ClaudeDo/tests/ClaudeDo.Ui.Tests/ViewModels/DiffDispatcherGuardTests.cs

using System.Collections.Concurrent;
using System.IO;
using System.Text;
using System.Threading;
using ClaudeDo.Ui.Services;
using ClaudeDo.Ui.ViewModels.Modals;
using Xunit;

namespace ClaudeDo.Ui.Tests.ViewModels;

/// <summary>
/// [B3] The one blockade-regression test in the operation-feedback effort. Proves that a large
/// diff no longer runs its parse work inline on the calling ("UI") thread — [B1]'s
/// <see cref="UnifiedDiffParser.Parse"/> offload via <see cref="DiffViewerViewModel"/>.
/// </summary>
/// <remarks>
/// <para>
/// <b>Option B chosen over Option A (see plan, Gruppe B / B3).</b> This test project has no
/// bootstrapped Avalonia dispatcher anywhere — <c>Avalonia.Headless.XUnit</c> / <c>[AvaloniaFact]</c>
/// is not referenced, and every existing test that would otherwise need one (e.g.
/// <c>Services/OperationStatusTests.cs</c>) sidesteps it via a synchronous seam instead.
/// </para>
/// <para>
/// Instead, a custom <see cref="QueueingSyncContext"/> is installed as
/// <see cref="SynchronizationContext.Current"/> on the calling thread before the property write
/// that triggers the parse. [B1] routes its offloaded work through <c>await Task.Run(...)</c>,
/// and the C# compiler only defers an <c>await</c>'s continuation to the ambient
/// <see cref="SynchronizationContext"/> when the awaited task is <i>not yet complete</i> at the
/// await point — which <c>Task.Run</c> guarantees (its delegate never runs inline on the calling
/// thread). So immediately after the triggering property set returns, the parsed result must
/// still be the previous (empty) one — not because of a timing race that could flip either way,
/// but because the continuation that would write it is sitting, unpumped, in the queue. Draining
/// that queue (<see cref="QueueingSyncContext.RunPending"/>) is what a real dispatcher's message
/// loop would do; this proves the calling thread stayed free to keep pumping while the work ran
/// elsewhere, which is the actual regression [B1] guards against. Unlike a bare
/// <c>ManagedThreadId</c> check, this also fails for the right reason if someone reintroduces a
/// <c>.Result</c>/<c>.Wait()</c> block that defeats the offload without technically running on the
/// calling thread's own stack frame.
/// </para>
/// <para>
/// <b>[B2]'s <see cref="DiffAlignment.Build"/> offload (in <see cref="ClaudeDo.Ui.Views.Controls.DiffTextView"/>)
/// is deliberately NOT covered here.</b> That offload only runs inside a real
/// <see cref="ClaudeDo.Ui.Views.Controls.DiffTextView"/> control, and constructing one at all —
/// even off-screen, never attached — requires a bootstrapped Avalonia <c>Application</c> (its
/// AXAML resolves <c>StaticResource</c>/<c>loc:Tr</c> bindings during <c>InitializeComponent</c>).
/// Three escalating attempts to make that safe in-process were each falsified empirically against
/// the FULL <c>ClaudeDo.Ui.Tests</c> suite (passing in isolation is not enough — see the project's
/// own "order-dependent flakiness" lesson): (1) a lazily-bootstrapped headless
/// <c>AppBuilder...UseHeadless()...SetupWithoutStarting()</c> call crashed with
/// <c>Dispatcher.VerifyAccess()</c> ("a different thread owns it") the first time <em>any</em> test
/// in the 500+ suite tried it, because Avalonia's <c>Dispatcher.UIThread</c> binds to whichever
/// thread touches it first and xUnit's worker pool gives no thread guarantee; (2) forcing that
/// bootstrap to run via <c>[ModuleInitializer]</c> (so it wins the race) moved the same failure
/// one level deeper — routine <c>AvaloniaObject.SetValue</c> property-change notifications on the
/// control itself also assert dispatcher-thread affinity, and the two <c>[Fact]</c>s in this class
/// still landed on different xUnit worker threads; (3) marshalling the whole test body onto one
/// dedicated background thread (a hand-rolled stand-in for what <c>Avalonia.Headless.XUnit</c>'s
/// custom test framework does) fixed the affinity crash, but bootstrapping a real, permanent,
/// process-wide <c>Application.Current</c> as a side effect then broke an unrelated, previously
/// green test (<c>WorkerLogLevelToBrushConverterTests</c>'s "no app" fallback case) purely by
/// having run earlier in the same test host process. Adding <c>Avalonia.Headless.XUnit</c> itself
/// (the mechanism that actually solves this, by owning test scheduling) was tried too: it pulls in
/// <c>xunit.v3.*</c>, which collides with this project's <c>xunit 2.9.3</c> — <c>FactAttribute</c>/
/// <c>TheoryAttribute</c>/<c>InlineDataAttribute</c> became ambiguous across ~30 existing test
/// files (CS0433), i.e. it would require migrating the whole test project to xUnit v3, far outside
/// a guard test's scope. [B2] is left to the plan's existing "visuelle Prüfung offen" item; a
/// follow-up to actually cover it would need either that migration or a purpose-built, isolated
/// (e.g. separate test assembly/process) Avalonia UI test harness.
/// </para>
/// </remarks>
public class DiffDispatcherGuardTests
{
    // 2 000 changed line-pairs (4 000 raw diff lines) — long enough that a synchronous
    // UnifiedDiffParser.Parse would be a visibly dropped frame (measured ~15ms on dev hardware for
    // parsing alone), short enough that this test stays well under a second even without offload.
    private const int LinePairCount = 2000;

    private static string BuildLargeDiff(int pairCount)
    {
        var sb = new StringBuilder();
        sb.Append("diff --git a/big.cs b/big.cs\n--- a/big.cs\n+++ b/big.cs\n");
        sb.Append($"@@ -1,{pairCount} +1,{pairCount} @@\n");
        for (var i = 0; i < pairCount; i++)
        {
            sb.Append($"-    var value{i} = ComputeSomething(alpha{i}, beta{i}, gamma{i});\n");
            sb.Append($"+    var value{i} = ComputeSomethingElse(alpha{i}, beta{i}, gamma{i});\n");
        }
        return sb.ToString();
    }

    /// Captures every <see cref="Post"/>ed continuation instead of running it, so a test can
    /// assert on state before it decides to drain the queue. <see cref="Send"/> is intentionally
    /// left unimplemented (base throws) — nothing under test uses synchronous marshalling.
    private sealed class QueueingSyncContext : SynchronizationContext
    {
        private readonly ConcurrentQueue<(SendOrPostCallback Callback, object? State)> _queue = new();

        public override void Post(SendOrPostCallback d, object? state) => _queue.Enqueue((d, state));

        public bool RunPending()
        {
            var ran = false;
            while (_queue.TryDequeue(out var item))
            {
                item.Callback(item.State);
                ran = true;
            }
            return ran;
        }
    }

    private sealed class GuardWorkerClient : StubWorkerClient
    {
        public IReadOnlyList<SubtaskDiffDto> AggregateResult { get; set; } = Array.Empty<SubtaskDiffDto>();
        public override Task<IReadOnlyList<SubtaskDiffDto>> GetPlanningAggregateAsync(string planningTaskId) =>
            Task.FromResult(AggregateResult);
    }

    [Fact]
    public async Task PlanningParse_OfLargeDiff_DoesNotRunOnCallingThread()
    {
        var ctx = new QueueingSyncContext();
        var previous = SynchronizationContext.Current;
        SynchronizationContext.SetSynchronizationContext(ctx);
        try
        {
            var raw = BuildLargeDiff(LinePairCount);
            var fake = new GuardWorkerClient
            {
                AggregateResult = new[] { new SubtaskDiffDto("s1", "First", "b1", "base1", "head1", null, raw) },
            };
            var vm = new DiffViewerViewModel(null!, fake, new AppSettings
            {
                ConfigPath = Path.Combine(Path.GetTempPath(), $"claudedo-uicfg-{Guid.NewGuid():N}.json"),
            });
            vm.ConfigurePlanning("plan-1", "main");

            // LoadAsync's own await chain only ever awaits already-completed tasks
            // (Task.FromResult, then synchronous property setters), so it never suspends and
            // never touches the sync context itself — only the fire-and-forget parse it
            // triggers as a side effect does. Auto-selecting the first (only) subtask sets
            // DisplayedDiff to the large diff, which kicks off that parse.
            await vm.LoadAsync();

            // Deterministic, not a race: the continuation that would populate PlanningFiles is
            // parked in `ctx`, unpumped. This is exactly what must be false pre-[B1], where
            // OnDisplayedDiffChanged parsed and wrote PlanningFiles synchronously and inline.
            Assert.Empty(vm.PlanningFiles);

            var pumped = false;
            var deadline = DateTime.UtcNow.AddSeconds(5);
            while (DateTime.UtcNow < deadline && !pumped)
            {
                pumped = ctx.RunPending();
                if (!pumped) Thread.Sleep(5);
            }

            Assert.True(pumped, "the offloaded parse never posted its completion back");
            var file = Assert.Single(vm.PlanningFiles);
            Assert.Equal("big.cs", file.Path);
        }
        finally
        {
            SynchronizationContext.SetSynchronizationContext(previous);
        }
    }
}