@tangle-network/agent-eval 0.92.0 → 0.93.0

package/dist/{chunk-RTWFUK6A.js → chunk-E4GH6USR.js}

@@ -1,7 +1,7 @@
 import {
   objectiveEval,
   runAgentControlLoop
-} from "./chunk-NCRFYPS3.js";
+} from "./chunk-YEHAEDUD.js";
 import {
   validateRunRecord
 } from "./chunk-KWRRMR3J.js";
@@ -610,4 +610,4 @@ export {
   runProposeReviewAsControlLoop,
   controlFailureClassFromVerification
 };
-//# sourceMappingURL=chunk-RTWFUK6A.js.map
+//# sourceMappingURL=chunk-E4GH6USR.js.map

package/dist/{chunk-3CKU6VGU.js → chunk-Q2JRAWRI.js}

@@ -1,6 +1,6 @@
 import {
   objectiveEval
-} from "./chunk-NCRFYPS3.js";
+} from "./chunk-YEHAEDUD.js";
 
 // src/knowledge/readiness.ts
 function scoreKnowledgeReadiness(options) {
@@ -193,4 +193,4 @@ export {
   userQuestionsForKnowledgeGaps,
   acquisitionPlansForKnowledgeGaps
 };
-//# sourceMappingURL=chunk-3CKU6VGU.js.map
+//# sourceMappingURL=chunk-Q2JRAWRI.js.map

package/dist/{chunk-NCRFYPS3.js → chunk-YEHAEDUD.js}

@@ -2,6 +2,112 @@ import {
   TraceEmitter
 } from "./chunk-TVVP3ZZQ.js";
 
+// src/detectors/index.ts
+function repeatedActionDetector(opts = {}) {
+  const max = opts.maxRepeated ?? 3;
+  const severity = opts.severity ?? "warn";
+  const failureClass = opts.failureClass ?? "tool_recovery_failure";
+  let last;
+  let streak = 0;
+  return {
+    id: "repeated-action",
+    get streak() {
+      return streak;
+    },
+    observe(event) {
+      const fp = event.actionFingerprint;
+      if (fp === void 0) return null;
+      streak = fp === last ? streak + 1 : 1;
+      last = fp;
+      if (max <= 0 || streak < max) return null;
+      return {
+        detector: "repeated-action",
+        severity,
+        failureClass,
+        reason: `stuck: repeated same action for ${streak} step(s)`,
+        streak,
+        ...event.label ? { evidence: { action: event.label } } : {}
+      };
+    },
+    reset() {
+      last = void 0;
+      streak = 0;
+    }
+  };
+}
+function noProgressDetector(opts = {}) {
+  const max = opts.maxNoProgress ?? 3;
+  const minScoreDelta = opts.minScoreDelta ?? 1e-3;
+  const severity = opts.severity ?? "warn";
+  const failureClass = opts.failureClass ?? "tool_recovery_failure";
+  let lastState;
+  let lastScore;
+  let streak = 0;
+  return {
+    id: "no-progress",
+    get streak() {
+      return streak;
+    },
+    observe(event) {
+      const stateUnchanged = lastState !== void 0 && lastState === event.stateFingerprint;
+      const scoreFlat = Math.abs((event.score ?? 0) - (lastScore ?? 0)) < minScoreDelta;
+      streak = stateUnchanged && scoreFlat ? streak + 1 : 0;
+      lastState = event.stateFingerprint;
+      lastScore = event.score;
+      if (max <= 0 || streak < max) return null;
+      return {
+        detector: "no-progress",
+        severity,
+        failureClass,
+        reason: `stuck: no state/score progress for ${streak} step(s)`,
+        streak,
+        ...event.label ? { evidence: { state: event.label } } : {}
+      };
+    },
+    reset() {
+      lastState = void 0;
+      lastScore = void 0;
+      streak = 0;
+    }
+  };
+}
+function errorStreakDetector(opts = {}) {
+  const max = opts.maxErrors ?? 3;
+  const severity = opts.severity ?? "warn";
+  const failureClass = opts.failureClass ?? "tool_recovery_failure";
+  let streak = 0;
+  return {
+    id: "error-streak",
+    get streak() {
+      return streak;
+    },
+    observe(event) {
+      if (event.status === void 0) return null;
+      streak = event.status === "error" ? streak + 1 : 0;
+      if (max <= 0 || streak < max) return null;
+      return {
+        detector: "error-streak",
+        severity,
+        failureClass,
+        reason: `stuck: ${streak} consecutive errored step(s)`,
+        streak,
+        ...event.label ? { evidence: { lastError: event.label } } : {}
+      };
+    },
+    reset() {
+      streak = 0;
+    }
+  };
+}
+function observeAll(detectors, event) {
+  const signals = [];
+  for (const d of detectors) {
+    const s = d.observe(event);
+    if (s) signals.push(s);
+  }
+  return signals;
+}
+
 // src/control-runtime.ts
 var DEFAULT_BUDGET = {
   maxSteps: 8,
@@ -25,10 +131,13 @@ async function runAgentControlLoop(config) {
   const emitter = config.store ? new TraceEmitter(config.store) : void 0;
   let spentCostUsd = 0;
   const runtimeErrors = [];
-  let lastStateFingerprint;
-  let lastActionFingerprint;
-  let noProgressStreak = 0;
-  let repeatedActionStreak = 0;
+  const repeatedDetector = repeatedActionDetector({
+    maxRepeated: config.stopPolicies?.maxRepeatedActions ?? 0
+  });
+  const progressDetector = noProgressDetector({
+    maxNoProgress: config.stopPolicies?.maxNoProgressSteps ?? 0,
+    minScoreDelta: config.stopPolicies?.minScoreDelta ?? 1e-3
+  });
   try {
     if (emitter) {
       await runTrace(
@@ -97,7 +206,10 @@ async function runAgentControlLoop(config) {
         stoppedBy: "runtime-error"
       });
     }
-    lastStateFingerprint = fingerprintState(state, config.stopPolicies);
+    progressDetector.observe({
+      stateFingerprint: fingerprintState(state, config.stopPolicies),
+      score: averageScore(evals)
+    });
     for (let stepIndex = 0; stepIndex < budget.maxSteps; stepIndex++) {
       if (controller.signal.aborted) {
         return finish(emitter, {
@@ -229,18 +341,14 @@ async function runAgentControlLoop(config) {
         });
       }
       const actionFingerprint = fingerprintAction(decision.action, config.stopPolicies);
-      repeatedActionStreak = actionFingerprint === lastActionFingerprint ? repeatedActionStreak + 1 : 1;
-      lastActionFingerprint = actionFingerprint;
-      const repeatedActionStop = repeatedActionStopDecision(
-        config.stopPolicies,
-        repeatedActionStreak
-      );
-      if (repeatedActionStop.stop) {
+      const repeatedActionSignal = repeatedDetector.observe({ actionFingerprint });
+      const repeatedActionStreak = repeatedDetector.streak;
+      if (repeatedActionSignal) {
         return finish(emitter, {
           intent: config.intent,
           pass: false,
           completed: true,
-          reason: repeatedActionStop.reason,
+          reason: repeatedActionSignal.reason,
           score: averageScore(evals),
           steps: history,
           finalState: state,
@@ -435,16 +543,8 @@ async function runAgentControlLoop(config) {
       }
       const scoreAfter = averageScore(evals);
       const stateFingerprint = fingerprintState(state, config.stopPolicies);
-      const noProgressStop = noProgressStopDecision({
-        policies: config.stopPolicies,
-        lastStateFingerprint,
-        stateFingerprint,
-        scoreBefore,
-        scoreAfter,
-        currentStreak: noProgressStreak
-      });
-      noProgressStreak = noProgressStop.streak;
-      lastStateFingerprint = stateFingerprint;
+      const noProgressSignal = progressDetector.observe({ stateFingerprint, score: scoreAfter });
+      const noProgressStreak = progressDetector.streak;
       const step = {
         index: stepIndex,
         decision,
@@ -484,12 +584,12 @@ async function runAgentControlLoop(config) {
         );
       }
       await runOnStep(config.onStep, step, runtimeErrors);
-      if (noProgressStop.stop) {
+      if (noProgressSignal) {
         return finish(emitter, {
           intent: config.intent,
           pass: false,
           completed: true,
-          reason: noProgressStop.reason,
+          reason: noProgressSignal.reason,
           score: scoreAfter,
           steps: history,
           finalState: state,
@@ -759,24 +859,6 @@ async function runTrace(runtimeErrors, stepIndex, write) {
     return void 0;
   }
 }
-function noProgressStopDecision(args) {
-  const max = args.policies?.maxNoProgressSteps;
-  if (!max || max <= 0) return { stop: false, reason: "", streak: 0 };
-  const minScoreDelta = args.policies?.minScoreDelta ?? 1e-3;
-  const scoreDelta = Math.abs((args.scoreAfter ?? 0) - (args.scoreBefore ?? 0));
-  const stateUnchanged = args.lastStateFingerprint !== void 0 && args.lastStateFingerprint === args.stateFingerprint;
-  const scoreFlat = scoreDelta < minScoreDelta;
-  const streak = stateUnchanged && scoreFlat ? args.currentStreak + 1 : 0;
-  return streak >= max ? { stop: true, reason: `stuck: no state/score progress for ${streak} step(s)`, streak } : { stop: false, reason: "", streak };
-}
-function repeatedActionStopDecision(policies, streak) {
-  const max = policies?.maxRepeatedActions;
-  if (!max || max <= 0 || streak < max) return { stop: false, reason: "" };
-  return {
-    stop: true,
-    reason: `stuck: repeated same action for ${streak} step(s)`
-  };
-}
 function fingerprintState(state, policies) {
   if (policies?.stateFingerprint) return policies.stateFingerprint(state);
   return stableFingerprint(state);
@@ -828,6 +910,10 @@ async function finish(emitter, result) {
 }
 
 export {
+  repeatedActionDetector,
+  noProgressDetector,
+  errorStreakDetector,
+  observeAll,
   runAgentControlLoop,
   stopOnNoProgress,
   stopOnRepeatedAction,
@@ -835,4 +921,4 @@ export {
   subjectiveEval,
   allCriticalPassed
 };
-//# sourceMappingURL=chunk-NCRFYPS3.js.map
+//# sourceMappingURL=chunk-YEHAEDUD.js.map

package/dist/chunk-YEHAEDUD.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/detectors/index.ts","../src/control-runtime.ts"],"sourcesContent":["/**\n * Streaming detectors — the canonical, online failure-mode kernel.\n *\n * A `StreamingDetector` is a pure incremental reducer: fold one normalized step, get a signal the\n * moment a threshold trips (else null). The SAME kernel runs in two places, so the logic lives once:\n *   - ONLINE, over a live agent pipe (a worker's tool-call stream) → raise a finding mid-run.\n *   - INSIDE the control loop (`control-runtime`) → its `stopOn*` policies fold these detectors.\n *\n * Detectors are stateful but self-contained; `streak` is exposed for telemetry and `reset` clears it.\n * Fingerprinting is the CALLER's job — the control loop hashes state/action with `stableFingerprint`,\n * a tool-call pipe hashes args with `argHash` — so a detector stays agnostic to what it's watching.\n */\n\nimport type { FailureClass } from '../trace/schema'\n\nexport type DetectorSeverity = 'info' | 'warn' | 'critical'\n\n/** A normalized step a detector folds over. Fields are optional; each detector reads only what it\n *  needs (repeated-action → `actionFingerprint`; no-progress → `stateFingerprint` + `score`;\n *  error-streak → `status`). */\nexport interface DetectorEvent {\n  /** Fingerprint of the action/tool-call this step took (caller pre-hashes). */\n  readonly actionFingerprint?: string\n  /** Fingerprint of observable state AFTER this step. */\n  readonly stateFingerprint?: string\n  /** Score after this step (the score-flat half of no-progress). */\n  readonly score?: number\n  /** Whether this step errored. */\n  readonly status?: 'ok' | 'error'\n  /** Free-form label carried into the signal evidence (e.g. the tool name). */\n  readonly label?: string\n}\n\nexport interface DetectorSignal {\n  readonly detector: string\n  readonly severity: DetectorSeverity\n  readonly failureClass?: FailureClass\n  readonly reason: string\n  /** Consecutive matching steps at the moment the signal fired. */\n  readonly streak: number\n  readonly evidence?: Record<string, unknown>\n}\n\nexport interface StreamingDetector {\n  readonly id: string\n  /** Current streak (consecutive matching steps) — for telemetry/introspection between observes. */\n  readonly streak: number\n  /** Fold one event; return a signal when the threshold trips, else null. */\n  observe(event: DetectorEvent): DetectorSignal | null\n  /** Clear all state. */\n  reset(): void\n}\n\nexport interface RepeatedActionOptions {\n  /** Signal once the same action fingerprint repeats this many CONSECUTIVE steps (default 3).\n   *  `<= 0` disables the signal (the streak is still tracked, for telemetry). */\n  readonly maxRepeated?: number\n  readonly severity?: DetectorSeverity\n  /** Failure class to stamp on the signal (default `tool_recovery_failure`). */\n  readonly failureClass?: FailureClass\n}\n\n/** Same action fingerprint N consecutive steps = a stuck loop (the #1 long-horizon failure mode). */\nexport function repeatedActionDetector(opts: RepeatedActionOptions = {}): StreamingDetector {\n  const max = opts.maxRepeated ?? 3\n  const severity = opts.severity ?? 'warn'\n  const failureClass: FailureClass = opts.failureClass ?? 'tool_recovery_failure'\n  let last: string | undefined\n  let streak = 0\n  return {\n    id: 'repeated-action',\n    get streak() {\n      return streak\n    },\n    observe(event) {\n      const fp = event.actionFingerprint\n      if (fp === undefined) return null\n      streak = fp === last ? streak + 1 : 1\n      last = fp\n      if (max <= 0 || streak < max) return null\n      return {\n        detector: 'repeated-action',\n        severity,\n        failureClass,\n        reason: `stuck: repeated same action for ${streak} step(s)`,\n        streak,\n        ...(event.label ? { evidence: { action: event.label } } : {}),\n      }\n    },\n    reset() {\n      last = undefined\n      streak = 0\n    },\n  }\n}\n\nexport interface NoProgressOptions {\n  /** Signal once state is unchanged AND score is flat for this many CONSECUTIVE steps (default 3).\n   *  `<= 0` disables the signal. */\n  readonly maxNoProgress?: number\n  /** Minimum |score change| that counts as progress (default 0.001). */\n  readonly minScoreDelta?: number\n  readonly severity?: DetectorSeverity\n  readonly failureClass?: FailureClass\n}\n\n/** State + score unchanged across N steps = spinning wheels. Compares each step to the previous one,\n *  so prime with the initial state (observe it once before the first real step) to detect on step 1. */\nexport function noProgressDetector(opts: NoProgressOptions = {}): StreamingDetector {\n  const max = opts.maxNoProgress ?? 3\n  const minScoreDelta = opts.minScoreDelta ?? 0.001\n  const severity = opts.severity ?? 'warn'\n  const failureClass: FailureClass = opts.failureClass ?? 'tool_recovery_failure'\n  let lastState: string | undefined\n  let lastScore: number | undefined\n  let streak = 0\n  return {\n    id: 'no-progress',\n    get streak() {\n      return streak\n    },\n    observe(event) {\n      const stateUnchanged = lastState !== undefined && lastState === event.stateFingerprint\n      const scoreFlat = Math.abs((event.score ?? 0) - (lastScore ?? 0)) < minScoreDelta\n      streak = stateUnchanged && scoreFlat ? streak + 1 : 0\n      lastState = event.stateFingerprint\n      lastScore = event.score\n      if (max <= 0 || streak < max) return null\n      return {\n        detector: 'no-progress',\n        severity,\n        failureClass,\n        reason: `stuck: no state/score progress for ${streak} step(s)`,\n        streak,\n        ...(event.label ? { evidence: { state: event.label } } : {}),\n      }\n    },\n    reset() {\n      lastState = undefined\n      lastScore = undefined\n      streak = 0\n    },\n  }\n}\n\nexport interface ErrorStreakOptions {\n  /** Signal once this many CONSECUTIVE steps errored (default 3). `<= 0` disables. */\n  readonly maxErrors?: number\n  readonly severity?: DetectorSeverity\n  readonly failureClass?: FailureClass\n}\n\n/** N consecutive tool errors = the worker is hammering a broken approach. */\nexport function errorStreakDetector(opts: ErrorStreakOptions = {}): StreamingDetector {\n  const max = opts.maxErrors ?? 3\n  const severity = opts.severity ?? 'warn'\n  const failureClass: FailureClass = opts.failureClass ?? 'tool_recovery_failure'\n  let streak = 0\n  return {\n    id: 'error-streak',\n    get streak() {\n      return streak\n    },\n    observe(event) {\n      if (event.status === undefined) return null\n      streak = event.status === 'error' ? streak + 1 : 0\n      if (max <= 0 || streak < max) return null\n      return {\n        detector: 'error-streak',\n        severity,\n        failureClass,\n        reason: `stuck: ${streak} consecutive errored step(s)`,\n        streak,\n        ...(event.label ? { evidence: { lastError: event.label } } : {}),\n      }\n    },\n    reset() {\n      streak = 0\n    },\n  }\n}\n\n/** Fold one event through many detectors at once; returns every signal that fired this step. The\n *  natural shape for an online pipe watching with a whole panel of detectors. */\nexport function observeAll(\n  detectors: ReadonlyArray<StreamingDetector>,\n  event: DetectorEvent,\n): DetectorSignal[] {\n  const signals: DetectorSignal[] = []\n  for (const d of detectors) {\n    const s = d.observe(event)\n    if (s) signals.push(s)\n  }\n  return signals\n}\n","/**\n * Policy-based agent control runtime.\n *\n * This is the minimal reusable loop behind driver-agent patterns:\n *\n *   observe state -> validate -> decide next action -> act -> observe -> ...\n *\n * It deliberately does not model named \"topologies\". Direct execution,\n * critic/revise, driver intervention, specialist calls, and human escalation\n * are all just actions chosen by the control policy.\n */\n\nimport { noProgressDetector, repeatedActionDetector } from './detectors'\nimport { type SpanHandle, TraceEmitter } from './trace/emitter'\nimport type { FailureClass } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport type ControlSeverity = 'info' | 'warning' | 'error' | 'critical'\nexport type ControlActionFailureMode = 'continue' | 'stop'\n\nexport interface ControlEvalResult {\n  /** Stable validator or judge id. */\n  id: string\n  /** Whether this check passed. */\n  passed: boolean\n  /** Optional normalized score. 1 = best, 0 = worst. */\n  score?: number\n  /** Objective validators should usually be \"error\" or \"critical\" when failed. */\n  severity?: ControlSeverity\n  /** Human-readable result. */\n  detail?: string\n  /** Small evidence string or pointer. Avoid large payloads. */\n  evidence?: string\n  /** True when the result came from deterministic state, not LLM judgment. */\n  objective?: boolean\n  /** Structured details for downstream control policies and reports. */\n  metadata?: Record<string, unknown>\n}\n\nexport interface ControlBudget {\n  maxSteps: number\n  maxWallMs?: number\n  maxCostUsd?: number\n}\n\nexport interface ControlStopPolicies<TState, TAction> {\n  /**\n   * Stop after N consecutive steps with no state fingerprint change and\n   * less than `minScoreDelta` score movement. Disabled when omitted.\n   */\n  maxNoProgressSteps?: number\n  /**\n   * Stop after the same action fingerprint is selected N consecutive\n   * times. Disabled when omitted.\n   */\n  maxRepeatedActions?: number\n  /** Minimum score movement that counts as progress. Default 0.001. */\n  minScoreDelta?: number\n  /** Override the default JSON/string fingerprint for state comparisons. */\n  stateFingerprint?: (state: TState) => string\n  /** Override the default JSON/string fingerprint for repeated-action checks. */\n  actionFingerprint?: (action: TAction) => string\n}\n\nexport interface ControlContext<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  intent: string\n  state: TState\n  evals: TEval[]\n  history: ControlStep<TState, TAction, TActionResult, TEval>[]\n  budget: ControlBudget\n  stepIndex: number\n  wallMs: number\n  spentCostUsd: number\n  remainingCostUsd?: number\n  abortSignal: AbortSignal\n  emitter?: TraceEmitter\n}\n\nexport type ControlDecision<TAction> =\n  | {\n      type: 'continue'\n      action: TAction\n      reason?: string\n    }\n  | {\n      type: 'stop'\n      reason: string\n      pass?: boolean\n      score?: number\n    }\n\nexport interface StopDecision {\n  stop: boolean\n  pass: boolean\n  reason: string\n  score?: number\n  failureClass?: FailureClass\n}\n\nexport interface ControlActionOutcome<TActionResult> {\n  ok: boolean\n  result?: TActionResult\n  error?: string\n  costUsd?: number\n  durationMs: number\n}\n\nexport interface ControlRuntimeError {\n  phase: 'observe' | 'validate' | 'decide' | 'act' | 'stop-policy' | 'on-step' | 'trace'\n  stepIndex: number\n  message: string\n}\n\nexport interface ControlStep<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  index: number\n  decision: ControlDecision<TAction>\n  beforeState: TState\n  afterState: TState\n  evalsBefore: TEval[]\n  evalsAfter: TEval[]\n  actionOutcome?: ControlActionOutcome<TActionResult>\n  startedAt: string\n  endedAt: string\n}\n\nexport interface ControlRunResult<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  intent: string\n  pass: boolean\n  completed: boolean\n  reason: string\n  score?: number\n  steps: ControlStep<TState, TAction, TActionResult, TEval>[]\n  finalState: TState | undefined\n  finalEvals: TEval[]\n  wallMs: number\n  spentCostUsd: number\n  /** null when the run executed without a TraceEmitter wired (no run record was persisted). */\n  runId: string | null\n  failureClass?: FailureClass\n  runtimeErrors: ControlRuntimeError[]\n  stoppedBy: 'policy' | 'stop-policy' | 'budget' | 'abort' | 'runtime-error'\n}\n\nexport interface ControlRuntimeConfig<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  intent: string\n  budget?: Partial<ControlBudget>\n  signal?: AbortSignal\n  /** Defaults to `continue`: action failures are recorded, then the policy gets another chance. */\n  actionFailure?: ControlActionFailureMode\n  /**\n   * Extract cost from an action result. Used for `maxCostUsd` budget\n   * enforcement and trace budget ledger emission.\n   */\n  getActionCostUsd?: (ctx: {\n    action: TAction\n    result: TActionResult\n    state: TState\n    evals: TEval[]\n    history: ControlStep<TState, TAction, TActionResult, TEval>[]\n  }) => number | undefined\n\n  /** Read typed task/product state. Prefer structured state over transcript-only context. */\n  observe: (ctx: {\n    history: ControlStep<TState, TAction, TActionResult, TEval>[]\n    abortSignal: AbortSignal\n  }) => Promise<TState> | TState\n\n  /** Objective validators first, subjective judges only where objective state is insufficient. */\n  validate: (ctx: {\n    intent: string\n    state: TState\n    history: ControlStep<TState, TAction, TActionResult, TEval>[]\n    abortSignal: AbortSignal\n  }) => Promise<TEval[]> | TEval[]\n\n  /** Choose the next control action. Can call a worker, ask user, run critic, inspect state, or stop. */\n  decide: (\n    ctx: ControlContext<TState, TAction, TActionResult, TEval>,\n  ) => Promise<ControlDecision<TAction>> | ControlDecision<TAction>\n\n  /** Execute the action selected by the policy. */\n  act: (\n    action: TAction,\n    ctx: ControlContext<TState, TAction, TActionResult, TEval>,\n  ) => Promise<TActionResult> | TActionResult\n\n  /** Final stopping policy. Called before decide and after each action. */\n  shouldStop?: (\n    ctx: ControlContext<TState, TAction, TActionResult, TEval>,\n  ) => Promise<StopDecision> | StopDecision\n\n  /** Optional hook for tracing or live progress updates. */\n  onStep?: (step: ControlStep<TState, TAction, TActionResult, TEval>) => Promise<void> | void\n\n  /** Optional generic stuck-loop policies. Custom `shouldStop` still runs first. */\n  stopPolicies?: ControlStopPolicies<TState, TAction>\n\n  /** Optional trace sink. Emits one run plus one span per control step. */\n  store?: TraceStore\n  scenarioId?: string\n  projectId?: string\n  variantId?: string\n}\n\nconst DEFAULT_BUDGET: ControlBudget = {\n  maxSteps: 8,\n  maxWallMs: 5 * 60 * 1000,\n}\n\nexport async function runAgentControlLoop<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n>(\n  config: ControlRuntimeConfig<TState, TAction, TActionResult, TEval>,\n): Promise<ControlRunResult<TState, TAction, TActionResult, TEval>> {\n  const budget = normalizeBudget(config.budget)\n  const actionFailure = config.actionFailure ?? 'continue'\n  const controller = new AbortController()\n  const upstreamAbort = () => controller.abort(config.signal?.reason)\n  if (config.signal) {\n    if (config.signal.aborted) controller.abort(config.signal.reason)\n    else config.signal.addEventListener('abort', upstreamAbort, { once: true })\n  }\n\n  const started = Date.now()\n  const wallTimer = budget.maxWallMs\n    ? setTimeout(\n        () => controller.abort(new Error('control runtime wall timeout')),\n        budget.maxWallMs,\n      )\n    : undefined\n  const history: ControlStep<TState, TAction, TActionResult, TEval>[] = []\n  const emitter = config.store ? new TraceEmitter(config.store) : undefined\n  let spentCostUsd = 0\n  const runtimeErrors: ControlRuntimeError[] = []\n  // The stuck-detection kernels (shared with the online detectors). Thresholds 0 = disabled; the\n  // streak is still tracked for telemetry. Both stops report `tool_recovery_failure`.\n  const repeatedDetector = repeatedActionDetector({\n    maxRepeated: config.stopPolicies?.maxRepeatedActions ?? 0,\n  })\n  const progressDetector = noProgressDetector({\n    maxNoProgress: config.stopPolicies?.maxNoProgressSteps ?? 0,\n    minScoreDelta: config.stopPolicies?.minScoreDelta ?? 0.001,\n  })\n\n  try {\n    if (emitter) {\n      await runTrace(runtimeErrors, 0, () =>\n        emitter.startRun({\n          scenarioId: config.scenarioId ?? 'agent-control-loop',\n          projectId: config.projectId,\n          variantId: config.variantId,\n          layer: 'meta',\n          tags: {\n            intent: config.intent.slice(0, 120),\n            maxSteps: String(budget.maxSteps),\n            ...(budget.maxCostUsd !== undefined ? { maxCostUsd: String(budget.maxCostUsd) } : {}),\n          },\n        }),\n      )\n    }\n\n    let state: TState\n    let evals: TEval[]\n    try {\n      state = await config.observe({ history, abortSignal: controller.signal })\n    } catch (err) {\n      const error = runtimeError('observe', 0, err)\n      runtimeErrors.push(error)\n      return finish(emitter, {\n        intent: config.intent,\n        pass: false,\n        completed: false,\n        reason: error.message,\n        steps: history,\n        finalState: undefined,\n        finalEvals: [],\n        wallMs: Date.now() - started,\n        spentCostUsd,\n        runId: emitter?.runId ?? null,\n        failureClass: 'unknown',\n        runtimeErrors,\n        stoppedBy: 'runtime-error',\n      })\n    }\n    try {\n      evals = await config.validate({\n        intent: config.intent,\n        state,\n        history,\n        abortSignal: controller.signal,\n      })\n      await recordEvalSpans(emitter, evals, 'initial', runtimeErrors, 0)\n    } catch (err) {\n      const error = runtimeError('validate', 0, err)\n      runtimeErrors.push(error)\n      return finish(emitter, {\n        intent: config.intent,\n        pass: false,\n        completed: false,\n        reason: error.message,\n        steps: history,\n        finalState: state,\n        finalEvals: [],\n        wallMs: Date.now() - started,\n        spentCostUsd,\n        runId: emitter?.runId ?? null,\n        failureClass: 'unknown',\n        runtimeErrors,\n        stoppedBy: 'runtime-error',\n      })\n    }\n    // Prime the no-progress detector with the initial state + score so step 0 can already detect a\n    // no-op (mirrors the pre-loop fingerprint baseline). The priming observe never signals (streak 0).\n    progressDetector.observe({\n      stateFingerprint: fingerprintState(state, config.stopPolicies),\n      score: averageScore(evals),\n    })\n\n    for (let stepIndex = 0; stepIndex < budget.maxSteps; stepIndex++) {\n      if (controller.signal.aborted) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: abortReason(controller.signal),\n          score: undefined,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'timeout',\n          runtimeErrors,\n          stoppedBy: 'abort',\n        })\n      }\n\n      const budgetStop = budgetStopDecision(budget, spentCostUsd)\n      if (budgetStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: budgetStop.reason,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'budget_exceeded',\n          runtimeErrors,\n          stoppedBy: 'budget',\n        })\n      }\n\n      const ctx = makeContext(\n        config.intent,\n        state,\n        evals,\n        history,\n        budget,\n        stepIndex,\n        started,\n        spentCostUsd,\n        controller.signal,\n        emitter,\n      )\n      let stop: StopDecision\n      try {\n        stop = config.shouldStop ? await config.shouldStop(ctx) : defaultStopDecision(evals)\n      } catch (err) {\n        runtimeErrors.push(runtimeError('stop-policy', stepIndex, err))\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      if (stop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: stop.pass,\n          completed: true,\n          reason: stop.reason,\n          score: stop.score,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: stop.failureClass,\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n\n      let decision: ControlDecision<TAction>\n      try {\n        decision = await config.decide(ctx)\n      } catch (err) {\n        runtimeErrors.push(runtimeError('decide', stepIndex, err))\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      if (decision.type === 'stop') {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: decision.pass ?? false,\n          completed: true,\n          reason: decision.reason,\n          score: decision.score,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: decision.pass === false ? 'unknown' : undefined,\n          runtimeErrors,\n          stoppedBy: 'policy',\n        })\n      }\n\n      const actionFingerprint = fingerprintAction(decision.action, config.stopPolicies)\n      const repeatedActionSignal = repeatedDetector.observe({ actionFingerprint })\n      const repeatedActionStreak = repeatedDetector.streak\n      if (repeatedActionSignal) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: true,\n          reason: repeatedActionSignal.reason,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'tool_recovery_failure',\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n\n      const beforeState = state\n      const evalsBefore = evals\n      const scoreBefore = averageScore(evals)\n      const actionStarted = Date.now()\n      const stepHandle = emitter\n        ? await runTrace(runtimeErrors, stepIndex, () =>\n            emitter.tool({\n              name: `control-step-${stepIndex}`,\n              toolName: 'agent-control-action',\n              args: decision.action,\n              attributes: {\n                decision: decision.reason ?? 'continue',\n                repeatedActionStreak,\n              },\n            }),\n          )\n        : undefined\n      let actionOutcome: ControlActionOutcome<TActionResult>\n      try {\n        const result = await config.act(decision.action, ctx)\n        const rawCostUsd = config.getActionCostUsd?.({\n          action: decision.action,\n          result,\n          state,\n          evals,\n          history,\n        })\n        const costUsd = normalizeActionCostUsd(rawCostUsd, runtimeErrors, stepIndex)\n        if (costUsd !== undefined && Number.isFinite(costUsd) && costUsd > 0) {\n          spentCostUsd += costUsd\n          await recordCostBudget(\n            emitter,\n            budget,\n            spentCostUsd,\n            stepHandle,\n            runtimeErrors,\n            stepIndex,\n          )\n        }\n        actionOutcome = {\n          ok: true,\n          result,\n          ...(costUsd !== undefined ? { costUsd } : {}),\n          durationMs: Date.now() - actionStarted,\n        }\n      } catch (err) {\n        runtimeErrors.push(runtimeError('act', stepIndex, err))\n        actionOutcome = {\n          ok: false,\n          error: runtimeErrors[runtimeErrors.length - 1]!.message,\n          durationMs: Date.now() - actionStarted,\n        }\n        if (actionFailure === 'stop') {\n          await runTrace(runtimeErrors, stepIndex, () =>\n            stepHandle?.fail(actionOutcome.error ?? 'action failed'),\n          )\n          const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n            index: stepIndex,\n            decision,\n            beforeState,\n            afterState: state,\n            evalsBefore,\n            evalsAfter: evals,\n            actionOutcome,\n            startedAt: new Date(actionStarted).toISOString(),\n            endedAt: new Date().toISOString(),\n          }\n          history.push(step)\n          await runOnStep(config.onStep, step, runtimeErrors)\n          return finish(emitter, {\n            intent: config.intent,\n            pass: false,\n            completed: false,\n            reason: actionOutcome.error ?? 'action failed',\n            score: averageScore(evals),\n            steps: history,\n            finalState: state,\n            finalEvals: evals,\n            wallMs: Date.now() - started,\n            spentCostUsd,\n            runId: emitter?.runId ?? null,\n            failureClass: 'unknown',\n            runtimeErrors,\n            stoppedBy: 'runtime-error',\n          })\n        }\n      }\n\n      try {\n        state = await config.observe({ history, abortSignal: controller.signal })\n      } catch (err) {\n        runtimeErrors.push(runtimeError('observe', stepIndex, err))\n        const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n          index: stepIndex,\n          decision,\n          beforeState,\n          afterState: beforeState,\n          evalsBefore,\n          evalsAfter: evals,\n          actionOutcome,\n          startedAt: new Date(actionStarted).toISOString(),\n          endedAt: new Date().toISOString(),\n        }\n        history.push(step)\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.fail(runtimeErrors[runtimeErrors.length - 1]!.message),\n        )\n        await runOnStep(config.onStep, step, runtimeErrors)\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: beforeState,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      try {\n        evals = await config.validate({\n          intent: config.intent,\n          state,\n          history,\n          abortSignal: controller.signal,\n        })\n        await recordEvalSpans(\n          emitter,\n          evals,\n          `step-${stepIndex}`,\n          runtimeErrors,\n          stepIndex,\n          stepHandle?.span.spanId,\n        )\n      } catch (err) {\n        runtimeErrors.push(runtimeError('validate', stepIndex, err))\n        const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n          index: stepIndex,\n          decision,\n          beforeState,\n          afterState: state,\n          evalsBefore,\n          evalsAfter: evals,\n          actionOutcome,\n          startedAt: new Date(actionStarted).toISOString(),\n          endedAt: new Date().toISOString(),\n        }\n        history.push(step)\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.fail(runtimeErrors[runtimeErrors.length - 1]!.message),\n        )\n        await runOnStep(config.onStep, step, runtimeErrors)\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      const scoreAfter = averageScore(evals)\n      const stateFingerprint = fingerprintState(state, config.stopPolicies)\n      const noProgressSignal = progressDetector.observe({ stateFingerprint, score: scoreAfter })\n      const noProgressStreak = progressDetector.streak\n\n      const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n        index: stepIndex,\n        decision,\n        beforeState,\n        afterState: state,\n        evalsBefore,\n        evalsAfter: evals,\n        actionOutcome,\n        startedAt: new Date(actionStarted).toISOString(),\n        endedAt: new Date().toISOString(),\n      }\n      history.push(step)\n      if (actionOutcome.ok) {\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.end({\n            attributes: {\n              actionCostUsd: actionOutcome.costUsd ?? null,\n              spentCostUsd,\n              scoreBefore: scoreBefore ?? null,\n              scoreAfter: scoreAfter ?? null,\n              noProgressStreak,\n            },\n          }),\n        )\n      } else {\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.fail(actionOutcome.error ?? 'action failed', {\n            attributes: {\n              spentCostUsd,\n              noProgressStreak,\n            },\n          }),\n        )\n      }\n      await runOnStep(config.onStep, step, runtimeErrors)\n\n      if (noProgressSignal) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: true,\n          reason: noProgressSignal.reason,\n          score: scoreAfter,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'tool_recovery_failure',\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n\n      const postStepBudgetStop = budgetStopDecision(budget, spentCostUsd)\n      if (postStepBudgetStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: postStepBudgetStop.reason,\n          score: scoreAfter,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'budget_exceeded',\n          runtimeErrors,\n          stoppedBy: 'budget',\n        })\n      }\n\n      const postStepCtx = makeContext(\n        config.intent,\n        state,\n        evals,\n        history,\n        budget,\n        stepIndex + 1,\n        started,\n        spentCostUsd,\n        controller.signal,\n        emitter,\n      )\n      let postStepStop: StopDecision\n      try {\n        postStepStop = config.shouldStop\n          ? await config.shouldStop(postStepCtx)\n          : defaultStopDecision(evals)\n      } catch (err) {\n        runtimeErrors.push(runtimeError('stop-policy', stepIndex + 1, err))\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      if (postStepStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: postStepStop.pass,\n          completed: true,\n          reason: postStepStop.reason,\n          score: postStepStop.score,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: postStepStop.failureClass,\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n    }\n\n    return finish(emitter, {\n      intent: config.intent,\n      pass: false,\n      completed: false,\n      reason: `budget exhausted: maxSteps=${budget.maxSteps}`,\n      steps: history,\n      finalState: state,\n      finalEvals: evals,\n      wallMs: Date.now() - started,\n      spentCostUsd,\n      runId: emitter?.runId ?? null,\n      failureClass: 'budget_exceeded',\n      runtimeErrors,\n      stoppedBy: 'budget',\n    })\n  } catch (err) {\n    runtimeErrors.push(runtimeError('act', history.length, err))\n    return finish(emitter, {\n      intent: config.intent,\n      pass: false,\n      completed: false,\n      reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n      steps: history,\n      finalState: undefined,\n      finalEvals: [],\n      wallMs: Date.now() - started,\n      spentCostUsd,\n      runId: emitter?.runId ?? null,\n      failureClass: 'unknown',\n      runtimeErrors,\n      stoppedBy: 'runtime-error',\n    })\n  } finally {\n    if (wallTimer) clearTimeout(wallTimer)\n    if (config.signal) config.signal.removeEventListener('abort', upstreamAbort)\n  }\n}\n\nexport function stopOnNoProgress<TState, TAction>(\n  maxNoProgressSteps: number,\n  options: Omit<ControlStopPolicies<TState, TAction>, 'maxNoProgressSteps'> = {},\n): ControlStopPolicies<TState, TAction> {\n  return { ...options, maxNoProgressSteps }\n}\n\nexport function stopOnRepeatedAction<TState, TAction>(\n  maxRepeatedActions: number,\n  options: Omit<ControlStopPolicies<TState, TAction>, 'maxRepeatedActions'> = {},\n): ControlStopPolicies<TState, TAction> {\n  return { ...options, maxRepeatedActions }\n}\n\nexport function objectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult {\n  return { ...input, objective: true }\n}\n\nexport function subjectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult {\n  return { ...input, objective: false }\n}\n\nfunction normalizeBudget(input: Partial<ControlBudget> | undefined): ControlBudget {\n  const raw = { ...DEFAULT_BUDGET, ...input } as Record<string, unknown>\n  if (!Number.isInteger(raw.maxSteps) || (raw.maxSteps as number) < 1) {\n    throw new RangeError(\n      `ControlRuntime budget.maxSteps must be an integer >= 1, got ${String(raw.maxSteps)}`,\n    )\n  }\n  const budget: ControlBudget = { maxSteps: raw.maxSteps as number }\n  if (raw.maxWallMs !== undefined) {\n    if (\n      typeof raw.maxWallMs !== 'number' ||\n      !Number.isFinite(raw.maxWallMs) ||\n      raw.maxWallMs <= 0\n    ) {\n      throw new RangeError(\n        `ControlRuntime budget.maxWallMs must be a positive finite number, got ${String(raw.maxWallMs)}`,\n      )\n    }\n    budget.maxWallMs = raw.maxWallMs\n  }\n  if (raw.maxCostUsd !== undefined) {\n    if (\n      typeof raw.maxCostUsd !== 'number' ||\n      !Number.isFinite(raw.maxCostUsd) ||\n      raw.maxCostUsd < 0\n    ) {\n      throw new RangeError(\n        `ControlRuntime budget.maxCostUsd must be a nonnegative finite number, got ${String(raw.maxCostUsd)}`,\n      )\n    }\n    budget.maxCostUsd = raw.maxCostUsd\n  }\n  return budget\n}\n\nfunction normalizeActionCostUsd(\n  costUsd: number | undefined,\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n): number | undefined {\n  if (costUsd === undefined) return undefined\n  if (!Number.isFinite(costUsd) || costUsd < 0) {\n    runtimeErrors.push(\n      runtimeError('act', stepIndex, new Error(`invalid action costUsd: ${String(costUsd)}`)),\n    )\n    return undefined\n  }\n  return costUsd\n}\n\nexport function allCriticalPassed(evals: ControlEvalResult[]): boolean {\n  return evals.every(\n    (result) => result.passed || (result.severity !== 'critical' && result.severity !== 'error'),\n  )\n}\n\nfunction makeContext<TState, TAction, TActionResult, TEval extends ControlEvalResult>(\n  intent: string,\n  state: TState,\n  evals: TEval[],\n  history: ControlStep<TState, TAction, TActionResult, TEval>[],\n  budget: ControlBudget,\n  stepIndex: number,\n  started: number,\n  spentCostUsd: number,\n  abortSignal: AbortSignal,\n  emitter?: TraceEmitter,\n): ControlContext<TState, TAction, TActionResult, TEval> {\n  return {\n    intent,\n    state,\n    evals,\n    history,\n    budget,\n    stepIndex,\n    wallMs: Date.now() - started,\n    spentCostUsd,\n    remainingCostUsd:\n      budget.maxCostUsd === undefined ? undefined : Math.max(0, budget.maxCostUsd - spentCostUsd),\n    abortSignal,\n    emitter,\n  }\n}\n\nfunction defaultStopDecision(evals: ControlEvalResult[]): StopDecision {\n  if (!evals.length) return { stop: false, pass: false, reason: 'no evals yet' }\n  const pass = allCriticalPassed(evals)\n  return pass\n    ? { stop: true, pass: true, reason: 'all critical evals passed', score: averageScore(evals) }\n    : {\n        stop: false,\n        pass: false,\n        reason: 'critical evals still failing',\n        score: averageScore(evals),\n      }\n}\n\nfunction averageScore(evals: ControlEvalResult[]): number | undefined {\n  const scored = evals\n    .map((result) => result.score)\n    .filter((score): score is number => typeof score === 'number')\n  if (!scored.length) return undefined\n  return Math.round((scored.reduce((sum, score) => sum + score, 0) / scored.length) * 1000) / 1000\n}\n\nfunction budgetStopDecision(\n  budget: ControlBudget,\n  spentCostUsd: number,\n): { stop: boolean; reason: string } {\n  if (budget.maxCostUsd !== undefined && spentCostUsd >= budget.maxCostUsd) {\n    return {\n      stop: true,\n      reason: `budget exhausted: maxCostUsd=${budget.maxCostUsd}`,\n    }\n  }\n  return { stop: false, reason: '' }\n}\n\nasync function recordCostBudget(\n  emitter: TraceEmitter | undefined,\n  budget: ControlBudget,\n  spentCostUsd: number,\n  handle: SpanHandle | undefined,\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n): Promise<void> {\n  if (!emitter || budget.maxCostUsd === undefined) return\n  const maxCostUsd = budget.maxCostUsd\n  await runTrace(runtimeErrors, stepIndex, () =>\n    emitter.recordBudget({\n      dimension: 'usd',\n      limit: maxCostUsd,\n      consumed: spentCostUsd,\n      remaining: Math.max(0, maxCostUsd - spentCostUsd),\n      breached: spentCostUsd >= maxCostUsd,\n      spanId: handle?.span.spanId,\n    }),\n  )\n}\n\nasync function recordEvalSpans(\n  emitter: TraceEmitter | undefined,\n  evals: ControlEvalResult[],\n  phase: string,\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n  targetSpanId?: string,\n): Promise<void> {\n  if (!emitter) return\n  for (const result of evals) {\n    await runTrace(runtimeErrors, stepIndex, () =>\n      emitter.recordJudge({\n        judgeId: result.objective ? 'objective-validator' : 'subjective-judge',\n        targetSpanId: targetSpanId ?? emitter.runId,\n        name: `control-eval/${result.id}`,\n        dimension: result.id,\n        score: typeof result.score === 'number' ? result.score : result.passed ? 1 : 0,\n        rationale: result.detail,\n        evidence: result.evidence,\n        attributes: {\n          phase,\n          passed: result.passed,\n          severity: result.severity,\n          objective: result.objective,\n        },\n      }),\n    )\n  }\n}\n\nasync function runOnStep<TState, TAction, TActionResult, TEval extends ControlEvalResult>(\n  onStep: ControlRuntimeConfig<TState, TAction, TActionResult, TEval>['onStep'] | undefined,\n  step: ControlStep<TState, TAction, TActionResult, TEval>,\n  runtimeErrors: ControlRuntimeError[],\n): Promise<void> {\n  if (!onStep) return\n  try {\n    await onStep(step)\n  } catch (err) {\n    runtimeErrors.push(runtimeError('on-step', step.index, err))\n  }\n}\n\nasync function runTrace<T>(\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n  write: () => Promise<T | undefined> | T | undefined,\n): Promise<T | undefined> {\n  try {\n    return await write()\n  } catch (err) {\n    runtimeErrors.push(runtimeError('trace', stepIndex, err))\n    return undefined\n  }\n}\n\nfunction fingerprintState<TState, TAction>(\n  state: TState,\n  policies?: ControlStopPolicies<TState, TAction>,\n): string {\n  if (policies?.stateFingerprint) return policies.stateFingerprint(state)\n  return stableFingerprint(state)\n}\n\nfunction fingerprintAction<TState, TAction>(\n  action: TAction,\n  policies?: ControlStopPolicies<TState, TAction>,\n): string {\n  if (policies?.actionFingerprint) return policies.actionFingerprint(action)\n  return stableFingerprint(action)\n}\n\nfunction stableFingerprint(value: unknown): string {\n  if (typeof value === 'string') return value\n  if (typeof value === 'number' || typeof value === 'boolean' || value == null) return String(value)\n  try {\n    return JSON.stringify(sortForFingerprint(value))\n  } catch {\n    return String(value)\n  }\n}\n\nfunction sortForFingerprint(value: unknown): unknown {\n  if (Array.isArray(value)) return value.map(sortForFingerprint)\n  if (!value || typeof value !== 'object') return value\n  const record = value as Record<string, unknown>\n  const sorted: Record<string, unknown> = {}\n  for (const key of Object.keys(record).sort()) {\n    sorted[key] = sortForFingerprint(record[key])\n  }\n  return sorted\n}\n\nfunction abortReason(signal: AbortSignal): string {\n  const reason = signal.reason\n  if (reason instanceof Error) return reason.message\n  return reason ? String(reason) : 'aborted'\n}\n\nfunction runtimeError(\n  phase: ControlRuntimeError['phase'],\n  stepIndex: number,\n  err: unknown,\n): ControlRuntimeError {\n  const message = err instanceof Error ? err.message : String(err)\n  return { phase, stepIndex, message }\n}\n\nasync function finish<TState, TAction, TActionResult, TEval extends ControlEvalResult>(\n  emitter: TraceEmitter | undefined,\n  result: ControlRunResult<TState, TAction, TActionResult, TEval>,\n): Promise<ControlRunResult<TState, TAction, TActionResult, TEval>> {\n  await runTrace(result.runtimeErrors, result.steps.length, () =>\n    emitter?.endRun({\n      pass: result.pass,\n      score: result.score ?? averageScore(result.finalEvals),\n      failureClass: result.failureClass,\n      notes: result.reason,\n    }),\n  )\n  return result\n}\n"],"mappings":";;;;;AA+DO,SAAS,uBAAuB,OAA8B,CAAC,GAAsB;AAC1F,QAAM,MAAM,KAAK,eAAe;AAChC,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,eAA6B,KAAK,gBAAgB;AACxD,MAAI;AACJ,MAAI,SAAS;AACb,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,IAAI,SAAS;AACX,aAAO;AAAA,IACT;AAAA,IACA,QAAQ,OAAO;AACb,YAAM,KAAK,MAAM;AACjB,UAAI,OAAO,OAAW,QAAO;AAC7B,eAAS,OAAO,OAAO,SAAS,IAAI;AACpC,aAAO;AACP,UAAI,OAAO,KAAK,SAAS,IAAK,QAAO;AACrC,aAAO;AAAA,QACL,UAAU;AAAA,QACV;AAAA,QACA;AAAA,QACA,QAAQ,mCAAmC,MAAM;AAAA,QACjD;AAAA,QACA,GAAI,MAAM,QAAQ,EAAE,UAAU,EAAE,QAAQ,MAAM,MAAM,EAAE,IAAI,CAAC;AAAA,MAC7D;AAAA,IACF;AAAA,IACA,QAAQ;AACN,aAAO;AACP,eAAS;AAAA,IACX;AAAA,EACF;AACF;AAcO,SAAS,mBAAmB,OAA0B,CAAC,GAAsB;AAClF,QAAM,MAAM,KAAK,iBAAiB;AAClC,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,eAA6B,KAAK,gBAAgB;AACxD,MAAI;AACJ,MAAI;AACJ,MAAI,SAAS;AACb,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,IAAI,SAAS;AACX,aAAO;AAAA,IACT;AAAA,IACA,QAAQ,OAAO;AACb,YAAM,iBAAiB,cAAc,UAAa,cAAc,MAAM;AACtE,YAAM,YAAY,KAAK,KAAK,MAAM,SAAS,MAAM,aAAa,EAAE,IAAI;AACpE,eAAS,kBAAkB,YAAY,SAAS,IAAI;AACpD,kBAAY,MAAM;AAClB,kBAAY,MAAM;AAClB,UAAI,OAAO,KAAK,SAAS,IAAK,QAAO;AACrC,aAAO;AAAA,QACL,UAAU;AAAA,QACV;AAAA,QACA;AAAA,QACA,QAAQ,sCAAsC,MAAM;AAAA,QACpD;AAAA,QACA,GAAI,MAAM,QAAQ,EAAE,UAAU,EAAE,OAAO,MAAM,MAAM,EAAE,IAAI,CAAC;AAAA,MAC5D;AAAA,IACF;AAAA,IACA,QAAQ;AACN,kBAAY;AACZ,kBAAY;AACZ,eAAS;AAAA,IACX;AAAA,EACF;AACF;AAUO,SAAS,oBAAoB,OAA2B,CAAC,GAAsB;AACpF,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,eAA6B,KAAK,gBAAgB;AACxD,MAAI,SAAS;AACb,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,IAAI,SAAS;AACX,aAAO;AAAA,IACT;AAAA,IACA,QAAQ,OAAO;AACb,UAAI,MAAM,WAAW,OAAW,QAAO;AACvC,eAAS,MAAM,WAAW,UAAU,SAAS,IAAI;AACjD,UAAI,OAAO,KAAK,SAAS,IAAK,QAAO;AACrC,aAAO;AAAA,QACL,UAAU;AAAA,QACV;AAAA,QACA;AAAA,QACA,QAAQ,UAAU,MAAM;AAAA,QACxB;AAAA,QACA,GAAI,MAAM,QAAQ,EAAE,UAAU,EAAE,WAAW,MAAM,MAAM,EAAE,IAAI,CAAC;AAAA,MAChE;AAAA,IACF;AAAA,IACA,QAAQ;AACN,eAAS;AAAA,IACX;AAAA,EACF;AACF;AAIO,SAAS,WACd,WACA,OACkB;AAClB,QAAM,UAA4B,CAAC;AACnC,aAAW,KAAK,WAAW;AACzB,UAAM,IAAI,EAAE,QAAQ,KAAK;AACzB,QAAI,EAAG,SAAQ,KAAK,CAAC;AAAA,EACvB;AACA,SAAO;AACT;;;AC8BA,IAAM,iBAAgC;AAAA,EACpC,UAAU;AAAA,EACV,WAAW,IAAI,KAAK;AACtB;AAEA,eAAsB,oBAMpB,QACkE;AAClE,QAAM,SAAS,gBAAgB,OAAO,MAAM;AAC5C,QAAM,gBAAgB,OAAO,iBAAiB;AAC9C,QAAM,aAAa,IAAI,gBAAgB;AACvC,QAAM,gBAAgB,MAAM,WAAW,MAAM,OAAO,QAAQ,MAAM;AAClE,MAAI,OAAO,QAAQ;AACjB,QAAI,OAAO,OAAO,QAAS,YAAW,MAAM,OAAO,OAAO,MAAM;AAAA,QAC3D,QAAO,OAAO,iBAAiB,SAAS,eAAe,EAAE,MAAM,KAAK,CAAC;AAAA,EAC5E;AAEA,QAAM,UAAU,KAAK,IAAI;AACzB,QAAM,YAAY,OAAO,YACrB;AAAA,IACE,MAAM,WAAW,MAAM,IAAI,MAAM,8BAA8B,CAAC;AAAA,IAChE,OAAO;AAAA,EACT,IACA;AACJ,QAAM,UAAgE,CAAC;AACvE,QAAM,UAAU,OAAO,QAAQ,IAAI,aAAa,OAAO,KAAK,IAAI;AAChE,MAAI,eAAe;AACnB,QAAM,gBAAuC,CAAC;AAG9C,QAAM,mBAAmB,uBAAuB;AAAA,IAC9C,aAAa,OAAO,cAAc,sBAAsB;AAAA,EAC1D,CAAC;AACD,QAAM,mBAAmB,mBAAmB;AAAA,IAC1C,eAAe,OAAO,cAAc,sBAAsB;AAAA,IAC1D,eAAe,OAAO,cAAc,iBAAiB;AAAA,EACvD,CAAC;AAED,MAAI;AACF,QAAI,SAAS;AACX,YAAM;AAAA,QAAS;AAAA,QAAe;AAAA,QAAG,MAC/B,QAAQ,SAAS;AAAA,UACf,YAAY,OAAO,cAAc;AAAA,UACjC,WAAW,OAAO;AAAA,UAClB,WAAW,OAAO;AAAA,UAClB,OAAO;AAAA,UACP,MAAM;AAAA,YACJ,QAAQ,OAAO,OAAO,MAAM,GAAG,GAAG;AAAA,YAClC,UAAU,OAAO,OAAO,QAAQ;AAAA,YAChC,GAAI,OAAO,eAAe,SAAY,EAAE,YAAY,OAAO,OAAO,UAAU,EAAE,IAAI,CAAC;AAAA,UACrF;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF;AAEA,QAAI;AACJ,QAAI;AACJ,QAAI;AACF,cAAQ,MAAM,OAAO,QAAQ,EAAE,SAAS,aAAa,WAAW,OAAO,CAAC;AAAA,IAC1E,SAAS,KAAK;AACZ,YAAM,QAAQ,aAAa,WAAW,GAAG,GAAG;AAC5C,oBAAc,KAAK,KAAK;AACxB,aAAO,OAAO,SAAS;AAAA,QACrB,QAAQ,OAAO;AAAA,QACf,MAAM;AAAA,QACN,WAAW;AAAA,QACX,QAAQ,MAAM;AAAA,QACd,OAAO;AAAA,QACP,YAAY;AAAA,QACZ,YAAY,CAAC;AAAA,QACb,QAAQ,KAAK,IAAI,IAAI;AAAA,QACrB;AAAA,QACA,OAAO,SAAS,SAAS;AAAA,QACzB,cAAc;AAAA,QACd;AAAA,QACA,WAAW;AAAA,MACb,CAAC;AAAA,IACH;AACA,QAAI;AACF,cAAQ,MAAM,OAAO,SAAS;AAAA,QAC5B,QAAQ,OAAO;AAAA,QACf;AAAA,QACA;AAAA,QACA,aAAa,WAAW;AAAA,MAC1B,CAAC;AACD,YAAM,gBAAgB,SAAS,OAAO,WAAW,eAAe,CAAC;AAAA,IACnE,SAAS,KAAK;AACZ,YAAM,QAAQ,aAAa,YAAY,GAAG,GAAG;AAC7C,oBAAc,KAAK,KAAK;AACxB,aAAO,OAAO,SAAS;AAAA,QACrB,QAAQ,OAAO;AAAA,QACf,MAAM;AAAA,QACN,WAAW;AAAA,QACX,QAAQ,MAAM;AAAA,QACd,OAAO;AAAA,QACP,YAAY;AAAA,QACZ,YAAY,CAAC;AAAA,QACb,QAAQ,KAAK,IAAI,IAAI;AAAA,QACrB;AAAA,QACA,OAAO,SAAS,SAAS;AAAA,QACzB,cAAc;AAAA,QACd;AAAA,QACA,WAAW;AAAA,MACb,CAAC;AAAA,IACH;AAGA,qBAAiB,QAAQ;AAAA,MACvB,kBAAkB,iBAAiB,OAAO,OAAO,YAAY;AAAA,MAC7D,OAAO,aAAa,KAAK;AAAA,IAC3B,CAAC;AAED,aAAS,YAAY,GAAG,YAAY,OAAO,UAAU,aAAa;AAChE,UAAI,WAAW,OAAO,SAAS;AAC7B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,YAAY,WAAW,MAAM;AAAA,UACrC,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,aAAa,mBAAmB,QAAQ,YAAY;AAC1D,UAAI,WAAW,MAAM;AACnB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,WAAW;AAAA,UACnB,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,MAAM;AAAA,QACV,OAAO;AAAA,QACP;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,WAAW;AAAA,QACX;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,eAAO,OAAO,aAAa,MAAM,OAAO,WAAW,GAAG,IAAI,oBAAoB,KAAK;AAAA,MACrF,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,eAAe,WAAW,GAAG,CAAC;AAC9D,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI,KAAK,MAAM;AACb,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,QAAQ,KAAK;AAAA,UACb,OAAO,KAAK;AAAA,UACZ,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc,KAAK;AAAA,UACnB;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,UAAI;AACJ,UAAI;AACF,mBAAW,MAAM,OAAO,OAAO,GAAG;AAAA,MACpC,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,UAAU,WAAW,GAAG,CAAC;AACzD,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI,SAAS,SAAS,QAAQ;AAC5B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM,SAAS,QAAQ;AAAA,UACvB,WAAW;AAAA,UACX,QAAQ,SAAS;AAAA,UACjB,OAAO,SAAS;AAAA,UAChB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc,SAAS,SAAS,QAAQ,YAAY;AAAA,UACpD;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,oBAAoB,kBAAkB,SAAS,QAAQ,OAAO,YAAY;AAChF,YAAM,uBAAuB,iBAAiB,QAAQ,EAAE,kBAAkB,CAAC;AAC3E,YAAM,uBAAuB,iBAAiB;AAC9C,UAAI,sBAAsB;AACxB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,qBAAqB;AAAA,UAC7B,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,cAAc;AACpB,YAAM,cAAc;AACpB,YAAM,cAAc,aAAa,KAAK;AACtC,YAAM,gBAAgB,KAAK,IAAI;AAC/B,YAAM,aAAa,UACf,MAAM;AAAA,QAAS;AAAA,QAAe;AAAA,QAAW,MACvC,QAAQ,KAAK;AAAA,UACX,MAAM,gBAAgB,SAAS;AAAA,UAC/B,UAAU;AAAA,UACV,MAAM,SAAS;AAAA,UACf,YAAY;AAAA,YACV,UAAU,SAAS,UAAU;AAAA,YAC7B;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,IACA;AACJ,UAAI;AACJ,UAAI;AACF,cAAM,SAAS,MAAM,OAAO,IAAI,SAAS,QAAQ,GAAG;AACpD,cAAM,aAAa,OAAO,mBAAmB;AAAA,UAC3C,QAAQ,SAAS;AAAA,UACjB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,QACF,CAAC;AACD,cAAM,UAAU,uBAAuB,YAAY,eAAe,SAAS;AAC3E,YAAI,YAAY,UAAa,OAAO,SAAS,OAAO,KAAK,UAAU,GAAG;AACpE,0BAAgB;AAChB,gBAAM;AAAA,YACJ;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,UACF;AAAA,QACF;AACA,wBAAgB;AAAA,UACd,IAAI;AAAA,UACJ;AAAA,UACA,GAAI,YAAY,SAAY,EAAE,QAAQ,IAAI,CAAC;AAAA,UAC3C,YAAY,KAAK,IAAI,IAAI;AAAA,QAC3B;AAAA,MACF,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,OAAO,WAAW,GAAG,CAAC;AACtD,wBAAgB;AAAA,UACd,IAAI;AAAA,UACJ,OAAO,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UAChD,YAAY,KAAK,IAAI,IAAI;AAAA,QAC3B;AACA,YAAI,kBAAkB,QAAQ;AAC5B,gBAAM;AAAA,YAAS;AAAA,YAAe;AAAA,YAAW,MACvC,YAAY,KAAK,cAAc,SAAS,eAAe;AAAA,UACzD;AACA,gBAAMA,QAA2D;AAAA,YAC/D,OAAO;AAAA,YACP;AAAA,YACA;AAAA,YACA,YAAY;AAAA,YACZ;AAAA,YACA,YAAY;AAAA,YACZ;AAAA,YACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,YAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,UAClC;AACA,kBAAQ,KAAKA,KAAI;AACjB,gBAAM,UAAU,OAAO,QAAQA,OAAM,aAAa;AAClD,iBAAO,OAAO,SAAS;AAAA,YACrB,QAAQ,OAAO;AAAA,YACf,MAAM;AAAA,YACN,WAAW;AAAA,YACX,QAAQ,cAAc,SAAS;AAAA,YAC/B,OAAO,aAAa,KAAK;AAAA,YACzB,OAAO;AAAA,YACP,YAAY;AAAA,YACZ,YAAY;AAAA,YACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,YACrB;AAAA,YACA,OAAO,SAAS,SAAS;AAAA,YACzB,cAAc;AAAA,YACd;AAAA,YACA,WAAW;AAAA,UACb,CAAC;AAAA,QACH;AAAA,MACF;AAEA,UAAI;AACF,gBAAQ,MAAM,OAAO,QAAQ,EAAE,SAAS,aAAa,WAAW,OAAO,CAAC;AAAA,MAC1E,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,WAAW,WAAW,GAAG,CAAC;AAC1D,cAAMA,QAA2D;AAAA,UAC/D,OAAO;AAAA,UACP;AAAA,UACA;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,UAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,QAClC;AACA,gBAAQ,KAAKA,KAAI;AACjB,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,KAAK,cAAc,cAAc,SAAS,CAAC,EAAG,OAAO;AAAA,QACnE;AACA,cAAM,UAAU,OAAO,QAAQA,OAAM,aAAa;AAClD,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI;AACF,gBAAQ,MAAM,OAAO,SAAS;AAAA,UAC5B,QAAQ,OAAO;AAAA,UACf;AAAA,UACA;AAAA,UACA,aAAa,WAAW;AAAA,QAC1B,CAAC;AACD,cAAM;AAAA,UACJ;AAAA,UACA;AAAA,UACA,QAAQ,SAAS;AAAA,UACjB;AAAA,UACA;AAAA,UACA,YAAY,KAAK;AAAA,QACnB;AAAA,MACF,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,YAAY,WAAW,GAAG,CAAC;AAC3D,cAAMA,QAA2D;AAAA,UAC/D,OAAO;AAAA,UACP;AAAA,UACA;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,UAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,QAClC;AACA,gBAAQ,KAAKA,KAAI;AACjB,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,KAAK,cAAc,cAAc,SAAS,CAAC,EAAG,OAAO;AAAA,QACnE;AACA,cAAM,UAAU,OAAO,QAAQA,OAAM,aAAa;AAClD,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,YAAM,aAAa,aAAa,KAAK;AACrC,YAAM,mBAAmB,iBAAiB,OAAO,OAAO,YAAY;AACpE,YAAM,mBAAmB,iBAAiB,QAAQ,EAAE,kBAAkB,OAAO,WAAW,CAAC;AACzF,YAAM,mBAAmB,iBAAiB;AAE1C,YAAM,OAA2D;AAAA,QAC/D,OAAO;AAAA,QACP;AAAA,QACA;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,QACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,QAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,MAClC;AACA,cAAQ,KAAK,IAAI;AACjB,UAAI,cAAc,IAAI;AACpB,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,IAAI;AAAA,YACd,YAAY;AAAA,cACV,eAAe,cAAc,WAAW;AAAA,cACxC;AAAA,cACA,aAAa,eAAe;AAAA,cAC5B,YAAY,cAAc;AAAA,cAC1B;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF,OAAO;AACL,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,KAAK,cAAc,SAAS,iBAAiB;AAAA,YACvD,YAAY;AAAA,cACV;AAAA,cACA;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,YAAM,UAAU,OAAO,QAAQ,MAAM,aAAa;AAElD,UAAI,kBAAkB;AACpB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,iBAAiB;AAAA,UACzB,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,qBAAqB,mBAAmB,QAAQ,YAAY;AAClE,UAAI,mBAAmB,MAAM;AAC3B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,mBAAmB;AAAA,UAC3B,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,cAAc;AAAA,QAClB,OAAO;AAAA,QACP;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,QACA;AAAA,QACA,WAAW;AAAA,QACX;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,uBAAe,OAAO,aAClB,MAAM,OAAO,WAAW,WAAW,IACnC,oBAAoB,KAAK;AAAA,MAC/B,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,eAAe,YAAY,GAAG,GAAG,CAAC;AAClE,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI,aAAa,MAAM;AACrB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM,aAAa;AAAA,UACnB,WAAW;AAAA,UACX,QAAQ,aAAa;AAAA,UACrB,OAAO,aAAa;AAAA,UACpB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc,aAAa;AAAA,UAC3B;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAAA,IACF;AAEA,WAAO,OAAO,SAAS;AAAA,MACrB,QAAQ,OAAO;AAAA,MACf,MAAM;AAAA,MACN,WAAW;AAAA,MACX,QAAQ,8BAA8B,OAAO,QAAQ;AAAA,MACrD,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,MACrB;AAAA,MACA,OAAO,SAAS,SAAS;AAAA,MACzB,cAAc;AAAA,MACd;AAAA,MACA,WAAW;AAAA,IACb,CAAC;AAAA,EACH,SAAS,KAAK;AACZ,kBAAc,KAAK,aAAa,OAAO,QAAQ,QAAQ,GAAG,CAAC;AAC3D,WAAO,OAAO,SAAS;AAAA,MACrB,QAAQ,OAAO;AAAA,MACf,MAAM;AAAA,MACN,WAAW;AAAA,MACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,MACjD,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY,CAAC;AAAA,MACb,QAAQ,KAAK,IAAI,IAAI;AAAA,MACrB;AAAA,MACA,OAAO,SAAS,SAAS;AAAA,MACzB,cAAc;AAAA,MACd;AAAA,MACA,WAAW;AAAA,IACb,CAAC;AAAA,EACH,UAAE;AACA,QAAI,UAAW,cAAa,SAAS;AACrC,QAAI,OAAO,OAAQ,QAAO,OAAO,oBAAoB,SAAS,aAAa;AAAA,EAC7E;AACF;AAEO,SAAS,iBACd,oBACA,UAA4E,CAAC,GACvC;AACtC,SAAO,EAAE,GAAG,SAAS,mBAAmB;AAC1C;AAEO,SAAS,qBACd,oBACA,UAA4E,CAAC,GACvC;AACtC,SAAO,EAAE,GAAG,SAAS,mBAAmB;AAC1C;AAEO,SAAS,cAAc,OAAgE;AAC5F,SAAO,EAAE,GAAG,OAAO,WAAW,KAAK;AACrC;AAEO,SAAS,eAAe,OAAgE;AAC7F,SAAO,EAAE,GAAG,OAAO,WAAW,MAAM;AACtC;AAEA,SAAS,gBAAgB,OAA0D;AACjF,QAAM,MAAM,EAAE,GAAG,gBAAgB,GAAG,MAAM;AAC1C,MAAI,CAAC,OAAO,UAAU,IAAI,QAAQ,KAAM,IAAI,WAAsB,GAAG;AACnE,UAAM,IAAI;AAAA,MACR,+DAA+D,OAAO,IAAI,QAAQ,CAAC;AAAA,IACrF;AAAA,EACF;AACA,QAAM,SAAwB,EAAE,UAAU,IAAI,SAAmB;AACjE,MAAI,IAAI,cAAc,QAAW;AAC/B,QACE,OAAO,IAAI,cAAc,YACzB,CAAC,OAAO,SAAS,IAAI,SAAS,KAC9B,IAAI,aAAa,GACjB;AACA,YAAM,IAAI;AAAA,QACR,yEAAyE,OAAO,IAAI,SAAS,CAAC;AAAA,MAChG;AAAA,IACF;AACA,WAAO,YAAY,IAAI;AAAA,EACzB;AACA,MAAI,IAAI,eAAe,QAAW;AAChC,QACE,OAAO,IAAI,eAAe,YAC1B,CAAC,OAAO,SAAS,IAAI,UAAU,KAC/B,IAAI,aAAa,GACjB;AACA,YAAM,IAAI;AAAA,QACR,6EAA6E,OAAO,IAAI,UAAU,CAAC;AAAA,MACrG;AAAA,IACF;AACA,WAAO,aAAa,IAAI;AAAA,EAC1B;AACA,SAAO;AACT;AAEA,SAAS,uBACP,SACA,eACA,WACoB;AACpB,MAAI,YAAY,OAAW,QAAO;AAClC,MAAI,CAAC,OAAO,SAAS,OAAO,KAAK,UAAU,GAAG;AAC5C,kBAAc;AAAA,MACZ,aAAa,OAAO,WAAW,IAAI,MAAM,2BAA2B,OAAO,OAAO,CAAC,EAAE,CAAC;AAAA,IACxF;AACA,WAAO;AAAA,EACT;AACA,SAAO;AACT;AAEO,SAAS,kBAAkB,OAAqC;AACrE,SAAO,MAAM;AAAA,IACX,CAAC,WAAW,OAAO,UAAW,OAAO,aAAa,cAAc,OAAO,aAAa;AAAA,EACtF;AACF;AAEA,SAAS,YACP,QACA,OACA,OACA,SACA,QACA,WACA,SACA,cACA,aACA,SACuD;AACvD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,QAAQ,KAAK,IAAI,IAAI;AAAA,IACrB;AAAA,IACA,kBACE,OAAO,eAAe,SAAY,SAAY,KAAK,IAAI,GAAG,OAAO,aAAa,YAAY;AAAA,IAC5F;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,oBAAoB,OAA0C;AACrE,MAAI,CAAC,MAAM,OAAQ,QAAO,EAAE,MAAM,OAAO,MAAM,OAAO,QAAQ,eAAe;AAC7E,QAAM,OAAO,kBAAkB,KAAK;AACpC,SAAO,OACH,EAAE,MAAM,MAAM,MAAM,MAAM,QAAQ,6BAA6B,OAAO,aAAa,KAAK,EAAE,IAC1F;AAAA,IACE,MAAM;AAAA,IACN,MAAM;AAAA,IACN,QAAQ;AAAA,IACR,OAAO,aAAa,KAAK;AAAA,EAC3B;AACN;AAEA,SAAS,aAAa,OAAgD;AACpE,QAAM,SAAS,MACZ,IAAI,CAAC,WAAW,OAAO,KAAK,EAC5B,OAAO,CAAC,UAA2B,OAAO,UAAU,QAAQ;AAC/D,MAAI,CAAC,OAAO,OAAQ,QAAO;AAC3B,SAAO,KAAK,MAAO,OAAO,OAAO,CAAC,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO,SAAU,GAAI,IAAI;AAC9F;AAEA,SAAS,mBACP,QACA,cACmC;AACnC,MAAI,OAAO,eAAe,UAAa,gBAAgB,OAAO,YAAY;AACxE,WAAO;AAAA,MACL,MAAM;AAAA,MACN,QAAQ,gCAAgC,OAAO,UAAU;AAAA,IAC3D;AAAA,EACF;AACA,SAAO,EAAE,MAAM,OAAO,QAAQ,GAAG;AACnC;AAEA,eAAe,iBACb,SACA,QACA,cACA,QACA,eACA,WACe;AACf,MAAI,CAAC,WAAW,OAAO,eAAe,OAAW;AACjD,QAAM,aAAa,OAAO;AAC1B,QAAM;AAAA,IAAS;AAAA,IAAe;AAAA,IAAW,MACvC,QAAQ,aAAa;AAAA,MACnB,WAAW;AAAA,MACX,OAAO;AAAA,MACP,UAAU;AAAA,MACV,WAAW,KAAK,IAAI,GAAG,aAAa,YAAY;AAAA,MAChD,UAAU,gBAAgB;AAAA,MAC1B,QAAQ,QAAQ,KAAK;AAAA,IACvB,CAAC;AAAA,EACH;AACF;AAEA,eAAe,gBACb,SACA,OACA,OACA,eACA,WACA,cACe;AACf,MAAI,CAAC,QAAS;AACd,aAAW,UAAU,OAAO;AAC1B,UAAM;AAAA,MAAS;AAAA,MAAe;AAAA,MAAW,MACvC,QAAQ,YAAY;AAAA,QAClB,SAAS,OAAO,YAAY,wBAAwB;AAAA,QACpD,cAAc,gBAAgB,QAAQ;AAAA,QACtC,MAAM,gBAAgB,OAAO,EAAE;AAAA,QAC/B,WAAW,OAAO;AAAA,QAClB,OAAO,OAAO,OAAO,UAAU,WAAW,OAAO,QAAQ,OAAO,SAAS,IAAI;AAAA,QAC7E,WAAW,OAAO;AAAA,QAClB,UAAU,OAAO;AAAA,QACjB,YAAY;AAAA,UACV;AAAA,UACA,QAAQ,OAAO;AAAA,UACf,UAAU,OAAO;AAAA,UACjB,WAAW,OAAO;AAAA,QACpB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAEA,eAAe,UACb,QACA,MACA,eACe;AACf,MAAI,CAAC,OAAQ;AACb,MAAI;AACF,UAAM,OAAO,IAAI;AAAA,EACnB,SAAS,KAAK;AACZ,kBAAc,KAAK,aAAa,WAAW,KAAK,OAAO,GAAG,CAAC;AAAA,EAC7D;AACF;AAEA,eAAe,SACb,eACA,WACA,OACwB;AACxB,MAAI;AACF,WAAO,MAAM,MAAM;AAAA,EACrB,SAAS,KAAK;AACZ,kBAAc,KAAK,aAAa,SAAS,WAAW,GAAG,CAAC;AACxD,WAAO;AAAA,EACT;AACF;AAEA,SAAS,iBACP,OACA,UACQ;AACR,MAAI,UAAU,iBAAkB,QAAO,SAAS,iBAAiB,KAAK;AACtE,SAAO,kBAAkB,KAAK;AAChC;AAEA,SAAS,kBACP,QACA,UACQ;AACR,MAAI,UAAU,kBAAmB,QAAO,SAAS,kBAAkB,MAAM;AACzE,SAAO,kBAAkB,MAAM;AACjC;AAEA,SAAS,kBAAkB,OAAwB;AACjD,MAAI,OAAO,UAAU,SAAU,QAAO;AACtC,MAAI,OAAO,UAAU,YAAY,OAAO,UAAU,aAAa,SAAS,KAAM,QAAO,OAAO,KAAK;AACjG,MAAI;AACF,WAAO,KAAK,UAAU,mBAAmB,KAAK,CAAC;AAAA,EACjD,QAAQ;AACN,WAAO,OAAO,KAAK;AAAA,EACrB;AACF;AAEA,SAAS,mBAAmB,OAAyB;AACnD,MAAI,MAAM,QAAQ,KAAK,EAAG,QAAO,MAAM,IAAI,kBAAkB;AAC7D,MAAI,CAAC,SAAS,OAAO,UAAU,SAAU,QAAO;AAChD,QAAM,SAAS;AACf,QAAM,SAAkC,CAAC;AACzC,aAAW,OAAO,OAAO,KAAK,MAAM,EAAE,KAAK,GAAG;AAC5C,WAAO,GAAG,IAAI,mBAAmB,OAAO,GAAG,CAAC;AAAA,EAC9C;AACA,SAAO;AACT;AAEA,SAAS,YAAY,QAA6B;AAChD,QAAM,SAAS,OAAO;AACtB,MAAI,kBAAkB,MAAO,QAAO,OAAO;AAC3C,SAAO,SAAS,OAAO,MAAM,IAAI;AACnC;AAEA,SAAS,aACP,OACA,WACA,KACqB;AACrB,QAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,SAAO,EAAE,OAAO,WAAW,QAAQ;AACrC;AAEA,eAAe,OACb,SACA,QACkE;AAClE,QAAM;AAAA,IAAS,OAAO;AAAA,IAAe,OAAO,MAAM;AAAA,IAAQ,MACxD,SAAS,OAAO;AAAA,MACd,MAAM,OAAO;AAAA,MACb,OAAO,OAAO,SAAS,aAAa,OAAO,UAAU;AAAA,MACrD,cAAc,OAAO;AAAA,MACrB,OAAO,OAAO;AAAA,IAChB,CAAC;AAAA,EACH;AACA,SAAO;AACT;","names":["step"]}

package/dist/control.js

@@ -4,7 +4,7 @@ import {
   runProposeReview,
   runProposeReviewAsControlLoop,
   scoreFromEvals
-} from "./chunk-RTWFUK6A.js";
+} from "./chunk-E4GH6USR.js";
 import {
   allCriticalPassed,
   objectiveEval,
@@ -12,7 +12,7 @@ import {
   stopOnNoProgress,
   stopOnRepeatedAction,
   subjectiveEval
-} from "./chunk-NCRFYPS3.js";
+} from "./chunk-YEHAEDUD.js";
 import "./chunk-KWRRMR3J.js";
 import "./chunk-TVVP3ZZQ.js";
 import "./chunk-VSMTAMNK.js";

package/dist/index.d.ts

@@ -10,6 +10,8 @@ import { TCloud } from '@tangle-network/tcloud';
 import { B as BenchmarkRunnerConfig, S as Scenario, c as BenchmarkReport, P as ProductClientConfig, C as CheckResult, T as TestResult, d as PersonaConfig, D as DriverResult, e as DriverState, b as JudgeFn, f as CollectedArtifacts, g as ScenarioResult, h as TurnMetrics, i as ScenarioFile, j as CompletionCriterion } from './types-Croy5h7V.js';
 export { A as ArtifactCheck, k as ArtifactResult, E as EvalResult, F as FeedbackPattern, l as JudgeConfig, a as JudgeInput, m as JudgeRubric, J as JudgeScore, n as PersonaRigor, R as RouteMap, o as RubricDimension, p as Turn, q as TurnResult } from './types-Croy5h7V.js';
 export { c as ControlActionFailureMode, d as ControlActionOutcome, e as ControlBudget, f as ControlContext, g as ControlDecision, C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig, i as ControlRuntimeError, j as ControlSeverity, b as ControlStep, k as ControlStopPolicies, S as StopDecision, l as allCriticalPassed, o as objectiveEval, r as runAgentControlLoop, s as stopOnNoProgress, m as stopOnRepeatedAction, n as subjectiveEval } from './control-runtime-DuFBYg7A.js';
+import { F as FailureClass, h as BudgetSpec, B as BudgetLedgerEntry, R as Run, L as LlmSpan } from './schema-m0gsnbt3.js';
+export { A as Artifact, E as EventKind, i as FAILURE_CLASSES, G as GenericSpan, J as JudgeSpan, M as Message, d as RetrievalSpan, g as RunLayer, f as RunStatus, e as SandboxSpan, S as Span, j as SpanBase, c as SpanKind, k as SpanStatus, l as TRACE_SCHEMA_VERSION, T as ToolSpan, a as TraceEvent, m as isJudgeSpan, n as isLlmSpan, o as isRetrievalSpan, p as isSandboxSpan, q as isToolSpan } from './schema-m0gsnbt3.js';
 import { A as AgentEvalError, J as JudgeError, a as ConfigError } from './errors-CzMUYo7b.js';
 export { b as AgentEvalErrorCode, C as CaptureIntegrityError, N as NotFoundError, R as ReplayError, V as ValidationError, c as VerificationError } from './errors-CzMUYo7b.js';
 import { b as FeedbackLabel, F as FeedbackTrajectoryStore, a as FeedbackTrajectory } from './feedback-trajectory-D9OVLrg9.js';
@@ -39,8 +41,6 @@ export { b as RunIntegrityError, R as RunIntegrityExpectations, c as RunIntegrit
 export { a as aggregateLlm, b as argHash, g as groupBy, j as judgeSpans, l as llmSpans, r as runFailureClass, c as runsForScenario, t as toolSpans } from './query-CqTxMwDw.js';
 export { F as FileSystemRawProviderSink, a as FileSystemRawProviderSinkOptions, I as InMemoryRawProviderSink, b as InMemoryRawProviderSinkOptions, N as NoopRawProviderSink, P as ProviderRedactor, c as RawProviderDirection, d as RawProviderEvent, R as RawProviderSink, e as RawProviderSinkFilter, f as defaultProviderRedactor, p as providerFromBaseUrl } from './raw-provider-sink-C46HDghv.js';
 export { D as DEFAULT_REDACTION_RULES, b as REDACTION_VERSION, a as RedactionReport, R as RedactionRule, r as redactString, c as redactValue } from './redact-B40YG2M_.js';
-import { h as BudgetSpec, B as BudgetLedgerEntry, R as Run, L as LlmSpan } from './schema-m0gsnbt3.js';
-export { A as Artifact, E as EventKind, i as FAILURE_CLASSES, F as FailureClass, G as GenericSpan, J as JudgeSpan, M as Message, d as RetrievalSpan, g as RunLayer, f as RunStatus, e as SandboxSpan, S as Span, j as SpanBase, c as SpanKind, k as SpanStatus, l as TRACE_SCHEMA_VERSION, T as ToolSpan, a as TraceEvent, m as isJudgeSpan, n as isLlmSpan, o as isRetrievalSpan, p as isSandboxSpan, q as isToolSpan } from './schema-m0gsnbt3.js';
 import { T as TraceStore, R as RunFilter } from './store-CKUAgsJz.js';
 export { E as EventFilter, F as FileSystemTraceStore, a as FileSystemTraceStoreOptions, I as InMemoryTraceStore, S as SpanFilter } from './store-CKUAgsJz.js';
 export { D as DEFAULT_FAILURE_RULES, b as FailureClassification, c as FailureContext, d as FailureRule, e as classifyFailure } from './failure-cluster-CL7IVgkJ.js';
@@ -276,6 +276,87 @@ declare class ProductClient {
  */
 declare function runE2EWorkflow(client: ProductClient, name: string, workflow: (client: ProductClient) => Promise<CheckResult[]>): Promise<TestResult>;
 
+/**
+ * Streaming detectors — the canonical, online failure-mode kernel.
+ *
+ * A `StreamingDetector` is a pure incremental reducer: fold one normalized step, get a signal the
+ * moment a threshold trips (else null). The SAME kernel runs in two places, so the logic lives once:
+ *   - ONLINE, over a live agent pipe (a worker's tool-call stream) → raise a finding mid-run.
+ *   - INSIDE the control loop (`control-runtime`) → its `stopOn*` policies fold these detectors.
+ *
+ * Detectors are stateful but self-contained; `streak` is exposed for telemetry and `reset` clears it.
+ * Fingerprinting is the CALLER's job — the control loop hashes state/action with `stableFingerprint`,
+ * a tool-call pipe hashes args with `argHash` — so a detector stays agnostic to what it's watching.
+ */
+
+type DetectorSeverity = 'info' | 'warn' | 'critical';
+/** A normalized step a detector folds over. Fields are optional; each detector reads only what it
+ *  needs (repeated-action → `actionFingerprint`; no-progress → `stateFingerprint` + `score`;
+ *  error-streak → `status`). */
+interface DetectorEvent {
+    /** Fingerprint of the action/tool-call this step took (caller pre-hashes). */
+    readonly actionFingerprint?: string;
+    /** Fingerprint of observable state AFTER this step. */
+    readonly stateFingerprint?: string;
+    /** Score after this step (the score-flat half of no-progress). */
+    readonly score?: number;
+    /** Whether this step errored. */
+    readonly status?: 'ok' | 'error';
+    /** Free-form label carried into the signal evidence (e.g. the tool name). */
+    readonly label?: string;
+}
+interface DetectorSignal {
+    readonly detector: string;
+    readonly severity: DetectorSeverity;
+    readonly failureClass?: FailureClass;
+    readonly reason: string;
+    /** Consecutive matching steps at the moment the signal fired. */
+    readonly streak: number;
+    readonly evidence?: Record<string, unknown>;
+}
+interface StreamingDetector {
+    readonly id: string;
+    /** Current streak (consecutive matching steps) — for telemetry/introspection between observes. */
+    readonly streak: number;
+    /** Fold one event; return a signal when the threshold trips, else null. */
+    observe(event: DetectorEvent): DetectorSignal | null;
+    /** Clear all state. */
+    reset(): void;
+}
+interface RepeatedActionOptions {
+    /** Signal once the same action fingerprint repeats this many CONSECUTIVE steps (default 3).
+     *  `<= 0` disables the signal (the streak is still tracked, for telemetry). */
+    readonly maxRepeated?: number;
+    readonly severity?: DetectorSeverity;
+    /** Failure class to stamp on the signal (default `tool_recovery_failure`). */
+    readonly failureClass?: FailureClass;
+}
+/** Same action fingerprint N consecutive steps = a stuck loop (the #1 long-horizon failure mode). */
+declare function repeatedActionDetector(opts?: RepeatedActionOptions): StreamingDetector;
+interface NoProgressOptions {
+    /** Signal once state is unchanged AND score is flat for this many CONSECUTIVE steps (default 3).
+     *  `<= 0` disables the signal. */
+    readonly maxNoProgress?: number;
+    /** Minimum |score change| that counts as progress (default 0.001). */
+    readonly minScoreDelta?: number;
+    readonly severity?: DetectorSeverity;
+    readonly failureClass?: FailureClass;
+}
+/** State + score unchanged across N steps = spinning wheels. Compares each step to the previous one,
+ *  so prime with the initial state (observe it once before the first real step) to detect on step 1. */
+declare function noProgressDetector(opts?: NoProgressOptions): StreamingDetector;
+interface ErrorStreakOptions {
+    /** Signal once this many CONSECUTIVE steps errored (default 3). `<= 0` disables. */
+    readonly maxErrors?: number;
+    readonly severity?: DetectorSeverity;
+    readonly failureClass?: FailureClass;
+}
+/** N consecutive tool errors = the worker is hammering a broken approach. */
+declare function errorStreakDetector(opts?: ErrorStreakOptions): StreamingDetector;
+/** Fold one event through many detectors at once; returns every signal that fired this step. The
+ *  natural shape for an online pipe watching with a whole panel of detectors. */
+declare function observeAll(detectors: ReadonlyArray<StreamingDetector>, event: DetectorEvent): DetectorSignal[];
+
 interface AgentDriverConfig {
     client: ProductClient;
     driverModel?: string;
@@ -317,6 +398,35 @@ declare class AgentDriver {
  * — exported so harness authors can inspect and regression-test it.
  */
 declare function buildDriverSystemPrompt(persona: PersonaConfig, state: DriverState, productContext?: string): string;
+interface WorkerDriverContext {
+    /** The goal (or sub-goal) the driven worker must actually accomplish. */
+    goal: string;
+    /** The worker's harness — e.g. 'claude-code' | 'codex' | 'opencode' | 'router-tools'.
+     *  Names which capability profile the driver should exploit. */
+    harness?: string;
+    /** A capability + caveat brief for THIS harness (parallel tool calls, sub-agents and
+     *  their depth/concurrency limits, autonomy/runaway profile, MCP, native tool-isolation)
+     *  — sourced from the harness-compat matrix. Free text so the substrate stays decoupled
+     *  from any runtime harness type. */
+    harnessBrief?: string;
+    /** What the worker has done so far — the trace/state summary the driver reasons over to
+     *  write its next instruction. Empty on the first turn. */
+    progress?: string;
+    /** Optional extra context (repo, constraints, the deliverable's acceptance check). */
+    context?: string;
+}
+/**
+ * Build the WORKER-DRIVER system prompt — the harness-aware sibling of
+ * `buildDriverSystemPrompt`. Where that one role-plays a demanding *user* of a
+ * product, this one is a meta-agent that DRIVES a capable coding *worker* to
+ * complete a goal: it writes rich, high-signal instructions that direct the
+ * worker to exploit its harness's full power (parallelize, sub-agents, run-to-
+ * completion, tools/MCP), matched to what THAT harness can actually do. The
+ * load-bearing contract: the driver never writes a thin steer — every message
+ * is the dense, specific directive a world-class engineering lead would write.
+ * Pure function; exported so harness authors can inspect and regression-test it.
+ */
+declare function buildWorkerDriverSystemPrompt(ctx: WorkerDriverContext): string;
 interface DecideNextUserTurnOpts {
     persona: PersonaConfig;
     state: DriverState;
@@ -5531,4 +5641,4 @@ type CachedJudge<TArtifact, TScenario extends Scenario$1 = Scenario$1> = JudgeCo
  */
 declare function cachedJudge<TArtifact, TScenario extends Scenario$1 = Scenario$1>(judge: JudgeConfig<TArtifact, TScenario>, store: VerdictCacheStore, options: CachedJudgeOptions): CachedJudge<TArtifact, TScenario>;
 
-export { ATTESTATION_ALGORITHM, type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, AgentEvalError, AgentProfile$1 as AgentProfile, type AgreementResult, type AlignmentOp, AnalyzeTracesInput, AnalyzeTracesOptions, AnalyzeTracesResult, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, type AssertCrossFamilyOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, type AutoPrClient, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, type BackendDescriptor, BaselineReport, BehaviorAssertion, BenchmarkReport, BenchmarkRunner, BenchmarkRunnerConfig, type BisectOptions, type BisectResult, type BisectStep, BudgetBreachError, BudgetGuard, BudgetLedgerEntry, BudgetSpec, type BuildAgreementJudgeOptions, type CachedJudge, type CachedJudgeOptions, CallExpectation, type CanaryAlert, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CanarySeverity, type CandidateComparison, type CandidateScenario, type CausalAttributionReport, type CellVerdict, ChannelRollup, ChatRequest, CheckResult, CollectedArtifacts, type CommandRunner, type CompareLabels, CompletionCriterion, ConfigError, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContractCheckResult, type ContractJudgeOptions, type ContractMetric, type ContractReport, type ContractRule, type ContractRuleKind, type ContractSpan, type ContractVerdict, type ContractViolation, ConvergenceTracker, CorrectnessChecker, type CostEntry, CostLedger, type CostReport, type CostSummary, CostTracker, CreateChatClientOpts, type CreateDefaultReviewerOptions, type CreateExperimentInput, type CreateSandboxPoolOpts, CrossFamilyError, type CrossTraceDiff, type CrossTraceDiffOptions, DEFAULT_AGENT_SLOS, DEFAULT_FINDERS, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_PR_REVIEW_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, Dataset, DatasetScenario, type DecideNextUserTurnOpts, DefaultVerdict, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DescriptionLengthCandidate, type DescriptionLengthConfig, type DescriptionLengthDecision, type DescriptionLengthEvidence, DescriptionLengthGate, type DescriptionLengthRejectionCode, type DiffScorecardOptions, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, DriverResult, DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, ERROR_COUNT_PATTERNS, type EnsembleAggregate, type EnsembleJudgeOptions, type ErrorCountPattern, type EvalToolDef, EvalTraceStore, type EvolutionRound, type ExecutorConfig, type Expectation, type Experiment, type ExperimentProvenance, type ExperimentRep, type ExperimentStats, type ExperimentStore, ExperimentTracker, type ExperimentTrackerOptions, type ExperimentVerdict, type ExportedRewardModel, type ExtractOptions, type ExtractResult, type FactorContribution, type FactorialCell, FeedbackLabel, FeedbackTrajectory, FeedbackTrajectoryStore, type FieldAgreementSpec, type FileChange, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, type GhCliClientOptions, type GoldScenario, type GoldSplit, type GoldenSeverity, type GoldenSpec, HarnessConfig, type HeldOutPartition, HoldoutAuditor, type HttpGithubClientOptions, INTENT_MATCH_JUDGE_VERSION, type ImageData, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, JudgeError, type JudgeFamily, type JudgeFleetOptions, JudgeFn, JudgeParseError, type JudgeReplayResult, type JudgeRetryOutcome, type JudgeRetryPolicy, JudgeRunner, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, Layer, LayerResult, type LiveProofArtifact, type LiveProofConfig, type LiveProofContext, type LiveProofResult, LlmClientOptions, LlmSpan, LockedJsonlAppender, MODEL_PRICING, type MakeEvalToolsConfig, type MatchResult, type MatcherResult, type MergeOptions, MetricsCollector, type ModelCostRollup, type ModelPreflight, type ModelSeats, ModelsUnreachableError, type MuffledFinder, type MuffledFinding, type MultiToolchainLayerConfig, type Mutator, Mutex, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, OtelExportConfig, OtelExporter, type OtelPipelineHandle, type OtelPipelineOptions, PairwiseSteeringOptimizer, type ParaphraseRobustnessScenarioInput, type ParaphraseRobustnessScenarioResult, type ParseStudentLabel, type PartitionHeldOutOptions, PersonaConfig, type Playbook, type PlaybookEntry, type PoolSlot, type PrReviewAuditCase, type PrReviewBenchmarkSummary, type PrReviewComment, type PrReviewMatchedFinding, type PrReviewOutcome, type PrReviewReferenceFinding, type PrReviewScore, type PrReviewScoreWeights, type PrReviewSeverity, type PrReviewSource, type PreflightModelsOptions, type PreflightOutcome, ProductClient, ProductClientConfig, type PromptHandle, PromptRegistry, type ProposeAutomatedPullRequestInput, type ProposeAutomatedPullRequestResult, type ProvenanceReader, type RecordRunsOptions, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, type ReflectionContext, type ReflectionProposal, ReleaseConfidenceScorecard, ReleaseConfidenceThresholds, type RenderStudentPrompt, type RepoRef, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RobustnessResult, Run, type RunCommandInput, type RunCommandResult, type RunDistillationOptions, type RunDistillationResult, RunFilter, RunRecord, type RunRecordBackend, type RunRecordFilter, RunScore, RunScoreWeights, RunSplitTag, SandboxDriver, SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxPool, type ScanOptions, Scenario, type ScenarioCost, ScenarioFile, ScenarioRegistry, ScenarioResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type ScoredTarget, type SeatName, type SeatPresetName, SeatUnsetError, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SerializedRegex, Severity, type SingleBackendDivergence, SingleBackendError, type SingleBackendField, type SingleBackendReport, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type SlotFactory, type SpanPredicate, type SplitGoldOptions, SteeringBundle, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type StepAttribution, type SynthesisReason, type SynthesisTarget, TestResult, type TextMatcher, type ThresholdContract, TokenCounter, type TokenSpec, type TraceContract, TraceContractBuilder, TraceEmitter, TraceStore, type TracedAnalystOptions, type TracedJudgeOptions, Trajectory, TrajectoryStep, type TrialTrace, TurnMetrics, UI_FINDING_SEVERITIES, UI_LENSES, UNIVERSAL_FINDERS, type UiFinding, type UiFindingScreenshot, type UiFindingSeverity, type UiLens, type VerdictCacheStats, type VerdictCacheStore, VerifyContext, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, type WranglerDeployRunnerInput, adversarialJudge, aggregateJudgeVerdicts, aggregatePrReviewScore, analyzeAntiSlop, appendScorecard, assertCrossFamily, assertModelsServed, assertSingleBackend, assignHeldOutTag, attachCostToReport, attest, bisect, buildAgreementJudge, buildDriverSystemPrompt, buildReflectionPrompt, buildReviewerPrompt, cachedJudge, canaryLeakView, canonicalJson, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, checkTraceContracts, codeExecutionJudge, coherenceJudge, collectionPreserved, commentsForSource, commitBisect, compareReferenceReplay, compilerJudge, computeExperimentStats, contentHash, contractJudge, costReport, createAntiSlopJudge, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createIntentMatchJudge, createSandboxPool, crossTraceDiff, dataDescriptionBits, decideNextUserTurn, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultJudges, defaultParseStudentLabel, defaultReferenceReplayMatcher, defaultRenderStudentPrompt, deployGateLayer, diffScorecard, discoverPersonas, distillPlaybook, ensembleJudge, estimateCost, estimateTokens, evaluateContract, evaluateOracles, evaluateTraceContract, executeScenario, expectAgent, exportRewardModel, extractAssetUrls, extractErrorCount, fieldAgreement, fileContains, fileExists, fileExperimentStore, fileVerdictCache, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findSkipCountsAsPass, flowLayer, fnv1a32, formatBenchmarkReport, formatDriverReport, formatFindings, formatScorecardDiff, ghCliClient, gitProvenanceReader, precision as goldenPrecision, hashContent, hashToUnit, htmlContainsElement, httpGithubClient, improvementVerdict, inMemoryExperimentStore, inMemoryReferenceReplayStore, inMemoryRunRecordBackend, inMemoryVerdictCache, isModelPriced, isOtelConfigured, jsonShape, jsonlReferenceReplayStore, jsonlRunRecordBackend, judgeFamily, keyPreserved, linterJudge, loadGoldScenarios, loadScorecard, loadScorerFromGrader, localCommandRunner, lowercaseMutator, makeEvalTools, matchGoldens, matchSpan, mergeLayerResults, modelDescriptionBits, multiToolchainLayer, notBlocked, paraphraseRobustness, paraphraseRobustnessScenarios, parseGoldJsonl, parseReflectionResponse, partitionHeldOut, passOrthogonality, pixelDeltaRatio, politenessPrefixMutator, preflightModels, printDriverSummary, index as profile, promptBisect, proposeAutomatedPullRequest, proposeSynthesisTargets, recordRuns, recordRunsToScorecard, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatches, renderMarkdownReport, renderPlaybookMarkdown, replayScorerOverCorpus, replayTraceThroughJudge, resetLockedAppendersForTesting, resolveModelPricing, resolveSeat, rowCount, rowWhere, runAssertions, runBehavioralCanaries, runCanaries, runDistillation, runE2EWorkflow, runExpectations, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runReferenceReplay, runScore, runSelfPlay, scanForMuffledGates, scoreContinuity, scorePrReviewComments, scorePrReviewSource, scoreReferenceReplay, seatPresets, securityJudge, sentenceReorderMutator, splitGold, statusAdvanced, summarizePrReviewBenchmark, testJudge, textInSnapshot, toLangfuseEnvelope, toOpenAiTool, toPrometheusText, traceContract, traceJudge, traceJudgeEnsemble, tracedAnalyzeTraces, typoMutator, urlContains, verifyAttestation, visualDiff, viteDeployRunner, weightedRecall, whitespaceCollapseMutator, withJudgeRetry, withOtelPipeline, wranglerDeployRunner };
+export { ATTESTATION_ALGORITHM, type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, AgentEvalError, AgentProfile$1 as AgentProfile, type AgreementResult, type AlignmentOp, AnalyzeTracesInput, AnalyzeTracesOptions, AnalyzeTracesResult, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, type AssertCrossFamilyOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, type AutoPrClient, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, type BackendDescriptor, BaselineReport, BehaviorAssertion, BenchmarkReport, BenchmarkRunner, BenchmarkRunnerConfig, type BisectOptions, type BisectResult, type BisectStep, BudgetBreachError, BudgetGuard, BudgetLedgerEntry, BudgetSpec, type BuildAgreementJudgeOptions, type CachedJudge, type CachedJudgeOptions, CallExpectation, type CanaryAlert, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CanarySeverity, type CandidateComparison, type CandidateScenario, type CausalAttributionReport, type CellVerdict, ChannelRollup, ChatRequest, CheckResult, CollectedArtifacts, type CommandRunner, type CompareLabels, CompletionCriterion, ConfigError, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContractCheckResult, type ContractJudgeOptions, type ContractMetric, type ContractReport, type ContractRule, type ContractRuleKind, type ContractSpan, type ContractVerdict, type ContractViolation, ConvergenceTracker, CorrectnessChecker, type CostEntry, CostLedger, type CostReport, type CostSummary, CostTracker, CreateChatClientOpts, type CreateDefaultReviewerOptions, type CreateExperimentInput, type CreateSandboxPoolOpts, CrossFamilyError, type CrossTraceDiff, type CrossTraceDiffOptions, DEFAULT_AGENT_SLOS, DEFAULT_FINDERS, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_PR_REVIEW_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, Dataset, DatasetScenario, type DecideNextUserTurnOpts, DefaultVerdict, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DescriptionLengthCandidate, type DescriptionLengthConfig, type DescriptionLengthDecision, type DescriptionLengthEvidence, DescriptionLengthGate, type DescriptionLengthRejectionCode, type DetectorEvent, type DetectorSeverity, type DetectorSignal, type DiffScorecardOptions, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, DriverResult, DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, ERROR_COUNT_PATTERNS, type EnsembleAggregate, type EnsembleJudgeOptions, type ErrorCountPattern, type ErrorStreakOptions, type EvalToolDef, EvalTraceStore, type EvolutionRound, type ExecutorConfig, type Expectation, type Experiment, type ExperimentProvenance, type ExperimentRep, type ExperimentStats, type ExperimentStore, ExperimentTracker, type ExperimentTrackerOptions, type ExperimentVerdict, type ExportedRewardModel, type ExtractOptions, type ExtractResult, type FactorContribution, type FactorialCell, FailureClass, FeedbackLabel, FeedbackTrajectory, FeedbackTrajectoryStore, type FieldAgreementSpec, type FileChange, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, type GhCliClientOptions, type GoldScenario, type GoldSplit, type GoldenSeverity, type GoldenSpec, HarnessConfig, type HeldOutPartition, HoldoutAuditor, type HttpGithubClientOptions, INTENT_MATCH_JUDGE_VERSION, type ImageData, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, JudgeError, type JudgeFamily, type JudgeFleetOptions, JudgeFn, JudgeParseError, type JudgeReplayResult, type JudgeRetryOutcome, type JudgeRetryPolicy, JudgeRunner, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, Layer, LayerResult, type LiveProofArtifact, type LiveProofConfig, type LiveProofContext, type LiveProofResult, LlmClientOptions, LlmSpan, LockedJsonlAppender, MODEL_PRICING, type MakeEvalToolsConfig, type MatchResult, type MatcherResult, type MergeOptions, MetricsCollector, type ModelCostRollup, type ModelPreflight, type ModelSeats, ModelsUnreachableError, type MuffledFinder, type MuffledFinding, type MultiToolchainLayerConfig, type Mutator, Mutex, type NoProgressOptions, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, OtelExportConfig, OtelExporter, type OtelPipelineHandle, type OtelPipelineOptions, PairwiseSteeringOptimizer, type ParaphraseRobustnessScenarioInput, type ParaphraseRobustnessScenarioResult, type ParseStudentLabel, type PartitionHeldOutOptions, PersonaConfig, type Playbook, type PlaybookEntry, type PoolSlot, type PrReviewAuditCase, type PrReviewBenchmarkSummary, type PrReviewComment, type PrReviewMatchedFinding, type PrReviewOutcome, type PrReviewReferenceFinding, type PrReviewScore, type PrReviewScoreWeights, type PrReviewSeverity, type PrReviewSource, type PreflightModelsOptions, type PreflightOutcome, ProductClient, ProductClientConfig, type PromptHandle, PromptRegistry, type ProposeAutomatedPullRequestInput, type ProposeAutomatedPullRequestResult, type ProvenanceReader, type RecordRunsOptions, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, type ReflectionContext, type ReflectionProposal, ReleaseConfidenceScorecard, ReleaseConfidenceThresholds, type RenderStudentPrompt, type RepeatedActionOptions, type RepoRef, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RobustnessResult, Run, type RunCommandInput, type RunCommandResult, type RunDistillationOptions, type RunDistillationResult, RunFilter, RunRecord, type RunRecordBackend, type RunRecordFilter, RunScore, RunScoreWeights, RunSplitTag, SandboxDriver, SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxPool, type ScanOptions, Scenario, type ScenarioCost, ScenarioFile, ScenarioRegistry, ScenarioResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type ScoredTarget, type SeatName, type SeatPresetName, SeatUnsetError, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SerializedRegex, Severity, type SingleBackendDivergence, SingleBackendError, type SingleBackendField, type SingleBackendReport, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type SlotFactory, type SpanPredicate, type SplitGoldOptions, SteeringBundle, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type StepAttribution, type StreamingDetector, type SynthesisReason, type SynthesisTarget, TestResult, type TextMatcher, type ThresholdContract, TokenCounter, type TokenSpec, type TraceContract, TraceContractBuilder, TraceEmitter, TraceStore, type TracedAnalystOptions, type TracedJudgeOptions, Trajectory, TrajectoryStep, type TrialTrace, TurnMetrics, UI_FINDING_SEVERITIES, UI_LENSES, UNIVERSAL_FINDERS, type UiFinding, type UiFindingScreenshot, type UiFindingSeverity, type UiLens, type VerdictCacheStats, type VerdictCacheStore, VerifyContext, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WorkerDriverContext, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, type WranglerDeployRunnerInput, adversarialJudge, aggregateJudgeVerdicts, aggregatePrReviewScore, analyzeAntiSlop, appendScorecard, assertCrossFamily, assertModelsServed, assertSingleBackend, assignHeldOutTag, attachCostToReport, attest, bisect, buildAgreementJudge, buildDriverSystemPrompt, buildReflectionPrompt, buildReviewerPrompt, buildWorkerDriverSystemPrompt, cachedJudge, canaryLeakView, canonicalJson, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, checkTraceContracts, codeExecutionJudge, coherenceJudge, collectionPreserved, commentsForSource, commitBisect, compareReferenceReplay, compilerJudge, computeExperimentStats, contentHash, contractJudge, costReport, createAntiSlopJudge, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createIntentMatchJudge, createSandboxPool, crossTraceDiff, dataDescriptionBits, decideNextUserTurn, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultJudges, defaultParseStudentLabel, defaultReferenceReplayMatcher, defaultRenderStudentPrompt, deployGateLayer, diffScorecard, discoverPersonas, distillPlaybook, ensembleJudge, errorStreakDetector, estimateCost, estimateTokens, evaluateContract, evaluateOracles, evaluateTraceContract, executeScenario, expectAgent, exportRewardModel, extractAssetUrls, extractErrorCount, fieldAgreement, fileContains, fileExists, fileExperimentStore, fileVerdictCache, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findSkipCountsAsPass, flowLayer, fnv1a32, formatBenchmarkReport, formatDriverReport, formatFindings, formatScorecardDiff, ghCliClient, gitProvenanceReader, precision as goldenPrecision, hashContent, hashToUnit, htmlContainsElement, httpGithubClient, improvementVerdict, inMemoryExperimentStore, inMemoryReferenceReplayStore, inMemoryRunRecordBackend, inMemoryVerdictCache, isModelPriced, isOtelConfigured, jsonShape, jsonlReferenceReplayStore, jsonlRunRecordBackend, judgeFamily, keyPreserved, linterJudge, loadGoldScenarios, loadScorecard, loadScorerFromGrader, localCommandRunner, lowercaseMutator, makeEvalTools, matchGoldens, matchSpan, mergeLayerResults, modelDescriptionBits, multiToolchainLayer, noProgressDetector, notBlocked, observeAll, paraphraseRobustness, paraphraseRobustnessScenarios, parseGoldJsonl, parseReflectionResponse, partitionHeldOut, passOrthogonality, pixelDeltaRatio, politenessPrefixMutator, preflightModels, printDriverSummary, index as profile, promptBisect, proposeAutomatedPullRequest, proposeSynthesisTargets, recordRuns, recordRunsToScorecard, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatches, renderMarkdownReport, renderPlaybookMarkdown, repeatedActionDetector, replayScorerOverCorpus, replayTraceThroughJudge, resetLockedAppendersForTesting, resolveModelPricing, resolveSeat, rowCount, rowWhere, runAssertions, runBehavioralCanaries, runCanaries, runDistillation, runE2EWorkflow, runExpectations, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runReferenceReplay, runScore, runSelfPlay, scanForMuffledGates, scoreContinuity, scorePrReviewComments, scorePrReviewSource, scoreReferenceReplay, seatPresets, securityJudge, sentenceReorderMutator, splitGold, statusAdvanced, summarizePrReviewBenchmark, testJudge, textInSnapshot, toLangfuseEnvelope, toOpenAiTool, toPrometheusText, traceContract, traceJudge, traceJudgeEnsemble, tracedAnalyzeTraces, typoMutator, urlContains, verifyAttestation, visualDiff, viteDeployRunner, weightedRecall, whitespaceCollapseMutator, withJudgeRetry, withOtelPipeline, wranglerDeployRunner };