The best way to control Claude Code, Codex and more from any computer or phone
| 1 | import type { UsageProviderKind } from "@t3tools/contracts"; |
| 2 | import type { |
| 3 | TranscriptUsageFormat, |
| 4 | UsageRecord, |
| 5 | UsageSpeed, |
| 6 | } from "@t3tools/provider-core/server/usage"; |
| 7 | import * as Schema from "effect/Schema"; |
| 8 | |
| 9 | import { CLAUDE_SYNTHETIC_MODEL } from "../provider/Drivers/claudeUsage.ts"; |
| 10 | import { mergeUsageCopies } from "./usageAggregation.ts"; |
| 11 | import { GUARD_LENGTH, type TranscriptParsePosition } from "./usageTranscriptReader.ts"; |
| 12 | |
| 13 | const USAGE_SCAN_CACHE_VERSION = 17 as const; |
| 14 | const SPEED_COMPATIBLE_SINCE_VERSION = 4; |
| 15 | |
| 16 | /** |
| 17 | * An older server sharing that directory cannot read a newer cache and would |
| 18 | * replace it, dropping saved usage for deleted transcripts. Separate files keep |
| 19 | * the newer one; older files are removed once it is saved. |
| 20 | */ |
| 21 | export const SCAN_CACHE_FILE_NAME = "usage-scan-cache-v17.json"; |
| 22 | export const LEGACY_SCAN_CACHE_FILE_NAMES = [ |
| 23 | "usage-scan-cache-v16.json", |
| 24 | "usage-scan-cache-v15.json", |
| 25 | "usage-scan-cache-v14.json", |
| 26 | "usage-scan-cache-v13.json", |
| 27 | "usage-scan-cache-v12.json", |
| 28 | "usage-scan-cache-v11.json", |
| 29 | "usage-scan-cache-v10.json", |
| 30 | "usage-scan-cache-v9.json", |
| 31 | "usage-scan-cache-v8.json", |
| 32 | "usage-scan-cache-v7.json", |
| 33 | "usage-scan-cache-v6.json", |
| 34 | "usage-scan-cache-v5.json", |
| 35 | "usage-scan-cache.json", |
| 36 | ] as const; |
| 37 | |
| 38 | const SPEEDS: readonly UsageSpeed[] = ["standard", "fast", "ultrafast"]; |
| 39 | |
| 40 | export interface CachedFile { |
| 41 | readonly size: number; |
| 42 | readonly mtimeMs: number; |
| 43 | readonly provider: UsageProviderKind; |
| 44 | readonly records: readonly UsageRecord[]; |
| 45 | /** Unterminated tail records stay separate because incremental parsing rereads them. */ |
| 46 | readonly tailRecords: readonly UsageRecord[]; |
| 47 | /** Whole-file count, including diagnostics from the tentative tail. */ |
| 48 | readonly malformedRecords: number; |
| 49 | readonly position: TranscriptParsePosition; |
| 50 | } |
| 51 | |
| 52 | export type ScanCache = Map<string, CachedFile>; |
| 53 | |
| 54 | type SerializedRecord = readonly [ |
| 55 | timestampMs: number, |
| 56 | modelIndex: number, |
| 57 | sessionIndex: number, |
| 58 | uncachedInputTokens: number, |
| 59 | cachedInputTokens: number, |
| 60 | cacheCreationTokens: number, |
| 61 | outputTokens: number, |
| 62 | reasoningTokens: number, |
| 63 | dedupeKey: string | null, |
| 64 | reportedCostUsd: number | null, |
| 65 | speed: number, |
| 66 | rateModelIndex?: number | null, |
| 67 | unresolvedServiceTierTurnId?: string | null, |
| 68 | cacheCreation1hTokens?: number, |
| 69 | ]; |
| 70 | |
| 71 | interface SerializedFile { |
| 72 | readonly s: number; |
| 73 | readonly m: number; |
| 74 | readonly p: UsageProviderKind; |
| 75 | readonly r: readonly SerializedRecord[]; |
| 76 | readonly t: readonly SerializedRecord[]; |
| 77 | readonly mr: number; |
| 78 | readonly o: number; |
| 79 | readonly gl: number; |
| 80 | readonly gh: number; |
| 81 | readonly cs: unknown; |
| 82 | } |
| 83 | |
| 84 | interface SerializedCache { |
| 85 | readonly version: number; |
| 86 | readonly models: readonly string[]; |
| 87 | readonly sessions: readonly string[]; |
| 88 | readonly files: Readonly<Record<string, SerializedFile>>; |
| 89 | } |
| 90 | |
| 91 | interface InternTables { |
| 92 | readonly models: string[]; |
| 93 | readonly sessions: string[]; |
| 94 | readonly modelIndex: Map<string, number>; |
| 95 | readonly sessionIndex: Map<string, number>; |
| 96 | } |
| 97 | |
| 98 | function makeInternTables(): InternTables { |
| 99 | return { models: [], sessions: [], modelIndex: new Map(), sessionIndex: new Map() }; |
| 100 | } |
| 101 | |
| 102 | function intern(table: string[], index: Map<string, number>, value: string): number { |
| 103 | const existing = index.get(value); |
| 104 | if (existing !== undefined) return existing; |
| 105 | const next = table.length; |
| 106 | table.push(value); |
| 107 | index.set(value, next); |
| 108 | return next; |
| 109 | } |
| 110 | |
| 111 | function serializeFile(entry: CachedFile, tables: InternTables): SerializedFile { |
| 112 | const serializeRecord = (record: UsageRecord): SerializedRecord => [ |
| 113 | record.timestampMs, |
| 114 | intern(tables.models, tables.modelIndex, record.model), |
| 115 | intern(tables.sessions, tables.sessionIndex, record.sessionId), |
| 116 | record.totals.uncachedInputTokens, |
| 117 | record.totals.cachedInputTokens, |
| 118 | record.totals.cacheCreationTokens, |
| 119 | record.totals.outputTokens, |
| 120 | record.totals.reasoningTokens, |
| 121 | record.dedupeKey, |
| 122 | record.reportedCostUsd, |
| 123 | SPEEDS.indexOf(record.speed), |
| 124 | record.rateModel === undefined |
| 125 | ? null |
| 126 | : intern(tables.models, tables.modelIndex, record.rateModel), |
| 127 | ...(record.cacheCreation1hTokens !== undefined |
| 128 | ? ([record.unresolvedServiceTierTurnId ?? null, record.cacheCreation1hTokens] as const) |
| 129 | : record.unresolvedServiceTierTurnId !== undefined |
| 130 | ? ([record.unresolvedServiceTierTurnId] as const) |
| 131 | : ([] as const)), |
| 132 | ]; |
| 133 | return { |
| 134 | s: entry.size, |
| 135 | m: entry.mtimeMs, |
| 136 | p: entry.provider, |
| 137 | r: entry.records.map(serializeRecord), |
| 138 | t: entry.tailRecords.map(serializeRecord), |
| 139 | mr: entry.malformedRecords, |
| 140 | o: entry.position.resumeOffset, |
| 141 | gl: entry.position.guardLength, |
| 142 | gh: entry.position.guardHash, |
| 143 | cs: entry.position.state, |
| 144 | }; |
| 145 | } |
| 146 | |
| 147 | /** Yields the document in pieces so large caches never become one string; growing intern tables preserve indexes, retaining unused strings until restart. */ |
| 148 | export function makeScanCacheWriter(): ( |
| 149 | cache: ScanCache, |
| 150 | extra: Readonly<Record<string, unknown>>, |
| 151 | ) => Iterable<string> { |
| 152 | const tables = makeInternTables(); |
| 153 | function* encode(entries: readonly (readonly [string, CachedFile])[], extra: object) { |
| 154 | yield '{"files":{'; |
| 155 | for (const [index, [path, entry]] of entries.entries()) { |
| 156 | yield `${index === 0 ? "" : ","}${JSON.stringify(path)}:${JSON.stringify(serializeFile(entry, tables))}`; |
| 157 | } |
| 158 | // Encoded after the files, which may have added to the intern tables. |
| 159 | const tail = JSON.stringify({ |
| 160 | ...extra, |
| 161 | version: USAGE_SCAN_CACHE_VERSION, |
| 162 | models: tables.models, |
| 163 | sessions: tables.sessions, |
| 164 | }); |
| 165 | yield `},${tail.slice(1)}`; |
| 166 | } |
| 167 | // Snapshot now: the encoding runs later, while the file is written. |
| 168 | return (cache, extra) => encode([...cache], extra); |
| 169 | } |
| 170 | |
| 171 | function isRecordArray(value: unknown): value is readonly unknown[] { |
| 172 | return Array.isArray(value); |
| 173 | } |
| 174 | |
| 175 | export function decodeScanCache( |
| 176 | document: unknown, |
| 177 | formats: ReadonlyMap<UsageProviderKind, TranscriptUsageFormat<unknown>>, |
| 178 | ): ScanCache { |
| 179 | const cache: ScanCache = new Map(); |
| 180 | const isValidState = makeStateValidators(formats); |
| 181 | if (typeof document !== "object" || document === null) return cache; |
| 182 | |
| 183 | const root = document as Partial<SerializedCache>; |
| 184 | const version = root.version; |
| 185 | if ( |
| 186 | typeof version !== "number" || |
| 187 | version < SPEED_COMPATIBLE_SINCE_VERSION || |
| 188 | version > USAGE_SCAN_CACHE_VERSION |
| 189 | ) { |
| 190 | return cache; |
| 191 | } |
| 192 | if (!isRecordArray(root.models) || !isRecordArray(root.sessions)) return cache; |
| 193 | if (typeof root.files !== "object" || root.files === null) return cache; |
| 194 | |
| 195 | if (!root.models.every((value) => typeof value === "string")) return cache; |
| 196 | if (!root.sessions.every((value) => typeof value === "string")) return cache; |
| 197 | const models = root.models as readonly string[]; |
| 198 | const sessions = root.sessions as readonly string[]; |
| 199 | |
| 200 | // Reject entire corrupt entries; partial survivors would appear warm and hide usage until the file changes. |
| 201 | const decodeRecords = ( |
| 202 | rows: readonly unknown[], |
| 203 | provider: UsageProviderKind, |
| 204 | ): UsageRecord[] | null => { |
| 205 | const records: UsageRecord[] = []; |
| 206 | for (const row of rows) { |
| 207 | if (!isRecordArray(row) || row.length < 11) return null; |
| 208 | const [ |
| 209 | timestampMs, |
| 210 | modelIndex, |
| 211 | sessionIndex, |
| 212 | uncached, |
| 213 | cached, |
| 214 | cacheCreation, |
| 215 | output, |
| 216 | reasoning, |
| 217 | dedupeKey, |
| 218 | reportedCostUsd, |
| 219 | speedIndex, |
| 220 | rateModelIndex, |
| 221 | unresolvedServiceTierTurnId, |
| 222 | cacheCreation1h = 0, |
| 223 | ] = row as SerializedRecord; |
| 224 | const speed = typeof speedIndex === "number" ? SPEEDS[speedIndex] : undefined; |
| 225 | |
| 226 | const model = typeof modelIndex === "number" ? models[modelIndex] : undefined; |
| 227 | const rateModel = typeof rateModelIndex === "number" ? models[rateModelIndex] : undefined; |
| 228 | if ( |
| 229 | typeof timestampMs !== "number" || |
| 230 | !Number.isFinite(timestampMs) || |
| 231 | model === undefined || |
| 232 | (unresolvedServiceTierTurnId != null && typeof unresolvedServiceTierTurnId !== "string") || |
| 233 | !Number.isFinite(cacheCreation1h) || |
| 234 | (rateModelIndex != null && rateModel === undefined) || |
| 235 | !Number.isFinite(uncached) || |
| 236 | !Number.isFinite(cached) || |
| 237 | !Number.isFinite(cacheCreation) || |
| 238 | !Number.isFinite(output) || |
| 239 | !Number.isFinite(reasoning) || |
| 240 | speed === undefined |
| 241 | ) { |
| 242 | return null; |
| 243 | } |
| 244 | |
| 245 | // Unchanged and deleted transcripts may only exist in the cache, so the |
| 246 | // parser cannot remove their previously saved local placeholders. |
| 247 | if (provider === "claude" && model === CLAUDE_SYNTHETIC_MODEL) continue; |
| 248 | |
| 249 | records.push({ |
| 250 | provider, |
| 251 | timestampMs, |
| 252 | model, |
| 253 | ...(rateModel === undefined ? {} : { rateModel }), |
| 254 | sessionId: (typeof sessionIndex === "number" ? sessions[sessionIndex] : undefined) ?? "", |
| 255 | totals: { |
| 256 | uncachedInputTokens: uncached, |
| 257 | cachedInputTokens: cached, |
| 258 | cacheCreationTokens: cacheCreation, |
| 259 | outputTokens: output, |
| 260 | reasoningTokens: reasoning, |
| 261 | }, |
| 262 | ...(cacheCreation1h > 0 ? { cacheCreation1hTokens: cacheCreation1h } : {}), |
| 263 | reportedCostUsd: typeof reportedCostUsd === "number" ? reportedCostUsd : null, |
| 264 | speed, |
| 265 | ...(version >= 7 && unresolvedServiceTierTurnId != null |
| 266 | ? { unresolvedServiceTierTurnId } |
| 267 | : {}), |
| 268 | dedupeKey: |
| 269 | typeof dedupeKey !== "string" |
| 270 | ? null |
| 271 | : // Saved copies of moved transcripts must still match the re-parsed v17 key. |
| 272 | provider === "grok" && |
| 273 | version < 17 && |
| 274 | dedupeKey.endsWith(`:${model}`) && |
| 275 | !dedupeKey.endsWith(`:${timestampMs}:${model}`) |
| 276 | ? `${dedupeKey.slice(0, -model.length)}${timestampMs}:${model}` |
| 277 | : dedupeKey, |
| 278 | }); |
| 279 | } |
| 280 | return records; |
| 281 | }; |
| 282 | |
| 283 | for (const [path, raw] of Object.entries(root.files)) { |
| 284 | if (typeof raw !== "object" || raw === null) continue; |
| 285 | const entry = raw as Partial<SerializedFile>; |
| 286 | if (typeof entry.s !== "number" || typeof entry.m !== "number") continue; |
| 287 | if (entry.p === undefined) continue; |
| 288 | const format = formats.get(entry.p); |
| 289 | if (format === undefined) continue; |
| 290 | if (!isRecordArray(entry.r) || !isRecordArray(entry.t)) continue; |
| 291 | const malformedRecords = entry.mr === undefined ? 0 : entry.mr; |
| 292 | if (!Number.isSafeInteger(malformedRecords) || malformedRecords < 0) continue; |
| 293 | // Position fields feed byte offsets and a Buffer allocation in the reader, |
| 294 | // so anything outside their real ranges must reject the entry: a bogus |
| 295 | // guard length would otherwise fail every parse of the file, silently |
| 296 | // dropping its usage instead of costing the documented cold re-parse. |
| 297 | if ( |
| 298 | typeof entry.o !== "number" || |
| 299 | !Number.isSafeInteger(entry.o) || |
| 300 | entry.o < 0 || |
| 301 | typeof entry.gl !== "number" || |
| 302 | !Number.isSafeInteger(entry.gl) || |
| 303 | entry.gl < 0 || |
| 304 | entry.gl > GUARD_LENGTH || |
| 305 | entry.gl > entry.o || |
| 306 | typeof entry.gh !== "number" || |
| 307 | !Number.isFinite(entry.gh) |
| 308 | ) { |
| 309 | continue; |
| 310 | } |
| 311 | // Keep history for deleted rollouts, but re-parse incompatible reducer state. |
| 312 | // v13 adds Claude 1-hour cache writes; v14 uses Codex request ledgers. |
| 313 | // v15 repairs skipped Codex requests whose counters rewound; clean v14 entries stay served. |
| 314 | // v17 keys Grok turns by time, restoring turns previously dropped as repeats. |
| 315 | const legacy = |
| 316 | (entry.p === "claude" && version < 13) || |
| 317 | (entry.p === "grok" && version < 17) || |
| 318 | (format.state !== undefined && |
| 319 | (version < 6 || |
| 320 | (entry.p === "codex" && (version < 14 || (version < 15 && malformedRecords > 0))))); |
| 321 | // A corrupt state disqualifies the entry: resuming with it would attach |
| 322 | // appended usage to the wrong model or replay fork-copied history. |
| 323 | if (!legacy && !isValidState.get(entry.p)?.(entry.cs)) continue; |
| 324 | // Before v16 a Codex baseline can lack the time of its latest move, or carry an older one, |
| 325 | // and that time keeps a re-emitted older snapshot from billing twice. A grown file re-parses |
| 326 | // whole instead of resuming. |
| 327 | const resumable = !legacy && !(entry.p === "codex" && version < 16); |
| 328 | |
| 329 | const provider = entry.p; |
| 330 | const records = decodeRecords(entry.r, provider); |
| 331 | const tailRecords = decodeRecords(entry.t, provider); |
| 332 | if (records === null || tailRecords === null) continue; |
| 333 | |
| 334 | cache.set(path, { |
| 335 | size: legacy ? -1 : entry.s, |
| 336 | mtimeMs: entry.m, |
| 337 | provider, |
| 338 | records, |
| 339 | tailRecords, |
| 340 | malformedRecords, |
| 341 | position: resumable |
| 342 | ? { |
| 343 | resumeOffset: entry.o, |
| 344 | guardLength: entry.gl, |
| 345 | guardHash: entry.gh, |
| 346 | state: entry.cs, |
| 347 | } |
| 348 | : { resumeOffset: 0, guardLength: 0, guardHash: 0, state: null }, |
| 349 | }); |
| 350 | } |
| 351 | |
| 352 | return cache; |
| 353 | } |
| 354 | |
| 355 | function makeStateValidators( |
| 356 | formats: ReadonlyMap<UsageProviderKind, TranscriptUsageFormat<unknown>>, |
| 357 | ): ReadonlyMap<UsageProviderKind, (value: unknown) => boolean> { |
| 358 | return new Map( |
| 359 | [...formats].map(([provider, format]) => { |
| 360 | const isState = format.state === undefined ? undefined : Schema.is(format.state.schema); |
| 361 | return [ |
| 362 | provider, |
| 363 | (value: unknown) => value === null || (isState !== undefined && isState(value)), |
| 364 | ] as const; |
| 365 | }), |
| 366 | ); |
| 367 | } |
| 368 | |
| 369 | export function pruneScanCache(cache: ScanCache, retentionCutoffMs: number): number { |
| 370 | let removed = 0; |
| 371 | for (const [path, entry] of cache) { |
| 372 | if (entry.mtimeMs < retentionCutoffMs) { |
| 373 | cache.delete(path); |
| 374 | removed += 1; |
| 375 | } |
| 376 | } |
| 377 | return removed; |
| 378 | } |
| 379 | |
| 380 | /** Pass only newline-terminated records: an unterminated tail can still change, so the aggregator merges it per scan. */ |
| 381 | export function dedupeWithinFile(records: readonly UsageRecord[]): readonly UsageRecord[] { |
| 382 | const kept: UsageRecord[] = []; |
| 383 | const firstIndex = new Map<string, number>(); |
| 384 | for (const record of records) { |
| 385 | if (record.dedupeKey !== null) { |
| 386 | const index = firstIndex.get(record.dedupeKey); |
| 387 | if (index !== undefined) { |
| 388 | kept[index] = mergeUsageCopies(kept[index]!, record); |
| 389 | continue; |
| 390 | } |
| 391 | firstIndex.set(record.dedupeKey, kept.length); |
| 392 | } |
| 393 | kept.push(record); |
| 394 | } |
| 395 | return kept; |
| 396 | } |