Repository navigation
fix(frontend): Finish classic, evaluations and data screens after a full audit - #8342
ashrafchowdury wants to merge 2 commits into
Conversation
…ull audit Two audit passes over /m found layout breaks, wrong data and unfinished copy. This fixes them on top of the release branch, with the copy in the i18n catalogs. - Layout: runs list, run overview and narrow pages no longer scroll sideways; the test set grid stops growing; trace drawer tabs no longer overlap. - Data: the test sets list shows the latest revision's message; an evaluator's Evaluations page lists its runs; Registry Compare diffs the committed revision; a new prompt in a folder is created in that folder; trace times are local. - Flows: leave guards push the route without the /m basePath; inline validation in the new evaluation sheet; one trace delete dialog that closes the drawer. - Copy: sentence case and product nouns across settings, test sets, queues, observability and the playground, with translations for every new key.
|
Important Review skippedToo many files! This PR contains 457 files, which is 157 over the limit of 300. To get a review, reduce the PR to 300 files or fewer by splitting it into smaller PRs or changing its base branch. Usage-priced reviews support at most 300 files.
|
Website previewPreview URL: https://pr-8342-agenta-website-preview.mahmoud-637.workers.dev Built from |
| message={`${testset.name} has no ${missing.length === 1 ? "column" : "columns"} for ${list}`} | ||
| description="Every row will fail with missing inputs. Pick another test set, or add the column to this one." |
There was a problem hiding this comment.
🟡 New evaluation warning remains English
With a non-English locale, TestsetColumnsWarning still displays English text. Its warning bypasses @agenta/i18n and assembles a sentence from fragments, contrary to the mobile translation rules.
Learn more
The mobile translation rules require user-visible text to come from @agenta/i18n. They also require complete sentences with placeholders instead of sentences assembled from fragments. Neither warning string uses a catalog.
Example: A user selects a non-English language and picks a test set without a required question column. The new warning heading and explanation remain English.
Recommended fix: Add full-sentence catalog entries for the singular warning, plural warning, and explanation. Render them with useT, using placeholders for the test set name and missing-column list. Add translations for the supported languages.
Was this helpful? React with 👍 or 👎 to provide feedback.
There was a problem hiding this comment.
Not changing this here. The whole new evaluation sheet (web/mobile/src/features/evaluations/new/) has not been moved to @agenta/i18n yet on this release branch; none of its files use useT. Translating one warning would leave it the only translated string on an English-only screen. It belongs with the migration of that screen.
- The evaluation revision picker offers a configured v0; useAppRevisions already drops empty seeds. - Filter Clear leaves an empty draft, so Apply can remove every filter. - Keep FilterDialog's onClearFilter as an optional unused prop; web/oss still passes it. - Update tests for the new copy, the SharedEditor-based JSON view and the ResizeObserver in the trace drawer.
Railway Preview Environment
Dependent tests finished. Post /preview for manual review. Environment absence verified; retained storage requires separate provider accounting. Post |
Context
Two audit passes over
/mfound about 100 problems in the classic screens (prompts, playground, registry), evaluations, test sets, queues, observability and settings. The worst ones:The rest were unfinished-looking details: Title Case labels, "testset" spelling and toast-only validation.
Changes
Layout
ListTable's scroll box is nowrelative, so its absolutely positioned sr-only header cells stay inside it and no longer widen the window. The chart's sr-only<table>sits in adiv.sr-onlyfor the same reason.flex-1 min-h-0down the chain). Before, the grid measured its own content and grew about 1.5px a second.VirtualTable: a column outside every header group spans the header rows, instead of leaving an empty cell above its title.Data
versionper test set./testsets/revisions/queryignores the per-ref limit and returns revisions in no order, so the list used to show a stale commit message./apps/<evaluatorId>/evaluations) the runs list filters byevaluator, notapplication.ArtifactCreatereads it:serverData, the committed revision, not the merged draft.metricDisplayLabel).Flows
TraceDrawer. It closes the drawer when the open trace is deleted./mbasePath; before, leaving opened/m/m/...and a 404.Copy
@agenta/i18nEnglish catalogs, and every new key has translations in all six languages (pnpm i18n:checkpasses).Not in this PR:
humanRunConfig.ts). The backend needs to accept a version in the run references first.palette.tsstill lives inweb/oss.Tests
web/mobileand the 12 changed packages. Lint is clean, andpnpm i18n:checkpasses.testsetsListDeveloperModepasses underTZ=UTC, the CI time zone; it formats a fixed time in local time.createPrompt"New prompt inside a folder…"testset-latest-revisionslistTableSrOnlyContainedevaluationRunsList(the evaluator scope)filterDialog.renderunsavedGuardappNavScopeSwitchvariantsData(the empty v0 seed)registryCompare.render/min production mode and checked the pages below in Chrome. At 400px wide, the runs list, run overview, observability, settings and app overview show no sideways scroll.Demo
Before: the app at the start of the audit. After: a production build of
/mwith these changes on the same local data.Evaluation runs list on the local EE stack, desktop width.
Test set
qa2-upload, left open for a minute.Run overview at 400px wide.
Settings, API keys tab.
Observability traces of an agent.
_agent.Observability time range menu.
Settings, Secrets tab, with the secrets request failing.
Trace drawer opened from a link to a deleted trace.
Trace drawer of an agent chat span, scrolled down.
Test set, Commit after editing one cell.
Test set grid, hover a long cell.
Playground of a new prompt, Create.
Playground with an edited draft, click Registry, then Confirm in the leave dialog.
Registry, Deployments tab, after deploying a newer revision to staging.
New evaluation, pick a test set that has no column for a prompt input.
Evaluation runs, empty project.
Run details, Configuration tab.
Run details, Overview tab of a run with one evaluator.
Annotation queues list, desktop width.
What to QA