Program.cs

using System.Text;
using Microsoft.AspNetCore.Http.Features;
using TinyNotebook.Core.Config;
using TinyNotebook.Core.Models;
using TinyNotebook.Core.Prompts;
using TinyNotebook.Core.Services;
using TinyNotebook.Core.Utilities;

namespace TinyNotebook;

public static class Program
{
    public static async Task Main(string[] args)
    {
        // Ensure Chinese characters in ASP.NET Core logs render correctly on Windows.
        Console.OutputEncoding = Encoding.UTF8;
        Console.InputEncoding  = Encoding.UTF8;
        var storagePaths = new RuntimeStoragePaths();
        storagePaths.EnsureCreated();

        var builder = WebApplication.CreateBuilder(args);
        builder.Configuration.AddJsonFile(Path.Combine(AppContext.BaseDirectory, "appsettings.web.json"), optional: true, reloadOnChange: true);

        builder.Services.Configure<LlmOptions>(builder.Configuration.GetSection(LlmOptions.SectionName));
        builder.Services.Configure<ModelDownloadOptions>(builder.Configuration.GetSection(ModelDownloadOptions.SectionName));
        builder.Services.AddSingleton<ModelPathResolver>();
        builder.Services.AddSingleton<RuntimeStoragePaths>();
        builder.Services.AddSingleton<PromptBuilder>();
        builder.Services.AddSingleton<IConversationStore, FileConversationStore>();
        builder.Services.AddSingleton<IModelDownloader, HuggingFaceModelDownloader>();
        builder.Services.AddSingleton<SessionManager>();
        builder.Services.AddSingleton<ILlamaCppBackend, LlamaCppBackend>();
        builder.Services.AddSingleton<ILlmService, LlmService>();
        builder.Services.AddHostedService<LlmWarmupHostedService>();
        builder.Services.AddRazorPages();

        var app = builder.Build();
        app.Services.GetRequiredService<RuntimeStoragePaths>().EnsureCreated();

        if (!app.Environment.IsDevelopment())
        {
            app.UseExceptionHandler("/Error");
            app.UseHsts();
        }

        app.UseHttpsRedirection();
        app.UseRouting();
        app.UseAuthorization();

        app.MapStaticAssets();
        app.MapRazorPages().WithStaticAssets();

        // ── Streaming endpoints as Minimal API ──────────────────────────────
        // Razor Pages handlers go through the full MVC pipeline which buffers
        // the response.  Minimal API endpoints write directly to the response
        // stream, so every token reaches the browser as soon as it is produced.

        app.MapPost("/api/chat", async (HttpContext ctx, ILlmService llmService) =>
        {
            // Disable all response buffering layers (Kestrel + any middleware).
            ctx.Features.Get<IHttpResponseBodyFeature>()?.DisableBuffering();

            ctx.Response.ContentType = "text/plain; charset=utf-8";
            ctx.Response.Headers.CacheControl = "no-cache";
            ctx.Response.Headers["X-Accel-Buffering"] = "no";

            var form = await ctx.Request.ReadFormAsync(ctx.RequestAborted);
            var message = form["message"].ToString().Trim();
            if (string.IsNullOrWhiteSpace(message))
            {
                ctx.Response.StatusCode = 400;
                await ctx.Response.WriteAsync("Message is required.", ctx.RequestAborted);
                return;
            }

            // Flush the response headers immediately so the browser's fetch()
            // promise resolves right away and the UI can show "connected" status
            // even before the first generated token arrives.
            await ctx.Response.Body.FlushAsync(ctx.RequestAborted);

            try
            {
                await foreach (var token in llmService.StreamAsync(
                    new ChatRequest { UserMessage = message }, ctx.RequestAborted))
                {
                    await ctx.Response.WriteAsync(token, ctx.RequestAborted);
                    await ctx.Response.Body.FlushAsync(ctx.RequestAborted);
                }
            }
            catch (OperationCanceledException) { /* browser aborted */ }
        }).DisableAntiforgery();

        app.Run();
    }
}