Budgetende, Stufe 2: laufende Laeufe reservieren ihr Budget (#716)

Der offene Rest aus #682: Jeder Lauf pruefte das Budget einzeln gegen den
bisher erfassten Verbrauch. Laufende Laeufe waren unsichtbar, gleichzeitig
startende passierten die Pruefung gemeinsam und ueberzogen gemeinsam.

- BudgetGuard.ReserveAsync prueft und reserviert unter einer Sperre. Die
  Reservierung ist der Mittelwert der letzten 20 Laeufe des Agenten (Laeufe
  ohne Verbrauch zaehlen nicht mit) und gilt, bis der Verbrauch des Laufs
  erfasst ist.
- Gegen die Grenze zaehlt: verbraucht + in laufenden Laeufen reserviert. Die
  eigene Schaetzung eines Laufs zaehlt bewusst nicht mit; ein einzelner Lauf
  wird weiter nur abgelehnt, wenn das Budget schon verbraucht ist. Sonst fiele
  eine kleine Aufgabe wegen eines grossen Durchschnitts aus.
- Gilt fuer Kosten- und Tokengrenze, je Agent und je Instanz.
- Die Ablehnungsmeldung nennt den reservierten Anteil, sonst wirkte "0,60 von
  1,00 USD verbraucht" als Ablehnungsgrund falsch.
- IUsageRepository.GetRecentForAgentAsync, mit leerer Vorgabe im Interface,
  damit bestehende Attrappen unberuehrt bleiben.

Tests: BudgetReservationTests (8 neu), darunter zehn gleichzeitige Starts, von
denen genau vier passieren. 1.246 gruen.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Richard
2026-09-24 12:06:07 +02:00
co-authored by Claude Opus 5.5
parent 5f3a85595b
commit 8a83a67fb4
5 changed files with 359 additions and 15 deletions
@@ -125,6 +125,29 @@ public sealed class SqliteUsageRepository : IUsageRepository
return results;
}
public async Task<IReadOnlyList<RunUsage>> GetRecentForAgentAsync(string agentId, int limit, CancellationToken ct)
{
await using var conn = await _storage.OpenConnectionAsync(ct);
using var cmd = conn.CreateCommand();
cmd.CommandText = """
SELECT Id, AgentId, Model, PromptTokens, CompletionTokens, CachedTokens,
CostUsd, CostIsKnown, Status, StepCount, DurationMs, OccurredAt
FROM RunUsage
WHERE AgentId = @agent COLLATE NOCASE
ORDER BY OccurredAt DESC
LIMIT @limit
""";
cmd.Parameters.AddWithValue("@agent", agentId);
cmd.Parameters.AddWithValue("@limit", Math.Clamp(limit, 1, 1000));
var results = new List<RunUsage>();
await using var reader = await cmd.ExecuteReaderAsync(ct);
while (await reader.ReadAsync(ct))
results.Add(Read(reader));
return results;
}
public Task<int> PurgeBeforeAsync(DateOnly date, CancellationToken ct)
=> _storage.WriteAsync(async conn =>
{
@@ -53,6 +53,14 @@ public interface IUsageRepository
Task<IReadOnlyList<RunUsage>> GetRecentAsync(int limit, CancellationToken ct);
/// <summary>
/// #716: The most recent runs of one agent — the basis for estimating what a run of
/// it will cost. Without an implementation there is no estimate, and a reservation
/// counts nothing.
/// </summary>
Task<IReadOnlyList<RunUsage>> GetRecentForAgentAsync(string agentId, int limit, CancellationToken ct)
=> Task.FromResult<IReadOnlyList<RunUsage>>([]);
/// <summary>Entfernt Einträge älter als das angegebene Datum. Gibt die Anzahl zurück.</summary>
Task<int> PurgeBeforeAsync(DateOnly date, CancellationToken ct);
}
+142 -7
View File
@@ -13,6 +13,13 @@ public enum BudgetLimit
InstanceTokens
}
/// <summary>#716: What a run is expected to use — the agent's average over recent runs.</summary>
public readonly record struct RunEstimate(decimal CostUsd, long Tokens)
{
public static RunEstimate operator +(RunEstimate a, RunEstimate b) => new(a.CostUsd + b.CostUsd, a.Tokens + b.Tokens);
public static RunEstimate operator -(RunEstimate a, RunEstimate b) => new(a.CostUsd - b.CostUsd, a.Tokens - b.Tokens);
}
public sealed record BudgetStatus(bool IsAllowed, BudgetLimit Limit, string? Reason)
{
public static BudgetStatus Allowed { get; } = new(true, BudgetLimit.None, null);
@@ -34,6 +41,19 @@ public sealed class BudgetGuard
private readonly IUsageRepository _usage;
private readonly Func<DateTime> _now;
/// <summary>How many recent runs of an agent make up its per-run estimate (#716).</summary>
public const int EstimateSample = 20;
/// <summary>
/// #716: Check and reserve happen under this lock. Without it, two runs starting at
/// the same moment would both read the same "used" figure and both pass.
/// </summary>
private readonly SemaphoreSlim _reserveGate = new(1, 1);
private readonly Lock _reservedLock = new();
private readonly Dictionary<string, RunEstimate> _reservedByAgent = new(StringComparer.OrdinalIgnoreCase);
private RunEstimate _reservedInstance;
public BudgetGuard(IUsageRepository usage, Func<DateTime>? now = null)
{
_usage = usage;
@@ -43,8 +63,112 @@ public sealed class BudgetGuard
/// <summary>
/// Prüft, ob der Agent heute noch laufen darf.
/// </summary>
public async Task<BudgetStatus> CheckAsync(
public Task<BudgetStatus> CheckAsync(
AgentConfig agent, InstanceBudget instanceBudget, CancellationToken ct)
=> CheckCoreAsync(agent, instanceBudget, default, default, ct);
/// <summary>
/// #716: Checks like <see cref="CheckAsync"/> and, if the run may start, reserves its
/// estimated usage until the returned handle is disposed — i.e. until the run's usage
/// is recorded.
///
/// <para><b>What changes.</b> Before, each run was checked on its own against what was
/// already recorded. Runs in flight were invisible, so several starting together all
/// passed and exceeded the budget together (#682 capped how many; this closes it).
/// Now the runs in flight count with their estimate.</para>
///
/// <para><b>What does not change.</b> A single run is still refused only when the
/// budget is already used up — its own estimate is not added. Otherwise a small task
/// would be refused because the agent's average run is large.</para>
/// </summary>
public async Task<(BudgetStatus Status, IDisposable? Reservation)> ReserveAsync(
AgentConfig agent, InstanceBudget instanceBudget, CancellationToken ct)
{
if (!agent.Budget.HasAnyLimit && !instanceBudget.HasAnyLimit)
return (BudgetStatus.Allowed, null);
var estimate = await EstimateRunAsync(agent.AgentId, ct);
await _reserveGate.WaitAsync(ct);
try
{
RunEstimate agentReserved, instanceReserved;
lock (_reservedLock)
{
agentReserved = _reservedByAgent.GetValueOrDefault(agent.AgentId);
instanceReserved = _reservedInstance;
}
var status = await CheckCoreAsync(agent, instanceBudget, agentReserved, instanceReserved, ct);
if (!status.IsAllowed)
return (status, null);
lock (_reservedLock)
{
_reservedByAgent[agent.AgentId] = _reservedByAgent.GetValueOrDefault(agent.AgentId) + estimate;
_reservedInstance += estimate;
}
return (status, new Reservation(this, agent.AgentId, estimate));
}
finally
{
_reserveGate.Release();
}
}
/// <summary>What is currently reserved by runs in flight — for the agent, or with an
/// empty id for the whole instance.</summary>
public RunEstimate Reserved(string agentId)
{
lock (_reservedLock)
return agentId.Length == 0 ? _reservedInstance : _reservedByAgent.GetValueOrDefault(agentId);
}
/// <summary>
/// The average of the agent's recent runs. Runs without usage (refusals, aborted
/// before the first call) do not count — they would drag the estimate towards zero.
/// </summary>
public async Task<RunEstimate> EstimateRunAsync(string agentId, CancellationToken ct)
{
var recent = await _usage.GetRecentForAgentAsync(agentId, EstimateSample, ct);
var runs = recent.Where(r => r.TotalTokens > 0).ToList();
if (runs.Count == 0)
return default;
return new RunEstimate(
runs.Average(r => r.CostUsd),
(long)Math.Ceiling(runs.Average(r => (double)r.TotalTokens)));
}
private void Release(string agentId, RunEstimate estimate)
{
lock (_reservedLock)
{
var left = _reservedByAgent.GetValueOrDefault(agentId) - estimate;
if (left.CostUsd <= 0 && left.Tokens <= 0)
_reservedByAgent.Remove(agentId);
else
_reservedByAgent[agentId] = left;
_reservedInstance -= estimate;
}
}
private sealed class Reservation(BudgetGuard guard, string agentId, RunEstimate estimate) : IDisposable
{
private int _disposed;
public void Dispose()
{
if (Interlocked.Exchange(ref _disposed, 1) == 0)
guard.Release(agentId, estimate);
}
}
private async Task<BudgetStatus> CheckCoreAsync(
AgentConfig agent, InstanceBudget instanceBudget,
RunEstimate agentReserved, RunEstimate instanceReserved, CancellationToken ct)
{
var today = DateOnly.FromDateTime(_now());
@@ -54,19 +178,21 @@ public sealed class BudgetGuard
{
var used = await _usage.GetDailyAsync(today, agent.AgentId, ct);
if (Exceeds(agentBudget.DailyCostUsd, used.CostUsd))
if (Exceeds(agentBudget.DailyCostUsd, used.CostUsd + agentReserved.CostUsd))
{
return new BudgetStatus(false, BudgetLimit.AgentCost,
$"Tagesbudget des Agenten erschöpft: {used.CostUsd:F4} von maximal " +
$"{agentBudget.DailyCostUsd:F2} USD verbraucht." +
ReservedHint(agentReserved.CostUsd) +
UnpricedHint(used));
}
if (Exceeds(agentBudget.DailyTokens, used.TotalTokens))
if (Exceeds(agentBudget.DailyTokens, used.TotalTokens + agentReserved.Tokens))
{
return new BudgetStatus(false, BudgetLimit.AgentTokens,
$"Token-Tageslimit des Agenten erschöpft: {used.TotalTokens:N0} von maximal " +
$"{agentBudget.DailyTokens:N0}.");
$"{agentBudget.DailyTokens:N0}." +
ReservedTokensHint(agentReserved.Tokens));
}
}
@@ -75,19 +201,21 @@ public sealed class BudgetGuard
{
var used = await _usage.GetDailyAsync(today, agentId: "", ct);
if (Exceeds(instanceBudget.DailyCostUsd, used.CostUsd))
if (Exceeds(instanceBudget.DailyCostUsd, used.CostUsd + instanceReserved.CostUsd))
{
return new BudgetStatus(false, BudgetLimit.InstanceCost,
$"Tagesbudget der Instanz erschöpft: {used.CostUsd:F4} von maximal " +
$"{instanceBudget.DailyCostUsd:F2} USD verbraucht." +
ReservedHint(instanceReserved.CostUsd) +
UnpricedHint(used));
}
if (Exceeds(instanceBudget.DailyTokens, used.TotalTokens))
if (Exceeds(instanceBudget.DailyTokens, used.TotalTokens + instanceReserved.Tokens))
{
return new BudgetStatus(false, BudgetLimit.InstanceTokens,
$"Token-Tageslimit der Instanz erschöpft: {used.TotalTokens:N0} von maximal " +
$"{instanceBudget.DailyTokens:N0}.");
$"{instanceBudget.DailyTokens:N0}." +
ReservedTokensHint(instanceReserved.Tokens));
}
}
@@ -109,6 +237,13 @@ public sealed class BudgetGuard
private static bool Exceeds(long limit, long used) => limit > 0 && used >= limit;
/// <summary>#716: Without this, a refusal at "3.20 of 5.00 USD used" would look wrong.</summary>
private static string ReservedHint(decimal reserved)
=> reserved > 0 ? $" Weitere {reserved:F4} USD sind für laufende Läufe reserviert." : "";
private static string ReservedTokensHint(long reserved)
=> reserved > 0 ? $" Weitere {reserved:N0} Tokens sind für laufende Läufe reserviert." : "";
/// <summary>
/// Ohne Preisangaben ist die verbrauchte Summe unvollständig — sonst wirkte ein
/// niedriger Wert wie ein noch offener Spielraum.
+13 -8
View File
@@ -163,8 +163,11 @@ public sealed class AgentEngine : IAgentMessageRouter, Staging.IFrozenCallExecut
return zu;
// Vor der ersten Anfrage prüfen — ein erschöpftes Budget soll gar nichts kosten.
if (await CheckBudgetAsync(agentConfig, source ?? AuditSource.Direct, taskId, externalCt) is { } denied)
// #716: Die Reservierung gilt, bis der Verbrauch erfasst ist (Ende der Methode).
var (denied, reservation) = await CheckBudgetAsync(agentConfig, source ?? AuditSource.Direct, taskId, externalCt);
if (denied is not null)
return denied;
using var budgetReservation = reservation;
using var lauf = TrackRun();
@@ -444,8 +447,10 @@ public sealed class AgentEngine : IAgentMessageRouter, Staging.IFrozenCallExecut
if (RefuseWhileShuttingDown(agentConfig) is { } zu)
return zu;
if (await CheckBudgetAsync(agentConfig, source, taskId, externalCt) is { } denied)
var (denied, reservation) = await CheckBudgetAsync(agentConfig, source, taskId, externalCt);
if (denied is not null)
return denied;
using var budgetReservation = reservation;
using var lauf = TrackRun();
@@ -1638,17 +1643,17 @@ public sealed class AgentEngine : IAgentMessageRouter, Staging.IFrozenCallExecut
/// <summary>
/// Prüft das Tagesbudget. Gibt ein Ergebnis zurück, wenn der Lauf nicht stattfinden
/// darf — sonst null.
/// darf — sonst null und, bei gesetzter Grenze, die Reservierung des Laufs (#716).
/// </summary>
private async Task<AgentRunResult?> CheckBudgetAsync(
private async Task<(AgentRunResult? Denied, IDisposable? Reservation)> CheckBudgetAsync(
AgentConfig agentConfig, string? source, string? taskId, CancellationToken ct)
{
if (_budgetGuard is null)
return null;
return (null, null);
var status = await _budgetGuard.CheckAsync(agentConfig, InstanceBudget, ct);
var (status, reservation) = await _budgetGuard.ReserveAsync(agentConfig, InstanceBudget, ct);
if (status.IsAllowed)
return null;
return (null, reservation);
// #608: Ist ein Budget erschoepft, lehnt es jeden weiteren Lauf des Tages ab. Einmal je
// Agent, Grenze und Tag als Warnung genuegt - danach steht jede Ablehnung als Beleg da.
@@ -1676,7 +1681,7 @@ public sealed class AgentEngine : IAgentMessageRouter, Staging.IFrozenCallExecut
await RecordReceiptAsync(Guid.NewGuid().ToString("N"), agentConfig, result, source, taskId);
OnRunCompleted?.Invoke(agentConfig.Model, result);
return result;
return (result, null);
}
/// <summary>Welche Budgetgrenze je Agent zuletzt an welchem Tag als Warnung gemeldet wurde.</summary>