🔎 Search Terms
checkJs, allowJs, .d.mts, .d.cts, declaration file adjacent to implementation, file silently dropped from program, extension priority
🕗 Version & Regression Information
Reproduces identically on 5.9.3, 6.0.3, and 7.1.0-dev.20260916.1. Long-standing, not a regression.
⏯ Playground Link
Not applicable: needs two sibling files on disk plus moduleResolution: nodenext.
💻 Code
Three directories, identical but for extensions. Same tsconfig.json in each; package.json has "type": "module" only for the .mjs case.
index.js / index.mjs / index.cjs:
export const n = 1;
const bad = null;
bad.a.b.c();
index.d.ts / index.d.mts / index.d.cts:
export declare const n: number;
{
"compilerOptions": {
"target": "esnext",
"module": "nodenext",
"moduleResolution": "nodenext",
"allowJs": true,
"checkJs": true,
"strict": true,
"noEmit": true
}
}
🙁 Actual behavior
index.js + index.d.ts -> error TS18047: 'bad' is possibly 'null'.
index.mjs + index.d.mts -> (no output)
index.cjs + index.d.cts -> (no output)
The same authoring relationship — a hand-written declaration file beside its implementation — produces opposite outcomes based only on the file extension.
In the .mjs/.cjs cases the implementation is not merely unchecked, it is absent from the program: --listFiles reports only index.d.mts. In the .js case --listFiles reports both index.d.ts and index.js.
🙂 Expected behavior
The three rows should agree. Either all of them check the implementation, or none do.
I'd expect the .js row to be the correct one — the declaration file describes the module for consumers, and checkJs still checks the implementation — but consistency matters more than which way it goes.
Additional information
#47796 reported the .mjs/.mts variant and was closed as "Working as Intended", on the rationale that the JS file is a build artifact of the TS file, and that "the compiler only automatically pulls in one copy of a file with the same basename (the most TS-y one)".
That rationale doesn't extend to this case:
- A
.d.mts is not something a .mjs is emitted from, so the build-artifact framing doesn't apply. .d.ts + .js is the same relationship, and there the compiler does pull in both.
- Neither does the one-copy-per-basename rule, for the same reason —
index.js and index.d.ts are both in the program.
The failure is silent, which is what makes it expensive: a package shipping .mjs/.cjs with adjacent hand-written declarations gets no checking of its implementation at all, while every sibling module without a declaration file is checked normally. tsc exits 0 and CI is green. I found this only by deliberately breaking a file and noticing the build still passed. Per #47796, VS Code does report these errors, so the editor and CI silently disagree.
#64098 is the same silent-drop shape from wildcard include (hasFileWithHigherPriorityExtension), if the two are related.
So, concretely:
- Make the three rows consistent, or
- if this is intended, document which sibling extensions suppress checking, and emit a diagnostic when an input file is dropped from the program for this reason.
🔎 Search Terms
checkJs, allowJs, .d.mts, .d.cts, declaration file adjacent to implementation, file silently dropped from program, extension priority
🕗 Version & Regression Information
Reproduces identically on 5.9.3, 6.0.3, and 7.1.0-dev.20260916.1. Long-standing, not a regression.
⏯ Playground Link
Not applicable: needs two sibling files on disk plus
moduleResolution: nodenext.💻 Code
Three directories, identical but for extensions. Same
tsconfig.jsonin each;package.jsonhas"type": "module"only for the.mjscase.index.js/index.mjs/index.cjs:index.d.ts/index.d.mts/index.d.cts:{ "compilerOptions": { "target": "esnext", "module": "nodenext", "moduleResolution": "nodenext", "allowJs": true, "checkJs": true, "strict": true, "noEmit": true } }🙁 Actual behavior
The same authoring relationship — a hand-written declaration file beside its implementation — produces opposite outcomes based only on the file extension.
In the
.mjs/.cjscases the implementation is not merely unchecked, it is absent from the program:--listFilesreports onlyindex.d.mts. In the.jscase--listFilesreports bothindex.d.tsandindex.js.🙂 Expected behavior
The three rows should agree. Either all of them check the implementation, or none do.
I'd expect the
.jsrow to be the correct one — the declaration file describes the module for consumers, andcheckJsstill checks the implementation — but consistency matters more than which way it goes.Additional information
#47796 reported the
.mjs/.mtsvariant and was closed as "Working as Intended", on the rationale that the JS file is a build artifact of the TS file, and that "the compiler only automatically pulls in one copy of a file with the same basename (the most TS-y one)".That rationale doesn't extend to this case:
.d.mtsis not something a.mjsis emitted from, so the build-artifact framing doesn't apply..d.ts+.jsis the same relationship, and there the compiler does pull in both.index.jsandindex.d.tsare both in the program.The failure is silent, which is what makes it expensive: a package shipping
.mjs/.cjswith adjacent hand-written declarations gets no checking of its implementation at all, while every sibling module without a declaration file is checked normally.tscexits 0 and CI is green. I found this only by deliberately breaking a file and noticing the build still passed. Per #47796, VS Code does report these errors, so the editor and CI silently disagree.#64098 is the same silent-drop shape from wildcard
include(hasFileWithHigherPriorityExtension), if the two are related.So, concretely: