[kennysarnoski-jarvis]

Note

🤖 Claude Mythos 5.1 writing on behalf of Kenny

Fixes #64757.

AI disclosure, per CONTRIBUTING.md. Kenny Sarnoski (kennysarnoski-jarvis, who filed #64757) chose this issue and is shepherding this PR himself. He and Claude (Mythos 5.1, Anthropic) worked out the cause and the fix together in a Claude Code session he directed; Claude wrote the patch, the test and this description. He has read the change and will answer review feedback. This is a single, targeted PR for his own issue, not part of any queue-driven workflow.

What breaks

Since 7.1.0-dev.20261001.1, inferring from an argument whose type is a generic type nested inside its own type arguments, such as Eff<() => Eff<A, R>, never>, drops the inferences for the inner type parameters when the parameter type comes from an instantiated generic alias or interface. R falls back to unknown. 7.0.2 and 7.1.0-dev.20260930.4 infer it. Three real-world shapes are in #64757: Effect's HttpRouter.middleware(...).layer (which becomes an error string), Prisma's nested select results, and Saga.toLayer from alchemy.run (137 errors and 285 false Effect diagnostics in one app).

Cause

#64553 routed the two "same generic type, infer from the type arguments" paths in inferFromTypes through invokeOnce so the new cache could cover them. invokeOnce also pushes the pair onto the recursion stacks and stops with a circularity when isDeeplyNestedType(..., 2) finds the same recursion identity twice on both sides. The outer Eff<...> and the inner Eff<A, R> share an identity, so the inner pair is reported as a cycle and inferFromTypeArguments never runs for it. Before #64553 those paths called inferFromTypeArguments directly, without the guard, as TypeScript 5.9 and 7.0 still do.

Fix

invokeOnce is split into a worker with a checkRecursion flag. The two type-argument paths go through invokeOnceWithoutRecursionCheck, which keeps the cache (and its circularity marker for true re-entry) but leaves the recursion stacks alone and never reports a circularity. Inference from type arguments is bounded by the arguments themselves; deferred type references were already excluded from this path, and object types reached on the way down still go through the guarded invokeOnce. The other three callers (conditional types, generic mapped types, object types) are unchanged.

Testing

• New compiler test inferFromNestedTypeArgumentsOfSameGeneric.ts with the three shapes from Regression in 7.1 (#64553): R inferred as unknown through a generic type alias call signature with a nested same-type argument #64757, the Prisma one as reported. On main it fails with four errors: three TS2322 ([unknown]/unknown where [Request], never and Request are expected) and one TS2571 on the Prisma viaAlias line; with this change the .types baseline shows the expected types and there is no .errors.txt.
• go test ./... in tsc passes (after npm ci, which internal/astnav needs).
• correlatedUnions.ts from Cache inferences made from type arguments #64553 keeps its current baseline.
• Real code: the one-file Effect 4.0.2 repro from 0.4.0 regression: Effect HttpRouter.middleware(...).layer types as string (requires inferred unknown); 0.3.0 and tsc 7.1.0-dev accept pingdotgg/ts-rust#79 and an Effect-based server (74 errors on main, 0 with this change, with no source changes) both typecheck.

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

The Prisma regression lacks demonstrated coverage because Shape 3 compares the same alias on both sides and was not shown to fail before the fix, unlike the reported repro’s distinct-alias nested-array inference path.

1 open issue

• Add a verified nested-array Prisma regression test · New

What changed in this PR

Fixes #64757 by allowing inference to continue through nested instances of the same generic type, while preserving the inference cache introduced in #64553.

Changes:

• Separates recursion checking from caching for same-generic type-argument inference.
• Adds compiler cases for nested inference through generic aliases and interfaces.
• Records expected inferred types and symbol bindings.

File	Description
tsc/​testdata/​tests/​cases/​compiler/​inferFromNestedTypeArgumentsOfSameGeneric.ts	Adds nested generic inference test cases.
tsc/​testdata/​baselines/​reference/​compiler/​inferFromNestedTypeArgumentsOfSameGeneric.types	Records expected inference results.
tsc/​testdata/​baselines/​reference/​compiler/​inferFromNestedTypeArgumentsOfSameGeneric.symbols	Records test symbol bindings.
tsc/​internal/​checker/​inference.go	Keeps type-argument caching without applying the nesting guard.

_{Review effort: Balanced}

[kennysarnoski-jarvis]

Note

🤖 Claude Mythos 5.1 responding on behalf of Kenny

Agreed, that reduction did not fail before the fix. Shape 3 is now the repro from the issue (generic alias as a parameter type, inference through nested arrays of a recursively computed type). Before the fix it fails with TS2571 on viaAlias; with it, both inline and viaAlias are string, and the baselines are regenerated (132d27c).

[kennysarnoski-jarvis]

@microsoft-github-policy-service agree