@posthog/wizard 1.34.0 → 1.35.0

package/dist/src/lib/middleware/benchmark.js

@@ -0,0 +1,49 @@
+"use strict";
+/**
+ * Benchmark tracking for wizard runs.
+ *
+ * Usage:
+ *   const pipeline = createBenchmarkPipeline(spinner, options);
+ *   pipeline.onMessage(message);
+ *   pipeline.finalize(resultMessage, durationMs);
+ */
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.createBenchmarkPipeline = createBenchmarkPipeline;
+const chalk_1 = __importDefault(require("chalk"));
+const clack_1 = __importDefault(require("../../utils/clack"));
+const debug_1 = require("../../utils/debug");
+const pipeline_1 = require("./pipeline");
+const phase_detector_1 = require("./phase-detector");
+const config_1 = require("./config");
+const benchmarks_1 = require("./benchmarks");
+const agent_interface_1 = require("../agent-interface");
+// ── Factory ────────────────────────────────────────────────────────────
+/**
+ * Create a middleware pipeline configured for benchmarking.
+ * Loads .benchmark-config.json from the install dir, falls back to defaults.
+ */
+function createBenchmarkPipeline(spinner, options, configOverride) {
+    const config = configOverride ?? (0, config_1.loadBenchmarkConfig)(options.installDir);
+    (0, debug_1.configureLogFile)({
+        path: config.output.logPath,
+        enabled: config.output.logEnabled,
+    });
+    const plugins = (0, benchmarks_1.createPluginsFromConfig)(config, {
+        spinner,
+        phased: false,
+        outputPath: config.output.benchmarkPath,
+    });
+    if (!config.output.suppressWizardLogs) {
+        clack_1.default.log.info(`${chalk_1.default.cyan(agent_interface_1.AgentSignals.BENCHMARK)} Verbose logs: ${(0, debug_1.getLogFilePath)()}`);
+        clack_1.default.log.info(`${chalk_1.default.cyan(agent_interface_1.AgentSignals.BENCHMARK)} Benchmark data will be written to: ${config.output.benchmarkPath}`);
+    }
+    (0, debug_1.logToFile)(`${agent_interface_1.AgentSignals.BENCHMARK} Tracking enabled, starting with setup phase`);
+    return new pipeline_1.MiddlewarePipeline(plugins, {
+        phaseDetector: new phase_detector_1.PhaseDetector(),
+        autoDetectPhases: true,
+    });
+}
+//# sourceMappingURL=benchmark.js.map

package/dist/src/lib/middleware/benchmark.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"benchmark.js","sourceRoot":"","sources":["../../../../src/lib/middleware/benchmark.ts"],"names":[],"mappings":";AAAA;;;;;;;GAOG;;;;;AA4DH,0DAqCC;AA/FD,kDAA0B;AAC1B,8DAAsC;AACtC,6CAAgF;AAChF,yCAAgD;AAChD,qDAAiD;AACjD,qCAA+C;AAC/C,6CAAuD;AAGvD,wDAAkD;AA2ClD,0EAA0E;AAE1E;;;GAGG;AACH,SAAgB,uBAAuB,CACrC,OAAyC,EACzC,OAAsB,EACtB,cAAgC;IAEhC,MAAM,MAAM,GAAG,cAAc,IAAI,IAAA,4BAAmB,EAAC,OAAO,CAAC,UAAU,CAAC,CAAC;IAEzE,IAAA,wBAAgB,EAAC;QACf,IAAI,EAAE,MAAM,CAAC,MAAM,CAAC,OAAO;QAC3B,OAAO,EAAE,MAAM,CAAC,MAAM,CAAC,UAAU;KAClC,CAAC,CAAC;IAEH,MAAM,OAAO,GAAG,IAAA,oCAAuB,EAAC,MAAM,EAAE;QAC9C,OAAO;QACP,MAAM,EAAE,KAAK;QACb,UAAU,EAAE,MAAM,CAAC,MAAM,CAAC,aAAa;KACxC,CAAC,CAAC;IAEH,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,kBAAkB,EAAE,CAAC;QACtC,eAAK,CAAC,GAAG,CAAC,IAAI,CACZ,GAAG,eAAK,CAAC,IAAI,CAAC,8BAAY,CAAC,SAAS,CAAC,kBAAkB,IAAA,sBAAc,GAAE,EAAE,CAC1E,CAAC;QACF,eAAK,CAAC,GAAG,CAAC,IAAI,CACZ,GAAG,eAAK,CAAC,IAAI,CACX,8BAAY,CAAC,SAAS,CACvB,uCAAuC,MAAM,CAAC,MAAM,CAAC,aAAa,EAAE,CACtE,CAAC;IACJ,CAAC;IAED,IAAA,iBAAS,EACP,GAAG,8BAAY,CAAC,SAAS,8CAA8C,CACxE,CAAC;IAEF,OAAO,IAAI,6BAAkB,CAAC,OAAO,EAAE;QACrC,aAAa,EAAE,IAAI,8BAAa,EAAE;QAClC,gBAAgB,EAAE,IAAI;KACvB,CAAC,CAAC;AACL,CAAC","sourcesContent":["/**\n * Benchmark tracking for wizard runs.\n *\n * Usage:\n *   const pipeline = createBenchmarkPipeline(spinner, options);\n *   pipeline.onMessage(message);\n *   pipeline.finalize(resultMessage, durationMs);\n */\n\nimport chalk from 'chalk';\nimport clack from '../../utils/clack';\nimport { logToFile, getLogFilePath, configureLogFile } from '../../utils/debug';\nimport { MiddlewarePipeline } from './pipeline';\nimport { PhaseDetector } from './phase-detector';\nimport { loadBenchmarkConfig } from './config';\nimport { createPluginsFromConfig } from './benchmarks';\nimport type { BenchmarkConfig } from './config';\nimport type { WizardOptions } from '../../utils/types';\nimport { AgentSignals } from '../agent-interface';\n\n// ── Types ──────────────────────────────────────────────────────────────\n\nexport interface StepUsage {\n  name: string;\n  usage: {\n    input_tokens: number;\n    output_tokens: number;\n    cache_creation_input_tokens: number;\n    cache_read_input_tokens: number;\n    cache_creation?: {\n      ephemeral_5m_input_tokens: number;\n      ephemeral_1h_input_tokens: number;\n    };\n  };\n  modelUsage: Record<string, unknown>;\n  totalCostUsd: number;\n  durationMs: number;\n  durationApiMs: number;\n  numTurns: number;\n  contextTokensIn?: number;\n  contextTokensOut?: number;\n  compactions?: number;\n  compactionPreTokens?: number[];\n}\n\nexport interface BenchmarkData {\n  timestamp: string;\n  steps: StepUsage[];\n  totals: {\n    totalCostUsd: number;\n    durationMs: number;\n    inputTokens: number;\n    outputTokens: number;\n    numTurns: number;\n    totalCompactions: number;\n    totalCacheReadTokens: number;\n    totalCacheCreation5mTokens: number;\n    totalCacheCreation1hTokens: number;\n  };\n}\n\n// ── Factory ────────────────────────────────────────────────────────────\n\n/**\n * Create a middleware pipeline configured for benchmarking.\n * Loads .benchmark-config.json from the install dir, falls back to defaults.\n */\nexport function createBenchmarkPipeline(\n  spinner: ReturnType<typeof clack.spinner>,\n  options: WizardOptions,\n  configOverride?: BenchmarkConfig,\n): MiddlewarePipeline {\n  const config = configOverride ?? loadBenchmarkConfig(options.installDir);\n\n  configureLogFile({\n    path: config.output.logPath,\n    enabled: config.output.logEnabled,\n  });\n\n  const plugins = createPluginsFromConfig(config, {\n    spinner,\n    phased: false,\n    outputPath: config.output.benchmarkPath,\n  });\n\n  if (!config.output.suppressWizardLogs) {\n    clack.log.info(\n      `${chalk.cyan(AgentSignals.BENCHMARK)} Verbose logs: ${getLogFilePath()}`,\n    );\n    clack.log.info(\n      `${chalk.cyan(\n        AgentSignals.BENCHMARK,\n      )} Benchmark data will be written to: ${config.output.benchmarkPath}`,\n    );\n  }\n\n  logToFile(\n    `${AgentSignals.BENCHMARK} Tracking enabled, starting with setup phase`,\n  );\n\n  return new MiddlewarePipeline(plugins, {\n    phaseDetector: new PhaseDetector(),\n    autoDetectPhases: true,\n  });\n}\n"]}

package/dist/src/lib/middleware/benchmarks/cache-tracker.d.ts

@@ -0,0 +1,44 @@
+/**
+ * Cache token tracking plugin (cache_read and cache_creation).
+ *
+ * Respects the dedup flag from TurnCounterPlugin.
+ */
+import type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';
+/** Matches SDK usage.cache_creation (ephemeral 5m vs 1h for pricing). */
+export interface CacheCreationBreakdown {
+    ephemeral_5m_input_tokens: number;
+    ephemeral_1h_input_tokens: number;
+}
+export interface CacheData {
+    phaseRead: number;
+    phaseCreation: number;
+    totalRead: number;
+    totalCreation: number;
+    totalCreation5m: number;
+    totalCreation1h: number;
+    phaseSnapshots: Array<{
+        phase: string;
+        cacheReadTokens: number;
+        cacheCreationTokens: number;
+        /** When present, use for pricing (5m vs 1h rates). */
+        cacheCreation5m: number;
+        cacheCreation1h: number;
+    }>;
+}
+export declare class CacheTrackerPlugin implements Middleware {
+    readonly name = "cache";
+    private phaseRead;
+    private phaseCreation;
+    private phaseCreation5m;
+    private phaseCreation1h;
+    private totalRead;
+    private totalCreation;
+    private totalCreation5m;
+    private totalCreation1h;
+    private phaseSnapshots;
+    private currentPhase;
+    onMessage(message: any, ctx: MiddlewareContext, store: MiddlewareStore): void;
+    onPhaseTransition(fromPhase: string, toPhase: string, _ctx: MiddlewareContext, store: MiddlewareStore): void;
+    onFinalize(_resultMessage: any, _totalDurationMs: number, _ctx: MiddlewareContext, store: MiddlewareStore): void;
+    private getData;
+}

package/dist/src/lib/middleware/benchmarks/cache-tracker.js

@@ -0,0 +1,81 @@
+"use strict";
+/**
+ * Cache token tracking plugin (cache_read and cache_creation).
+ *
+ * Respects the dedup flag from TurnCounterPlugin.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.CacheTrackerPlugin = void 0;
+class CacheTrackerPlugin {
+    name = 'cache';
+    phaseRead = 0;
+    phaseCreation = 0;
+    phaseCreation5m = 0;
+    phaseCreation1h = 0;
+    totalRead = 0;
+    totalCreation = 0;
+    totalCreation5m = 0;
+    totalCreation1h = 0;
+    phaseSnapshots = [];
+    currentPhase = 'setup';
+    onMessage(message, ctx, store) {
+        if (message.type !== 'assistant')
+            return;
+        const turns = ctx.get('turns');
+        if (turns?.isDuplicate)
+            return;
+        const usage = message.message?.usage;
+        if (usage) {
+            const read = Number(usage.cache_read_input_tokens ?? 0);
+            const creation = Number(usage.cache_creation_input_tokens ?? 0);
+            const cc = usage.cache_creation;
+            const creation5m = Number(cc?.ephemeral_5m_input_tokens ?? 0);
+            const creation1h = Number(cc?.ephemeral_1h_input_tokens ?? 0);
+            this.phaseRead += read;
+            this.phaseCreation += creation;
+            this.phaseCreation5m += creation5m;
+            this.phaseCreation1h += creation1h;
+            this.totalRead += read;
+            this.totalCreation += creation;
+        }
+        store.set('cache', this.getData());
+    }
+    onPhaseTransition(fromPhase, toPhase, _ctx, store) {
+        this.phaseSnapshots.push({
+            phase: fromPhase,
+            cacheReadTokens: this.phaseRead,
+            cacheCreationTokens: this.phaseCreation,
+            cacheCreation5m: this.phaseCreation5m,
+            cacheCreation1h: this.phaseCreation1h,
+        });
+        this.currentPhase = toPhase;
+        this.phaseRead = 0;
+        this.phaseCreation = 0;
+        this.phaseCreation5m = 0;
+        this.phaseCreation1h = 0;
+        store.set('cache', this.getData());
+    }
+    onFinalize(_resultMessage, _totalDurationMs, _ctx, store) {
+        this.phaseSnapshots.push({
+            phase: this.currentPhase,
+            cacheReadTokens: this.phaseRead,
+            cacheCreationTokens: this.phaseCreation,
+            cacheCreation5m: this.phaseCreation5m,
+            cacheCreation1h: this.phaseCreation1h,
+        });
+        store.set('cache', this.getData());
+    }
+    getData() {
+        return {
+            phaseRead: this.phaseRead,
+            phaseCreation: this.phaseCreation,
+            totalRead: this.totalRead,
+            totalCreation: this.totalCreation,
+            totalCreation5m: this.totalCreation5m,
+            totalCreation1h: this.totalCreation1h,
+            phaseSnapshots: [...this.phaseSnapshots],
+        };
+    }
+}
+exports.CacheTrackerPlugin = CacheTrackerPlugin;
+//# sourceMappingURL=cache-tracker.js.map

package/dist/src/lib/middleware/benchmarks/cache-tracker.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cache-tracker.js","sourceRoot":"","sources":["../../../../../src/lib/middleware/benchmarks/cache-tracker.ts"],"names":[],"mappings":";AAAA;;;;GAIG;;;AA4BH,MAAa,kBAAkB;IACpB,IAAI,GAAG,OAAO,CAAC;IAEhB,SAAS,GAAG,CAAC,CAAC;IACd,aAAa,GAAG,CAAC,CAAC;IAClB,eAAe,GAAG,CAAC,CAAC;IACpB,eAAe,GAAG,CAAC,CAAC;IACpB,SAAS,GAAG,CAAC,CAAC;IACd,aAAa,GAAG,CAAC,CAAC;IAClB,eAAe,GAAG,CAAC,CAAC;IACpB,eAAe,GAAG,CAAC,CAAC;IACpB,cAAc,GAMjB,EAAE,CAAC;IACA,YAAY,GAAG,OAAO,CAAC;IAE/B,SAAS,CACP,OAAY,EACZ,GAAsB,EACtB,KAAsB;QAEtB,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW;YAAE,OAAO;QAEzC,MAAM,KAAK,GAAG,GAAG,CAAC,GAAG,CAAW,OAAO,CAAC,CAAC;QACzC,IAAI,KAAK,EAAE,WAAW;YAAE,OAAO;QAE/B,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,EAAE,KAAK,CAAC;QACrC,IAAI,KAAK,EAAE,CAAC;YACV,MAAM,IAAI,GAAG,MAAM,CAAC,KAAK,CAAC,uBAAuB,IAAI,CAAC,CAAC,CAAC;YACxD,MAAM,QAAQ,GAAG,MAAM,CAAC,KAAK,CAAC,2BAA2B,IAAI,CAAC,CAAC,CAAC;YAChE,MAAM,EAAE,GAAG,KAAK,CAAC,cAAc,CAAC;YAChC,MAAM,UAAU,GAAG,MAAM,CAAC,EAAE,EAAE,yBAAyB,IAAI,CAAC,CAAC,CAAC;YAC9D,MAAM,UAAU,GAAG,MAAM,CAAC,EAAE,EAAE,yBAAyB,IAAI,CAAC,CAAC,CAAC;YAC9D,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC;YACvB,IAAI,CAAC,aAAa,IAAI,QAAQ,CAAC;YAC/B,IAAI,CAAC,eAAe,IAAI,UAAU,CAAC;YACnC,IAAI,CAAC,eAAe,IAAI,UAAU,CAAC;YACnC,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC;YACvB,IAAI,CAAC,aAAa,IAAI,QAAQ,CAAC;QACjC,CAAC;QAED,KAAK,CAAC,GAAG,CAAC,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IACrC,CAAC;IAED,iBAAiB,CACf,SAAiB,EACjB,OAAe,EACf,IAAuB,EACvB,KAAsB;QAEtB,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC;YACvB,KAAK,EAAE,SAAS;YAChB,eAAe,EAAE,IAAI,CAAC,SAAS;YAC/B,mBAAmB,EAAE,IAAI,CAAC,aAAa;YACvC,eAAe,EAAE,IAAI,CAAC,eAAe;YACrC,eAAe,EAAE,IAAI,CAAC,eAAe;SACtC,CAAC,CAAC;QACH,IAAI,CAAC,YAAY,GAAG,OAAO,CAAC;QAC5B,IAAI,CAAC,SAAS,GAAG,CAAC,CAAC;QACnB,IAAI,CAAC,aAAa,GAAG,CAAC,CAAC;QACvB,IAAI,CAAC,eAAe,GAAG,CAAC,CAAC;QACzB,IAAI,CAAC,eAAe,GAAG,CAAC,CAAC;QACzB,KAAK,CAAC,GAAG,CAAC,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IACrC,CAAC;IAED,UAAU,CACR,cAAmB,EACnB,gBAAwB,EACxB,IAAuB,EACvB,KAAsB;QAEtB,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC;YACvB,KAAK,EAAE,IAAI,CAAC,YAAY;YACxB,eAAe,EAAE,IAAI,CAAC,SAAS;YAC/B,mBAAmB,EAAE,IAAI,CAAC,aAAa;YACvC,eAAe,EAAE,IAAI,CAAC,eAAe;YACrC,eAAe,EAAE,IAAI,CAAC,eAAe;SACtC,CAAC,CAAC;QACH,KAAK,CAAC,GAAG,CAAC,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IACrC,CAAC;IAEO,OAAO;QACb,OAAO;YACL,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,aAAa,EAAE,IAAI,CAAC,aAAa;YACjC,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,aAAa,EAAE,IAAI,CAAC,aAAa;YACjC,eAAe,EAAE,IAAI,CAAC,eAAe;YACrC,eAAe,EAAE,IAAI,CAAC,eAAe;YACrC,cAAc,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC;SACzC,CAAC;IACJ,CAAC;CACF;AAhGD,gDAgGC","sourcesContent":["/**\n * Cache token tracking plugin (cache_read and cache_creation).\n *\n * Respects the dedup flag from TurnCounterPlugin.\n */\n\nimport type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';\nimport type { TurnData } from './turn-counter';\n\n/** Matches SDK usage.cache_creation (ephemeral 5m vs 1h for pricing). */\nexport interface CacheCreationBreakdown {\n  ephemeral_5m_input_tokens: number;\n  ephemeral_1h_input_tokens: number;\n}\n\nexport interface CacheData {\n  phaseRead: number;\n  phaseCreation: number;\n  totalRead: number;\n  totalCreation: number;\n  totalCreation5m: number;\n  totalCreation1h: number;\n  phaseSnapshots: Array<{\n    phase: string;\n    cacheReadTokens: number;\n    cacheCreationTokens: number;\n    /** When present, use for pricing (5m vs 1h rates). */\n    cacheCreation5m: number;\n    cacheCreation1h: number;\n  }>;\n}\n\nexport class CacheTrackerPlugin implements Middleware {\n  readonly name = 'cache';\n\n  private phaseRead = 0;\n  private phaseCreation = 0;\n  private phaseCreation5m = 0;\n  private phaseCreation1h = 0;\n  private totalRead = 0;\n  private totalCreation = 0;\n  private totalCreation5m = 0;\n  private totalCreation1h = 0;\n  private phaseSnapshots: Array<{\n    phase: string;\n    cacheReadTokens: number;\n    cacheCreationTokens: number;\n    cacheCreation5m: number;\n    cacheCreation1h: number;\n  }> = [];\n  private currentPhase = 'setup';\n\n  onMessage(\n    message: any,\n    ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    if (message.type !== 'assistant') return;\n\n    const turns = ctx.get<TurnData>('turns');\n    if (turns?.isDuplicate) return;\n\n    const usage = message.message?.usage;\n    if (usage) {\n      const read = Number(usage.cache_read_input_tokens ?? 0);\n      const creation = Number(usage.cache_creation_input_tokens ?? 0);\n      const cc = usage.cache_creation;\n      const creation5m = Number(cc?.ephemeral_5m_input_tokens ?? 0);\n      const creation1h = Number(cc?.ephemeral_1h_input_tokens ?? 0);\n      this.phaseRead += read;\n      this.phaseCreation += creation;\n      this.phaseCreation5m += creation5m;\n      this.phaseCreation1h += creation1h;\n      this.totalRead += read;\n      this.totalCreation += creation;\n    }\n\n    store.set('cache', this.getData());\n  }\n\n  onPhaseTransition(\n    fromPhase: string,\n    toPhase: string,\n    _ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    this.phaseSnapshots.push({\n      phase: fromPhase,\n      cacheReadTokens: this.phaseRead,\n      cacheCreationTokens: this.phaseCreation,\n      cacheCreation5m: this.phaseCreation5m,\n      cacheCreation1h: this.phaseCreation1h,\n    });\n    this.currentPhase = toPhase;\n    this.phaseRead = 0;\n    this.phaseCreation = 0;\n    this.phaseCreation5m = 0;\n    this.phaseCreation1h = 0;\n    store.set('cache', this.getData());\n  }\n\n  onFinalize(\n    _resultMessage: any,\n    _totalDurationMs: number,\n    _ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    this.phaseSnapshots.push({\n      phase: this.currentPhase,\n      cacheReadTokens: this.phaseRead,\n      cacheCreationTokens: this.phaseCreation,\n      cacheCreation5m: this.phaseCreation5m,\n      cacheCreation1h: this.phaseCreation1h,\n    });\n    store.set('cache', this.getData());\n  }\n\n  private getData(): CacheData {\n    return {\n      phaseRead: this.phaseRead,\n      phaseCreation: this.phaseCreation,\n      totalRead: this.totalRead,\n      totalCreation: this.totalCreation,\n      totalCreation5m: this.totalCreation5m,\n      totalCreation1h: this.totalCreation1h,\n      phaseSnapshots: [...this.phaseSnapshots],\n    };\n  }\n}\n"]}

package/dist/src/lib/middleware/benchmarks/compaction-tracker.d.ts

@@ -0,0 +1,29 @@
+/**
+ * Compaction event tracking plugin.
+ *
+ * Tracks context compaction events (compact_boundary system messages)
+ * including pre-compaction token counts per phase.
+ */
+import type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';
+export interface CompactionData {
+    phaseCompactions: number;
+    phasePreTokens: number[];
+    totalCompactions: number;
+    phaseSnapshots: Array<{
+        phase: string;
+        compactions: number;
+        preTokens: number[];
+    }>;
+}
+export declare class CompactionTrackerPlugin implements Middleware {
+    readonly name = "compactions";
+    private phaseCompactions;
+    private phasePreTokens;
+    private totalCompactions;
+    private phaseSnapshots;
+    private currentPhase;
+    onMessage(message: any, ctx: MiddlewareContext, store: MiddlewareStore): void;
+    onPhaseTransition(fromPhase: string, toPhase: string, _ctx: MiddlewareContext, store: MiddlewareStore): void;
+    onFinalize(_resultMessage: any, _totalDurationMs: number, _ctx: MiddlewareContext, store: MiddlewareStore): void;
+    private getData;
+}

package/dist/src/lib/middleware/benchmarks/compaction-tracker.js

@@ -0,0 +1,60 @@
+"use strict";
+/**
+ * Compaction event tracking plugin.
+ *
+ * Tracks context compaction events (compact_boundary system messages)
+ * including pre-compaction token counts per phase.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.CompactionTrackerPlugin = void 0;
+const debug_1 = require("../../../utils/debug");
+const agent_interface_1 = require("../../agent-interface");
+class CompactionTrackerPlugin {
+    name = 'compactions';
+    phaseCompactions = 0;
+    phasePreTokens = [];
+    totalCompactions = 0;
+    phaseSnapshots = [];
+    currentPhase = 'setup';
+    onMessage(message, ctx, store) {
+        if (message.type !== 'system' || message.subtype !== 'compact_boundary') {
+            return;
+        }
+        const preTokens = message.compact_metadata?.pre_tokens ?? 0;
+        const trigger = message.compact_metadata?.trigger ?? 'unknown';
+        this.phaseCompactions++;
+        this.totalCompactions++;
+        this.phasePreTokens.push(preTokens);
+        (0, debug_1.logToFile)(`${agent_interface_1.AgentSignals.BENCHMARK} [COMPACTION] Context compacted during "${ctx.currentPhase}" (trigger: ${trigger}, pre_tokens: ${preTokens})`);
+        store.set('compactions', this.getData());
+    }
+    onPhaseTransition(fromPhase, toPhase, _ctx, store) {
+        this.phaseSnapshots.push({
+            phase: fromPhase,
+            compactions: this.phaseCompactions,
+            preTokens: [...this.phasePreTokens],
+        });
+        this.currentPhase = toPhase;
+        this.phaseCompactions = 0;
+        this.phasePreTokens = [];
+        store.set('compactions', this.getData());
+    }
+    onFinalize(_resultMessage, _totalDurationMs, _ctx, store) {
+        this.phaseSnapshots.push({
+            phase: this.currentPhase,
+            compactions: this.phaseCompactions,
+            preTokens: [...this.phasePreTokens],
+        });
+        store.set('compactions', this.getData());
+    }
+    getData() {
+        return {
+            phaseCompactions: this.phaseCompactions,
+            phasePreTokens: [...this.phasePreTokens],
+            totalCompactions: this.totalCompactions,
+            phaseSnapshots: [...this.phaseSnapshots],
+        };
+    }
+}
+exports.CompactionTrackerPlugin = CompactionTrackerPlugin;
+//# sourceMappingURL=compaction-tracker.js.map

package/dist/src/lib/middleware/benchmarks/compaction-tracker.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"compaction-tracker.js","sourceRoot":"","sources":["../../../../../src/lib/middleware/benchmarks/compaction-tracker.ts"],"names":[],"mappings":";AAAA;;;;;GAKG;;;AAGH,gDAAiD;AACjD,2DAAqD;AAarD,MAAa,uBAAuB;IACzB,IAAI,GAAG,aAAa,CAAC;IAEtB,gBAAgB,GAAG,CAAC,CAAC;IACrB,cAAc,GAAa,EAAE,CAAC;IAC9B,gBAAgB,GAAG,CAAC,CAAC;IACrB,cAAc,GAIjB,EAAE,CAAC;IACA,YAAY,GAAG,OAAO,CAAC;IAE/B,SAAS,CACP,OAAY,EACZ,GAAsB,EACtB,KAAsB;QAEtB,IAAI,OAAO,CAAC,IAAI,KAAK,QAAQ,IAAI,OAAO,CAAC,OAAO,KAAK,kBAAkB,EAAE,CAAC;YACxE,OAAO;QACT,CAAC;QAED,MAAM,SAAS,GAAG,OAAO,CAAC,gBAAgB,EAAE,UAAU,IAAI,CAAC,CAAC;QAC5D,MAAM,OAAO,GAAG,OAAO,CAAC,gBAAgB,EAAE,OAAO,IAAI,SAAS,CAAC;QAC/D,IAAI,CAAC,gBAAgB,EAAE,CAAC;QACxB,IAAI,CAAC,gBAAgB,EAAE,CAAC;QACxB,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;QAEpC,IAAA,iBAAS,EACP,GAAG,8BAAY,CAAC,SAAS,2CAA2C,GAAG,CAAC,YAAY,eAAe,OAAO,iBAAiB,SAAS,GAAG,CACxI,CAAC;QAEF,KAAK,CAAC,GAAG,CAAC,aAAa,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IAC3C,CAAC;IAED,iBAAiB,CACf,SAAiB,EACjB,OAAe,EACf,IAAuB,EACvB,KAAsB;QAEtB,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC;YACvB,KAAK,EAAE,SAAS;YAChB,WAAW,EAAE,IAAI,CAAC,gBAAgB;YAClC,SAAS,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC;SACpC,CAAC,CAAC;QACH,IAAI,CAAC,YAAY,GAAG,OAAO,CAAC;QAC5B,IAAI,CAAC,gBAAgB,GAAG,CAAC,CAAC;QAC1B,IAAI,CAAC,cAAc,GAAG,EAAE,CAAC;QACzB,KAAK,CAAC,GAAG,CAAC,aAAa,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IAC3C,CAAC;IAED,UAAU,CACR,cAAmB,EACnB,gBAAwB,EACxB,IAAuB,EACvB,KAAsB;QAEtB,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC;YACvB,KAAK,EAAE,IAAI,CAAC,YAAY;YACxB,WAAW,EAAE,IAAI,CAAC,gBAAgB;YAClC,SAAS,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC;SACpC,CAAC,CAAC;QACH,KAAK,CAAC,GAAG,CAAC,aAAa,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IAC3C,CAAC;IAEO,OAAO;QACb,OAAO;YACL,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;YACvC,cAAc,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC;YACxC,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;YACvC,cAAc,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC;SACzC,CAAC;IACJ,CAAC;CACF;AA1ED,0DA0EC","sourcesContent":["/**\n * Compaction event tracking plugin.\n *\n * Tracks context compaction events (compact_boundary system messages)\n * including pre-compaction token counts per phase.\n */\n\nimport type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';\nimport { logToFile } from '../../../utils/debug';\nimport { AgentSignals } from '../../agent-interface';\n\nexport interface CompactionData {\n  phaseCompactions: number;\n  phasePreTokens: number[];\n  totalCompactions: number;\n  phaseSnapshots: Array<{\n    phase: string;\n    compactions: number;\n    preTokens: number[];\n  }>;\n}\n\nexport class CompactionTrackerPlugin implements Middleware {\n  readonly name = 'compactions';\n\n  private phaseCompactions = 0;\n  private phasePreTokens: number[] = [];\n  private totalCompactions = 0;\n  private phaseSnapshots: Array<{\n    phase: string;\n    compactions: number;\n    preTokens: number[];\n  }> = [];\n  private currentPhase = 'setup';\n\n  onMessage(\n    message: any,\n    ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    if (message.type !== 'system' || message.subtype !== 'compact_boundary') {\n      return;\n    }\n\n    const preTokens = message.compact_metadata?.pre_tokens ?? 0;\n    const trigger = message.compact_metadata?.trigger ?? 'unknown';\n    this.phaseCompactions++;\n    this.totalCompactions++;\n    this.phasePreTokens.push(preTokens);\n\n    logToFile(\n      `${AgentSignals.BENCHMARK} [COMPACTION] Context compacted during \"${ctx.currentPhase}\" (trigger: ${trigger}, pre_tokens: ${preTokens})`,\n    );\n\n    store.set('compactions', this.getData());\n  }\n\n  onPhaseTransition(\n    fromPhase: string,\n    toPhase: string,\n    _ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    this.phaseSnapshots.push({\n      phase: fromPhase,\n      compactions: this.phaseCompactions,\n      preTokens: [...this.phasePreTokens],\n    });\n    this.currentPhase = toPhase;\n    this.phaseCompactions = 0;\n    this.phasePreTokens = [];\n    store.set('compactions', this.getData());\n  }\n\n  onFinalize(\n    _resultMessage: any,\n    _totalDurationMs: number,\n    _ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    this.phaseSnapshots.push({\n      phase: this.currentPhase,\n      compactions: this.phaseCompactions,\n      preTokens: [...this.phasePreTokens],\n    });\n    store.set('compactions', this.getData());\n  }\n\n  private getData(): CompactionData {\n    return {\n      phaseCompactions: this.phaseCompactions,\n      phasePreTokens: [...this.phasePreTokens],\n      totalCompactions: this.totalCompactions,\n      phaseSnapshots: [...this.phaseSnapshots],\n    };\n  }\n}\n"]}

package/dist/src/lib/middleware/benchmarks/context-size-tracker.d.ts

@@ -0,0 +1,26 @@
+/**
+ * Context size tracking plugin (context tokens in/out per phase).
+ *
+ * Context tokens out = sum of input + cache_read + cache_creation from the
+ * last assistant message's usage (per-turn, NOT aggregate).
+ * Context tokens in = previous phase's context tokens out.
+ */
+import type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';
+export interface ContextSizeData {
+    /** Per-phase context size snapshots */
+    phaseSnapshots: Array<{
+        phase: string;
+        contextTokensIn?: number;
+        contextTokensOut?: number;
+        freshContext: boolean;
+    }>;
+}
+export declare class ContextSizeTrackerPlugin implements Middleware {
+    readonly name = "contextSize";
+    private phaseSnapshots;
+    private lastContextTokensOut?;
+    onPhaseTransition(fromPhase: string, _toPhase: string, ctx: MiddlewareContext, store: MiddlewareStore): void;
+    onFinalize(_resultMessage: any, _totalDurationMs: number, ctx: MiddlewareContext, store: MiddlewareStore): void;
+    private computeContextTokensOut;
+    private getData;
+}

package/dist/src/lib/middleware/benchmarks/context-size-tracker.js

@@ -0,0 +1,56 @@
+"use strict";
+/**
+ * Context size tracking plugin (context tokens in/out per phase).
+ *
+ * Context tokens out = sum of input + cache_read + cache_creation from the
+ * last assistant message's usage (per-turn, NOT aggregate).
+ * Context tokens in = previous phase's context tokens out.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.ContextSizeTrackerPlugin = void 0;
+class ContextSizeTrackerPlugin {
+    name = 'contextSize';
+    phaseSnapshots = [];
+    lastContextTokensOut;
+    onPhaseTransition(fromPhase, _toPhase, ctx, store) {
+        const tokens = ctx.get('tokens');
+        const contextTokensOut = this.computeContextTokensOut(tokens?.lastUsage);
+        this.phaseSnapshots.push({
+            phase: fromPhase,
+            contextTokensIn: ctx.currentPhaseFreshContext
+                ? undefined
+                : this.lastContextTokensOut,
+            contextTokensOut,
+            freshContext: ctx.currentPhaseFreshContext,
+        });
+        this.lastContextTokensOut = contextTokensOut;
+        store.set('contextSize', this.getData());
+    }
+    onFinalize(_resultMessage, _totalDurationMs, ctx, store) {
+        const tokens = ctx.get('tokens');
+        const contextTokensOut = this.computeContextTokensOut(tokens?.lastUsage);
+        this.phaseSnapshots.push({
+            phase: ctx.currentPhase,
+            contextTokensIn: ctx.currentPhaseFreshContext
+                ? undefined
+                : this.lastContextTokensOut,
+            contextTokensOut,
+            freshContext: ctx.currentPhaseFreshContext,
+        });
+        store.set('contextSize', this.getData());
+    }
+    computeContextTokensOut(usage) {
+        if (!usage)
+            return undefined;
+        return (Number(usage.input_tokens ?? 0) +
+            Number(usage.cache_read_input_tokens ?? 0) +
+            Number(usage.cache_creation_input_tokens ?? 0));
+    }
+    getData() {
+        return {
+            phaseSnapshots: [...this.phaseSnapshots],
+        };
+    }
+}
+exports.ContextSizeTrackerPlugin = ContextSizeTrackerPlugin;
+//# sourceMappingURL=context-size-tracker.js.map

package/dist/src/lib/middleware/benchmarks/context-size-tracker.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"context-size-tracker.js","sourceRoot":"","sources":["../../../../../src/lib/middleware/benchmarks/context-size-tracker.ts"],"names":[],"mappings":";AAAA;;;;;;GAMG;;;AAeH,MAAa,wBAAwB;IAC1B,IAAI,GAAG,aAAa,CAAC;IAEtB,cAAc,GAKjB,EAAE,CAAC;IACA,oBAAoB,CAAU;IAEtC,iBAAiB,CACf,SAAiB,EACjB,QAAgB,EAChB,GAAsB,EACtB,KAAsB;QAEtB,MAAM,MAAM,GAAG,GAAG,CAAC,GAAG,CAAY,QAAQ,CAAC,CAAC;QAC5C,MAAM,gBAAgB,GAAG,IAAI,CAAC,uBAAuB,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;QAEzE,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC;YACvB,KAAK,EAAE,SAAS;YAChB,eAAe,EAAE,GAAG,CAAC,wBAAwB;gBAC3C,CAAC,CAAC,SAAS;gBACX,CAAC,CAAC,IAAI,CAAC,oBAAoB;YAC7B,gBAAgB;YAChB,YAAY,EAAE,GAAG,CAAC,wBAAwB;SAC3C,CAAC,CAAC;QAEH,IAAI,CAAC,oBAAoB,GAAG,gBAAgB,CAAC;QAC7C,KAAK,CAAC,GAAG,CAAC,aAAa,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IAC3C,CAAC;IAED,UAAU,CACR,cAAmB,EACnB,gBAAwB,EACxB,GAAsB,EACtB,KAAsB;QAEtB,MAAM,MAAM,GAAG,GAAG,CAAC,GAAG,CAAY,QAAQ,CAAC,CAAC;QAC5C,MAAM,gBAAgB,GAAG,IAAI,CAAC,uBAAuB,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;QAEzE,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC;YACvB,KAAK,EAAE,GAAG,CAAC,YAAY;YACvB,eAAe,EAAE,GAAG,CAAC,wBAAwB;gBAC3C,CAAC,CAAC,SAAS;gBACX,CAAC,CAAC,IAAI,CAAC,oBAAoB;YAC7B,gBAAgB;YAChB,YAAY,EAAE,GAAG,CAAC,wBAAwB;SAC3C,CAAC,CAAC;QAEH,KAAK,CAAC,GAAG,CAAC,aAAa,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IAC3C,CAAC;IAEO,uBAAuB,CAAC,KAAU;QACxC,IAAI,CAAC,KAAK;YAAE,OAAO,SAAS,CAAC;QAC7B,OAAO,CACL,MAAM,CAAC,KAAK,CAAC,YAAY,IAAI,CAAC,CAAC;YAC/B,MAAM,CAAC,KAAK,CAAC,uBAAuB,IAAI,CAAC,CAAC;YAC1C,MAAM,CAAC,KAAK,CAAC,2BAA2B,IAAI,CAAC,CAAC,CAC/C,CAAC;IACJ,CAAC;IAEO,OAAO;QACb,OAAO;YACL,cAAc,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC;SACzC,CAAC;IACJ,CAAC;CACF;AApED,4DAoEC","sourcesContent":["/**\n * Context size tracking plugin (context tokens in/out per phase).\n *\n * Context tokens out = sum of input + cache_read + cache_creation from the\n * last assistant message's usage (per-turn, NOT aggregate).\n * Context tokens in = previous phase's context tokens out.\n */\n\nimport type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';\nimport type { TokenData } from './token-tracker';\n\nexport interface ContextSizeData {\n  /** Per-phase context size snapshots */\n  phaseSnapshots: Array<{\n    phase: string;\n    contextTokensIn?: number;\n    contextTokensOut?: number;\n    freshContext: boolean;\n  }>;\n}\n\nexport class ContextSizeTrackerPlugin implements Middleware {\n  readonly name = 'contextSize';\n\n  private phaseSnapshots: Array<{\n    phase: string;\n    contextTokensIn?: number;\n    contextTokensOut?: number;\n    freshContext: boolean;\n  }> = [];\n  private lastContextTokensOut?: number;\n\n  onPhaseTransition(\n    fromPhase: string,\n    _toPhase: string,\n    ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    const tokens = ctx.get<TokenData>('tokens');\n    const contextTokensOut = this.computeContextTokensOut(tokens?.lastUsage);\n\n    this.phaseSnapshots.push({\n      phase: fromPhase,\n      contextTokensIn: ctx.currentPhaseFreshContext\n        ? undefined\n        : this.lastContextTokensOut,\n      contextTokensOut,\n      freshContext: ctx.currentPhaseFreshContext,\n    });\n\n    this.lastContextTokensOut = contextTokensOut;\n    store.set('contextSize', this.getData());\n  }\n\n  onFinalize(\n    _resultMessage: any,\n    _totalDurationMs: number,\n    ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    const tokens = ctx.get<TokenData>('tokens');\n    const contextTokensOut = this.computeContextTokensOut(tokens?.lastUsage);\n\n    this.phaseSnapshots.push({\n      phase: ctx.currentPhase,\n      contextTokensIn: ctx.currentPhaseFreshContext\n        ? undefined\n        : this.lastContextTokensOut,\n      contextTokensOut,\n      freshContext: ctx.currentPhaseFreshContext,\n    });\n\n    store.set('contextSize', this.getData());\n  }\n\n  private computeContextTokensOut(usage: any): number | undefined {\n    if (!usage) return undefined;\n    return (\n      Number(usage.input_tokens ?? 0) +\n      Number(usage.cache_read_input_tokens ?? 0) +\n      Number(usage.cache_creation_input_tokens ?? 0)\n    );\n  }\n\n  private getData(): ContextSizeData {\n    return {\n      phaseSnapshots: [...this.phaseSnapshots],\n    };\n  }\n}\n"]}

package/dist/src/lib/middleware/benchmarks/cost-tracker.d.ts

@@ -0,0 +1,16 @@
+import type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';
+export interface CostData {
+    totalCost: number;
+    phaseCosts: Array<{
+        phase: string;
+        cost: number;
+    }>;
+}
+export declare class CostTrackerPlugin implements Middleware {
+    readonly name = "cost";
+    private phaseCosts;
+    private totalCost;
+    onPhaseTransition(fromPhase: string, _toPhase: string, ctx: MiddlewareContext, store: MiddlewareStore): void;
+    onFinalize(resultMessage: any, _totalDurationMs: number, ctx: MiddlewareContext, store: MiddlewareStore): void;
+    private getData;
+}

package/dist/src/lib/middleware/benchmarks/cost-tracker.js

@@ -0,0 +1,76 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.CostTrackerPlugin = void 0;
+/** Claude Sonnet 4.6 pricing (USD per 1M tokens) */
+const PRICE_PER_MTOK = {
+    input: 3,
+    output: 15,
+    cacheRead: 0.3,
+    cacheCreation5m: 3.75,
+    cacheCreation1h: 6,
+};
+function computeCost(inputTokens, outputTokens, cacheReadTokens, cacheCreation5m, cacheCreation1h, cacheCreationFallback) {
+    const hasBreakdown = cacheCreation5m > 0 || cacheCreation1h > 0;
+    return (inputTokens * (PRICE_PER_MTOK.input / 1e6) +
+        outputTokens * (PRICE_PER_MTOK.output / 1e6) +
+        cacheReadTokens * (PRICE_PER_MTOK.cacheRead / 1e6) +
+        (hasBreakdown
+            ? cacheCreation5m * (PRICE_PER_MTOK.cacheCreation5m / 1e6) +
+                cacheCreation1h * (PRICE_PER_MTOK.cacheCreation1h / 1e6)
+            : cacheCreationFallback * (PRICE_PER_MTOK.cacheCreation5m / 1e6)));
+}
+class CostTrackerPlugin {
+    name = 'cost';
+    phaseCosts = [];
+    totalCost = 0;
+    onPhaseTransition(fromPhase, _toPhase, ctx, store) {
+        const tokens = ctx.get('tokens');
+        const cache = ctx.get('cache');
+        const tokenSnap = tokens?.phaseSnapshots.at(-1);
+        const cacheSnap = cache?.phaseSnapshots.at(-1);
+        const totalIn = tokenSnap?.inputTokens ?? 0;
+        const read = cacheSnap?.cacheReadTokens ?? 0;
+        const creation = cacheSnap?.cacheCreationTokens ?? 0;
+        const c5m = cacheSnap?.cacheCreation5m ?? 0;
+        const c1h = cacheSnap?.cacheCreation1h ?? 0;
+        const baseIn = Math.max(0, totalIn - read - creation);
+        const phaseCost = computeCost(baseIn, tokenSnap?.outputTokens ?? 0, read, c5m, c1h, creation);
+        this.phaseCosts.push({ phase: fromPhase, cost: phaseCost });
+        this.totalCost += phaseCost;
+        store.set('cost', this.getData());
+    }
+    onFinalize(resultMessage, _totalDurationMs, ctx, store) {
+        const tokens = ctx.get('tokens');
+        const cache = ctx.get('cache');
+        const tokenSnap = tokens?.phaseSnapshots.at(-1);
+        const cacheSnap = cache?.phaseSnapshots.at(-1);
+        const totalIn = tokenSnap?.inputTokens ?? 0;
+        const read = cacheSnap?.cacheReadTokens ?? 0;
+        const creation = cacheSnap?.cacheCreationTokens ?? 0;
+        const c5m = cacheSnap?.cacheCreation5m ?? 0;
+        const c1h = cacheSnap?.cacheCreation1h ?? 0;
+        const baseIn = Math.max(0, totalIn - read - creation);
+        const lastPhaseCost = computeCost(baseIn, tokenSnap?.outputTokens ?? 0, read, c5m, c1h, creation);
+        this.phaseCosts.push({ phase: ctx.currentPhase, cost: lastPhaseCost });
+        this.totalCost += lastPhaseCost;
+        const sdkTotal = Number(resultMessage?.usage?.total_cost_usd ?? 0) ||
+            Number(resultMessage?.total_cost_usd ?? 0);
+        if (sdkTotal > 0 && this.totalCost > 0) {
+            const scale = sdkTotal / this.totalCost;
+            this.phaseCosts = this.phaseCosts.map((p) => ({
+                phase: p.phase,
+                cost: p.cost * scale,
+            }));
+            this.totalCost = sdkTotal;
+        }
+        store.set('cost', this.getData());
+    }
+    getData() {
+        return {
+            totalCost: this.totalCost,
+            phaseCosts: [...this.phaseCosts],
+        };
+    }
+}
+exports.CostTrackerPlugin = CostTrackerPlugin;
+//# sourceMappingURL=cost-tracker.js.map

package/dist/src/lib/middleware/benchmarks/cost-tracker.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cost-tracker.js","sourceRoot":"","sources":["../../../../../src/lib/middleware/benchmarks/cost-tracker.ts"],"names":[],"mappings":";;;AASA,oDAAoD;AACpD,MAAM,cAAc,GAAG;IACrB,KAAK,EAAE,CAAC;IACR,MAAM,EAAE,EAAE;IACV,SAAS,EAAE,GAAG;IACd,eAAe,EAAE,IAAI;IACrB,eAAe,EAAE,CAAC;CACV,CAAC;AAEX,SAAS,WAAW,CAClB,WAAmB,EACnB,YAAoB,EACpB,eAAuB,EACvB,eAAuB,EACvB,eAAuB,EACvB,qBAA6B;IAE7B,MAAM,YAAY,GAAG,eAAe,GAAG,CAAC,IAAI,eAAe,GAAG,CAAC,CAAC;IAChE,OAAO,CACL,WAAW,GAAG,CAAC,cAAc,CAAC,KAAK,GAAG,GAAG,CAAC;QAC1C,YAAY,GAAG,CAAC,cAAc,CAAC,MAAM,GAAG,GAAG,CAAC;QAC5C,eAAe,GAAG,CAAC,cAAc,CAAC,SAAS,GAAG,GAAG,CAAC;QAClD,CAAC,YAAY;YACX,CAAC,CAAC,eAAe,GAAG,CAAC,cAAc,CAAC,eAAe,GAAG,GAAG,CAAC;gBACxD,eAAe,GAAG,CAAC,cAAc,CAAC,eAAe,GAAG,GAAG,CAAC;YAC1D,CAAC,CAAC,qBAAqB,GAAG,CAAC,cAAc,CAAC,eAAe,GAAG,GAAG,CAAC,CAAC,CACpE,CAAC;AACJ,CAAC;AAED,MAAa,iBAAiB;IACnB,IAAI,GAAG,MAAM,CAAC;IAEf,UAAU,GAA2C,EAAE,CAAC;IACxD,SAAS,GAAG,CAAC,CAAC;IAEtB,iBAAiB,CACf,SAAiB,EACjB,QAAgB,EAChB,GAAsB,EACtB,KAAsB;QAEtB,MAAM,MAAM,GAAG,GAAG,CAAC,GAAG,CAAY,QAAQ,CAAC,CAAC;QAC5C,MAAM,KAAK,GAAG,GAAG,CAAC,GAAG,CAAY,OAAO,CAAC,CAAC;QAC1C,MAAM,SAAS,GAAG,MAAM,EAAE,cAAc,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAChD,MAAM,SAAS,GAAG,KAAK,EAAE,cAAc,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAE/C,MAAM,OAAO,GAAG,SAAS,EAAE,WAAW,IAAI,CAAC,CAAC;QAC5C,MAAM,IAAI,GAAG,SAAS,EAAE,eAAe,IAAI,CAAC,CAAC;QAC7C,MAAM,QAAQ,GAAG,SAAS,EAAE,mBAAmB,IAAI,CAAC,CAAC;QACrD,MAAM,GAAG,GAAG,SAAS,EAAE,eAAe,IAAI,CAAC,CAAC;QAC5C,MAAM,GAAG,GAAG,SAAS,EAAE,eAAe,IAAI,CAAC,CAAC;QAC5C,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,IAAI,GAAG,QAAQ,CAAC,CAAC;QAEtD,MAAM,SAAS,GAAG,WAAW,CAC3B,MAAM,EACN,SAAS,EAAE,YAAY,IAAI,CAAC,EAC5B,IAAI,EACJ,GAAG,EACH,GAAG,EACH,QAAQ,CACT,CAAC;QAEF,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,EAAE,SAAS,EAAE,CAAC,CAAC;QAC5D,IAAI,CAAC,SAAS,IAAI,SAAS,CAAC;QAC5B,KAAK,CAAC,GAAG,CAAC,MAAM,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IACpC,CAAC;IAED,UAAU,CACR,aAAkB,EAClB,gBAAwB,EACxB,GAAsB,EACtB,KAAsB;QAEtB,MAAM,MAAM,GAAG,GAAG,CAAC,GAAG,CAAY,QAAQ,CAAC,CAAC;QAC5C,MAAM,KAAK,GAAG,GAAG,CAAC,GAAG,CAAY,OAAO,CAAC,CAAC;QAC1C,MAAM,SAAS,GAAG,MAAM,EAAE,cAAc,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAChD,MAAM,SAAS,GAAG,KAAK,EAAE,cAAc,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAE/C,MAAM,OAAO,GAAG,SAAS,EAAE,WAAW,IAAI,CAAC,CAAC;QAC5C,MAAM,IAAI,GAAG,SAAS,EAAE,eAAe,IAAI,CAAC,CAAC;QAC7C,MAAM,QAAQ,GAAG,SAAS,EAAE,mBAAmB,IAAI,CAAC,CAAC;QACrD,MAAM,GAAG,GAAG,SAAS,EAAE,eAAe,IAAI,CAAC,CAAC;QAC5C,MAAM,GAAG,GAAG,SAAS,EAAE,eAAe,IAAI,CAAC,CAAC;QAC5C,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,GAAG,IAAI,GAAG,QAAQ,CAAC,CAAC;QAEtD,MAAM,aAAa,GAAG,WAAW,CAC/B,MAAM,EACN,SAAS,EAAE,YAAY,IAAI,CAAC,EAC5B,IAAI,EACJ,GAAG,EACH,GAAG,EACH,QAAQ,CACT,CAAC;QAEF,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,GAAG,CAAC,YAAY,EAAE,IAAI,EAAE,aAAa,EAAE,CAAC,CAAC;QACvE,IAAI,CAAC,SAAS,IAAI,aAAa,CAAC;QAEhC,MAAM,QAAQ,GACZ,MAAM,CAAC,aAAa,EAAE,KAAK,EAAE,cAAc,IAAI,CAAC,CAAC;YACjD,MAAM,CAAC,aAAa,EAAE,cAAc,IAAI,CAAC,CAAC,CAAC;QAE7C,IAAI,QAAQ,GAAG,CAAC,IAAI,IAAI,CAAC,SAAS,GAAG,CAAC,EAAE,CAAC;YACvC,MAAM,KAAK,GAAG,QAAQ,GAAG,IAAI,CAAC,SAAS,CAAC;YACxC,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;gBAC5C,KAAK,EAAE,CAAC,CAAC,KAAK;gBACd,IAAI,EAAE,CAAC,CAAC,IAAI,GAAG,KAAK;aACrB,CAAC,CAAC,CAAC;YACJ,IAAI,CAAC,SAAS,GAAG,QAAQ,CAAC;QAC5B,CAAC;QAED,KAAK,CAAC,GAAG,CAAC,MAAM,EAAE,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC;IACpC,CAAC;IAEO,OAAO;QACb,OAAO;YACL,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,UAAU,EAAE,CAAC,GAAG,IAAI,CAAC,UAAU,CAAC;SACjC,CAAC;IACJ,CAAC;CACF;AA1FD,8CA0FC","sourcesContent":["import type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';\nimport type { TokenData } from './token-tracker';\nimport type { CacheData } from './cache-tracker';\n\nexport interface CostData {\n  totalCost: number;\n  phaseCosts: Array<{ phase: string; cost: number }>;\n}\n\n/** Claude Sonnet 4.6 pricing (USD per 1M tokens) */\nconst PRICE_PER_MTOK = {\n  input: 3,\n  output: 15,\n  cacheRead: 0.3,\n  cacheCreation5m: 3.75,\n  cacheCreation1h: 6,\n} as const;\n\nfunction computeCost(\n  inputTokens: number,\n  outputTokens: number,\n  cacheReadTokens: number,\n  cacheCreation5m: number,\n  cacheCreation1h: number,\n  cacheCreationFallback: number,\n): number {\n  const hasBreakdown = cacheCreation5m > 0 || cacheCreation1h > 0;\n  return (\n    inputTokens * (PRICE_PER_MTOK.input / 1e6) +\n    outputTokens * (PRICE_PER_MTOK.output / 1e6) +\n    cacheReadTokens * (PRICE_PER_MTOK.cacheRead / 1e6) +\n    (hasBreakdown\n      ? cacheCreation5m * (PRICE_PER_MTOK.cacheCreation5m / 1e6) +\n        cacheCreation1h * (PRICE_PER_MTOK.cacheCreation1h / 1e6)\n      : cacheCreationFallback * (PRICE_PER_MTOK.cacheCreation5m / 1e6))\n  );\n}\n\nexport class CostTrackerPlugin implements Middleware {\n  readonly name = 'cost';\n\n  private phaseCosts: Array<{ phase: string; cost: number }> = [];\n  private totalCost = 0;\n\n  onPhaseTransition(\n    fromPhase: string,\n    _toPhase: string,\n    ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    const tokens = ctx.get<TokenData>('tokens');\n    const cache = ctx.get<CacheData>('cache');\n    const tokenSnap = tokens?.phaseSnapshots.at(-1);\n    const cacheSnap = cache?.phaseSnapshots.at(-1);\n\n    const totalIn = tokenSnap?.inputTokens ?? 0;\n    const read = cacheSnap?.cacheReadTokens ?? 0;\n    const creation = cacheSnap?.cacheCreationTokens ?? 0;\n    const c5m = cacheSnap?.cacheCreation5m ?? 0;\n    const c1h = cacheSnap?.cacheCreation1h ?? 0;\n    const baseIn = Math.max(0, totalIn - read - creation);\n\n    const phaseCost = computeCost(\n      baseIn,\n      tokenSnap?.outputTokens ?? 0,\n      read,\n      c5m,\n      c1h,\n      creation,\n    );\n\n    this.phaseCosts.push({ phase: fromPhase, cost: phaseCost });\n    this.totalCost += phaseCost;\n    store.set('cost', this.getData());\n  }\n\n  onFinalize(\n    resultMessage: any,\n    _totalDurationMs: number,\n    ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    const tokens = ctx.get<TokenData>('tokens');\n    const cache = ctx.get<CacheData>('cache');\n    const tokenSnap = tokens?.phaseSnapshots.at(-1);\n    const cacheSnap = cache?.phaseSnapshots.at(-1);\n\n    const totalIn = tokenSnap?.inputTokens ?? 0;\n    const read = cacheSnap?.cacheReadTokens ?? 0;\n    const creation = cacheSnap?.cacheCreationTokens ?? 0;\n    const c5m = cacheSnap?.cacheCreation5m ?? 0;\n    const c1h = cacheSnap?.cacheCreation1h ?? 0;\n    const baseIn = Math.max(0, totalIn - read - creation);\n\n    const lastPhaseCost = computeCost(\n      baseIn,\n      tokenSnap?.outputTokens ?? 0,\n      read,\n      c5m,\n      c1h,\n      creation,\n    );\n\n    this.phaseCosts.push({ phase: ctx.currentPhase, cost: lastPhaseCost });\n    this.totalCost += lastPhaseCost;\n\n    const sdkTotal =\n      Number(resultMessage?.usage?.total_cost_usd ?? 0) ||\n      Number(resultMessage?.total_cost_usd ?? 0);\n\n    if (sdkTotal > 0 && this.totalCost > 0) {\n      const scale = sdkTotal / this.totalCost;\n      this.phaseCosts = this.phaseCosts.map((p) => ({\n        phase: p.phase,\n        cost: p.cost * scale,\n      }));\n      this.totalCost = sdkTotal;\n    }\n\n    store.set('cost', this.getData());\n  }\n\n  private getData(): CostData {\n    return {\n      totalCost: this.totalCost,\n      phaseCosts: [...this.phaseCosts],\n    };\n  }\n}\n"]}

package/dist/src/lib/middleware/benchmarks/duration-tracker.d.ts

@@ -0,0 +1,20 @@
+/**
+ * Duration tracking plugin (per-phase and total).
+ */
+import type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';
+export interface DurationData {
+    phaseSnapshots: Array<{
+        phase: string;
+        startTime: number;
+        endTime: number;
+        durationMs: number;
+    }>;
+    totalDurationMs: number;
+}
+export declare class DurationTrackerPlugin implements Middleware {
+    readonly name = "duration";
+    private phaseStartTime;
+    private phaseSnapshots;
+    onPhaseTransition(fromPhase: string, _toPhase: string, _ctx: MiddlewareContext, store: MiddlewareStore): void;
+    onFinalize(_resultMessage: any, totalDurationMs: number, ctx: MiddlewareContext, store: MiddlewareStore): void;
+}

package/dist/src/lib/middleware/benchmarks/duration-tracker.js

@@ -0,0 +1,40 @@
+"use strict";
+/**
+ * Duration tracking plugin (per-phase and total).
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.DurationTrackerPlugin = void 0;
+class DurationTrackerPlugin {
+    name = 'duration';
+    phaseStartTime = Date.now();
+    phaseSnapshots = [];
+    onPhaseTransition(fromPhase, _toPhase, _ctx, store) {
+        const now = Date.now();
+        this.phaseSnapshots.push({
+            phase: fromPhase,
+            startTime: this.phaseStartTime,
+            endTime: now,
+            durationMs: now - this.phaseStartTime,
+        });
+        this.phaseStartTime = now;
+        store.set('duration', {
+            phaseSnapshots: [...this.phaseSnapshots],
+            totalDurationMs: 0,
+        });
+    }
+    onFinalize(_resultMessage, totalDurationMs, ctx, store) {
+        const now = Date.now();
+        this.phaseSnapshots.push({
+            phase: ctx.currentPhase,
+            startTime: this.phaseStartTime,
+            endTime: now,
+            durationMs: now - this.phaseStartTime,
+        });
+        store.set('duration', {
+            phaseSnapshots: [...this.phaseSnapshots],
+            totalDurationMs,
+        });
+    }
+}
+exports.DurationTrackerPlugin = DurationTrackerPlugin;
+//# sourceMappingURL=duration-tracker.js.map

package/dist/src/lib/middleware/benchmarks/duration-tracker.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"duration-tracker.js","sourceRoot":"","sources":["../../../../../src/lib/middleware/benchmarks/duration-tracker.ts"],"names":[],"mappings":";AAAA;;GAEG;;;AAcH,MAAa,qBAAqB;IACvB,IAAI,GAAG,UAAU,CAAC;IAEnB,cAAc,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;IAC5B,cAAc,GAKjB,EAAE,CAAC;IAER,iBAAiB,CACf,SAAiB,EACjB,QAAgB,EAChB,IAAuB,EACvB,KAAsB;QAEtB,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QACvB,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC;YACvB,KAAK,EAAE,SAAS;YAChB,SAAS,EAAE,IAAI,CAAC,cAAc;YAC9B,OAAO,EAAE,GAAG;YACZ,UAAU,EAAE,GAAG,GAAG,IAAI,CAAC,cAAc;SACtC,CAAC,CAAC;QACH,IAAI,CAAC,cAAc,GAAG,GAAG,CAAC;QAC1B,KAAK,CAAC,GAAG,CAAC,UAAU,EAAE;YACpB,cAAc,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC;YACxC,eAAe,EAAE,CAAC;SACI,CAAC,CAAC;IAC5B,CAAC;IAED,UAAU,CACR,cAAmB,EACnB,eAAuB,EACvB,GAAsB,EACtB,KAAsB;QAEtB,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QACvB,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC;YACvB,KAAK,EAAE,GAAG,CAAC,YAAY;YACvB,SAAS,EAAE,IAAI,CAAC,cAAc;YAC9B,OAAO,EAAE,GAAG;YACZ,UAAU,EAAE,GAAG,GAAG,IAAI,CAAC,cAAc;SACtC,CAAC,CAAC;QAEH,KAAK,CAAC,GAAG,CAAC,UAAU,EAAE;YACpB,cAAc,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC;YACxC,eAAe;SACO,CAAC,CAAC;IAC5B,CAAC;CACF;AAlDD,sDAkDC","sourcesContent":["/**\n * Duration tracking plugin (per-phase and total).\n */\n\nimport type { Middleware, MiddlewareContext, MiddlewareStore } from '../types';\n\nexport interface DurationData {\n  phaseSnapshots: Array<{\n    phase: string;\n    startTime: number;\n    endTime: number;\n    durationMs: number;\n  }>;\n  totalDurationMs: number;\n}\n\nexport class DurationTrackerPlugin implements Middleware {\n  readonly name = 'duration';\n\n  private phaseStartTime = Date.now();\n  private phaseSnapshots: Array<{\n    phase: string;\n    startTime: number;\n    endTime: number;\n    durationMs: number;\n  }> = [];\n\n  onPhaseTransition(\n    fromPhase: string,\n    _toPhase: string,\n    _ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    const now = Date.now();\n    this.phaseSnapshots.push({\n      phase: fromPhase,\n      startTime: this.phaseStartTime,\n      endTime: now,\n      durationMs: now - this.phaseStartTime,\n    });\n    this.phaseStartTime = now;\n    store.set('duration', {\n      phaseSnapshots: [...this.phaseSnapshots],\n      totalDurationMs: 0,\n    } satisfies DurationData);\n  }\n\n  onFinalize(\n    _resultMessage: any,\n    totalDurationMs: number,\n    ctx: MiddlewareContext,\n    store: MiddlewareStore,\n  ): void {\n    const now = Date.now();\n    this.phaseSnapshots.push({\n      phase: ctx.currentPhase,\n      startTime: this.phaseStartTime,\n      endTime: now,\n      durationMs: now - this.phaseStartTime,\n    });\n\n    store.set('duration', {\n      phaseSnapshots: [...this.phaseSnapshots],\n      totalDurationMs,\n    } satisfies DurationData);\n  }\n}\n"]}

package/dist/src/lib/middleware/benchmarks/index.d.ts

@@ -0,0 +1,9 @@
+/**
+ * Plugin registry and factory.
+ *
+ * Maps plugin names to constructors and creates the ordered plugin list
+ * from a BenchmarkConfig.
+ */
+import type { Middleware, MiddlewareFactoryOptions } from '../types';
+import type { BenchmarkConfig } from '../config';
+export declare function createPluginsFromConfig(config: BenchmarkConfig, opts: MiddlewareFactoryOptions): Middleware[];