@tangle-network/agent-eval 0.23.0 → 0.24.0

package/dist/chunk-LSH4MMOZ.js

@@ -0,0 +1,838 @@
+import {
+  TraceEmitter
+} from "./chunk-TVVP3ZZQ.js";
+
+// src/control-runtime.ts
+var DEFAULT_BUDGET = {
+  maxSteps: 8,
+  maxWallMs: 5 * 60 * 1e3
+};
+async function runAgentControlLoop(config) {
+  const budget = normalizeBudget(config.budget);
+  const actionFailure = config.actionFailure ?? "continue";
+  const controller = new AbortController();
+  const upstreamAbort = () => controller.abort(config.signal?.reason);
+  if (config.signal) {
+    if (config.signal.aborted) controller.abort(config.signal.reason);
+    else config.signal.addEventListener("abort", upstreamAbort, { once: true });
+  }
+  const started = Date.now();
+  const wallTimer = budget.maxWallMs ? setTimeout(
+    () => controller.abort(new Error("control runtime wall timeout")),
+    budget.maxWallMs
+  ) : void 0;
+  const history = [];
+  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;
+  try {
+    if (emitter) {
+      await runTrace(
+        runtimeErrors,
+        0,
+        () => emitter.startRun({
+          scenarioId: config.scenarioId ?? "agent-control-loop",
+          projectId: config.projectId,
+          variantId: config.variantId,
+          layer: "meta",
+          tags: {
+            intent: config.intent.slice(0, 120),
+            maxSteps: String(budget.maxSteps),
+            ...budget.maxCostUsd !== void 0 ? { maxCostUsd: String(budget.maxCostUsd) } : {}
+          }
+        })
+      );
+    }
+    let state;
+    let evals;
+    try {
+      state = await config.observe({ history, abortSignal: controller.signal });
+    } catch (err) {
+      const error = runtimeError("observe", 0, err);
+      runtimeErrors.push(error);
+      return finish(emitter, {
+        intent: config.intent,
+        pass: false,
+        completed: false,
+        reason: error.message,
+        steps: history,
+        finalState: void 0,
+        finalEvals: [],
+        wallMs: Date.now() - started,
+        spentCostUsd,
+        runId: emitter?.runId ?? null,
+        failureClass: "unknown",
+        runtimeErrors,
+        stoppedBy: "runtime-error"
+      });
+    }
+    try {
+      evals = await config.validate({
+        intent: config.intent,
+        state,
+        history,
+        abortSignal: controller.signal
+      });
+      await recordEvalSpans(emitter, evals, "initial", runtimeErrors, 0);
+    } catch (err) {
+      const error = runtimeError("validate", 0, err);
+      runtimeErrors.push(error);
+      return finish(emitter, {
+        intent: config.intent,
+        pass: false,
+        completed: false,
+        reason: error.message,
+        steps: history,
+        finalState: state,
+        finalEvals: [],
+        wallMs: Date.now() - started,
+        spentCostUsd,
+        runId: emitter?.runId ?? null,
+        failureClass: "unknown",
+        runtimeErrors,
+        stoppedBy: "runtime-error"
+      });
+    }
+    lastStateFingerprint = fingerprintState(state, config.stopPolicies);
+    for (let stepIndex = 0; stepIndex < budget.maxSteps; stepIndex++) {
+      if (controller.signal.aborted) {
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: false,
+          reason: abortReason(controller.signal),
+          score: void 0,
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "timeout",
+          runtimeErrors,
+          stoppedBy: "abort"
+        });
+      }
+      const budgetStop = budgetStopDecision(budget, spentCostUsd);
+      if (budgetStop.stop) {
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: false,
+          reason: budgetStop.reason,
+          score: averageScore(evals),
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "budget_exceeded",
+          runtimeErrors,
+          stoppedBy: "budget"
+        });
+      }
+      const ctx = makeContext(
+        config.intent,
+        state,
+        evals,
+        history,
+        budget,
+        stepIndex,
+        started,
+        spentCostUsd,
+        controller.signal,
+        emitter
+      );
+      let stop;
+      try {
+        stop = config.shouldStop ? await config.shouldStop(ctx) : defaultStopDecision(evals);
+      } catch (err) {
+        runtimeErrors.push(runtimeError("stop-policy", stepIndex, err));
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: false,
+          reason: runtimeErrors[runtimeErrors.length - 1].message,
+          score: averageScore(evals),
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "unknown",
+          runtimeErrors,
+          stoppedBy: "runtime-error"
+        });
+      }
+      if (stop.stop) {
+        return finish(emitter, {
+          intent: config.intent,
+          pass: stop.pass,
+          completed: true,
+          reason: stop.reason,
+          score: stop.score,
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: stop.failureClass,
+          runtimeErrors,
+          stoppedBy: "stop-policy"
+        });
+      }
+      let decision;
+      try {
+        decision = await config.decide(ctx);
+      } catch (err) {
+        runtimeErrors.push(runtimeError("decide", stepIndex, err));
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: false,
+          reason: runtimeErrors[runtimeErrors.length - 1].message,
+          score: averageScore(evals),
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "unknown",
+          runtimeErrors,
+          stoppedBy: "runtime-error"
+        });
+      }
+      if (decision.type === "stop") {
+        return finish(emitter, {
+          intent: config.intent,
+          pass: decision.pass ?? false,
+          completed: true,
+          reason: decision.reason,
+          score: decision.score,
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: decision.pass === false ? "unknown" : void 0,
+          runtimeErrors,
+          stoppedBy: "policy"
+        });
+      }
+      const actionFingerprint = fingerprintAction(decision.action, config.stopPolicies);
+      repeatedActionStreak = actionFingerprint === lastActionFingerprint ? repeatedActionStreak + 1 : 1;
+      lastActionFingerprint = actionFingerprint;
+      const repeatedActionStop = repeatedActionStopDecision(
+        config.stopPolicies,
+        repeatedActionStreak
+      );
+      if (repeatedActionStop.stop) {
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: true,
+          reason: repeatedActionStop.reason,
+          score: averageScore(evals),
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "tool_recovery_failure",
+          runtimeErrors,
+          stoppedBy: "stop-policy"
+        });
+      }
+      const beforeState = state;
+      const evalsBefore = evals;
+      const scoreBefore = averageScore(evals);
+      const actionStarted = Date.now();
+      const stepHandle = emitter ? await runTrace(
+        runtimeErrors,
+        stepIndex,
+        () => emitter.tool({
+          name: `control-step-${stepIndex}`,
+          toolName: "agent-control-action",
+          args: decision.action,
+          attributes: {
+            decision: decision.reason ?? "continue",
+            repeatedActionStreak
+          }
+        })
+      ) : void 0;
+      let actionOutcome;
+      try {
+        const result = await config.act(decision.action, ctx);
+        const rawCostUsd = config.getActionCostUsd?.({
+          action: decision.action,
+          result,
+          state,
+          evals,
+          history
+        });
+        const costUsd = normalizeActionCostUsd(rawCostUsd, runtimeErrors, stepIndex);
+        if (costUsd !== void 0 && Number.isFinite(costUsd) && costUsd > 0) {
+          spentCostUsd += costUsd;
+          await recordCostBudget(
+            emitter,
+            budget,
+            spentCostUsd,
+            stepHandle,
+            runtimeErrors,
+            stepIndex
+          );
+        }
+        actionOutcome = {
+          ok: true,
+          result,
+          ...costUsd !== void 0 ? { costUsd } : {},
+          durationMs: Date.now() - actionStarted
+        };
+      } catch (err) {
+        runtimeErrors.push(runtimeError("act", stepIndex, err));
+        actionOutcome = {
+          ok: false,
+          error: runtimeErrors[runtimeErrors.length - 1].message,
+          durationMs: Date.now() - actionStarted
+        };
+        if (actionFailure === "stop") {
+          await runTrace(
+            runtimeErrors,
+            stepIndex,
+            () => stepHandle?.fail(actionOutcome.error ?? "action failed")
+          );
+          const step2 = {
+            index: stepIndex,
+            decision,
+            beforeState,
+            afterState: state,
+            evalsBefore,
+            evalsAfter: evals,
+            actionOutcome,
+            startedAt: new Date(actionStarted).toISOString(),
+            endedAt: (/* @__PURE__ */ new Date()).toISOString()
+          };
+          history.push(step2);
+          await runOnStep(config.onStep, step2, runtimeErrors);
+          return finish(emitter, {
+            intent: config.intent,
+            pass: false,
+            completed: false,
+            reason: actionOutcome.error ?? "action failed",
+            score: averageScore(evals),
+            steps: history,
+            finalState: state,
+            finalEvals: evals,
+            wallMs: Date.now() - started,
+            spentCostUsd,
+            runId: emitter?.runId ?? null,
+            failureClass: "unknown",
+            runtimeErrors,
+            stoppedBy: "runtime-error"
+          });
+        }
+      }
+      try {
+        state = await config.observe({ history, abortSignal: controller.signal });
+      } catch (err) {
+        runtimeErrors.push(runtimeError("observe", stepIndex, err));
+        const step2 = {
+          index: stepIndex,
+          decision,
+          beforeState,
+          afterState: beforeState,
+          evalsBefore,
+          evalsAfter: evals,
+          actionOutcome,
+          startedAt: new Date(actionStarted).toISOString(),
+          endedAt: (/* @__PURE__ */ new Date()).toISOString()
+        };
+        history.push(step2);
+        await runTrace(
+          runtimeErrors,
+          stepIndex,
+          () => stepHandle?.fail(runtimeErrors[runtimeErrors.length - 1].message)
+        );
+        await runOnStep(config.onStep, step2, runtimeErrors);
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: false,
+          reason: runtimeErrors[runtimeErrors.length - 1].message,
+          score: averageScore(evals),
+          steps: history,
+          finalState: beforeState,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "unknown",
+          runtimeErrors,
+          stoppedBy: "runtime-error"
+        });
+      }
+      try {
+        evals = await config.validate({
+          intent: config.intent,
+          state,
+          history,
+          abortSignal: controller.signal
+        });
+        await recordEvalSpans(
+          emitter,
+          evals,
+          `step-${stepIndex}`,
+          runtimeErrors,
+          stepIndex,
+          stepHandle?.span.spanId
+        );
+      } catch (err) {
+        runtimeErrors.push(runtimeError("validate", stepIndex, err));
+        const step2 = {
+          index: stepIndex,
+          decision,
+          beforeState,
+          afterState: state,
+          evalsBefore,
+          evalsAfter: evals,
+          actionOutcome,
+          startedAt: new Date(actionStarted).toISOString(),
+          endedAt: (/* @__PURE__ */ new Date()).toISOString()
+        };
+        history.push(step2);
+        await runTrace(
+          runtimeErrors,
+          stepIndex,
+          () => stepHandle?.fail(runtimeErrors[runtimeErrors.length - 1].message)
+        );
+        await runOnStep(config.onStep, step2, runtimeErrors);
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: false,
+          reason: runtimeErrors[runtimeErrors.length - 1].message,
+          score: averageScore(evals),
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "unknown",
+          runtimeErrors,
+          stoppedBy: "runtime-error"
+        });
+      }
+      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 step = {
+        index: stepIndex,
+        decision,
+        beforeState,
+        afterState: state,
+        evalsBefore,
+        evalsAfter: evals,
+        actionOutcome,
+        startedAt: new Date(actionStarted).toISOString(),
+        endedAt: (/* @__PURE__ */ new Date()).toISOString()
+      };
+      history.push(step);
+      if (actionOutcome.ok) {
+        await runTrace(
+          runtimeErrors,
+          stepIndex,
+          () => stepHandle?.end({
+            attributes: {
+              actionCostUsd: actionOutcome.costUsd ?? null,
+              spentCostUsd,
+              scoreBefore: scoreBefore ?? null,
+              scoreAfter: scoreAfter ?? null,
+              noProgressStreak
+            }
+          })
+        );
+      } else {
+        await runTrace(
+          runtimeErrors,
+          stepIndex,
+          () => stepHandle?.fail(actionOutcome.error ?? "action failed", {
+            attributes: {
+              spentCostUsd,
+              noProgressStreak
+            }
+          })
+        );
+      }
+      await runOnStep(config.onStep, step, runtimeErrors);
+      if (noProgressStop.stop) {
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: true,
+          reason: noProgressStop.reason,
+          score: scoreAfter,
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "tool_recovery_failure",
+          runtimeErrors,
+          stoppedBy: "stop-policy"
+        });
+      }
+      const postStepBudgetStop = budgetStopDecision(budget, spentCostUsd);
+      if (postStepBudgetStop.stop) {
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: false,
+          reason: postStepBudgetStop.reason,
+          score: scoreAfter,
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "budget_exceeded",
+          runtimeErrors,
+          stoppedBy: "budget"
+        });
+      }
+      const postStepCtx = makeContext(
+        config.intent,
+        state,
+        evals,
+        history,
+        budget,
+        stepIndex + 1,
+        started,
+        spentCostUsd,
+        controller.signal,
+        emitter
+      );
+      let postStepStop;
+      try {
+        postStepStop = config.shouldStop ? await config.shouldStop(postStepCtx) : defaultStopDecision(evals);
+      } catch (err) {
+        runtimeErrors.push(runtimeError("stop-policy", stepIndex + 1, err));
+        return finish(emitter, {
+          intent: config.intent,
+          pass: false,
+          completed: false,
+          reason: runtimeErrors[runtimeErrors.length - 1].message,
+          score: averageScore(evals),
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: "unknown",
+          runtimeErrors,
+          stoppedBy: "runtime-error"
+        });
+      }
+      if (postStepStop.stop) {
+        return finish(emitter, {
+          intent: config.intent,
+          pass: postStepStop.pass,
+          completed: true,
+          reason: postStepStop.reason,
+          score: postStepStop.score,
+          steps: history,
+          finalState: state,
+          finalEvals: evals,
+          wallMs: Date.now() - started,
+          spentCostUsd,
+          runId: emitter?.runId ?? null,
+          failureClass: postStepStop.failureClass,
+          runtimeErrors,
+          stoppedBy: "stop-policy"
+        });
+      }
+    }
+    return finish(emitter, {
+      intent: config.intent,
+      pass: false,
+      completed: false,
+      reason: `budget exhausted: maxSteps=${budget.maxSteps}`,
+      steps: history,
+      finalState: state,
+      finalEvals: evals,
+      wallMs: Date.now() - started,
+      spentCostUsd,
+      runId: emitter?.runId ?? null,
+      failureClass: "budget_exceeded",
+      runtimeErrors,
+      stoppedBy: "budget"
+    });
+  } catch (err) {
+    runtimeErrors.push(runtimeError("act", history.length, err));
+    return finish(emitter, {
+      intent: config.intent,
+      pass: false,
+      completed: false,
+      reason: runtimeErrors[runtimeErrors.length - 1].message,
+      steps: history,
+      finalState: void 0,
+      finalEvals: [],
+      wallMs: Date.now() - started,
+      spentCostUsd,
+      runId: emitter?.runId ?? null,
+      failureClass: "unknown",
+      runtimeErrors,
+      stoppedBy: "runtime-error"
+    });
+  } finally {
+    if (wallTimer) clearTimeout(wallTimer);
+    if (config.signal) config.signal.removeEventListener("abort", upstreamAbort);
+  }
+}
+function stopOnNoProgress(maxNoProgressSteps, options = {}) {
+  return { ...options, maxNoProgressSteps };
+}
+function stopOnRepeatedAction(maxRepeatedActions, options = {}) {
+  return { ...options, maxRepeatedActions };
+}
+function objectiveEval(input) {
+  return { ...input, objective: true };
+}
+function subjectiveEval(input) {
+  return { ...input, objective: false };
+}
+function normalizeBudget(input) {
+  const raw = { ...DEFAULT_BUDGET, ...input };
+  if (!Number.isInteger(raw.maxSteps) || raw.maxSteps < 1) {
+    throw new RangeError(
+      `ControlRuntime budget.maxSteps must be an integer >= 1, got ${String(raw.maxSteps)}`
+    );
+  }
+  const budget = { maxSteps: raw.maxSteps };
+  if (raw.maxWallMs !== void 0) {
+    if (typeof raw.maxWallMs !== "number" || !Number.isFinite(raw.maxWallMs) || raw.maxWallMs <= 0) {
+      throw new RangeError(
+        `ControlRuntime budget.maxWallMs must be a positive finite number, got ${String(raw.maxWallMs)}`
+      );
+    }
+    budget.maxWallMs = raw.maxWallMs;
+  }
+  if (raw.maxCostUsd !== void 0) {
+    if (typeof raw.maxCostUsd !== "number" || !Number.isFinite(raw.maxCostUsd) || raw.maxCostUsd < 0) {
+      throw new RangeError(
+        `ControlRuntime budget.maxCostUsd must be a nonnegative finite number, got ${String(raw.maxCostUsd)}`
+      );
+    }
+    budget.maxCostUsd = raw.maxCostUsd;
+  }
+  return budget;
+}
+function normalizeActionCostUsd(costUsd, runtimeErrors, stepIndex) {
+  if (costUsd === void 0) return void 0;
+  if (!Number.isFinite(costUsd) || costUsd < 0) {
+    runtimeErrors.push(
+      runtimeError("act", stepIndex, new Error(`invalid action costUsd: ${String(costUsd)}`))
+    );
+    return void 0;
+  }
+  return costUsd;
+}
+function allCriticalPassed(evals) {
+  return evals.every(
+    (result) => result.passed || result.severity !== "critical" && result.severity !== "error"
+  );
+}
+function makeContext(intent, state, evals, history, budget, stepIndex, started, spentCostUsd, abortSignal, emitter) {
+  return {
+    intent,
+    state,
+    evals,
+    history,
+    budget,
+    stepIndex,
+    wallMs: Date.now() - started,
+    spentCostUsd,
+    remainingCostUsd: budget.maxCostUsd === void 0 ? void 0 : Math.max(0, budget.maxCostUsd - spentCostUsd),
+    abortSignal,
+    emitter
+  };
+}
+function defaultStopDecision(evals) {
+  if (!evals.length) return { stop: false, pass: false, reason: "no evals yet" };
+  const pass = allCriticalPassed(evals);
+  return pass ? { stop: true, pass: true, reason: "all critical evals passed", score: averageScore(evals) } : {
+    stop: false,
+    pass: false,
+    reason: "critical evals still failing",
+    score: averageScore(evals)
+  };
+}
+function averageScore(evals) {
+  const scored = evals.map((result) => result.score).filter((score) => typeof score === "number");
+  if (!scored.length) return void 0;
+  return Math.round(scored.reduce((sum, score) => sum + score, 0) / scored.length * 1e3) / 1e3;
+}
+function budgetStopDecision(budget, spentCostUsd) {
+  if (budget.maxCostUsd !== void 0 && spentCostUsd >= budget.maxCostUsd) {
+    return {
+      stop: true,
+      reason: `budget exhausted: maxCostUsd=${budget.maxCostUsd}`
+    };
+  }
+  return { stop: false, reason: "" };
+}
+async function recordCostBudget(emitter, budget, spentCostUsd, handle, runtimeErrors, stepIndex) {
+  if (!emitter || budget.maxCostUsd === void 0) return;
+  const maxCostUsd = budget.maxCostUsd;
+  await runTrace(
+    runtimeErrors,
+    stepIndex,
+    () => emitter.recordBudget({
+      dimension: "usd",
+      limit: maxCostUsd,
+      consumed: spentCostUsd,
+      remaining: Math.max(0, maxCostUsd - spentCostUsd),
+      breached: spentCostUsd >= maxCostUsd,
+      spanId: handle?.span.spanId
+    })
+  );
+}
+async function recordEvalSpans(emitter, evals, phase, runtimeErrors, stepIndex, targetSpanId) {
+  if (!emitter) return;
+  for (const result of evals) {
+    await runTrace(
+      runtimeErrors,
+      stepIndex,
+      () => emitter.recordJudge({
+        judgeId: result.objective ? "objective-validator" : "subjective-judge",
+        targetSpanId: targetSpanId ?? emitter.runId,
+        name: `control-eval/${result.id}`,
+        dimension: result.id,
+        score: typeof result.score === "number" ? result.score : result.passed ? 1 : 0,
+        rationale: result.detail,
+        evidence: result.evidence,
+        attributes: {
+          phase,
+          passed: result.passed,
+          severity: result.severity,
+          objective: result.objective
+        }
+      })
+    );
+  }
+}
+async function runOnStep(onStep, step, runtimeErrors) {
+  if (!onStep) return;
+  try {
+    await onStep(step);
+  } catch (err) {
+    runtimeErrors.push(runtimeError("on-step", step.index, err));
+  }
+}
+async function runTrace(runtimeErrors, stepIndex, write) {
+  try {
+    return await write();
+  } catch (err) {
+    runtimeErrors.push(runtimeError("trace", stepIndex, err));
+    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);
+}
+function fingerprintAction(action, policies) {
+  if (policies?.actionFingerprint) return policies.actionFingerprint(action);
+  return stableFingerprint(action);
+}
+function stableFingerprint(value) {
+  if (typeof value === "string") return value;
+  if (typeof value === "number" || typeof value === "boolean" || value == null) return String(value);
+  try {
+    return JSON.stringify(sortForFingerprint(value));
+  } catch {
+    return String(value);
+  }
+}
+function sortForFingerprint(value) {
+  if (Array.isArray(value)) return value.map(sortForFingerprint);
+  if (!value || typeof value !== "object") return value;
+  const record = value;
+  const sorted = {};
+  for (const key of Object.keys(record).sort()) {
+    sorted[key] = sortForFingerprint(record[key]);
+  }
+  return sorted;
+}
+function abortReason(signal) {
+  const reason = signal.reason;
+  if (reason instanceof Error) return reason.message;
+  return reason ? String(reason) : "aborted";
+}
+function runtimeError(phase, stepIndex, err) {
+  const message = err instanceof Error ? err.message : String(err);
+  return { phase, stepIndex, message };
+}
+async function finish(emitter, result) {
+  await runTrace(
+    result.runtimeErrors,
+    result.steps.length,
+    () => emitter?.endRun({
+      pass: result.pass,
+      score: result.score ?? averageScore(result.finalEvals),
+      failureClass: result.failureClass,
+      notes: result.reason
+    })
+  );
+  return result;
+}
+
+export {
+  runAgentControlLoop,
+  stopOnNoProgress,
+  stopOnRepeatedAction,
+  objectiveEval,
+  subjectiveEval,
+  allCriticalPassed
+};
+//# sourceMappingURL=chunk-LSH4MMOZ.js.map