not exist". On JavaScript's Map, Intent-X's analytics loop over dictionary.keys().length ran zero times and Reserve's RSR_CHAINLINK_CONTRACT_ADDRESS[network] read undefined. The transform turns new Map / new Set (when the module doesn't bind its own) into __envio_Map / __envio_Set, which the runtime installs: Map and Set subclasses with those methods, behind a proxy that serves a property access that isn't a member as the [] operator (a numeric spelling as a number key when the collection holds numbers). Test: SubgraphAsMap_test.res, cargo test builds_assemblyscript_maps_and_sets. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Clean: the diff tool's rate-limit spacing, --wait, --config and timeseries/aggregation comparison, and event.logIndex in graph-node's minimal bytes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…n up as missing A file data source whose gateways all answered 429 was skipped after ENVIO_SUBGRAPH_FILE_ATTEMPTS rounds, as if the file weren't there (HandProtocol lost three RoundMetadata files while public gateways rate-limited the run). graph-node's polling monitor backs off and asks again after an error as after a miss (core/src/polling_monitor/mod.rs). Only rounds in which every gateway answered 404 now count toward ENVIO_SUBGRAPH_FILE_ATTEMPTS. A failure asking again can mend (429, 5xx, timeout) is asked again with backoff for ENVIO_SUBGRAPH_IPFS_TIMEOUT_MS, the ipfs.cat budget; a file still out of reach then skips its handler uncached, with a warning that says so. Test: SubgraphFileRateLimit_test (stub gateway on an OS-assigned port). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…t to InternalTestIndexer.fromSubgraph wrote each fixture's project to os.tmpdir()/envio-subgraph-* and never removed it (25.7k directories, 478 MB on one machine). A parse-only fixture's workspace is now removed once the parse is done or has failed; one that runs tests is removed by an afterAll on the suite that declared it, after its tests. Test: SubgraphFixtureWorkspace_test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs/subgraph-compat.md: the register pass moved ahead of §4a on integration; the block handlers' paragraph and the §7 row are re-applied there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

ENVIO_SUBGRAPH_RPC, ENVIO_SUBGRAPH_NETWORK and ENVIO_API_TOKEN come from the process environment over the project's .env (ProjectEnv), as every envio setting does; the batch size and the block handler header lag were read with std::env::var alone, so a .env line setting them did nothing, and a fixture's own environment couldn't set them either (SubgraphEventConstructor set process.env instead). translate() now takes the environment map parse_subgraph has, and the project-files path fills it from ProjectEnv with TRANSLATION_ENV_VARS. Test: SubgraphEnvSettings_test (.env and the fixture env both set them). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…'s minimal bytes graph-node builds every BigInt it hands a mapping from chain data with BigInt::from(u64) or BigInt::from_unsigned_u256 and writes it as to_signed_bytes_le (chain/ethereum/src/runtime/abi.rs: Log's transaction_index/log_index via asc_new_or_null_u64, EthereumBlockData's number), so 189 is bd00 and 258 is 0201. Following 79a64fd (event.logIndex), two sites still used fromI32's four bytes: receipt.logs entries' transactionIndex, logIndex and transactionLogIndex, and a block handler's block.number. A mapping reading them as bytes (Bytes.fromBigInt, byteLength, changetype<Bytes>) built other ids. Test: SubgraphChainBigIntWidth reads all 60 chain BigInts an event, its receipt and receipt logs, a block handler's block and a call hand over (an RPC-only chain, stub RPC on an OS-assigned port) and checks each one's bytes against num_bigint's to_signed_bytes_le; only these 8 differed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The CLI (check_generated_code) and the runtime (a mapping importing the tree it couldn't build) each wrote the message, and they had come apart: "npx graph codegen" in Rust, "pnpm exec graph codegen" in TypeScript. Neither works when the message shows, with graph-cli not installed: npx reads "graph" as a package of that name. The runtime now asks the CLI's code for it (napi subgraph_missing_graph_cli_message), and the manual command names the package: npx @graphprotocol/graph-cli codegen. Test: SubgraphMissingGraphCliMessage.test.ts pins both entries to one text. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…chain createTestIndexer() reads the project synchronously, and reading a subgraph asks ENVIO_SUBGRAPH_RPC its eth_chainId. A test's own RPC stub answers on the thread that read blocks, so the ask waited out its 15 s deadline (scenarios/subgraph_test, and any user test with a stub beside createTestIndexer()). getConfigJson now passes the ports servers in the calling process listen on (process._getActiveHandles; nothing when Node lacks it), and an endpoint on a local address and one of those ports isn't asked: it counts as an endpoint that doesn't answer, so a known network name keeps its table chain, and an unknown one's error says why it wasn't asked. An endpoint in another process is asked as before, so its chain still wins over an unknown or mismatched name. Tests: cargo an_endpoint_this_process_serves_is_not_asked (a silent listener, 30 s -> 0 s); subgraph_test's createTestIndexer() 15 s -> <1 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…r a sure create d8d46d8 gave a data source's block handler a register pass at each of its blocks whenever the create analysis said its code can reach a create, and that analysis counts anything it can't follow as one. kiln's vsuite handleBlock (unfiltered) verifies BLS12-381 signatures in a module the analysis can't follow and creates nothing; its register pass, reading a store without the entities its handler pass reads, threw before advancing its tracker and redid the pairings at every block: the indexer stopped at 101 events where integration syncs the window in 22 s. A data source's unfiltered or polling handler that runs on more than 1,000 blocks now gets a pass only when its code surely reaches a template's generated class (creates::sure_creates: tinlake's handleBlock reaches all six). A once handler, or one on a data source of at most 1,000 blocks (beefy-balances' one-block binder), keeps its pass on suspicion. A create from a block handler left without a pass is refused where it runs, as a template's unfiltered handler's is, rather than lost. Tests: SubgraphBlockHandlerUnsureCreate_test (an unfiltered handler's raw DataSourceTemplate.create is refused at the create; it got a pass before); SubgraphPollingBlockHandlerCreate_test now reaches the template through its generated-style class. kiln-vsuite: ok 101 events -> synced 5,465 in 27 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

6548d15 gave a data source's unfiltered or polling handler on more than 1,000 blocks a pass only for a sure create; §4 and the §7 row say so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The test parsed its project with the addon at ENVIO_DEV_ADDON, the debug build globalSetup compiles. Where envio's platform package is installed (CI installs the release build of the commit under test), every other test loads that one, and without the debug build this test failed with "Cannot find module target/debug/envio.node". The child process now imports envio's Core and parses through getAddon(), which picks the platform package first and the dev build otherwise, as the CLI does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

globalSetup ran `cargo build --lib` before every vitest run. envio loads its platform package before any dev build (Core.loadAddon), and CI installs that package from the release build of the commit under test, so on CI the build produced a debug addon no test loads. On a fresh runner with no cargo cache it was a cold build of the whole crate: 3m14s of the envio-tests job at 76a793f and 5m00s at d3e211a, which put the job past its 9-minute limit. globalSetup now resolves envio's platform package the way Core.loadAddon does (envio-<platform>-<arch>, then -musl) and builds only when neither is installed, which is the case where workers fall back to the dev build. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

InternalTestIndexer imports TypeChecker.ts, so every test file that uses it (294 of 440) loaded the compiler while vitest collected it: about 250 ms per forked process. Only the fromUserApi fixtures with handlers or a test, and completionsAt, type-check; the subgraph fixtures and parse-only fixtures never do. The compiler, its host and the language service are now built on the first check, and the TypeScript version pin is checked there. Full `pnpm test` on Postgres, 4 CPUs (taskset -c 0-3): vitest's summed import time 221 s -> 145 s, wall 153 s -> 117 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

solana.hypersync.xyz resolves to five addresses and one of them, 193.32.220.135, accepts no TCP connection (checked with curl --resolve against each). A test whose first connection lands there waits about 30 s before the source retries on another address. SvmWhereLive and SvmBytesLive allow 300 s for that; SvmAccountKeysLive ran under vitest's 30 s default and failed with "Test timed out in 30000ms" in 2 of 6 local runs of the three live SVM files. CI hit the same stall on SvmWhereLive (31.4 s) in the cancelled run at d3e211a. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ainStuckAtHead (#1690) Cherry-pick of main's 04798e0, which this branch's base (merge-base 04e33b9, 5 Oct) predates: - globalSetup skips the dev addon cargo build when envio's platform package is installed, as CI installs it. - SvmAccountKeysLive gets a 300 s timeout. - MultichainStuckAtHead runs the scenario with a 2 s realtime stall timeout (newBlockStallTimeoutRealtime threaded through IndexerState and ChainState to SourceManager, as reducedPollingInterval is) instead of waiting out the real 10..20 s twice: 34 s -> 3.9 s locally. globalSetup.ts and SvmAccountKeysLive_test.res conflicted with this branch's own equivalents (face3e6, d587bda) and are resolved to main's text, so a later merge of main finds the same lines on both sides. (cherry picked from commit 04798e0) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…new fork's address is stored Array.includes compared the rows by reference, so its loop always ran to the 20 s deadline (the assertion after it compares deeply and passed). 20 s -> 80 ms. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Don't start a write once the indexer has stopped Chain metadata writes are throttled: staging arms a timer that later schedules the write fiber. A restage after stop()'s flush - finalizing the backfill commits ready_at, and the loop restages the chain's metadata - left that timer armed, so the write started after stop() had returned. In envio-tests the schema is dropped by then, the write fails, and the harness's default onError calls process.exit, which kills the vitest worker with no test named (seen as "Worker exited unexpectedly" on ResponseBlocksMetric_test). Writing.schedule now starts nothing once the indexer is stopped. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ETueLEEQ8tbFKY69q1ys9k * Fail the test on an indexer error instead of exiting the worker IndexerRunner's default onError logged the error and called process.exit, and test logs are silenced, so any indexer failure in a test without its own onError killed the vitest fork with only "Worker exited unexpectedly" - no test name, no error. The default now records the first failure and fails the run with it. The body is raced against the failure so a test waiting on progress the failed indexer will never make fails right away instead of timing out. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ETueLEEQ8tbFKY69q1ys9k * Add Somnia Testnet (50312) as a HyperSync chain The HyperSync API now lists it, which failed the hypersync health check's drift check against the Network enum. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ETueLEEQ8tbFKY69q1ys9k * Refuse a restart once the test run is tearing down The run stops waiting on the body once the indexer fails, but the body keeps running. A restart() it made after that started an indexer and a Postgres client that teardown had already passed, so nothing stopped or closed them. make() now refuses to start, or to start the loop after its awaits, once teardown has begun. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ETueLEEQ8tbFKY69q1ys9k --------- Co-authored-by: Claude <noreply@anthropic.com> (cherry picked from commit 773752a)

FinalizeIndexLogging_test and WorkerResumeLogging_test each read a pino log file while its writer thread was still appending. A read landing mid-line handed the partial line to JSON.parseOrThrow, which threw ("Unexpected end of JSON input") instead of letting the poll read again - seen failing envio-tests (clickhouse) on #1693. The two copies of the reader are now one helper, LogFile.messages, which parses only lines that end in a newline. Claude-Session: https://claude.ai/code/session_01ETueLEEQ8tbFKY69q1ys9k Co-authored-by: Claude <noreply@anthropic.com> (cherry picked from commit b0bd566)

moose-code deployed to internal October 20, 2026 14:22 — with GitHub Actions Active

moose-code and others added 28 commits October 20, 2026 15:35

docs: case-matched mapping imports in subgraph-compat §6d

0b0fe84

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

moose-code and others added 3 commits October 21, 2026 06:14

Merge moose/sg2-hosted (79a64fd) into sg2-integration

6fe34bd

Clean: the diff tool's rate-limit spacing, --wait, --config and timeseries/aggregation comparison, and event.logIndex in graph-node's minimal bytes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

moose-code deployed to internal October 21, 2026 06:37 — with GitHub Actions Active

moose-code and others added 6 commits October 21, 2026 06:38

Merge moose/sg2-integration (6fe34bd) into moose/sg2-blockcreate

9e8603c

docs/subgraph-compat.md: the register pass moved ahead of §4a on integration; the block handlers' paragraph and the §7 row are re-applied there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Merge moose/sg2-integration (6fe34bd) into moose/sg2-polish

629f296

moose-code deployed to internal October 21, 2026 06:53 — with GitHub Actions Active

moose-code and others added 8 commits October 21, 2026 06:53

Merge moose/sg2-integration (d3e211a) into moose/sg2-blockcreate

3545f25

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs: which block handlers get a register pass

0bd5a3f

6548d15 gave a data source's unfiltered or polling handler on more than 1,000 blocks a pass only for a sure create; §4 and the §7 row say so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

moose-code deployed to internal October 21, 2026 08:51 — with GitHub Actions Active

moose-code and others added 2 commits October 21, 2026 08:51

Merge moose/sg2-integration (26d0b54) into moose/sg2-blockcreate

f3bc4fc

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

moose-code deployed to internal October 21, 2026 08:56 — with GitHub Actions Active

DZakh and others added 2 commits October 21, 2026 09:07

moose-code deployed to internal October 21, 2026 09:09 — with GitHub Actions Active

moose-code changed the title ~~Subgraph mode: every compatibility fix in one PR (20 → 89 of 130 real subgraphs)~~ Subgraph mode: compatibility fixes, 30% → 83% of 860 real subgraphs (85% of 2026 repos) Oct 21, 2026

moose-code requested a review from DZakh October 21, 2026 09:47

DZakh merged commit fdca61e into claude/hyperindex-subgraph-mode-bs67pq Oct 22, 2026
9 of 10 checks passed

DZakh deleted the moose/subgraph-compat-all branch October 22, 2026 11:33

Navigation Menu

Repository navigation

Subgraph mode: compatibility fixes, 30% → 83% of 860 real subgraphs (85% of 2026 repos) - #1759

Subgraph mode: compatibility fixes, 30% → 83% of 860 real subgraphs (85% of 2026 repos)#1759
DZakh merged 447 commits into
claude/hyperindex-subgraph-mode-bs67pqfrom
moose/subgraph-compat-all

moose-code commented Oct 20, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Oct 20, 2026 •

edited

Loading

Review skipped

Uh oh!

Uh oh!

This branch was successfully deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

moose-code commented Oct 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Results: unmodified real subgraphs

What changed, by area

Please review closely

Known gaps

Tests and CI

What's in it

Uh oh!

coderabbitai Bot commented Oct 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Uh oh!

Uh oh!

This branch was successfully deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

moose-code commented Oct 20, 2026 •

edited

Loading

coderabbitai Bot commented Oct 20, 2026 •

edited

Loading