chore(claude-do): UsageGate: Parallelitaet stufenweise drosseln statt erst bei

## Kontext: Limits sind Fenster, nicht Summen

Die Runs laufen ueber das Claude-Abo. Limits greifen pro 5h-Fenster und pro 7 Tage. Nicht die Wochensumme tut weh, sondern dass ein Agent-Burst ein Fenster leerraeumt, in dem Mika selbst interaktiv arbeiten will.

## Messgrundlage (alle Transcripts unter ~/.claude/projects)

Agent-Runs sind ueber die ganze Historie nur **18,4 %** des Account-Verbrauch

ClaudeDo-Task: 87105f5e-c4f4-4af4-ae60-89cd2e153e3c
This commit is contained in:
mika kuns
2026-08-05 15:53:15 +02:00
parent 83ea429b8a
commit a201d3f43d
24 changed files with 1466 additions and 19 deletions
@@ -46,6 +46,17 @@ public sealed class UsageSnapshotBuilder
.Select(l => new UsageLimitDto(l.Kind, l.Group, l.Percent, l.Severity, l.ResetsAt, l.ScopeModelDisplayName, l.IsActive))
.ToList();
var configuredSlots = Math.Max(1, settings.MaxParallelExecutions);
var effectiveSlots = snapshot is null || lastError is not null
? configuredSlots
: UsageThrottle.EffectiveSlots(
configuredSlots, snapshot.FiveHour?.Utilization, snapshot.SevenDay?.Utilization,
settings.UsageThrottleSoftPct, settings.UsageThrottleHardPct,
settings.UsageGateFiveHourPct, settings.UsageGateSevenDayPct);
var throttleBucket = effectiveSlots < configuredSlots
? DecisiveBucket(snapshot?.FiveHour?.Utilization, snapshot?.SevenDay?.Utilization)
: null;
return new UsageSnapshotDto(
snapshot?.FiveHour?.Utilization,
snapshot?.FiveHour?.ResetsAt,
@@ -58,6 +69,15 @@ public sealed class UsageSnapshotBuilder
decision.Reason,
snapshot?.FetchedAtUtc,
isStale,
lastError);
lastError,
configuredSlots,
effectiveSlots,
throttleBucket);
}
private static string? DecisiveBucket(double? fiveHourPct, double? sevenDayPct)
{
if (fiveHourPct is null && sevenDayPct is null) return null;
return (fiveHourPct ?? 0) >= (sevenDayPct ?? 0) ? "five_hour" : "seven_day";
}
}
@@ -0,0 +1,38 @@
namespace ClaudeDo.Worker.Usage;
/// <summary>
/// Pure staged throttle sitting ahead of <see cref="UsageGate"/>'s hard pause: as the 5h/7d usage
/// window fills up, the queue's effective parallelism steps down before it hits zero, instead of
/// running at full tilt right up to the gate threshold. Whichever of the two buckets is more
/// utilized decides the stage. A missing bucket (null) is treated as 0% for that bucket only —
/// callers with no snapshot at all should skip this and use <paramref name="configuredSlots"/>
/// directly (fail-open), same policy as <see cref="UsageGate"/>.
/// </summary>
public static class UsageThrottle
{
public static int EffectiveSlots(
int configuredSlots,
double? fiveHourPct,
double? sevenDayPct,
int softPct,
int hardPct,
int gateFiveHourPct,
int gateSevenDayPct)
{
var slots = Math.Max(1, configuredSlots);
if (gateFiveHourPct > 0 && fiveHourPct is { } five && five >= gateFiveHourPct)
return 0;
if (gateSevenDayPct > 0 && sevenDayPct is { } seven && seven >= gateSevenDayPct)
return 0;
var worst = Math.Max(fiveHourPct ?? 0, sevenDayPct ?? 0);
if (hardPct > 0 && worst >= hardPct)
return Math.Min(slots, 1);
if (softPct > 0 && worst >= softPct)
return Math.Min(slots, 2);
return slots;
}
}