@tangle-network/agent-eval 0.221.8 → 0.222.0

package/dist/trace-repair/index.js

@@ -1,17 +1,17 @@
-import { c as errorMessage, n as CaptureIntegrityError, s as ValidationError } from "../errors-BcOs5MKW.js";
-import { i as compareCodeUnits } from "../canonical-Dq2hOQ8L.js";
-import { i as pairedBootstrap, t as mean } from "../mean-BbA_JCzW.js";
-import { n as contentHash } from "../verdict-cache-CL_7tePD.js";
-import { n as isRecord } from "../json-BhdsJaOQ.js";
-import { t as certificationEvidenceDigest } from "../verdict-BGQmxQ8A.js";
-import { t as packageVersion } from "../package-version-D7lQHt_-.js";
-import { t as assertEqualDeclarativeTerms } from "../equal-terms-DbPGcVtI.js";
-import { c as runPrimeExchange, n as buildPrimePrompt, t as analystUsageReceiptFromPrimeUsage } from "../prime-protocol-BFgGrEMP.js";
-import { i as TIMEOUT_OBSERVATION_MARKER, l as isRecordedTimeout, n as FORMAT_ERROR_OBSERVATION_PREFIX, p as parseRecordedReturncode, r as SUBMIT_ACTION_SIGNATURE, s as deriveFailureSignature, t as wrapActionForExec } from "../exec-CpbtOXVt.js";
+import { c as errorMessage, n as CaptureIntegrityError, s as ValidationError } from "../errors-DeEEV9Vx.js";
+import { i as compareCodeUnits } from "../canonical-CicJlINA.js";
+import { i as pairedBootstrap, t as mean } from "../mean-CdDYG-us.js";
+import { n as contentHash } from "../verdict-cache-BVH-fl18.js";
+import { n as isRecord } from "../json-DtvZNAU9.js";
+import { t as certificationEvidenceDigest } from "../verdict-C8TcdwfD.js";
+import { t as packageVersion } from "../package-version-CJl_pY9W.js";
+import { t as assertEqualDeclarativeTerms } from "../equal-terms-K11VHmR3.js";
+import { c as runPrimeExchange, n as buildPrimePrompt, t as analystUsageReceiptFromPrimeUsage } from "../prime-protocol-CrhpXRz9.js";
+import { i as TIMEOUT_OBSERVATION_MARKER, l as isRecordedTimeout, n as FORMAT_ERROR_OBSERVATION_PREFIX, p as parseRecordedReturncode, r as SUBMIT_ACTION_SIGNATURE, s as deriveFailureSignature, t as wrapActionForExec } from "../exec-C0_AkLUv.js";
 import { createHash } from "node:crypto";
 import { constants } from "node:os";
 import { spawn } from "node:child_process";
-//#region src/trace-repair/admission-records.ts
+// src/trace-repair/admission-records.ts
 /**
 * The vocabulary of the TB-Repair admission pre-pass: what a corpus row is,
 * which failure population it belongs to, why it can leave the funnel, and the
@@ -156,8 +156,7 @@ function assertChainReconciles(artifact) {
 	const stratumInputs = artifact.byStratum.reduce((total, chain) => total + chain.input, 0);
 	if (stratumInputs + artifact.unstratified !== artifact.overall.input) throw new AdmissionDenominatorError(`stratum inputs ${stratumInputs} plus ${artifact.unstratified} unstratified do not cover ${artifact.overall.input} input rows`);
 }
-//#endregion
-//#region src/trace-repair/mini-swe-scaffold.ts
+// src/trace-repair/mini-swe-scaffold.ts
 /**
 * The mini-swe-agent scaffold as the Terminal-Bench-2 trajectory corpus
 * recorded it: one bash block per turn, one observation per command, and a
@@ -342,8 +341,7 @@ function submissionOf(output) {
 	if (first === void 0 || first.trim() !== SUBMIT_SENTINEL) return null;
 	return lines.slice(1).join("");
 }
-//#endregion
-//#region src/trace-repair/continuation-policy.ts
+// src/trace-repair/continuation-policy.ts
 /** A rollout ran outside the pinned policy, so its evidence cannot be used. */
 var ContinuationPolicyViolationError = class extends CaptureIntegrityError {};
 /** Two arms did not run the same policy, so their difference is not the intervention. */
@@ -707,8 +705,7 @@ function assertArmSymmetry(rollouts) {
 		if (seen.seed !== rollout.seed) throw new ContinuationSymmetryError(`paired rollouts ${key} drew different seeds: ${seen.arm}=${seen.seed}, ${rollout.arm}=${rollout.seed}`);
 	}
 }
-//#endregion
-//#region src/trace-repair/control-policy.ts
+// src/trace-repair/control-policy.ts
 /**
 * The control policy admission screens under, declared rather than assumed.
 *
@@ -774,8 +771,7 @@ function assertControlCalibrated(policy, screening) {
 	if (screening === "enforced" && !canRescue) throw new UncalibratedControlError(`control policy ${policy.id} (digest ${policy.digest}) allows ${policy.stepBudget} model call(s) per rollout, so a control rollout executes no command and grades the same bytes the end-state check already graded as failing. Under it conditions 3 and 4 can only fire on a grader that disagrees with itself, so they screen nothing. Give the control a step budget of at least 1, or set controlScreening to 'declared-inert' and read a control pass as the oracle flip it is.`);
 	if (screening === "declared-inert" && canRescue) throw new UncalibratedControlError(`control policy ${policy.id} (digest ${policy.digest}) allows ${policy.stepBudget} model call(s) per rollout, so it can rescue a row, but the criteria declare it inert. Set controlScreening to 'enforced' so a control pass is recorded as a rescue.`);
 }
-//#endregion
-//#region src/trace-repair/admission.ts
+// src/trace-repair/admission.ts
 const ADMISSION_CONFIG_DEFAULTS = Object.freeze({
 	maxPrefixDivergence: .1,
 	controlRollouts: 3,
@@ -1154,8 +1150,7 @@ function preview(rowIds) {
 	const shown = rowIds.slice(0, 5).join(", ");
 	return rowIds.length > 5 ? `${shown}, +${rowIds.length - 5} more` : shown;
 }
-//#endregion
-//#region src/trace-repair/admission-contract.ts
+// src/trace-repair/admission-contract.ts
 /**
 * Admission: the executed checks a row must pass before an analyst is allowed
 * to see it, and the brand that proves it did.
@@ -1274,8 +1269,7 @@ function assertCriteria(criteria) {
 	if (!(maxPrefixDivergenceRatio >= 0 && maxPrefixDivergenceRatio <= 1)) throw new ValidationError(`admission maxPrefixDivergenceRatio must be within [0,1], got ${maxPrefixDivergenceRatio}`);
 	if (!Number.isInteger(controlRollouts) || controlRollouts <= 0) throw new ValidationError(`admission controlRollouts must be a positive integer, got ${controlRollouts}`);
 }
-//#endregion
-//#region src/trace-repair/admission-report.ts
+// src/trace-repair/admission-report.ts
 /** Plain JSON. `JSON.stringify` of this object is the machine-readable artifact. */
 function admissionArtifact(report) {
 	assertChainReconciles(report.chain);
@@ -1403,8 +1397,7 @@ function rate(part, total) {
 function formatShare(share) {
 	return `${(share * 100).toFixed(2)}%`;
 }
-//#endregion
-//#region src/trace-repair/action-budget.ts
+// src/trace-repair/action-budget.ts
 /**
 * The action budget an intervention must fit inside.
 *
@@ -1765,8 +1758,7 @@ function assertBudget(budget) {
 function normalizeActionForComparison(action) {
 	return action.split("\n").map((line) => line.trimEnd()).filter((line, index, lines) => line.length > 0 || index > 0 && index < lines.length - 1).join("\n").trim();
 }
-//#endregion
-//#region src/trace-repair/blinding.ts
+// src/trace-repair/blinding.ts
 /**
 * What the analyst is shown.
 *
@@ -1812,8 +1804,7 @@ Object.freeze([
 	"image",
 	"cwd"
 ]);
-//#endregion
-//#region src/trace-repair/repair-prompt.ts
+// src/trace-repair/repair-prompt.ts
 /**
 * The one question every repair arm answers.
 *
@@ -1932,8 +1923,7 @@ function repairArmPromptSha256(budget, contract) {
 		contract
 	})).digest("hex");
 }
-//#endregion
-//#region src/trace-repair/analyst-arm.ts
+// src/trace-repair/analyst-arm.ts
 /**
 * What an analyst arm is, and what every arm owes the comparison.
 *
@@ -2074,8 +2064,7 @@ function assertReplyWithinDeclaration(declaration, reply) {
 	if (reply.status === "failed" && !reply.failure.trim()) throw new ValidationError(`arm '${declaration.id}' failed without stating why`);
 	if (reply.status === "finding" && !reply.intervention.action) throw new ValidationError(`arm '${declaration.id}' returned a finding with an empty intervention`);
 }
-//#endregion
-//#region src/trace-repair/analyst-response.ts
+// src/trace-repair/analyst-response.ts
 /** The literal an analyst returns when no single step carries the failure. */
 const NO_DECISIVE_FAILURE = "no-decisive-failure";
 /**
@@ -2221,8 +2210,7 @@ function topLevelObjectSlices(text) {
 	}
 	return slices;
 }
-//#endregion
-//#region src/trace-repair/arm-completion.ts
+// src/trace-repair/arm-completion.ts
 function createCompletionRepairArm(options) {
 	if (!Number.isSafeInteger(options.timeoutMs) || options.timeoutMs <= 0) throw new ValidationError(`repair arm '${options.id}' timeoutMs must be a positive safe integer, got ${options.timeoutMs}`);
 	const budget = options.budget ?? SCAFFOLD_INTERVENTION_BUDGET;
@@ -2361,8 +2349,7 @@ function describe$1(value) {
 	if (value === void 0) return "undefined";
 	return typeof value === "string" ? JSON.stringify(value) : String(value);
 }
-//#endregion
-//#region src/trace-repair/arm-dspy.ts
+// src/trace-repair/arm-dspy.ts
 /**
 * Token the bridge reads to select the typed repair signature.
 *
@@ -2669,8 +2656,7 @@ function describe(value) {
 function assertDspyRepairEngine(engine) {
 	if (engine.id !== "dspy-rlm") throw new ValidationError(`the DSPy repair arm needs the dspy-rlm engine, got '${engine.id}'; another engine does not run the typed repair signature and would answer a different contract`);
 }
-//#endregion
-//#region src/trace-repair/continuation-records.ts
+// src/trace-repair/continuation-records.ts
 /** Corpus steps for the continuation alone, excluding the prefix it inherited. */
 function rolloutRecordedSteps(rollout) {
 	return rollout.steps.map((step) => ({
@@ -2717,8 +2703,7 @@ function rolloutDigest(rollout) {
 		}))
 	});
 }
-//#endregion
-//#region src/trace-repair/degenerate-strategies.ts
+// src/trace-repair/degenerate-strategies.ts
 const DEGENERATE_STRATEGIES = [
 	{
 		id: "point-at-any-nonzero-exit-step",
@@ -2777,8 +2762,7 @@ const DEGENERATE_STRATEGIES = [
 		enforcedIn: "grade.ts (the intervention is applied at the named k only)"
 	}
 ];
-//#endregion
-//#region src/trace-repair/funnel.ts
+// src/trace-repair/funnel.ts
 const ZERO_CREDIT = Object.freeze({
 	executes: 0,
 	localFlip: 0,
@@ -2833,8 +2817,7 @@ function countFunnel(grades) {
 		t4RepairFlipAll: t4All
 	};
 }
-//#endregion
-//#region src/trace-repair/delta-repair.ts
+// src/trace-repair/delta-repair.ts
 /**
 * The headline metric.
 *
@@ -3039,8 +3022,7 @@ function signed(value) {
 	const points = value * 100;
 	return `${points >= 0 ? "+" : ""}${points.toFixed(1)} pp`;
 }
-//#endregion
-//#region src/trace-repair/docker-environment.ts
+// src/trace-repair/docker-environment.ts
 /**
 * Docker-backed continuation environment.
 *
@@ -3187,8 +3169,7 @@ function createDockerContinuationEnvironment(options) {
 		}
 	};
 }
-//#endregion
-//#region src/trace-repair/grade.ts
+// src/trace-repair/grade.ts
 /**
 * Grade one analyst answer about one admitted row.
 *
@@ -3515,8 +3496,7 @@ function excerpt(text) {
 	const half = Math.floor(OUTPUT_EXCERPT_CHARS / 2);
 	return `${text.slice(0, half)}\n… [${text.length - OUTPUT_EXCERPT_CHARS} chars elided] …\n${text.slice(-2e3)}`;
 }
-//#endregion
-//#region src/trace-repair/oracle-determinism.ts
+// src/trace-repair/oracle-determinism.ts
 /**
 * Whether a task's own grader is a function of the container state.
 *
@@ -3658,8 +3638,7 @@ function assertDeterministicOracle(verdict) {
 	if (verdict.stable) return;
 	throw new NondeterministicOracleError(`task ${verdict.taskName} graded byte-identical state inconsistently: flip rate ${(verdict.flipRate * 100).toFixed(1)} % over ${verdict.replicates} replicates (${verdict.detail}). Its verdict is not a function of the state, so it cannot carry ground truth for an intervention study.`);
 }
-//#endregion
-//#region src/trace-repair/test-oracle.ts
+// src/trace-repair/test-oracle.ts
 /**
 * The held-out suite, injected from outside the box at grade time.
 *
@@ -3763,7 +3742,6 @@ function parentDirectory(path) {
 function shellQuote(value) {
 	return `'${value.replaceAll("'", `'\\''`)}'`;
 }
-//#endregion
 export { ADMISSION_CONFIG_DEFAULTS, DEGENERATE_STRATEGIES, DSPY_REPAIR_SIGNATURE, TestSuiteTamperedError, UncalibratedControlError, admissionArtifact, admitRow, askRepairArm, assertAnalystIndependent, assertArmSymmetry, assertChainReconciles, assertControlCalibrated, assertDenominatorIntact, assertDeterministicOracle, blindTrajectory, buildDenominatorChain, continuationPolicyDigest, continuationSeed, createCompletionRepairArm, createDockerContinuationEnvironment, createDspyRepairArm, defineControlPolicy, definePinnedContinuationPolicy, deltaRepair, gradeRepairRow, injectedTestOracle, nodeProcessRunner, parseAction, parseAnalystResponse, parseTaskOracleRegistry, renderAdmissionReport, renderDeltaRepairReport, repairArmAsymmetries, repairArmPromptSha256, repairArmResponse, repairQuestionSha256, rolloutDigest, rolloutRecordedSteps, runAdmission, runContinuation, totalCost, totalUsage };
 
 //# sourceMappingURL=index.js.map