Nothing catches a `vi.mock()` whose relative specifier resolves to no file — vitest no-ops it and the suite stays green

[os-sales]

Observation filed while implementing #5577 (PR #5645). No live instance exists today — the repo is currently clean, measured below. Filed because the failure mode is invisible by construction and nothing prevents the next one.

The mechanism

vi.mock('<relative path>', factory) with a specifier that resolves to no file on disk does not error. Vitest registers the mock against a module id nothing imports, and the run proceeds with the real module everywhere. The suite passes.

That is the worst available failure direction: the test does not fail, does not warn, and does not report a smaller number of assertions. It reports exactly the same green as a correct one, so nothing in the output distinguishes "this stand-in is installed" from "this stand-in is inert".

How it surfaced

A new suite in PR #5645 mocked the runtime-config module to install a deliberately partial snapshot. The file lives at packages/app-shell/src/layout/__tests__/, and the mock was written vi.mock('../runtime-config', …) — one .. short. That resolves to packages/app-shell/src/layout/runtime-config, which does not exist; the correct specifier is '../../runtime-config'.

The suite passed. It also passed when the code under test was reverted to the exact pre-fix shape the suite was written to catch — a TypeError that, with the mock correctly installed, fails all four cases. The probe was inert and green, and only an ablation leg (mutate the subject, confirm the test goes red) exposed it.

Neighbouring mocks in the same file hid it further: '../AiUsageIndicator' (one level) and '../../hooks/surfaceAgent' (two levels) are both correct from a __tests__/ subdirectory, because the targets sit at different depths. So a reader scanning the mock block sees a plausible mix of ../ and ../../ with no way to tell which is wrong.

Current state — measured, not assumed

Scanned every vi.mock / vi.doMock with a relative specifier across the repository:

scanned 1807 test files, 577 relative vi.mock specifiers
UNRESOLVABLE: 0

(Resolution tried the bare path plus .ts/.tsx/.js/.jsx/.mjs/.cjs and the /index.* forms, and stripped a trailing .js to account for the NodeNext specifier style used throughout src/.)

So this is not a cleanup card. Every one of those 577 is load-bearing today, and the point is that if one of them silently stopped being load-bearing, the only signal would be a test that keeps passing.

Shape of a fix

A scripts/check-*.mjs gate in the existing family. The scan above is the whole check — walk the tracked test files, regex the relative vi.mock/vi.doMock specifiers, resolve each against the file's own directory, fail on any that resolve to nothing. It needs no build, no install and no vitest run, and it completed over all 1807 files in about a second, so it fits the cheap-gate tier rather than the CI-heavy one.

Worth deciding at implementation time, not assumed here:

• whether to cover bare specifiers too (vi.mock('@object-ui/…')) — those can also be misspelled, but resolving them needs the workspace map rather than the filesystem, so it is a larger check with a different failure mode;
• whether an inert mock should be an error or a warning — there is no legitimate reason to mock a path that does not exist, which argues for error;
• whether the same treatment is owed to vi.mock calls written as vi.mock(import('…')), which the regex above already matches but which no current call site uses.

Why this is the same family as an already-accepted gate

#4347 (closed) was "check-type-check-coverage.mjs reports green when a type-check script chains a tsconfig project that does not exist" — the same shape one layer up: a declaration pointing at nothing, reported as a pass. This is that failure at the module-mock layer, and it is currently ungated.

Refs

• features.aiStudio is read inline at two call sites while features.marketplace has an accessor — one fail-open doctrine, two spellings #5577 / PR refactor(app-shell): one isAiStudioEnabled() accessor for features.aiStudio, replacing two inline spellings #5645 — where the inert mock was written, caught by ablation, and corrected

[huangyiirene]

Concentration-round first-touch (triage seat): promoted finding → pm:queue, Task S, routed domain:devx (a scripts/check-*.mjs gate over test infrastructure — the devx face of this repo).

Queued despite zero live instances because the class is invisible by construction (the one occurrence was caught only by an ablation leg, and 577 live specifiers have no other guard) and the check is already written and measured cheap (~1s, no build). Scope ruled: relative vi.mock/vi.doMock specifiers only, error not warning (the card is right — no legitimate reason exists to mock a nonexistent path); bare-specifier coverage deferred until a real instance shows (different resolver, different failure mode — don't build it speculatively). Self-test both ways: a planted unresolvable specifier must fail, the 577 real ones must pass. Precedent lineage #4347 noted.

---

Generated by Claude Code