festbench-web/tests/Feature/ServerConfigAndSummaryTest.php

<?php

use App\Models\BenchmarkRun;
use App\Models\DepthsweepResult;
use App\Models\ServerConfig;
use App\Models\SoftbenchResult;
use App\Models\User;
use Illuminate\Support\Facades\Artisan;
use Illuminate\Support\Facades\DB;

test('engine and server config split otherwise identical session bands', function () {
    $owner = User::factory()->create();
    $first = ServerConfig::create([
        'user_id' => $owner->id,
        'name' => 'DFlash Q4 draft',
        'engine' => 'llama.cpp',
        'config_yaml' => '--spec-draft-hf z-lab/Qwen-DFlash:Q4_K_M',
    ]);
    $second = ServerConfig::create([
        'user_id' => $owner->id,
        'name' => 'No draft',
        'engine' => 'llama.cpp',
        'config_yaml' => '--no-speculative',
    ]);

    foreach ([$first, $second] as $config) {
        $run = BenchmarkRun::factory()->create([
            'suite' => 'softbench',
            'user_id' => $owner->id,
            'label' => 'same test label',
            'model_id' => 'same/model',
            'model_quant' => 'Q5_K_M',
            'gpu' => 'AMD Radeon AI PRO R9700',
            'gpu_count' => 1,
            'engine' => 'llama.cpp',
            'config_id' => $config->id,
        ]);
        SoftbenchResult::factory()->create(['benchmark_run_id' => $run->id]);
    }

    Artisan::call('festbench:summary');

    expect(DB::table('session_summaries')->count())->toBe(2)
        ->and(DB::table('session_summaries')->pluck('engine')->unique()->all())->toBe(['llama.cpp'])
        ->and(DB::table('session_summaries')->pluck('config_name')->sort()->values()->all())
        ->toBe(['DFlash Q4 draft', 'No draft']);
});

test('depthsweep summary expands one run into a row for each thinking level', function () {
    $run = BenchmarkRun::factory()->create(['suite' => 'depthsweep', 'visibility' => 'public']);

    foreach (['off', 'low', 'medium', 'xhigh'] as $level) {
        DepthsweepResult::factory()->count(2)->create([
            'benchmark_run_id' => $run->id,
            'level' => $level,
            'total' => null,
            'user_tps' => null,
        ]);
    }

    Artisan::call('festbench:summary');

    $rows = DB::table('session_suite_rows')->where('suite', 'depthsweep')->orderBy('sequence')->get();

    expect($rows)->toHaveCount(4)
        ->and($rows->pluck('level')->all())->toBe(['off', 'low', 'medium', 'xhigh']);
    expect($rows->pluck('avg_user_tps')->filter())->toBeEmpty();
});

test('depthsweep run detail honors a thinking level query filter', function () {
    $run = BenchmarkRun::factory()->create(['suite' => 'depthsweep', 'visibility' => 'public']);
    DepthsweepResult::factory()->create([
        'benchmark_run_id' => $run->id,
        'level' => 'off',
        'task' => 'off-only-task',
    ]);
    DepthsweepResult::factory()->create([
        'benchmark_run_id' => $run->id,
        'level' => 'xhigh',
        'task' => 'xhigh-only-task',
    ]);

    $this->get(route('runs.show', $run->run_id).'?level=off')
        ->assertOk()
        ->assertSee('Thinking level')
        ->assertSee('off-only-task')
        ->assertDontSee('xhigh-only-task');
});

test('server configs settings page is available to authenticated users', function () {
    $this->actingAs(User::factory()->create())
        ->get(route('settings.server-configs'))
        ->assertOk()
        ->assertSee('Inference Args');
});