Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.
Model Picker CI Readiness
Cause and Scope
Master 616aa5df5 failed UI job 101333936578 in lint run 33976469074:
the picker title was present while the model request was still loading, but the
test immediately queried a group that only exists after that request completes.
The loading DOM contained a disabled search field and no candidate groups.
The same test file had three instances of this class: waiting for the static
title, an unchanged confirmation label, or a search field that exists while
disabled. They originated in #1839 (77c1fd8a4), not the unrelated #1890 change
whose master run exposed the race.
The initial scope was the picker test and this record. During full local UI
validation, the unchanged catalog test also failed at its title-to-row wait.
Its loading DOM showed the same class: title present, no rows, disabled inputs.
The orchestrator inspected the catalog consumer and all its asynchronous read
boundaries before extending scope to BackendModelCatalogDialog.test.tsx.
Product behavior, workflow gates, timeouts, dependencies, and test selection
remain unchanged.
A controlled 250 ms delay on the catalog and candidate reads reproduced twelve catalog failures: one title-to-row assertion, nine clicks on the still-disabled Add models button, and two attempts to type into still-disabled search inputs. The complete class is addressed with a local enabled-button query and explicit row/input readiness waits, not test reruns or implicit synchronization in render. Error and legacy-read tests retain their own readiness boundaries.
Invariant
Assertions and interactions that consume candidate data must wait for its observable readiness, not for the dialog shell. The group test awaits its real group, the selection test awaits its candidate, and the search test awaits an enabled input. All original assertions remain.
A deferred-response regression holds the real component in loading state, checks that the title and confirmation label already exist while search is disabled, then completes the request explicitly. It verifies the actual group and enabled search after completion. It uses no sleeps, retries, or fake timers.
Evidence
- Before editing, the original eight tests passed with immediate responses.
- An isolated in-memory mock with a 250 ms response delay reproduced exactly the three premature waits; the group failure matched the archived CI failure.
- Under that same controlled delay, readiness waits made all eight pass.
- The final picker suite passed all nine cases normally and with that delay.
- The catalog suite passed all 35 cases with controlled delayed reads, after the original source failed twelve of its 34 cases under identical stimulus.
- Both focused suites passed 44 cases. Full UI validation passed all 287 files and 3816 cases; test typecheck, theme validation, and production build passed.
- An AST assertion inventory retained all 207 original assertions across 39
test definitions (42 expanded cases). The locked-row query alone changes from
synchronous
getByTextto awaitedfindByText; no expected value changes. - A lint run concurrent with full UI tests observed the lint-integrity test's
deliberately invalid temporary source. That test removes it in
finally. Lint must run after that suite, as CI already does, not concurrently with it. - Exact-head review and CI results are recorded in the PR.
The controlled delay is diagnostic stimulus, not a production or committed-test sleep. No model API calls or live user state are used. This fixes test ordering; it does not claim a production model-loading defect or a CI speedup.