feat: show Terminal Bench model details

[lambertjosh]

Summary

• Carry optional Terminal Bench 2.0 completion and per-attempt cost metadata through the Kilo Gateway model catalog and generated SDK model type.
• Show the benchmark rows as a separate group in VS Code model previews, the CLI wide model picker, and the CLI session sidebar for the currently selected prompt model.
• Omit malformed or missing optional benchmark metadata so display enrichment cannot block model selection.

Rollout

• Companion Cloud catalog enrichment: feat(ai-gateway): expose Terminal Bench model stats cloud#3703
• Prefer deploying the Cloud PR first, but either order is safe because the field is optional and older clients ignore it.

Notes

• Benchmark cost stays separate from token pricing because it represents one complete Terminal Bench attempt.
• Models without a publishable result omit the entire benchmark group.

Validation

CLI TUI

VSCode

[kilo-code-bot]

Code Review Summary

Status: 1 Issue Carried Forward | Recommendation: Minor suggestion, safe to merge

Overview

Severity	Count
CRITICAL	0
WARNING	0
SUGGESTION	1

Issue Details (click to expand)

SUGGESTION (carried forward — unchanged file)

File	Line	Issue
packages/kilo-vscode/webview-ui/src/components/shared/ModelPreview.tsx	149–158	bench()! non-null assertions inside unkeyed <Show> — inconsistent with model-info-panel.tsx which uses the keyed accessor form

Incremental Change Notes (commit 210f587)

The second commit ("test: focus Terminal Bench coverage") reorganises tests:

• kilo-gateway/test/api/models.test.ts: The terminalBench undefined assertion was moved from the generic success test into a new dedicated test ("omits malformed Terminal Bench metadata without rejecting the catalog"). Coverage is equivalent or better.

• model-preview-data-line.test.ts: Removed the describe("model preview Terminal Bench group") block, which tested source code structure via string-matching (checking for specific variable names and JSX strings in ModelPreview.tsx). This is the right call — those tests were asserting implementation details rather than behavior.

• model-price-format.test.ts / model-info-panel-utils.test.ts: Removed unit tests for fmtTerminalBenchScore, fmtScore, and fmtAttemptCost. These are simple one-liner formatters ((n * 100).toFixed(1) + "%" and "$" + n.toFixed(2)) and are exercised indirectly by the gateway integration tests, so the risk is low. The commit message says "focus Terminal Bench coverage" but the refocus is on integration-level tests rather than unit-level formatting tests — a reasonable tradeoff for functions this simple.

Other Observations

Keyed vs unkeyed <Show> pattern inconsistency (unchanged from previous review):

In ModelPreview.tsx, the Terminal Bench section uses <Show when={bench()}> and accesses bench()!.overallScore / bench()!.avgAttemptCostUsd with non-null assertions:

<Show when={bench()}>
  ...
  <span>{fmtTerminalBenchScore(bench()!.overallScore)}</span>
  <span>{fmtAttemptCost(bench()!.avgAttemptCostUsd)}</span>
</Show>

The CLI equivalent in model-info-panel.tsx uses the keyed accessor form, which is idiomatic SolidJS:

<Show when={m().terminalBench}>
  {(bench) => (
    ...
    <text>{fmtScore(bench().overallScore)}</text>
    ...
  )}
</Show>

This is not a bug — Show gates the render so bench()! is always non-null when the content renders. But the keyed form is preferred and the existing cost / caps sections in the same component use it consistently. Consider:

<Show when={bench()}>
  {(b) => (
    <div class="model-preview-group">
      ...
      <span class="model-preview-value">{fmtTerminalBenchScore(b().overallScore)}</span>
      ...
      <span class="model-preview-value">{fmtAttemptCost(b().avgAttemptCostUsd)}</span>
    </div>
  )}
</Show>

Files Reviewed

First commit (b0a4f03) — full review:

• .changeset/brisk-terminal-bench-details.md
• packages/kilo-gateway/src/api/models.ts
• packages/kilo-gateway/test/api/models.test.ts
• packages/kilo-vscode/tests/unit/model-preview-data-line.test.ts
• packages/kilo-vscode/tests/unit/model-price-format.test.ts
• packages/kilo-vscode/webview-ui/src/components/shared/ModelPreview.tsx
• packages/kilo-vscode/webview-ui/src/components/shared/model-preview-utils.ts
• packages/kilo-vscode/webview-ui/src/styles/model-selector.css
• packages/kilo-vscode/webview-ui/src/types/messages/providers.ts
• packages/opencode/src/kilocode/components/model-info-panel-utils.ts
• packages/opencode/src/kilocode/components/model-info-panel.tsx
• packages/opencode/src/kilocode/plugins/sidebar-usage.tsx
• packages/opencode/src/kilocode/provider/provider.ts
• packages/opencode/test/kilocode/model-info-panel-utils.test.ts
• packages/sdk/js/src/v2/gen/types.gen.ts
• packages/sdk/openapi.json
• i18n files (17 languages)

Second commit (210f587) — incremental review:

• packages/kilo-gateway/test/api/models.test.ts — test reorganisation, no new issues
• packages/kilo-vscode/tests/unit/model-preview-data-line.test.ts — string-match tests removed, no new issues
• packages/kilo-vscode/tests/unit/model-price-format.test.ts — formatting tests removed, no new issues
• packages/opencode/test/kilocode/model-info-panel-utils.test.ts — formatting tests removed, no new issues

Fix these issues in Kilo Cloud

---

_{Reviewed by claude-4.6-sonnet-20260217 · 415,059 tokens}

_{Review guidance: REVIEW.md from base branch main}