feat(admin): pick auto-routing benchmark models with the shared picker - #5144
Conversation
Code Review SummaryStatus: No Issues Found | Recommendation: Merge Executive SummaryIncremental re-review of Files Reviewed (0 files)
Previous Review Summary (commit 8e4d0dd)Current summary above is authoritative. Previous snapshots are kept for context only. Previous review (commit 8e4d0dd)Status: 1 Issue Found | Recommendation: Address before merge Overview
Issue Details (click to expand)SUGGESTION
Files Reviewed (20 files)
Verified: the nullable D1 Reviewed by kimi-k3 · Input: 55.7K · Output: 3K · Cached: 78.4K Review guidance: REVIEW.md from base branch |
Users can select benchmark classifier and manual decider models with the shared model picker. Decider variants come from each model catalog and persist across saves and reloads.
Product managers get consistent model selection for both benchmark tables and variant-aware published routing data. Existing server validation still rejects ineligible or duplicate decider models.
Maintainers get a nullable D1 variant column, legacy effort fallback, canonical variant propagation to the CLI, and publish fidelity for non-enum variants. The benchmark worker must deploy with migration 0009 before or with the web app; its predeploy applies the migration. The services/auto-routing worker gains variant-accurate serving after its deployment, but no rollout order is required because older readers keep today’s default behavior.
Shape decision: the variant picker has no clear item by design. Removing and re-adding a row resets its variant, while a saved catalog key remains visible and editable.
Human steps: before merge, deploy the benchmark worker with migration 0009 before or with the web app. After merge, deploy services/auto-routing to enable variant-accurate serving for non-enum keys.
Visual Changes: Benchmark Config picker card
E2E: bot-e2e —
VERIFICATION PASSED.from the fresh local iOS browser round.