@tangle-network/agent-eval 0.92.0 → 0.94.0

package/dist/{chunk-CY6U5S3X.js → chunk-URWVKRCS.js}

@@ -5,7 +5,7 @@ import {
   pairedBootstrap,
   pairedMde,
   wilcoxonSignedRank
-} from "./chunk-LMZQ2Z4U.js";
+} from "./chunk-56GC6VYO.js";
 import {
   canonicalize,
   hashJson
@@ -877,4 +877,4 @@ export {
   RESEARCH_REPORT_HARD_PAIR_FLOOR,
   researchReport
 };
-//# sourceMappingURL=chunk-CY6U5S3X.js.map
+//# sourceMappingURL=chunk-URWVKRCS.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/{chunk-T375SUOZ.js → chunk-ZOPLCJSY.js}

@@ -67,48 +67,101 @@ var SubprocessSandboxDriver = class {
     const start = Date.now();
     const effectiveCwd = config.cwd ?? this.defaultCwd;
     const effectiveEnv = { ...process.env, ...this.defaultEnv ?? {}, ...config.env ?? {} };
+    const maxOutputBytes = config.maxOutputBytes ?? 16 * 1024 * 1024;
     return await new Promise((resolve) => {
       const child = spawn(command, {
         shell: true,
         cwd: effectiveCwd,
-        env: effectiveEnv
+        env: effectiveEnv,
+        // Own process group so a timeout can SIGKILL the whole tree, not
+        // just the shell — otherwise a runaway grandchild keeps the pipes
+        // open and `close` never fires (the timeout silently does nothing).
+        detached: process.platform !== "win32"
       });
       let stdout = "";
       let stderr = "";
-      child.stdout?.on("data", (d) => {
-        stdout += String(d);
-      });
-      child.stderr?.on("data", (d) => {
-        stderr += String(d);
-      });
-      const timeout = setTimeout(
-        () => {
-          try {
-            child.kill("SIGKILL");
-          } catch (err) {
-            console.warn("[sandbox-harness] SIGKILL on timeout failed:", err);
-          }
-        },
-        config.timeoutMs ?? 10 * 6e4
-      );
-      child.on("close", (code) => {
+      let outputBytes = 0;
+      let outputTruncated = false;
+      let killedByTimeout = false;
+      let settled = false;
+      const onStdout = (d) => {
+        const chunk = capture(String(d));
+        if (chunk) stdout += chunk;
+      };
+      const onStderr = (d) => {
+        const chunk = capture(String(d));
+        if (chunk) stderr += chunk;
+      };
+      function capture(s) {
+        if (outputBytes >= maxOutputBytes) {
+          outputTruncated = true;
+          return "";
+        }
+        const room = maxOutputBytes - outputBytes;
+        if (s.length > room) {
+          outputBytes = maxOutputBytes;
+          outputTruncated = true;
+          return s.slice(0, room);
+        }
+        outputBytes += s.length;
+        return s;
+      }
+      child.stdout?.on("data", onStdout);
+      child.stderr?.on("data", onStderr);
+      const cleanup = () => {
         clearTimeout(timeout);
+        if (forceResolve) clearTimeout(forceResolve);
+        child.stdout?.off("data", onStdout);
+        child.stderr?.off("data", onStderr);
+        child.removeAllListeners("close");
+        child.removeAllListeners("error");
+      };
+      const finish = (outcome) => {
+        if (settled) return;
+        settled = true;
+        cleanup();
         const wallMs = Date.now() - start;
-        const parsed = phase === "test" && config.testParser ? config.testParser.parse(stdout, stderr, code ?? 1) : void 0;
+        const exitCode = killedByTimeout ? outcome.code && outcome.code !== 0 ? outcome.code : 124 : outcome.code ?? 1;
+        const parsed = !killedByTimeout && phase === "test" && config.testParser ? config.testParser.parse(stdout, stderr, exitCode) : void 0;
         resolve({
           phase,
-          exitCode: code ?? 1,
+          exitCode,
           stdout,
-          stderr,
+          stderr: outcome.runnerError ? stderr + outcome.runnerError : stderr,
           wallMs,
           testsTotal: parsed?.testsTotal,
-          testsPassed: parsed?.testsPassed
+          testsPassed: parsed?.testsPassed,
+          killedByTimeout: killedByTimeout || void 0,
+          outputTruncated: outputTruncated || void 0
         });
+      };
+      const killTree = () => {
+        try {
+          if (process.platform !== "win32" && typeof child.pid === "number") {
+            process.kill(-child.pid, "SIGKILL");
+          } else {
+            child.kill("SIGKILL");
+          }
+        } catch (err) {
+          console.warn("[sandbox-harness] SIGKILL on timeout failed:", err);
+        }
+      };
+      let forceResolve;
+      const timeout = setTimeout(
+        () => {
+          killedByTimeout = true;
+          killTree();
+          forceResolve = setTimeout(() => finish({ code: null }), 2e3);
+        },
+        config.timeoutMs ?? 10 * 6e4
+      );
+      child.on("close", (code) => {
+        if (forceResolve) clearTimeout(forceResolve);
+        finish({ code });
       });
       child.on("error", (err) => {
-        clearTimeout(timeout);
-        const wallMs = Date.now() - start;
-        resolve({ phase, exitCode: 127, stdout, stderr: stderr + String(err), wallMs });
+        if (forceResolve) clearTimeout(forceResolve);
+        finish({ code: 127, runnerError: String(err) });
       });
     });
   }
@@ -143,8 +196,9 @@ var SandboxHarness = class {
       if (config.setupCommand) {
         result.setup = await this.driver.exec("setup", config.setupCommand, config);
         result.totalWallMs += result.setup.wallMs;
-        if (result.setup.exitCode !== 0) {
-          await handle.fail(`setup failed (exit ${result.setup.exitCode})`, {
+        if (result.setup.killedByTimeout || result.setup.exitCode !== 0) {
+          const reason = result.setup.killedByTimeout ? `setup timed out after ${result.setup.wallMs}ms` : `setup failed (exit ${result.setup.exitCode})`;
+          await handle.fail(reason, {
             exitCode: result.setup.exitCode,
             wallMs: result.totalWallMs
           });
@@ -154,8 +208,9 @@ var SandboxHarness = class {
       if (config.runCommand) {
         result.run = await this.driver.exec("run", config.runCommand, config);
         result.totalWallMs += result.run.wallMs;
-        if (result.run.exitCode !== 0) {
-          await handle.fail(`run failed (exit ${result.run.exitCode})`, {
+        if (result.run.killedByTimeout || result.run.exitCode !== 0) {
+          const reason = result.run.killedByTimeout ? `run timed out after ${result.run.wallMs}ms` : `run failed (exit ${result.run.exitCode})`;
+          await handle.fail(reason, {
             exitCode: result.run.exitCode,
             wallMs: result.totalWallMs
           });
@@ -165,7 +220,7 @@ var SandboxHarness = class {
       if (config.testCommand) {
         result.test = await this.driver.exec("test", config.testCommand, config);
         result.totalWallMs += result.test.wallMs;
-        const passed = result.test.exitCode === 0;
+        const passed = !result.test.killedByTimeout && result.test.exitCode === 0;
         result.passed = passed;
         if (result.test.testsTotal !== void 0 && result.test.testsTotal > 0) {
           result.score = (result.test.testsPassed ?? 0) / result.test.testsTotal;
@@ -248,4 +303,4 @@ export {
   SandboxHarness,
   runTestGradedScenario
 };
-//# sourceMappingURL=chunk-T375SUOZ.js.map
+//# sourceMappingURL=chunk-ZOPLCJSY.js.map

package/dist/chunk-ZOPLCJSY.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/sandbox-harness.ts","../src/test-graded-scenario.ts"],"sourcesContent":["/**\n * SandboxHarness — executes a scenario in an isolated environment and\n * emits a rich SandboxSpan into the trace.\n *\n * Two built-in drivers:\n *   - `SubprocessSandboxDriver` — spawn in a local cwd with env vars.\n *     Fast, no dependencies, fine for unit tests and most CI gates.\n *   - `DockerSandboxDriver` — lifted from tangle-router's sandbox path;\n *     shells out to `docker run`. Stronger isolation, slower startup.\n *\n * Consumers implement `SandboxDriver` for custom backends (Firecracker,\n * Cloudflare sandbox product, etc.). The harness doesn't care which.\n */\n\nimport { ConfigError } from './errors'\nimport type { TraceEmitter } from './trace/emitter'\nimport type { SandboxSpan } from './trace/schema'\n\nexport interface HarnessConfig {\n  /** Setup command (e.g. \"pnpm install\"). Non-zero exit fails the run. */\n  setupCommand?: string\n  /** Run command (e.g. \"pnpm build\"). */\n  runCommand?: string\n  /** Test command (e.g. \"pnpm test --run\"). Drives the test count + pass count. */\n  testCommand?: string\n  /** Absolute cwd for the subprocess driver. Ignored by docker driver. */\n  cwd?: string\n  /** Max wall-clock per phase in ms. Default 10 minutes. */\n  timeoutMs?: number\n  /**\n   * Cap on captured stdout+stderr bytes per phase. A runaway process can\n   * otherwise grow the in-memory buffer without bound. Once hit, further\n   * output is dropped and `outputTruncated` is set. Default 16 MiB.\n   */\n  maxOutputBytes?: number\n  /** Image for the docker driver. */\n  image?: string\n  /** Extra env vars (validated; shell-escaped). */\n  env?: Record<string, string>\n  /** Parser for the test output — maps stdout/stderr/exit code → pass count. */\n  testParser?: TestOutputParser\n}\n\nexport interface TestOutputParser {\n  id: string\n  parse(\n    stdout: string,\n    stderr: string,\n    exitCode: number,\n  ): { testsTotal: number; testsPassed: number } | undefined\n}\n\nexport interface SandboxResult {\n  phase: 'setup' | 'run' | 'test'\n  exitCode: number\n  stdout: string\n  stderr: string\n  wallMs: number\n  testsTotal?: number\n  testsPassed?: number\n  /**\n   * True when the process was killed because it exceeded `timeoutMs`. A\n   * SIGKILLed child can still close with exit code 0; callers MUST treat\n   * a timed-out phase as a hard failure regardless of `exitCode`, never\n   * as a pass. `undefined`/`false` means the process completed on its own.\n   */\n  killedByTimeout?: boolean\n  /**\n   * True when captured stdout/stderr hit `maxOutputBytes` and further\n   * output was dropped. The result is still returned (the process was\n   * not killed for this), but downstream parsers see truncated text.\n   */\n  outputTruncated?: boolean\n}\n\nexport interface SandboxDriver {\n  id: string\n  exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult>\n}\n\n// ── Parsers ──────────────────────────────────────────────────────────\n\n/** Vitest default summary line: \"Tests  X passed | Y failed\". */\nexport const vitestTestParser: TestOutputParser = {\n  id: 'vitest',\n  parse(stdout) {\n    const m = stdout.match(/Tests\\s+(\\d+)\\s+(passed|failed)(?:\\s*\\|\\s*(\\d+)\\s+(passed|failed))?/i)\n    if (!m) return undefined\n    let passed = 0\n    let failed = 0\n    const a = parseInt(m[1]!, 10)\n    const aLabel = m[2]!.toLowerCase()\n    if (aLabel === 'passed') passed += a\n    else failed += a\n    if (m[3] && m[4]) {\n      const b = parseInt(m[3], 10)\n      if (m[4].toLowerCase() === 'passed') passed += b\n      else failed += b\n    }\n    return { testsTotal: passed + failed, testsPassed: passed }\n  },\n}\n\n/** Pytest default: \"collected N items\" + \" X passed, Y failed\". */\nexport const pytestTestParser: TestOutputParser = {\n  id: 'pytest',\n  parse(stdout) {\n    const total = stdout.match(/collected\\s+(\\d+)\\s+items?/i)\n    const passed = stdout.match(/(\\d+)\\s+passed/)\n    if (!total || !passed) return undefined\n    return { testsTotal: parseInt(total[1]!, 10), testsPassed: parseInt(passed[1]!, 10) }\n  },\n}\n\n/** Jest: \"Tests: X passed, Y total\" (and optional failed). */\nexport const jestTestParser: TestOutputParser = {\n  id: 'jest',\n  parse(stdout) {\n    const m = stdout.match(/Tests:\\s+(?:(\\d+)\\s+failed[^,]*,\\s*)?(\\d+)\\s+passed,\\s+(\\d+)\\s+total/i)\n    if (!m) return undefined\n    return { testsTotal: parseInt(m[3]!, 10), testsPassed: parseInt(m[2]!, 10) }\n  },\n}\n\n/** Composite parser — tries a list of parsers in order. */\nexport function composeParsers(...parsers: TestOutputParser[]): TestOutputParser {\n  return {\n    id: parsers.map((p) => p.id).join('|'),\n    parse(stdout, stderr, exitCode) {\n      for (const p of parsers) {\n        const res = p.parse(stdout, stderr, exitCode)\n        if (res) return res\n      }\n      return undefined\n    },\n  }\n}\n\n// ── Drivers ──────────────────────────────────────────────────────────\n\nexport interface SubprocessSandboxDriverOptions {\n  /**\n   * Default cwd for all `exec` calls. Used when the per-call `HarnessConfig`\n   * does not set its own `cwd`. Lets callers bind the driver to a working\n   * directory once instead of spreading cwd into every harness config —\n   * useful when the harness config is constructed far from the call site\n   * (e.g. starter-foundry's promoter passes a static HarnessConfig per\n   * family taxonomy but needs a per-run composed-scaffold cwd).\n   */\n  cwd?: string\n  /**\n   * Default env merged into every `exec` call's env (per-call `HarnessConfig.env`\n   * still wins on key collision). Same ergonomic rationale as `cwd` above.\n   */\n  env?: Record<string, string>\n}\n\nexport class SubprocessSandboxDriver implements SandboxDriver {\n  id = 'subprocess'\n  private defaultCwd?: string\n  private defaultEnv?: Record<string, string>\n\n  constructor(options: SubprocessSandboxDriverOptions = {}) {\n    this.defaultCwd = options.cwd\n    this.defaultEnv = options.env\n  }\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    const { spawn } = await import('node:child_process')\n    const start = Date.now()\n    // Per-call config wins; fall back to constructor defaults. Honoring\n    // the constructor `cwd` keeps the subprocess from inheriting Node's\n    // cwd when only the constructor arg is supplied.\n    const effectiveCwd = config.cwd ?? this.defaultCwd\n    const effectiveEnv = { ...process.env, ...(this.defaultEnv ?? {}), ...(config.env ?? {}) }\n    const maxOutputBytes = config.maxOutputBytes ?? 16 * 1024 * 1024\n    return await new Promise<SandboxResult>((resolve) => {\n      const child = spawn(command, {\n        shell: true,\n        cwd: effectiveCwd,\n        env: effectiveEnv,\n        // Own process group so a timeout can SIGKILL the whole tree, not\n        // just the shell — otherwise a runaway grandchild keeps the pipes\n        // open and `close` never fires (the timeout silently does nothing).\n        detached: process.platform !== 'win32',\n      })\n      let stdout = ''\n      let stderr = ''\n      let outputBytes = 0\n      let outputTruncated = false\n      let killedByTimeout = false\n      let settled = false\n\n      const onStdout = (d: unknown) => {\n        const chunk = capture(String(d))\n        if (chunk) stdout += chunk\n      }\n      const onStderr = (d: unknown) => {\n        const chunk = capture(String(d))\n        if (chunk) stderr += chunk\n      }\n      // Bound the in-memory buffer so a runaway process can't OOM the\n      // harness. Once the cap is hit we keep draining the streams (to let\n      // the child reach `close`) but discard the bytes.\n      function capture(s: string): string {\n        if (outputBytes >= maxOutputBytes) {\n          outputTruncated = true\n          return ''\n        }\n        const room = maxOutputBytes - outputBytes\n        if (s.length > room) {\n          outputBytes = maxOutputBytes\n          outputTruncated = true\n          return s.slice(0, room)\n        }\n        outputBytes += s.length\n        return s\n      }\n\n      child.stdout?.on('data', onStdout)\n      child.stderr?.on('data', onStderr)\n\n      const cleanup = () => {\n        clearTimeout(timeout)\n        if (forceResolve) clearTimeout(forceResolve)\n        child.stdout?.off('data', onStdout)\n        child.stderr?.off('data', onStderr)\n        child.removeAllListeners('close')\n        child.removeAllListeners('error')\n      }\n\n      // Resolve exactly once. `code`/`signal` come from `close`; on the\n      // timeout path we force a non-zero exit so a SIGKILLed child that\n      // reports exit 0 can never read as a clean pass downstream.\n      const finish = (outcome: { code: number | null; runnerError?: string }) => {\n        if (settled) return\n        settled = true\n        cleanup()\n        const wallMs = Date.now() - start\n        const exitCode = killedByTimeout\n          ? outcome.code && outcome.code !== 0\n            ? outcome.code\n            : 124\n          : (outcome.code ?? 1)\n        const parsed =\n          !killedByTimeout && phase === 'test' && config.testParser\n            ? config.testParser.parse(stdout, stderr, exitCode)\n            : undefined\n        resolve({\n          phase,\n          exitCode,\n          stdout,\n          stderr: outcome.runnerError ? stderr + outcome.runnerError : stderr,\n          wallMs,\n          testsTotal: parsed?.testsTotal,\n          testsPassed: parsed?.testsPassed,\n          killedByTimeout: killedByTimeout || undefined,\n          outputTruncated: outputTruncated || undefined,\n        })\n      }\n\n      // Kill the whole process group on win32-less platforms (we spawned\n      // detached). On a clean close this is a no-op.\n      const killTree = () => {\n        try {\n          if (process.platform !== 'win32' && typeof child.pid === 'number') {\n            process.kill(-child.pid, 'SIGKILL')\n          } else {\n            child.kill('SIGKILL')\n          }\n        } catch (err) {\n          console.warn('[sandbox-harness] SIGKILL on timeout failed:', err)\n        }\n      }\n\n      let forceResolve: ReturnType<typeof setTimeout> | undefined\n      const timeout = setTimeout(\n        () => {\n          killedByTimeout = true\n          killTree()\n          // Give `close` a brief window to fire (flushing any final output)\n          // after the group dies; if an orphaned grandchild keeps the pipes\n          // open, force-resolve so the harness can't hang on a runaway.\n          forceResolve = setTimeout(() => finish({ code: null }), 2000)\n        },\n        config.timeoutMs ?? 10 * 60_000,\n      )\n\n      child.on('close', (code) => {\n        if (forceResolve) clearTimeout(forceResolve)\n        finish({ code })\n      })\n      child.on('error', (err) => {\n        if (forceResolve) clearTimeout(forceResolve)\n        finish({ code: 127, runnerError: String(err) })\n      })\n    })\n  }\n}\n\nexport class DockerSandboxDriver implements SandboxDriver {\n  id = 'docker'\n\n  async exec(\n    phase: SandboxResult['phase'],\n    command: string,\n    config: HarnessConfig,\n  ): Promise<SandboxResult> {\n    if (!config.image) throw new ConfigError('DockerSandboxDriver requires config.image')\n    const sub = new SubprocessSandboxDriver()\n    const envArgs = Object.entries(config.env ?? {})\n      .map(([k, v]) => `-e ${shellQuote(k)}=${shellQuote(v)}`)\n      .join(' ')\n    const wrapped = `docker run --rm ${envArgs} ${shellQuote(config.image)} sh -c ${shellQuote(command)}`\n    return sub.exec(phase, wrapped, { ...config, env: undefined })\n  }\n}\n\nfunction shellQuote(v: string): string {\n  if (/^[A-Za-z0-9_\\-/.@:=]+$/.test(v)) return v\n  return `'${v.replace(/'/g, `'\\\\''`)}'`\n}\n\n// ── Harness orchestration ────────────────────────────────────────────\n\nexport interface SandboxHarnessResult {\n  passed: boolean\n  setup?: SandboxResult\n  run?: SandboxResult\n  test?: SandboxResult\n  totalWallMs: number\n  /** Final score — 0 when no tests; otherwise testsPassed/testsTotal. */\n  score: number\n}\n\nexport class SandboxHarness {\n  private driver: SandboxDriver\n  constructor(driver: SandboxDriver = new SubprocessSandboxDriver()) {\n    this.driver = driver\n  }\n\n  async run(config: HarnessConfig, emitter: TraceEmitter): Promise<SandboxHarnessResult> {\n    const handle = await emitter.sandbox({\n      name: `sandbox(${this.driver.id})`,\n      image: config.image,\n      command: [config.setupCommand, config.runCommand, config.testCommand]\n        .filter(Boolean)\n        .join(' && '),\n    })\n    const result: SandboxHarnessResult = { passed: false, totalWallMs: 0, score: 0 }\n    try {\n      if (config.setupCommand) {\n        result.setup = await this.driver.exec('setup', config.setupCommand, config)\n        result.totalWallMs += result.setup.wallMs\n        if (result.setup.killedByTimeout || result.setup.exitCode !== 0) {\n          const reason = result.setup.killedByTimeout\n            ? `setup timed out after ${result.setup.wallMs}ms`\n            : `setup failed (exit ${result.setup.exitCode})`\n          await handle.fail(reason, {\n            exitCode: result.setup.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.runCommand) {\n        result.run = await this.driver.exec('run', config.runCommand, config)\n        result.totalWallMs += result.run.wallMs\n        if (result.run.killedByTimeout || result.run.exitCode !== 0) {\n          const reason = result.run.killedByTimeout\n            ? `run timed out after ${result.run.wallMs}ms`\n            : `run failed (exit ${result.run.exitCode})`\n          await handle.fail(reason, {\n            exitCode: result.run.exitCode,\n            wallMs: result.totalWallMs,\n          } as Partial<SandboxSpan>)\n          return result\n        }\n      }\n      if (config.testCommand) {\n        result.test = await this.driver.exec('test', config.testCommand, config)\n        result.totalWallMs += result.test.wallMs\n        // A timed-out test phase is a hard fail regardless of the exit code\n        // a SIGKILLed child happens to report — never let it score as a pass.\n        const passed = !result.test.killedByTimeout && result.test.exitCode === 0\n        result.passed = passed\n        if (result.test.testsTotal !== undefined && result.test.testsTotal > 0) {\n          result.score = (result.test.testsPassed ?? 0) / result.test.testsTotal\n        } else {\n          result.score = passed ? 1 : 0\n        }\n        await handle.end({\n          exitCode: result.test.exitCode,\n          testsTotal: result.test.testsTotal,\n          testsPassed: result.test.testsPassed,\n          wallMs: result.totalWallMs,\n          status: passed ? 'ok' : 'error',\n        } as Partial<SandboxSpan>)\n      } else {\n        result.passed = true\n        result.score = 1\n        await handle.end({ wallMs: result.totalWallMs } as Partial<SandboxSpan>)\n      }\n    } catch (err) {\n      await handle.fail(err instanceof Error ? err : String(err))\n      throw err\n    }\n    return result\n  }\n}\n","/**\n * TestGradedScenario — a scenario whose score comes from a test suite.\n *\n * This is the SWE-bench pattern generalized. The scenario ships:\n *   - fixture data (setup instructions)\n *   - a test command the harness runs\n *   - optional assertion overrides\n *\n * The runner emits a run, delegates to SandboxHarness, records the\n * outcome, and returns a structured verdict. Consumers bind their own\n * agent execution to this contract.\n */\n\nimport type { HarnessConfig, SandboxDriver, SandboxHarnessResult } from './sandbox-harness'\nimport { SandboxHarness } from './sandbox-harness'\nimport { TraceEmitter } from './trace/emitter'\nimport type { FailureClass, Run } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface TestGradedScenario {\n  id: string\n  description?: string\n  harness: HarnessConfig\n  /** Optional pass threshold in 0..1 (default 1.0 = all tests must pass). */\n  passThreshold?: number\n  /** Provenance for dataset tracking. */\n  datasetVersion?: string\n  /** Free-form tags (difficulty, category, etc.). */\n  tags?: Record<string, string>\n}\n\nexport interface TestGradedRunOptions {\n  variantId?: string\n  driver?: SandboxDriver\n  /** Metadata recorded on the Run (codeSha, promptSha, modelFingerprint, seed). */\n  provenance?: Pick<Run, 'codeSha' | 'promptSha' | 'modelFingerprint' | 'seed' | 'envFingerprint'>\n}\n\nexport interface TestGradedRunResult {\n  runId: string\n  scenario: TestGradedScenario\n  harness: SandboxHarnessResult\n  pass: boolean\n  score: number\n  failureClass?: FailureClass\n}\n\nexport async function runTestGradedScenario(\n  scenario: TestGradedScenario,\n  store: TraceStore,\n  options: TestGradedRunOptions = {},\n): Promise<TestGradedRunResult> {\n  const emitter = new TraceEmitter(store)\n  await emitter.startRun({\n    scenarioId: scenario.id,\n    variantId: options.variantId,\n    datasetVersion: scenario.datasetVersion,\n    tags: scenario.tags,\n    ...options.provenance,\n  })\n  const harness = new SandboxHarness(options.driver)\n  const result = await harness.run(scenario.harness, emitter)\n  const threshold = scenario.passThreshold ?? 1.0\n  const pass = result.passed && result.score >= threshold\n  const setupFailed = result.setup !== undefined && result.setup.exitCode !== 0\n  const runFailed = result.run !== undefined && result.run.exitCode !== 0\n  const testFailed = result.test !== undefined && result.test.exitCode !== 0\n  const failureClass: FailureClass | undefined = pass\n    ? 'success'\n    : setupFailed || runFailed\n      ? 'sandbox_failure'\n      : testFailed\n        ? 'format_drift'\n        : 'unknown'\n  await emitter.endRun({\n    pass,\n    score: result.score,\n    failureClass,\n    notes: pass ? undefined : reasonForFailure(result),\n  })\n  return {\n    runId: emitter.runId,\n    scenario,\n    harness: result,\n    pass,\n    score: result.score,\n    failureClass,\n  }\n}\n\nfunction reasonForFailure(result: SandboxHarnessResult): string {\n  if (result.setup && result.setup.exitCode !== 0)\n    return `setup failed: exit ${result.setup.exitCode}`\n  if (result.run && result.run.exitCode !== 0) return `run failed: exit ${result.run.exitCode}`\n  if (result.test) {\n    if (result.test.testsTotal !== undefined) {\n      return `tests: ${result.test.testsPassed ?? 0}/${result.test.testsTotal}`\n    }\n    return `test exit ${result.test.exitCode}`\n  }\n  return 'no test command'\n}\n"],"mappings":";;;;;;;;AAuFO,IAAM,mBAAqC;AAAA,EAChD,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,IAAI,OAAO,MAAM,sEAAsE;AAC7F,QAAI,CAAC,EAAG,QAAO;AACf,QAAI,SAAS;AACb,QAAI,SAAS;AACb,UAAM,IAAI,SAAS,EAAE,CAAC,GAAI,EAAE;AAC5B,UAAM,SAAS,EAAE,CAAC,EAAG,YAAY;AACjC,QAAI,WAAW,SAAU,WAAU;AAAA,QAC9B,WAAU;AACf,QAAI,EAAE,CAAC,KAAK,EAAE,CAAC,GAAG;AAChB,YAAM,IAAI,SAAS,EAAE,CAAC,GAAG,EAAE;AAC3B,UAAI,EAAE,CAAC,EAAE,YAAY,MAAM,SAAU,WAAU;AAAA,UAC1C,WAAU;AAAA,IACjB;AACA,WAAO,EAAE,YAAY,SAAS,QAAQ,aAAa,OAAO;AAAA,EAC5D;AACF;AAGO,IAAM,mBAAqC;AAAA,EAChD,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,QAAQ,OAAO,MAAM,6BAA6B;AACxD,UAAM,SAAS,OAAO,MAAM,gBAAgB;AAC5C,QAAI,CAAC,SAAS,CAAC,OAAQ,QAAO;AAC9B,WAAO,EAAE,YAAY,SAAS,MAAM,CAAC,GAAI,EAAE,GAAG,aAAa,SAAS,OAAO,CAAC,GAAI,EAAE,EAAE;AAAA,EACtF;AACF;AAGO,IAAM,iBAAmC;AAAA,EAC9C,IAAI;AAAA,EACJ,MAAM,QAAQ;AACZ,UAAM,IAAI,OAAO,MAAM,uEAAuE;AAC9F,QAAI,CAAC,EAAG,QAAO;AACf,WAAO,EAAE,YAAY,SAAS,EAAE,CAAC,GAAI,EAAE,GAAG,aAAa,SAAS,EAAE,CAAC,GAAI,EAAE,EAAE;AAAA,EAC7E;AACF;AAGO,SAAS,kBAAkB,SAA+C;AAC/E,SAAO;AAAA,IACL,IAAI,QAAQ,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE,KAAK,GAAG;AAAA,IACrC,MAAM,QAAQ,QAAQ,UAAU;AAC9B,iBAAW,KAAK,SAAS;AACvB,cAAM,MAAM,EAAE,MAAM,QAAQ,QAAQ,QAAQ;AAC5C,YAAI,IAAK,QAAO;AAAA,MAClB;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAqBO,IAAM,0BAAN,MAAuD;AAAA,EAC5D,KAAK;AAAA,EACG;AAAA,EACA;AAAA,EAER,YAAY,UAA0C,CAAC,GAAG;AACxD,SAAK,aAAa,QAAQ;AAC1B,SAAK,aAAa,QAAQ;AAAA,EAC5B;AAAA,EAEA,MAAM,KACJ,OACA,SACA,QACwB;AACxB,UAAM,EAAE,MAAM,IAAI,MAAM,OAAO,eAAoB;AACnD,UAAM,QAAQ,KAAK,IAAI;AAIvB,UAAM,eAAe,OAAO,OAAO,KAAK;AACxC,UAAM,eAAe,EAAE,GAAG,QAAQ,KAAK,GAAI,KAAK,cAAc,CAAC,GAAI,GAAI,OAAO,OAAO,CAAC,EAAG;AACzF,UAAM,iBAAiB,OAAO,kBAAkB,KAAK,OAAO;AAC5D,WAAO,MAAM,IAAI,QAAuB,CAAC,YAAY;AACnD,YAAM,QAAQ,MAAM,SAAS;AAAA,QAC3B,OAAO;AAAA,QACP,KAAK;AAAA,QACL,KAAK;AAAA;AAAA;AAAA;AAAA,QAIL,UAAU,QAAQ,aAAa;AAAA,MACjC,CAAC;AACD,UAAI,SAAS;AACb,UAAI,SAAS;AACb,UAAI,cAAc;AAClB,UAAI,kBAAkB;AACtB,UAAI,kBAAkB;AACtB,UAAI,UAAU;AAEd,YAAM,WAAW,CAAC,MAAe;AAC/B,cAAM,QAAQ,QAAQ,OAAO,CAAC,CAAC;AAC/B,YAAI,MAAO,WAAU;AAAA,MACvB;AACA,YAAM,WAAW,CAAC,MAAe;AAC/B,cAAM,QAAQ,QAAQ,OAAO,CAAC,CAAC;AAC/B,YAAI,MAAO,WAAU;AAAA,MACvB;AAIA,eAAS,QAAQ,GAAmB;AAClC,YAAI,eAAe,gBAAgB;AACjC,4BAAkB;AAClB,iBAAO;AAAA,QACT;AACA,cAAM,OAAO,iBAAiB;AAC9B,YAAI,EAAE,SAAS,MAAM;AACnB,wBAAc;AACd,4BAAkB;AAClB,iBAAO,EAAE,MAAM,GAAG,IAAI;AAAA,QACxB;AACA,uBAAe,EAAE;AACjB,eAAO;AAAA,MACT;AAEA,YAAM,QAAQ,GAAG,QAAQ,QAAQ;AACjC,YAAM,QAAQ,GAAG,QAAQ,QAAQ;AAEjC,YAAM,UAAU,MAAM;AACpB,qBAAa,OAAO;AACpB,YAAI,aAAc,cAAa,YAAY;AAC3C,cAAM,QAAQ,IAAI,QAAQ,QAAQ;AAClC,cAAM,QAAQ,IAAI,QAAQ,QAAQ;AAClC,cAAM,mBAAmB,OAAO;AAChC,cAAM,mBAAmB,OAAO;AAAA,MAClC;AAKA,YAAM,SAAS,CAAC,YAA2D;AACzE,YAAI,QAAS;AACb,kBAAU;AACV,gBAAQ;AACR,cAAM,SAAS,KAAK,IAAI,IAAI;AAC5B,cAAM,WAAW,kBACb,QAAQ,QAAQ,QAAQ,SAAS,IAC/B,QAAQ,OACR,MACD,QAAQ,QAAQ;AACrB,cAAM,SACJ,CAAC,mBAAmB,UAAU,UAAU,OAAO,aAC3C,OAAO,WAAW,MAAM,QAAQ,QAAQ,QAAQ,IAChD;AACN,gBAAQ;AAAA,UACN;AAAA,UACA;AAAA,UACA;AAAA,UACA,QAAQ,QAAQ,cAAc,SAAS,QAAQ,cAAc;AAAA,UAC7D;AAAA,UACA,YAAY,QAAQ;AAAA,UACpB,aAAa,QAAQ;AAAA,UACrB,iBAAiB,mBAAmB;AAAA,UACpC,iBAAiB,mBAAmB;AAAA,QACtC,CAAC;AAAA,MACH;AAIA,YAAM,WAAW,MAAM;AACrB,YAAI;AACF,cAAI,QAAQ,aAAa,WAAW,OAAO,MAAM,QAAQ,UAAU;AACjE,oBAAQ,KAAK,CAAC,MAAM,KAAK,SAAS;AAAA,UACpC,OAAO;AACL,kBAAM,KAAK,SAAS;AAAA,UACtB;AAAA,QACF,SAAS,KAAK;AACZ,kBAAQ,KAAK,gDAAgD,GAAG;AAAA,QAClE;AAAA,MACF;AAEA,UAAI;AACJ,YAAM,UAAU;AAAA,QACd,MAAM;AACJ,4BAAkB;AAClB,mBAAS;AAIT,yBAAe,WAAW,MAAM,OAAO,EAAE,MAAM,KAAK,CAAC,GAAG,GAAI;AAAA,QAC9D;AAAA,QACA,OAAO,aAAa,KAAK;AAAA,MAC3B;AAEA,YAAM,GAAG,SAAS,CAAC,SAAS;AAC1B,YAAI,aAAc,cAAa,YAAY;AAC3C,eAAO,EAAE,KAAK,CAAC;AAAA,MACjB,CAAC;AACD,YAAM,GAAG,SAAS,CAAC,QAAQ;AACzB,YAAI,aAAc,cAAa,YAAY;AAC3C,eAAO,EAAE,MAAM,KAAK,aAAa,OAAO,GAAG,EAAE,CAAC;AAAA,MAChD,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AACF;AAEO,IAAM,sBAAN,MAAmD;AAAA,EACxD,KAAK;AAAA,EAEL,MAAM,KACJ,OACA,SACA,QACwB;AACxB,QAAI,CAAC,OAAO,MAAO,OAAM,IAAI,YAAY,2CAA2C;AACpF,UAAM,MAAM,IAAI,wBAAwB;AACxC,UAAM,UAAU,OAAO,QAAQ,OAAO,OAAO,CAAC,CAAC,EAC5C,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,MAAM,WAAW,CAAC,CAAC,IAAI,WAAW,CAAC,CAAC,EAAE,EACtD,KAAK,GAAG;AACX,UAAM,UAAU,mBAAmB,OAAO,IAAI,WAAW,OAAO,KAAK,CAAC,UAAU,WAAW,OAAO,CAAC;AACnG,WAAO,IAAI,KAAK,OAAO,SAAS,EAAE,GAAG,QAAQ,KAAK,OAAU,CAAC;AAAA,EAC/D;AACF;AAEA,SAAS,WAAW,GAAmB;AACrC,MAAI,yBAAyB,KAAK,CAAC,EAAG,QAAO;AAC7C,SAAO,IAAI,EAAE,QAAQ,MAAM,OAAO,CAAC;AACrC;AAcO,IAAM,iBAAN,MAAqB;AAAA,EAClB;AAAA,EACR,YAAY,SAAwB,IAAI,wBAAwB,GAAG;AACjE,SAAK,SAAS;AAAA,EAChB;AAAA,EAEA,MAAM,IAAI,QAAuB,SAAsD;AACrF,UAAM,SAAS,MAAM,QAAQ,QAAQ;AAAA,MACnC,MAAM,WAAW,KAAK,OAAO,EAAE;AAAA,MAC/B,OAAO,OAAO;AAAA,MACd,SAAS,CAAC,OAAO,cAAc,OAAO,YAAY,OAAO,WAAW,EACjE,OAAO,OAAO,EACd,KAAK,MAAM;AAAA,IAChB,CAAC;AACD,UAAM,SAA+B,EAAE,QAAQ,OAAO,aAAa,GAAG,OAAO,EAAE;AAC/E,QAAI;AACF,UAAI,OAAO,cAAc;AACvB,eAAO,QAAQ,MAAM,KAAK,OAAO,KAAK,SAAS,OAAO,cAAc,MAAM;AAC1E,eAAO,eAAe,OAAO,MAAM;AACnC,YAAI,OAAO,MAAM,mBAAmB,OAAO,MAAM,aAAa,GAAG;AAC/D,gBAAM,SAAS,OAAO,MAAM,kBACxB,yBAAyB,OAAO,MAAM,MAAM,OAC5C,sBAAsB,OAAO,MAAM,QAAQ;AAC/C,gBAAM,OAAO,KAAK,QAAQ;AAAA,YACxB,UAAU,OAAO,MAAM;AAAA,YACvB,QAAQ,OAAO;AAAA,UACjB,CAAyB;AACzB,iBAAO;AAAA,QACT;AAAA,MACF;AACA,UAAI,OAAO,YAAY;AACrB,eAAO,MAAM,MAAM,KAAK,OAAO,KAAK,OAAO,OAAO,YAAY,MAAM;AACpE,eAAO,eAAe,OAAO,IAAI;AACjC,YAAI,OAAO,IAAI,mBAAmB,OAAO,IAAI,aAAa,GAAG;AAC3D,gBAAM,SAAS,OAAO,IAAI,kBACtB,uBAAuB,OAAO,IAAI,MAAM,OACxC,oBAAoB,OAAO,IAAI,QAAQ;AAC3C,gBAAM,OAAO,KAAK,QAAQ;AAAA,YACxB,UAAU,OAAO,IAAI;AAAA,YACrB,QAAQ,OAAO;AAAA,UACjB,CAAyB;AACzB,iBAAO;AAAA,QACT;AAAA,MACF;AACA,UAAI,OAAO,aAAa;AACtB,eAAO,OAAO,MAAM,KAAK,OAAO,KAAK,QAAQ,OAAO,aAAa,MAAM;AACvE,eAAO,eAAe,OAAO,KAAK;AAGlC,cAAM,SAAS,CAAC,OAAO,KAAK,mBAAmB,OAAO,KAAK,aAAa;AACxE,eAAO,SAAS;AAChB,YAAI,OAAO,KAAK,eAAe,UAAa,OAAO,KAAK,aAAa,GAAG;AACtE,iBAAO,SAAS,OAAO,KAAK,eAAe,KAAK,OAAO,KAAK;AAAA,QAC9D,OAAO;AACL,iBAAO,QAAQ,SAAS,IAAI;AAAA,QAC9B;AACA,cAAM,OAAO,IAAI;AAAA,UACf,UAAU,OAAO,KAAK;AAAA,UACtB,YAAY,OAAO,KAAK;AAAA,UACxB,aAAa,OAAO,KAAK;AAAA,UACzB,QAAQ,OAAO;AAAA,UACf,QAAQ,SAAS,OAAO;AAAA,QAC1B,CAAyB;AAAA,MAC3B,OAAO;AACL,eAAO,SAAS;AAChB,eAAO,QAAQ;AACf,cAAM,OAAO,IAAI,EAAE,QAAQ,OAAO,YAAY,CAAyB;AAAA,MACzE;AAAA,IACF,SAAS,KAAK;AACZ,YAAM,OAAO,KAAK,eAAe,QAAQ,MAAM,OAAO,GAAG,CAAC;AAC1D,YAAM;AAAA,IACR;AACA,WAAO;AAAA,EACT;AACF;;;AClXA,eAAsB,sBACpB,UACA,OACA,UAAgC,CAAC,GACH;AAC9B,QAAM,UAAU,IAAI,aAAa,KAAK;AACtC,QAAM,QAAQ,SAAS;AAAA,IACrB,YAAY,SAAS;AAAA,IACrB,WAAW,QAAQ;AAAA,IACnB,gBAAgB,SAAS;AAAA,IACzB,MAAM,SAAS;AAAA,IACf,GAAG,QAAQ;AAAA,EACb,CAAC;AACD,QAAM,UAAU,IAAI,eAAe,QAAQ,MAAM;AACjD,QAAM,SAAS,MAAM,QAAQ,IAAI,SAAS,SAAS,OAAO;AAC1D,QAAM,YAAY,SAAS,iBAAiB;AAC5C,QAAM,OAAO,OAAO,UAAU,OAAO,SAAS;AAC9C,QAAM,cAAc,OAAO,UAAU,UAAa,OAAO,MAAM,aAAa;AAC5E,QAAM,YAAY,OAAO,QAAQ,UAAa,OAAO,IAAI,aAAa;AACtE,QAAM,aAAa,OAAO,SAAS,UAAa,OAAO,KAAK,aAAa;AACzE,QAAM,eAAyC,OAC3C,YACA,eAAe,YACb,oBACA,aACE,iBACA;AACR,QAAM,QAAQ,OAAO;AAAA,IACnB;AAAA,IACA,OAAO,OAAO;AAAA,IACd;AAAA,IACA,OAAO,OAAO,SAAY,iBAAiB,MAAM;AAAA,EACnD,CAAC;AACD,SAAO;AAAA,IACL,OAAO,QAAQ;AAAA,IACf;AAAA,IACA,SAAS;AAAA,IACT;AAAA,IACA,OAAO,OAAO;AAAA,IACd;AAAA,EACF;AACF;AAEA,SAAS,iBAAiB,QAAsC;AAC9D,MAAI,OAAO,SAAS,OAAO,MAAM,aAAa;AAC5C,WAAO,sBAAsB,OAAO,MAAM,QAAQ;AACpD,MAAI,OAAO,OAAO,OAAO,IAAI,aAAa,EAAG,QAAO,oBAAoB,OAAO,IAAI,QAAQ;AAC3F,MAAI,OAAO,MAAM;AACf,QAAI,OAAO,KAAK,eAAe,QAAW;AACxC,aAAO,UAAU,OAAO,KAAK,eAAe,CAAC,IAAI,OAAO,KAAK,UAAU;AAAA,IACzE;AACA,WAAO,aAAa,OAAO,KAAK,QAAQ;AAAA,EAC1C;AACA,SAAO;AACT;","names":[]}

package/dist/cli.js

@@ -5,8 +5,8 @@ import {
   runRpcBatch,
   runRpcOnce,
   startServer
-} from "./chunk-QS3RBQPI.js";
-import "./chunk-CVVHBFGN.js";
+} from "./chunk-D5KBVULY.js";
+import "./chunk-CWNP4DV4.js";
 import "./chunk-PC4UYEBM.js";
 import "./chunk-3BFEG2F6.js";
 import "./chunk-PZ5AY32C.js";