[oscarlius]

Decode, Prefill, Quality and Tool Eval workflows now localize configuration, progress, results, history, comparisons and generated share cards. Stable option and metric IDs remain separate from labels; model names, commands, scenario data and raw output are preserved.

Scope: benchmarks, tooleval, toolMessages, shareCards namespaces. This is one of six complete-feature migrations reusing the reviewed wording/design from #194, implemented on current main (b96244a). #194 remains unchanged as the historical reference. Settings #310 is now on main via the maintainer's author-preserving port; #315–#319 remain separate reviews.

Validation:

• Combined candidate 2a16e5a: 800/800 frontend tests, typecheck, catalog checks and production build passed. Follow-up 48dbf89 passed the affected static-key/raw-response regressions and type/catalog/build checks.
• Extracted implementation 10c93d8: 2 targeted frontend regressions plus typecheck/catalog validation passed.
• Linux server gap closed: npm run test:server passed 1,433/1,433, 0 failures, 0 skips, on the combined 48dbf89 source in an isolated Debian Bookworm / Node 24.19.0 container. Its server tree is identical to feat(i18n): add bounded server message metadata and rendering #325; the other five feature reviews do not modify the server. The earlier Windows npm test attempt remains incomplete.
• Browser acceptance: Tool Eval setup and settled command preview; custom URL/model drafts and raw command were retained, with no additional preview request after switching language. Checked in English and Simplified Chinese at 1440 x 1000 and 390 x 1000 on Chrome 156, using real components/styles with local mocked responses. Across eight representative surfaces (32 captures), no page-level horizontal overflow or browser exceptions were observed; pending form values were preserved. Token tables retain intentional internal scrolling.
• Remaining boundary: representative pages, not every dialog, theme, error state or real-device workflow. No real connection test, benchmark, model launch, production traffic, merge or deployment.

Integration: merge the loader resources/catalogs and documentation additively, retaining every already-merged namespace and older pending review. The server-message group should be integrated after fleet reports and Spark details because it adds metadata rendering to the same EventRow/SparkActions surfaces; its HealthFindings integration also overlaps #316. PR creation does not authorize merging.

Progress: Unified localization index. Settings #310 (completed on main); Shell #315; Overview #316; resource panels #317; LLM live/sessions #318; LLM history/totals #319; fleet reports/activity/kiosk #320; device setup #321; remaining Spark panels/model operations #322; benchmarks/share cards #323; Showcase #324; server messages #325. All remain separate reviews. #194 stays open as the historical reference. Source is included in the private local showcase; no upstream PR was merged.

Maintainer feedback follow-up (2026-11-01): author review.

c356923 restores the pre-migration STOPPED label for stopped Tool Eval runs, with a separate localized label. CANCELLED remains reserved for cancelled benchmark runs; wire status values are unchanged. The new regression failed before the fix; 47 focused share-card/locale tests, typecheck, catalog validation and production build passed.

Current PR head: c356923. Unified progress and feedback record. Previous no-feedback/CI snapshot is superseded for review status; local test evidence is separate from GitHub CI. No PR merge, rebase or .9 deployment was performed by this follow-up.

Current-main maintainer follow-up (2026-11-03)

Addressed latest maintainer feedback at f86a14c, rebased onto main 7c2b963. Keep landing order #315, then #316–#325 in numeric order; separate PRs still target main and shared catalogs/loaders must be combined additively.

Preserved context-based Prefill limits/custom sizes and localized their new feedback. STOPPED remains distinct from CANCELLED. Share-card user strings remain Canvas fillText input, with no HTML sink.

Local validation completed before this push/PR update: per-PR catalog/type checks and affected regressions; final integrated source cf08c32 passed 888 frontend tests (136 files), typecheck, catalog/checker tests and production build. Linux server validation passed 1,463 tests, 0 failures/skips on 3f61343, whose server tree is identical to cf08c32. Fresh Overview/Kiosk acceptance at 1440×1000 and 390×1000 in English/zh-CN found no page overflow or browser errors. Actual Canvas exports were checked at both viewport widths (2400×1350 backing image), with markup-like raw user strings remaining literal text. These are local checks, not GitHub CI: audited heads had no check runs/statuses or PR Actions runs.

The local showcase includes this submitted source and retains native Maya monitoring. Unified current progress and evidence. No upstream PR merge was performed; #194 remains untouched as the maintainer-requested reference.

[MiaAI-Lab]

First review, at b85d6428 against main (5e7be05): clean. A disciplined label migration — request payloads, validation rules, row/history computations, run cancellation and the share-card painting pipeline are unchanged, the canvas draws every localized string through the same narrow context with no new HTML sink, and keys are complete across all four namespaces' catalogs. One deliberate English-visible change to confirm: the tool-eval share-card status label for a stopped run changes from "STOPPED" to "CANCELLED" — intended as part of the unified status vocabulary, or should the old word stay? Please confirm and this is next in line after the open page slices. Leaving it open.

[MiaAI-Lab]

Benchmark workflows and share cards (largest diff of the chain). Share cards render user data into images — confirm the localized templates still escape user strings the same way, and that the card canvas layout is verified at both widths as described. Chain before 324.

(Review notes prepared with Claude Code on behalf of the maintainer.)