fix(go): a type alias type A = B is indexed and links to the type it names

[colbymchenry]

Problem

A Go alias declaration (type Event = mvccpb.Event, type ( A = B )) was in neither extractor's graph. tree-sitter-go parses it as a type_alias node beside the type_spec of a defined type, and both the wasm extractor (languages/go.ts listed only typeAliasTypes: ['type_spec']) and the kernel (go.rs matched only type_spec) read type_spec alone. So etcd's client/v3/watch.go:52 type Event = mvccpb.Event had no node while line 54's type WatchChan <-chan WatchResponse did.

What that did downstream, measured on main (fe95183, and again on 31c3328):

• A use written through the alias's package (clientv3.Event, prometheus's 390 storage.AOptions{…} literals) linked to nothing.
• A use inside the package linked by name to whichever type of that name the project had: prometheus's &PrometheusVersion{} in web/ went to web/api/testhelpers's struct, and []Type in util/compression to tsdb/compression's Type.
• A method called on an alias-typed value fell back to a lookup by name, which can pick another package's same-named method.
• A generic alias (type Items[T any] = List[T]) has no rule in tree-sitter-go 0.23. It parses as a type_spec around an ERROR for the =, so the kernel defers the file and wasm's error recovery already made a type_alias node. That node named nothing, and method resolution then stopped at it: xs.Len() on an *Items[int] resolved to nothing.

Fix

Extraction (TS and kernel, kept at parity)

• Go's typeAliasTypes is ['type_spec', 'type_alias']. The kernel's walker routes type_alias the same way.
• An alias of a struct or interface literal becomes a struct / interface node, exactly as a type_spec does.
• Otherwise it is a type_alias node with a references ref for each type its right-hand side names. Each ref sits on the name, so resolution reads a package qualifier back (Go: cross-package calls are not resolved when go.mod is not at the project root #2322). The alias's own type parameters and Go's predeclared types are skipped.
• goAliasTypeNames (go.ts) and alias_type_names (go.rs) also recognize the generic-alias error shape on the wasm side. The kernel never sees it, since that file defers.

Resolution

• resolveGoMethodInPackage treats an alias as the type it names. A method called on an *Event is looked up on mvccpb.Event, through any chain of aliases and pointers (type Ptr = *Local, type Local = WatchResponse).
• An alias whose target can't be placed in the project (type Ctx = context.Context, func(…), map[…]…) resolves exactly as before, by name. Defined types (type Dur int) are not followed.
• gateTargetKind: an alias's target written through a package that is none of the file's imports as indexed (clientv3 under an unaliased go.etcd.io/etcd/client/v3, which the index knows as v3 or, since fix(go): an unaliased import is also known by the name goimports assumes for its package #2410, client) stays unresolved. This is the rule fix(go): an interface links the interfaces it embeds #2397 applies to embedded types. Without it, name matching linked the alias to itself or to a namesake.

Validation

Tests: __tests__/go-type-alias.test.ts (14 tests).

• Extraction on both backends, LF and CRLF.
• Generic aliases on wasm, asserting the kernel defers that file.
• An indexed module: alias edges, a bare Event{} / *Event reaching the alias, method calls through one, two and generic aliases, and an alias written through an import the index doesn't know. That last case fails with the rule disabled.
• All 14 failed on fe95183 and pass here. The module's namesake package sorts before api/, so on main the call through *Event went to its Event::IsCreate.
• torture.go gains every alias shape; kernel-tsjs-parity (92 tests) passes.
• After merging main at 31c3328 (fix(kernel): a Go value like wm.new or raft.None keeps its function ref #2401, fix(sync): a Liquid section or snippet reference links its file when the file appears later #2403 and fix(cpp): a base class links to the class C++ name lookup finds, not a namesake #2407 landed meanwhile, fix(kernel): a Go value like wm.new or raft.None keeps its function ref #2401 editing go.rs), the kernel was rebuilt from the merged source. Then 16 files, 926 tests, pass: these tests, every Go test file, kernel-tsjs-parity, kernel-grammar-parity, kernel-node-id-parity, and the files main's new commits touched (extraction, cpp-base-class-lookup, sync-path-reference-retry).
• After merging main again at d8a7f86 (fix(go): an unaliased import is also known by the name goimports assumes for its package #2410 makes go.yaml.in/yaml/v3 known as yaml, so the test's unknown-package case moved to clientv3, the shape fix(go): an unaliased import is also known by the name goimports assumes for its package #2410 keeps unknown; test(sync): compare synced graphs with an index rebuilt from scratch #2411), 14 files, 251 tests, pass. These are these tests, go-import-package-names, go-interface-embedding, sync, sync-import-retry, kernel-tsjs-parity and the other Go test files.
• After merging main a third time at 855eca2 (fix(go): a call through an interface reaches a method the struct gets by embedding #2414, fix(dead-code): a type that declare module or declare global merges outward is not dead #2418, fix(c,cpp): a file named like a test is visible to the code that includes it #2421, fix(react-router): a JSX index route is the page at its parent's address, inside the layout around it #2423 and fix(resolution): a large minified JS bundle no longer stalls resolving on Node 22 #2424; a CHANGELOG conflict was resolved by keeping both), the kernel was rebuilt. 19 files, 294 tests, pass. These include go-promoted-dispatch, dead-code, cpp-included-test-suite, react-router, ts-ambient-exports, js-local-binding-backtracking, jsx-render-work and kernel-tsjs-parity.
• Main was merged again at 9d07864 (fix(js,ts): import { default as X } is the module's default import #2412, fix(frameworks): a framework resolves only references written in its own languages #2415, fix(sync): a route links the module it lazily loads when the module appears later #2422, fix(cpp): a class keeps its namespaces and enclosing class past code the parser misreads #2426) and at d348d30 (fix(go): a type in a signature or composite literal links a type, never a method of its name #2416, fix(sync): a navigation call links its route when the route appears later #2425).
  • fix(go): a type in a signature or composite literal links a type, never a method of its name #2416's Go type-position gate conflicted with this PR in gateTargetKind. The alias rule now runs first: that gate returns early for every Go type position, and it keeps a qualified ref that landed on any type even when the qualifier is unknown.
  • 17 files, 278 tests, pass, including go-type-position-kinds, framework-resolve-language and sync-navigation-retry.
• Merged once more at 2f2afea (fix(go): a defined type implements the interfaces its methods satisfy #2419, Go defined types as implementers). go-type-alias gains a KumaSDConfig-shaped case, alias methods implementing an interface, and 14 files, 185 tests, pass, including go-implements-defined-types (its true-alias case still holds: that alias owns no methods).
• Merged at 61f8b09 (fix(cpp): a receiver declared as a std or other outside type with a lowercase name calls that type's own member #2413, C++ receivers, gated on ref.language === 'cpp'). 10 files, 172 tests, pass, including cpp-library-type-receiver.

Full suite (before the merge, on fe95183 + this change): 26 of 6496 failed in the parallel run on this shared Windows box. All were daemon / MCP / sync / watcher / perf-threshold tests: hook timeouts, EBUSY on teardown, and two timing bounds (2341 ms vs 2000, 101.7 ms vs 100).

• Run serially with longer timeouts, those 18 files passed but for 6.
• Of those 6, mcp-daemon (21/21) and mpeg-ts-not-typescript (11/11) pass alone.
• The last is function-ref's Caller/impact graph misses methods passed as first-class references (callbacks), e.g. executor.submit(obj.method, …) #1820 "a module global never binds…" case, which overruns its own 60 s timeout on this box. It is Python-only, and every resolver path this PR changes is behind ref.language === 'go'.

Kernel parity sweep (scripts/kernel-parity.mjs --lang go), main 31c3328 vs this branch:

repo	files at parity	deferred to wasm	diffs
etcd	1065/1105 → 1065/1105	40 → 40	0 → 0
prometheus	736/736 → 736/736	0 → 0	0 → 0
gin	99/99 → 99/99	0 → 0	0 → 0

The wasm totals move by exactly the aliases: etcd +13 nodes / +26 refs, prometheus +9 / +8. (Before #2401 landed, the sweep showed its 6 / 202 raft.None / compression.None diffs, the same count in both arms.)

A/B on real repos (shallow clones: etcd 76d58e3, prometheus 39c878f, gin 43fe48e). Each repo was indexed with real builds of main 31c3328 and of this branch, on the kernel path and, for etcd and prometheus, the wasm path too. Kernel and wasm graph dumps are byte-identical in each arm, and the diffs are the same on both paths, and the same as the first A/B against fe95183. Re-run against each later main, the delta was identical edge for edge every time, with no dispatch or implements edge changed: d8a7f86 (after #2410), 855eca2 (after #2414's promoted-method dispatch), 9d07864 (after #2415) and d348d30 (after #2416). Only the baselines moved: at d348d30, etcd went 63105 → 63215 and prometheus 96813 → 97294. Against 2f2afea (after #2419), prometheus adds the 4 implements/dispatch edges described below (+485), and etcd and gin are unchanged. The table shows the 31c3328 run.

Edges are diffed site by site (source, line, column, ref name), and each target is judged by Go's scoping: a bare name belongs to its own package, pkg.X to the import's package.

repo	aliases	edges	retargeted	added	dropped
etcd	13	63209 → 63319	2	97 + 13 contains	0
prometheus	9	96824 → 97305	10	467 + 9 contains + 5 go-method-contains	0
gin	0	7674 → 7674	0	0	0

Every retargeted and added edge was checked:

• Retargeted (12): all moved from another package's namesake to the alias in the reference's own package:
  • etcd's &Event{ ×2 in client/v3/watch_test.go, from mvccpb's struct;
  • prometheus's &PrometheusVersion{} ×7 in web/web_test.go, from web/api/testhelpers's struct;
  • []Type ×3 in util/compression, from tsdb/compression's Type.
• Added (564): every one lands in the package Go reads the name from, and none is a self-loop:
  • uses through the alias's package that resolved to nothing before (clientv3.Event, pb.Watch_WatchServer, storage.AOptions{…}, compression.Type);
  • uses inside the package (KeyPredicate, recCase, the test files' sample);
  • the aliases' own refs to their targets, e.g. Event → mvccpb.Event, and the grpc stream aliases → their request/response structs, while grpc.ServerStreamingClient stays unresolved;
  • one embedding, pb.KV_RangeStreamServer.
• go-method-contains (5): prometheus's func (c *KumaSDConfig) Name() and four more are written on the alias KumaSDConfig = SDConfig. The existing synthesizer already accepts type_alias owners, the same path that gives a defined type its methods, so these five methods, orphaned before, now hang off the alias.
• With fix(go): a defined type implements the interfaces its methods satisfy #2419 merged (defined types as implementers, which also walks type_alias), the alias that owns those methods implements what they satisfy, standing in for SDConfig. On prometheus that is 1 go-implements edge, KumaSDConfig → discovery.Config, whose methods are Name, NewDiscoverer and NewDiscovererMetrics, and 3 interface-impl edges, so a call through discovery.Config now reaches the Kuma config's methods. An alias without methods of its own, the rest of them, implements nothing. The goImplementsEdges doc comment said a true alias is not extracted, so it now says what the alias owns.
• Metadata only (1): etcd's p.Config.RevisionOffsetRange.Rand() still calls Range::Rand, at 0.7 instead of 0.9, because a second Rand (the antithesis package's type Rand = mrand.Rand) now exists.
• Dead code: the type_alias rows of the dead-code report are unchanged on etcd and prometheus, with exported, test and generated code included.

The same module also reproduces the issue on the built CLI. On main, isCreate's e.IsCreate() and newEvent's Event{} go to the namesake package, and p.Err() (through two aliases) and xs.Len() (generic alias) resolve to nothing. This branch gives the alias nodes, Event → mvccpb.Event, and the calls on mvccpb.Event, WatchResponse and List.

Overlap

#1954 (open, @danusha2345) also teaches resolveGoMethodInPackage about aliases, reading type Name = Target from the package's files when the package has no node of that name. Once this lands, aliases have nodes, so that branch no longer fires and the alias-following here covers it. #1954's other parts are untouched by this PR:

• standard-library-named variables;
• qualified parameter types;
• call-result receivers;
• no edge for outside types.

Here, an alias of an outside type still resolves by name, as on main, which leaves #1954's "outside types get no edge" rule to it.

Not in this PR

• A defined type (type WatchChan <-chan WatchResponse, gin's type HandlerFunc func(*Context)) still references nothing: before this PR no Go type declaration had a references edge to the type it is defined from. Aliases now do. Doing the same for defined types is a separate, larger graph change.
• A doc comment above an ungrouped Go type (// Foo … then type Foo struct{}) is dropped by both extractors; only types inside a type ( … ) group keep theirs. Aliases behave like other types here.
• Methods written on an alias receiver (func (c *KumaSDConfig) Name() with KumaSDConfig = SDConfig) belong to SDConfig's method set in Go. The graph keeps them under the alias, which implements discovery.Config in SDConfig's place. An interface whose methods are split between the two (some written on SDConfig, some on the alias) is satisfied by neither node.

🤖 Generated with Claude Code