Files
ClawdDotNet/tests/ClawdDotNet.Core.Tests/Engine/DelegationDepthTests.cs
T
RichardandClaude Opus 5.5 77062facc9 Stufe 4c (#764): Bezeichner auf Englisch - Typen, Methoden, Felder, Parameter, Locals, Testnamen
Mit Roslyn umbenannt (Verweise folgen, Zeichenketten bleiben unberuehrt).
Nicht umbenannt, weil gespeichert oder ueber die API gebunden (Stufe 4e):
Properties, Record-Parameter, Enum-Member sowie die Parameter der
Endpunkt-Handler (Routen-/Query-Bindung ueber den Namen). Die JSON-Felder
anonymer Objekte bleiben ausdruecklich benannt (JsonFilter: pfad).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-04 13:18:43 +02:00

308 lines
13 KiB
C#

using ClawdDotNet.Core.Engine;
using ClawdDotNet.Core.Tests.Infrastructure;
using ClawdDotNet.Core.Tools;
using Shouldly;
namespace ClawdDotNet.Core.Tests.Engine;
/// <summary>
/// Test A13 of the test strategy — regression tests for B8 (#206).
///
/// <para><c>send_message</c> is a synchronous call: A holds its gate while it
/// waits for B. If B then called A, B waited for a gate that A only releases once B
/// has answered. Both stood still until the timeout of the outer run resolved it.</para>
///
/// <para>The tests check the three promises from <see cref="DelegationChain"/>: no
/// agent twice in a chain, at most <c>MaxDelegationDepth</c> hand-offs, and
/// the chain belongs to the call instead of to the agent.</para>
///
/// <para>A time limit is deliberately on every test that could deadlock: without the
/// check it would not turn red, but not continue at all.</para>
/// </summary>
public sealed class DelegationDepthTests
{
private static readonly TimeSpan Deadline = TimeSpan.FromSeconds(20);
/// <summary>Waits for the run, but fails instead of hanging.</summary>
private static async Task<AgentRunResult> WithDeadline(Task<AgentRunResult> run)
{
var done = await Task.WhenAny(run, Task.Delay(Deadline));
done.ShouldBe(run, "Der Lauf ist nicht zurückgekommen — das ist die Verklemmung aus B8.");
return await run;
}
// ═══════════════════════════════════════════════════════════
// A13 — the reported case
// ═══════════════════════════════════════════════════════════
[Fact]
public async Task A_to_B_to_A_is_aborted_instead_of_deadlocking()
{
var tool = new DelegatingTool(new Dictionary<string, string>
{
["agent-a"] = "agent-b",
["agent-b"] = "agent-a" // back to A — the cycle
});
var fixture = new EngineFixture().WithTool(tool);
var a = fixture.AddAgent("agent-a", "TestTool");
fixture.AddAgent("agent-b", "TestTool");
fixture.Client.RespondsContextually();
var result = await WithDeadline(
fixture.Engine.ChatAsync(a, "Bitte klären", "test-instance", default));
result.Status.ShouldBe(AgentRunStatus.Completed);
// The way back B → A was rejected, the way there A → B was not.
var wayBack = tool.Handoffs.Single(h => h is { From: "agent-b", To: "agent-a" });
wayBack.Delivered.ShouldBeFalse();
wayBack.Error.ShouldNotBeNull().ShouldContain("Zyklus");
wayBack.Error.ShouldNotBeNull().ShouldContain("agent-a -> agent-b -> agent-a");
tool.Handoffs.Single(h => h is { From: "agent-a", To: "agent-b" }).Delivered.ShouldBeTrue();
}
[Fact]
public async Task The_rejection_comes_back_as_a_result_not_as_a_throw()
{
// The rejected agent should be able to continue working — it gets a reason,
// not an abort. Otherwise the cycle check would only lead to a different failure.
var tool = new DelegatingTool(new Dictionary<string, string>
{
["agent-a"] = "agent-b",
["agent-b"] = "agent-a"
});
var fixture = new EngineFixture().WithTool(tool);
var a = fixture.AddAgent("agent-a", "TestTool");
fixture.AddAgent("agent-b", "TestTool");
fixture.Client.RespondsContextually();
var result = await WithDeadline(
fixture.Engine.ChatAsync(a, "Bitte klären", "test-instance", default));
result.Status.ShouldBe(AgentRunStatus.Completed);
result.FinalMessage.ShouldBe("Fertig");
}
// ═══════════════════════════════════════════════════════════
// The relay that does not form a cycle
// ═══════════════════════════════════════════════════════════
[Fact]
public async Task A_relay_ends_at_the_depth_limit()
{
// A → B → C → D would be depth 3. With MaxDelegationDepth = 2 it stops at C.
var tool = new DelegatingTool(new Dictionary<string, string>
{
["agent-a"] = "agent-b",
["agent-b"] = "agent-c",
["agent-c"] = "agent-d"
});
var fixture = new EngineFixture().WithTool(tool);
fixture.Engine.MaxDelegationDepth = 2;
var a = fixture.AddAgent("agent-a", "TestTool");
fixture.AddAgent("agent-b", "TestTool");
fixture.AddAgent("agent-c", "TestTool");
fixture.AddAgent("agent-d", "TestTool");
fixture.Client.RespondsContextually();
await WithDeadline(fixture.Engine.ChatAsync(a, "Bitte klären", "test-instance", default));
tool.Handoffs.Single(h => h.To == "agent-b").Delivered.ShouldBeTrue();
tool.Handoffs.Single(h => h.To == "agent-c").Delivered.ShouldBeTrue();
var tooDeep = tool.Handoffs.Single(h => h.To == "agent-d");
tooDeep.Delivered.ShouldBeFalse();
tooDeep.Error.ShouldNotBeNull().ShouldContain("Delegationstiefe");
}
[Fact]
public async Task Within_the_limit_the_relay_goes_through()
{
var tool = new DelegatingTool(new Dictionary<string, string>
{
["agent-a"] = "agent-b",
["agent-b"] = "agent-c"
});
var fixture = new EngineFixture().WithTool(tool);
fixture.Engine.MaxDelegationDepth = 2;
var a = fixture.AddAgent("agent-a", "TestTool");
fixture.AddAgent("agent-b", "TestTool");
fixture.AddAgent("agent-c", "TestTool");
fixture.Client.RespondsContextually();
var result = await WithDeadline(
fixture.Engine.ChatAsync(a, "Bitte klären", "test-instance", default));
result.Status.ShouldBe(AgentRunStatus.Completed);
tool.Handoffs.ShouldAllBe(h => h.Delivered);
}
// ═══════════════════════════════════════════════════════════
// The chain belongs to the call, not to the agent
// ═══════════════════════════════════════════════════════════
[Fact]
public async Task Two_consecutive_runs_do_not_inherit_each_others_chain()
{
// Otherwise an agent could be permanently "used up" after a deep chain.
var tool = new DelegatingTool(new Dictionary<string, string> { ["agent-a"] = "agent-b" });
var fixture = new EngineFixture().WithTool(tool);
var a = fixture.AddAgent("agent-a", "TestTool");
fixture.AddAgent("agent-b", "TestTool");
fixture.Client.RespondsContextually();
await WithDeadline(fixture.Engine.ChatAsync(a, "Erste Frage", "test-instance", default));
await WithDeadline(fixture.Engine.ChatAsync(a, "Zweite Frage", "test-instance", default));
tool.Handoffs.Count(h => h.To == "agent-b").ShouldBe(2);
tool.Handoffs.ShouldAllBe(h => h.Delivered);
}
// ═══════════════════════════════════════════════════════════
// spawn counts too
// ═══════════════════════════════════════════════════════════
[Fact]
public async Task The_chain_cannot_be_bypassed_via_spawn_either()
{
var tool = new DelegatingTool(new Dictionary<string, string>
{
["agent-a"] = "agent-b",
["agent-b"] = "agent-a"
}, spawn: true);
var fixture = new EngineFixture().WithTool(tool);
var a = fixture.AddAgent("agent-a", "TestTool");
fixture.AddAgent("agent-b", "TestTool");
fixture.Client.RespondsContextually();
await WithDeadline(fixture.Engine.ChatAsync(a, "Bitte klären", "test-instance", default));
var wayBack = tool.Handoffs.Single(h => h is { From: "agent-b", To: "agent-a" });
wayBack.Delivered.ShouldBeFalse();
wayBack.Error.ShouldNotBeNull().ShouldContain("Zyklus");
}
// ═══════════════════════════════════════════════════════════
// #461 — what the chain fundamentally cannot see
// ═══════════════════════════════════════════════════════════
/// <summary>
/// Two independent chains crossing over: run 1 starts at A and commissions B, run 2
/// starts at the same time at B and commissions A. Neither chain contains the
/// other's agent — both are cycle-free on their own, and yet run 1 holds
/// gate(A) and waits for gate(B), while run 2 holds gate(B) and waits for gate(A).
///
/// <para>The delegation chain fundamentally cannot recognize that; only a deadline
/// resolves it. A delegated run therefore waits at most
/// <c>DelegationGateTimeout</c> for a foreign gate and refuses afterwards.</para>
///
/// <para>The pause before the hand-off is the core of the setup: the fake client
/// answers synchronously; without it the first chat would run through completely before the
/// second begins — the chains could not overlap at all, and the test
/// would be green without showing anything.</para>
/// </summary>
[Fact]
public async Task Two_independent_crossing_chains_do_not_deadlock()
{
var tool = new DelegatingTool(new Dictionary<string, string>
{
["agent-a"] = "agent-b",
["agent-b"] = "agent-a"
}, delayBeforeHandoff: TimeSpan.FromMilliseconds(300));
var fixture = new EngineFixture().WithTool(tool);
fixture.Engine.DelegationGateTimeout = TimeSpan.FromSeconds(1);
var a = fixture.AddAgent("agent-a", "TestTool");
var b = fixture.AddAgent("agent-b", "TestTool");
fixture.Client.RespondsContextually();
var runA = fixture.Engine.ChatAsync(a, "A fragt B", "test-instance", default);
var runB = fixture.Engine.ChatAsync(b, "B fragt A", "test-instance", default);
var both = Task.WhenAll(runA, runB);
var done = await Task.WhenAny(both, Task.Delay(Deadline));
done.ShouldBe(both, "Beide Laeufe warten auf das Gate des jeweils anderen (#461).");
// Both outer runs come to an end. At least one hand-off failed at the
// deadline — which one is decided by the order and is not
// fixed; that both runs come back is.
(await runA).Status.ShouldBe(AgentRunStatus.Completed);
(await runB).Status.ShouldBe(AgentRunStatus.Completed);
tool.Handoffs.ShouldContain(h => !h.Delivered && h.Error!.Contains("belegt"));
}
[Fact]
public async Task A_chat_from_a_human_keeps_waiting_as_long_as_it_takes()
{
// The counter-check to the deadline: it only applies to delegated runs. A second
// chat of the same agent queues up and is served, even if the first one
// takes longer than the delegation deadline.
var fixture = new EngineFixture().WithTool(FakeTool.Slow(TimeSpan.FromMilliseconds(600)));
fixture.Engine.DelegationGateTimeout = TimeSpan.FromMilliseconds(50);
var a = fixture.AddAgent("agent-a", "TestTool");
fixture.Client.RespondsContextually();
var firstOne = fixture.Engine.ChatAsync(a, "Erste Frage", "test-instance", default);
var secondOne = fixture.Engine.ChatAsync(a, "Zweite Frage", "test-instance", default);
(await WithDeadline(firstOne)).Status.ShouldBe(AgentRunStatus.Completed);
(await WithDeadline(secondOne)).Status.ShouldBe(AgentRunStatus.Completed);
}
// ═══════════════════════════════════════════════════════════
// The chain itself
// ═══════════════════════════════════════════════════════════
[Fact]
public void The_chain_counts_handoffs_not_agents()
{
var chain = DelegationChain.Root("a");
chain.Depth.ShouldBe(0);
chain.Append("b").Depth.ShouldBe(1);
chain.Append("b").Append("c").Depth.ShouldBe(2);
}
[Fact]
public void Append_does_not_change_the_callers_chain()
{
// Immutable: two hand-offs from the same run must not see
// each other — otherwise the second would count the first as a predecessor.
var chain = DelegationChain.Root("a");
var toB = chain.Append("b");
var toC = chain.Append("c");
chain.Agents.ShouldBe(["a"]);
toB.Agents.ShouldBe(["a", "b"]);
toC.Agents.ShouldBe(["a", "c"]);
toB.Contains("c").ShouldBeFalse();
}
[Fact]
public void The_chain_reads_as_a_path()
{
DelegationChain.Root("a").Append("b").Append("c").ToString()
.ShouldBe("a -> b -> c");
}
}