What happened (found during #356)
Withdrawing BatchAction.DELETE (#331) revealed two transcriptions of the batch action table:
frontend/ui-core/src/testing/wire.fixtures.ts — typed BatchAction; removing the member
was an instant tsc error.frontend/app/e2e/_wire.ts — typed readonly string[]; compiled clean and kept serving
"delete".
The generated checks.ts rejected the stale payload inside a hook, which doesn't raise — the
query never resolves, and every gallery spec failed with element(s) not found and 20-second
timeouts reading as a UI regression. All 2590 pytest, both vitest suites, mypy, ruff,
lint-imports, generate:client:check and typecheck:e2e were green with the broken fixture in
place. Only the browser suite could see it, and it pointed nowhere near the cause.
Values were fixed in #356; the structural gap was not (recorded there under "Found, not fixed").
Why it wasn't fixed inline
Typing _wire.ts against the real roster means importing from @visionset/ui-core internals,
and the package exports only its root — typing the fixture widens the public API, which is a
decision, not a diff.
Options
- Export a
@visionset/ui-core/testing subpath carrying the wire fixtures and their types;
_wire.ts imports and narrows. Widens the API deliberately, kills the second roster.
- Generate
_wire.ts from the same source as wire.fixtures.ts (the generated/ client
already sets the precedent: regenerate, never hand-edit).
- A comparison test asserting the two rosters agree — cheapest, keeps both copies, and
turns the silent drift into a named red test.
Whichever lands, the acceptance test is the same: re-adding "delete" to _wire.ts alone must
turn a named check red before any browser spec runs.
cf. #331, #356.
What happened (found during #356)
Withdrawing
BatchAction.DELETE(#331) revealed two transcriptions of the batch action table:frontend/ui-core/src/testing/wire.fixtures.ts— typedBatchAction; removing the memberwas an instant
tscerror.frontend/app/e2e/_wire.ts— typedreadonly string[]; compiled clean and kept serving"delete".The generated
checks.tsrejected the stale payload inside a hook, which doesn't raise — thequery never resolves, and every gallery spec failed with
element(s) not foundand 20-secondtimeouts reading as a UI regression. All 2590 pytest, both vitest suites, mypy, ruff,
lint-imports,
generate:client:checkandtypecheck:e2ewere green with the broken fixture inplace. Only the browser suite could see it, and it pointed nowhere near the cause.
Values were fixed in #356; the structural gap was not (recorded there under "Found, not fixed").
Why it wasn't fixed inline
Typing
_wire.tsagainst the real roster means importing from@visionset/ui-coreinternals,and the package exports only its root — typing the fixture widens the public API, which is a
decision, not a diff.
Options
@visionset/ui-core/testingsubpath carrying the wire fixtures and their types;_wire.tsimports and narrows. Widens the API deliberately, kills the second roster._wire.tsfrom the same source aswire.fixtures.ts(thegenerated/clientalready sets the precedent: regenerate, never hand-edit).
turns the silent drift into a named red test.
Whichever lands, the acceptance test is the same: re-adding
"delete"to_wire.tsalone mustturn a named check red before any browser spec runs.
cf. #331, #356.