| 1 | import type { |
| 2 | OrchestrationV2ProviderTurn, |
| 3 | OrchestrationV2ProviderTurnTokenUsage, |
| 4 | OrchestrationV2ProviderThread, |
| 5 | OrchestrationV2TurnItem, |
| 6 | ThreadTokenUsageSnapshot, |
| 7 | } from "@t3tools/contracts"; |
| 8 | import * as DateTime from "effect/DateTime"; |
| 9 | |
| 10 | function asFiniteNumber(value: unknown): number | null { |
| 11 | return typeof value === "number" && Number.isFinite(value) ? value : null; |
| 12 | } |
| 13 | |
| 14 | type NullableContextWindowUsage = { |
| 15 | readonly [Key in keyof ThreadTokenUsageSnapshot]: undefined extends ThreadTokenUsageSnapshot[Key] |
| 16 | ? Exclude<ThreadTokenUsageSnapshot[Key], undefined> | null |
| 17 | : ThreadTokenUsageSnapshot[Key]; |
| 18 | }; |
| 19 | |
| 20 | export type ContextWindowSnapshot = NullableContextWindowUsage & { |
| 21 | readonly remainingTokens: number | null; |
| 22 | readonly usedPercentage: number | null; |
| 23 | readonly remainingPercentage: number | null; |
| 24 | readonly updatedAt: string; |
| 25 | }; |
| 26 | |
| 27 | /** Newest usage report across a thread's provider turns. Snapshots order turns |
| 28 | by provider thread, not time, so pick by timestamp. */ |
| 29 | export function latestProviderTurnTokenUsage( |
| 30 | turns: ReadonlyArray<Pick<OrchestrationV2ProviderTurn, "tokenUsage">>, |
| 31 | ): OrchestrationV2ProviderTurnTokenUsage | null { |
| 32 | let latest: OrchestrationV2ProviderTurnTokenUsage | null = null; |
| 33 | for (const { tokenUsage } of turns) { |
| 34 | if (tokenUsage === undefined) continue; |
| 35 | // Ties go to the later entry, since live upserts append. |
| 36 | if (latest === null || Date.parse(tokenUsage.updatedAt) >= Date.parse(latest.updatedAt)) { |
| 37 | latest = tokenUsage; |
| 38 | } |
| 39 | } |
| 40 | return latest; |
| 41 | } |
| 42 | |
| 43 | /** Prefers the provider's live usage report (#8144); falls back to the last compaction item. */ |
| 44 | export function deriveLatestContextWindowSnapshot( |
| 45 | entries: ReadonlyArray<{ |
| 46 | readonly item: OrchestrationV2TurnItem; |
| 47 | }>, |
| 48 | liveUsage?: OrchestrationV2ProviderTurnTokenUsage | null, |
| 49 | providerThread?: Pick<OrchestrationV2ProviderThread, "contextUsage" | "updatedAt"> | null, |
| 50 | ): ContextWindowSnapshot | null { |
| 51 | if (liveUsage != null) { |
| 52 | const usedTokens = Math.max(0, liveUsage.usedTokens); |
| 53 | const maxTokens = liveUsage.maxTokens ?? null; |
| 54 | const usedPercentage = |
| 55 | maxTokens !== null && maxTokens > 0 ? Math.min(100, (usedTokens / maxTokens) * 100) : null; |
| 56 | const remainingTokens = |
| 57 | maxTokens !== null ? Math.max(0, Math.round(maxTokens - usedTokens)) : null; |
| 58 | const remainingPercentage = usedPercentage !== null ? Math.max(0, 100 - usedPercentage) : null; |
| 59 | return { |
| 60 | usedTokens, |
| 61 | totalProcessedTokens: null, |
| 62 | maxTokens, |
| 63 | remainingTokens, |
| 64 | usedPercentage, |
| 65 | remainingPercentage, |
| 66 | inputTokens: liveUsage.inputTokens ?? null, |
| 67 | cachedInputTokens: liveUsage.cachedInputTokens ?? null, |
| 68 | outputTokens: liveUsage.outputTokens ?? null, |
| 69 | reasoningOutputTokens: liveUsage.reasoningOutputTokens ?? null, |
| 70 | lastUsedTokens: null, |
| 71 | lastInputTokens: null, |
| 72 | lastCachedInputTokens: null, |
| 73 | lastOutputTokens: null, |
| 74 | lastReasoningOutputTokens: null, |
| 75 | toolUses: null, |
| 76 | durationMs: null, |
| 77 | compactsAutomatically: true, |
| 78 | autoCompactThreshold: null, |
| 79 | cost: null, |
| 80 | updatedAt: liveUsage.updatedAt, |
| 81 | }; |
| 82 | } |
| 83 | const providerUsage = providerThread?.contextUsage; |
| 84 | const providerUsageUpdatedAt = providerThread?.updatedAt; |
| 85 | if ( |
| 86 | providerUsage !== null && |
| 87 | providerUsage !== undefined && |
| 88 | providerUsageUpdatedAt !== undefined |
| 89 | ) { |
| 90 | const maxTokens = asFiniteNumber(providerUsage.maxTokens); |
| 91 | const usedTokens = providerUsage.usedTokens; |
| 92 | const usedPercentage = |
| 93 | maxTokens !== null && maxTokens > 0 ? Math.min(100, (usedTokens / maxTokens) * 100) : null; |
| 94 | return { |
| 95 | usedTokens, |
| 96 | totalProcessedTokens: asFiniteNumber(providerUsage.totalProcessedTokens), |
| 97 | maxTokens, |
| 98 | remainingTokens: maxTokens === null ? null : Math.max(0, Math.round(maxTokens - usedTokens)), |
| 99 | usedPercentage, |
| 100 | remainingPercentage: usedPercentage === null ? null : Math.max(0, 100 - usedPercentage), |
| 101 | inputTokens: asFiniteNumber(providerUsage.inputTokens), |
| 102 | cachedInputTokens: asFiniteNumber(providerUsage.cachedInputTokens), |
| 103 | outputTokens: asFiniteNumber(providerUsage.outputTokens), |
| 104 | reasoningOutputTokens: asFiniteNumber(providerUsage.reasoningOutputTokens), |
| 105 | lastUsedTokens: asFiniteNumber(providerUsage.lastUsedTokens), |
| 106 | lastInputTokens: asFiniteNumber(providerUsage.lastInputTokens), |
| 107 | lastCachedInputTokens: asFiniteNumber(providerUsage.lastCachedInputTokens), |
| 108 | lastOutputTokens: asFiniteNumber(providerUsage.lastOutputTokens), |
| 109 | lastReasoningOutputTokens: asFiniteNumber(providerUsage.lastReasoningOutputTokens), |
| 110 | toolUses: asFiniteNumber(providerUsage.toolUses), |
| 111 | durationMs: asFiniteNumber(providerUsage.durationMs), |
| 112 | compactsAutomatically: providerUsage.compactsAutomatically ?? null, |
| 113 | cost: providerUsage.cost ?? null, |
| 114 | autoCompactThreshold: providerUsage.autoCompactThreshold ?? null, |
| 115 | updatedAt: DateTime.formatIso(providerUsageUpdatedAt), |
| 116 | }; |
| 117 | } |
| 118 | for (let index = entries.length - 1; index >= 0; index -= 1) { |
| 119 | const entry = entries[index]; |
| 120 | if (!entry || entry.item.type !== "compaction") { |
| 121 | continue; |
| 122 | } |
| 123 | const payload = entry.item; |
| 124 | const usedTokens = asFiniteNumber(payload.afterTokenCount); |
| 125 | if (usedTokens === null || usedTokens < 0) { |
| 126 | continue; |
| 127 | } |
| 128 | |
| 129 | const maxTokens = null; |
| 130 | const usedPercentage = |
| 131 | maxTokens !== null && maxTokens > 0 ? Math.min(100, (usedTokens / maxTokens) * 100) : null; |
| 132 | const remainingTokens = |
| 133 | maxTokens !== null ? Math.max(0, Math.round(maxTokens - usedTokens)) : null; |
| 134 | const remainingPercentage = usedPercentage !== null ? Math.max(0, 100 - usedPercentage) : null; |
| 135 | |
| 136 | return { |
| 137 | usedTokens, |
| 138 | totalProcessedTokens: asFiniteNumber(payload.beforeTokenCount), |
| 139 | maxTokens, |
| 140 | remainingTokens, |
| 141 | usedPercentage, |
| 142 | remainingPercentage, |
| 143 | inputTokens: null, |
| 144 | cachedInputTokens: null, |
| 145 | outputTokens: null, |
| 146 | reasoningOutputTokens: null, |
| 147 | lastUsedTokens: null, |
| 148 | lastInputTokens: null, |
| 149 | lastCachedInputTokens: null, |
| 150 | lastOutputTokens: null, |
| 151 | lastReasoningOutputTokens: null, |
| 152 | toolUses: null, |
| 153 | durationMs: null, |
| 154 | compactsAutomatically: true, |
| 155 | autoCompactThreshold: null, |
| 156 | cost: null, |
| 157 | updatedAt: DateTime.formatIso(payload.startedAt ?? payload.updatedAt), |
| 158 | }; |
| 159 | } |
| 160 | |
| 161 | return null; |
| 162 | } |
| 163 | |
| 164 | export function formatContextWindowTokens(value: number | null): string { |
| 165 | if (value === null || !Number.isFinite(value)) { |
| 166 | return "0"; |
| 167 | } |
| 168 | if (value < 1_000) { |
| 169 | return `${Math.round(value)}`; |
| 170 | } |
| 171 | if (value < 10_000) { |
| 172 | return `${(value / 1_000).toFixed(1).replace(/\.0$/, "")}k`; |
| 173 | } |
| 174 | // Anything from 999,500 rounds to 1000k, so it belongs to the next unit. |
| 175 | if (value < 999_500) { |
| 176 | return `${Math.round(value / 1_000)}k`; |
| 177 | } |
| 178 | return `${(value / 1_000_000).toFixed(1).replace(/\.0$/, "")}m`; |
| 179 | } |