@tangle-network/agent-eval 0.216.1 → 0.218.1

package/dist/contract/index.js

@@ -1,29 +1,30 @@
-import { a as hashCanonical } from "../canonical-B2RqLofv.js";
-import { a as parseRunRecordSafe } from "../run-record-Boc5IUTp.js";
-import { i as pairedBootstrap } from "../mean-DHoMiudD.js";
-import { t as decidePairedPromotion } from "../paired-promotion-decision-DsAPYGCm.js";
-import { a as heldoutSignificance } from "../statistical-heldout-42NlCNDe.js";
-import { r as summarizeExecution, t as analyzeRuns } from "../analyze-runs-C84irm4i.js";
-import { G as defaultProductionGate, a as transientDispatchFailure, m as runImprovementLoop, nt as runEval, rt as runCampaign, t as llmJudge } from "../llm-judge-CG12SQJq.js";
+import { a as hashCanonical } from "../canonical-Dq2hOQ8L.js";
+import "../run-record-DhEJgoPz.js";
+import { r as summarizeExecution, t as analyzeRuns } from "../analyze-runs-DqPJaJ4X.js";
+import { i as pairedBootstrap } from "../mean-BbA_JCzW.js";
+import { t as decidePairedPromotion } from "../paired-promotion-decision-D0SHb-ZL.js";
+import { a as heldoutSignificance } from "../statistical-heldout-Bksv92Th.js";
+import { B as defaultProductionGate, Q as runCampaign, Z as runEval, i as transientDispatchFailure, o as clamp01, p as runImprovementLoop, t as llmJudge } from "../llm-judge-Djxh-Uhs.js";
 import { i as isModelPriced, n as estimateCost } from "../metrics-BKaPBtIG.js";
-import { t as CostLedger } from "../cost-ledger-BP8V4kUQ.js";
-import { G as campaignSplitDigest } from "../campaign-evidence-Clrn4hBV.js";
-import { h as mapConcurrentRange } from "../journal-DfRz9kps.js";
-import { F as inMemoryCampaignStorage, P as fsCampaignStorage } from "../external-optimizer-subprocess-BqW0hHcI.js";
-import { o as makeProposalFinding } from "../types-IXD2HJzu.js";
-import { n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-CNTPzIkp.js";
-import { t as buildDefaultAnalystRegistry } from "../default-registry-COhk9g36.js";
-import { t as extractUsage } from "../extract-usage-BrnzRJ6H.js";
-import { t as createChatClient } from "../chat-client-l-MVDlDk.js";
+import { t as CostLedger } from "../cost-ledger-D0352Z4d.js";
+import { z as campaignSplitDigest } from "../campaign-evidence-DunzcJGv.js";
+import { o as mapConcurrentRange } from "../journal-C7om8Gum.js";
+import { M as fsCampaignStorage, N as inMemoryCampaignStorage } from "../external-optimizer-subprocess-B5gdLRZv.js";
+import { a as makeProposalFinding } from "../types-DqFMNTgK.js";
+import { n as isRecord } from "../json-BhdsJaOQ.js";
+import { n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-BGdFKlrc.js";
+import { t as buildDefaultAnalystRegistry } from "../default-registry-BJ-Pq6yL.js";
+import { t as extractUsage } from "../extract-usage-BZLU01kR.js";
+import { t as createChatClient } from "../chat-client-B2pXlPdx.js";
 import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-DUcMj-3D.js";
-import { _ as runReferenceEquivalenceJudge, d as externalTextOptimizationMethod, g as createReferenceEquivalenceJudge, h as REFERENCE_EQUIVALENCE_JUDGE_VERSION, l as heldOutGate, m as REFERENCE_EQUIVALENCE_INPUT_LIMITS, n as compareOptimizationMethods, r as gepaOptimizationMethod, t as skillOptOptimizationMethod, u as composeGate } from "../skillopt-optimization-method-BiwA8nfX.js";
-import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-KO57gbyV.js";
-import { t as fromOtelSpans } from "../otel-spans-B9tUXRHF.js";
+import { c as composeGate, l as externalTextOptimizationMethod, n as compareOptimizationMethods, r as gepaOptimizationMethod, s as heldOutGate, t as skillOptOptimizationMethod } from "../skillopt-optimization-method-5dlex7MJ.js";
+import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-6A8osGr_.js";
+import { t as fromOtelSpans } from "../otel-spans-CdN909k1.js";
 import { createHash } from "node:crypto";
 import { agentCandidateBenchmarkSuiteSchema, agentCandidateBenchmarkTaskSchema, agentCandidateBundleSchema, agentCandidateEvaluationPolicySchema, agentCandidateExperimentSchema, agentImprovementMeasuredComparisonSchema, agentProfileImprovementExperimentSchema, agentProfileImprovementMeasuredComparisonSchema, agentProfileImprovementRunCellSchema, agentProfileImprovementRunReceiptSchema, agentProfileImprovementSuiteInputsSchema, agentProfileImprovementSuiteSchema, agentProfileImprovementTaskSchema, candidateExecutionEvidenceSchema, canonicalCandidateDigest, canonicalCandidateJson, numbersApproximatelyEqual, omitTopLevelDigest } from "@tangle-network/agent-interface";
 import { createReadStream } from "node:fs";
-import { dirname, isAbsolute, join, relative, resolve, sep } from "node:path";
-import { mkdir, readFile, readdir, stat, writeFile } from "node:fs/promises";
+import "node:path";
+import "node:fs/promises";
 import { readSessionInput, summarizeSessionInput } from "@tangle-network/harness-sessions";
 //#region src/contract/fixed-spend.ts
 function addFixedSpend(left, right) {
@@ -368,6 +369,11 @@ function assertKnownCost(cost, label) {
 * only validates the projected measurements and derives their shared decision.
 */
 function evaluatePairedMeasurements(options) {
+	return pairedVerdict(budgetAndChecks(runTotals(dimensionRegressions(pairedEstimates(validatePairedInput({ options }))))));
+}
+/** The checked policy and measurements, their cell ids, shared dimensions and score arrays. */
+function validatePairedInput(prev) {
+	const { options } = prev;
 	if (options.measurements.length === 0) throw new Error("paired measurement evaluation requires at least one paired cell");
 	const preparationCost = options.preparationCost ?? {
 		usd: 0,
@@ -383,6 +389,28 @@ function evaluatePairedMeasurements(options) {
 	const baselineScores = measurements.map((measurement) => measurement.baseline.score);
 	const candidateScores = measurements.map((measurement) => measurement.candidate.score);
 	const { confidenceLevel: confidence, resamples, bootstrapSeed, deltaThreshold, minProductiveRuns, budgetUsd, criticalDimensions, regressionTolerance } = policy;
+	return {
+		...prev,
+		preparationCost,
+		policy,
+		measurements,
+		cellIds,
+		dimensions,
+		baselineScores,
+		candidateScores,
+		confidence,
+		resamples,
+		bootstrapSeed,
+		deltaThreshold,
+		minProductiveRuns,
+		budgetUsd,
+		criticalDimensions,
+		regressionTolerance
+	};
+}
+/** Held-out significance, the overall, cost and latency estimates, and the observed power. */
+function pairedEstimates(prev) {
+	const { options, measurements, cellIds, dimensions, baselineScores, candidateScores, confidence, resamples, bootstrapSeed, deltaThreshold, minProductiveRuns } = prev;
 	const significance = heldoutSignificance({
 		before: baselineScores,
 		after: candidateScores,
@@ -455,6 +483,19 @@ function evaluatePairedMeasurements(options) {
 		confidence,
 		sharedScorerChannel: options.sharedScorerChannel
 	});
+	return {
+		...prev,
+		significance,
+		overall,
+		objectives,
+		cost,
+		latency,
+		power
+	};
+}
+/** Critical dimensions that regressed, on either burden of proof. */
+function dimensionRegressions(prev) {
+	const { measurements, dimensions, confidence, resamples, bootstrapSeed, minProductiveRuns, criticalDimensions, regressionTolerance, objectives, power } = prev;
 	const powerSufficient = power.sufficient;
 	const guardedDimensions = new Set(criticalDimensions);
 	const missingCriticalDimensions = criticalDimensions.filter((dimension) => !dimensions.includes(dimension));
@@ -472,9 +513,22 @@ function evaluatePairedMeasurements(options) {
 		}).promote;
 	}));
 	const regressions = objectives.filter((objective) => objective.kind === "dimension" && guardedDimensions.has(objective.name) && objective.availability === "measured" && (objective.confidenceInterval.lower < -regressionTolerance || regressionProven.has(objective.name)));
+	return {
+		...prev,
+		powerSufficient,
+		guardedDimensions,
+		missingCriticalDimensions,
+		regressionProven,
+		regressions
+	};
+}
+/** Run completion, per-task regressions and the measurement cost. */
+function runTotals(prev) {
+	const { options, measurements } = prev;
 	const measurementCostUsd = measurements.reduce((sum, measurement) => sum + measurement.baseline.costUsd + measurement.candidate.costUsd, 0);
 	const measurementWorkDurationMs = measurements.reduce((sum, measurement) => sum + measurement.baseline.latencyMs + measurement.candidate.latencyMs, 0);
-	const incompleteRuns = measurements.flatMap((measurement) => [measurement.baseline, measurement.candidate]).filter((run) => !run.completed);
+	const completedRuns = measurements.flatMap((measurement) => [measurement.baseline, measurement.candidate]);
+	const incompleteRuns = completedRuns.filter((run) => !run.completed);
 	const regressedTasks = measurements.filter((measurement) => !measurement.candidate.passed && measurement.baseline.passed);
 	const sharedTaskFailures = measurements.filter((measurement) => !measurement.candidate.passed && !measurement.baseline.passed);
 	const repairedTasks = measurements.filter((measurement) => measurement.candidate.passed && !measurement.baseline.passed);
@@ -483,7 +537,22 @@ function evaluatePairedMeasurements(options) {
 		provenance: measurements.every((measurement) => measurement.baseline.costProvenance === "observed" && measurement.candidate.costProvenance === "observed") ? "observed" : "estimated"
 	};
 	const measurementCost = options.measurementCost ?? derivedMeasurementCost;
-	assertKnownCost(measurementCost, "paired measurement cost");
+	return {
+		...prev,
+		measurementCostUsd,
+		measurementWorkDurationMs,
+		completedRuns,
+		incompleteRuns,
+		regressedTasks,
+		sharedTaskFailures,
+		repairedTasks,
+		derivedMeasurementCost,
+		measurementCost
+	};
+}
+/** Total cost against the budget, and the checks the verdict is drawn from. */
+function budgetAndChecks(prev) {
+	const { options, preparationCost, budgetUsd, significance, powerSufficient, missingCriticalDimensions, regressions, incompleteRuns, regressedTasks, derivedMeasurementCost, measurementCost } = prev;
 	if (options.measurementCost !== void 0 && (measurementCost.provenance !== derivedMeasurementCost.provenance || !numbersApproximatelyEqual(measurementCost.usd, derivedMeasurementCost.usd))) throw new Error("paired measurement cost does not match its signed receipts");
 	const totalCost = {
 		usd: preparationCost.usd + measurementCost.usd,
@@ -518,6 +587,18 @@ function evaluatePairedMeasurements(options) {
 	];
 	const shipped = checks.every((check) => check.passed);
 	const hardFailure = incompleteRuns.length > 0 || regressedTasks.length > 0 || regressions.length > 0 || missingCriticalDimensions.length > 0 || !budgetPassed;
+	return {
+		...prev,
+		totalCost,
+		budgetPassed,
+		checks,
+		shipped,
+		hardFailure
+	};
+}
+/** The verdict, its reasons and the evaluation record. */
+function pairedVerdict(prev) {
+	const { options, baselineScores, confidence, deltaThreshold, budgetUsd, significance, overall, objectives, power, powerSufficient, missingCriticalDimensions, regressions, measurementWorkDurationMs, incompleteRuns, regressedTasks, sharedTaskFailures, repairedTasks, measurementCost, totalCost, budgetPassed, checks, shipped, hardFailure } = prev;
 	const reasons = [
 		...significance.significant ? [] : [significanceReason(significance, deltaThreshold)],
 		...powerSufficient || significance.fewRuns ? [] : [power.reason],
@@ -1128,182 +1209,6 @@ const profileReceiptAdapter = {
 	passed: (receipt) => receipt.grading.passed
 };
 //#endregion
-//#region src/contract/intake/run-record-dir.ts
-/**
-* # `intake/run-record-dir` — load a directory or file of `RunRecord`s.
-*
-* The on-disk counterpart to the in-memory intake adapters: point it at a
-* single `.json` (array) / `.jsonl` (one record per line) file or at a
-* directory of such files, and it returns the substrate-canonical
-* `RunRecord[]` ready for `analyzeRuns({ runs })`.
-*
-* Validation is at the boundary: each parsed object goes through
-* `parseRunRecordSafe`. By default an invalid record fails loud with its
-* file + index; pass `onInvalid: 'collect'` to keep the valid records and
-* receive the rejects as structured diagnostics instead.
-*/
-const ANALYSIS_ARTIFACT$1 = "analysis.json";
-function defaultInclude(fileName) {
-	if (fileName === ANALYSIS_ARTIFACT$1) return false;
-	return fileName.endsWith(".json") || fileName.endsWith(".jsonl");
-}
-/**
-* Resolve a file or directory path into validated `RunRecord[]`.
-*
-* A `.json` file must parse to a top-level array; a `.jsonl` file is one
-* record per non-empty line. Directories are read shallowly by default
-* (set `recursive` to descend); the `analysis.json` output artifact is
-* always excluded.
-*/
-async function fromRunRecordDir(path, options = {}) {
-	const onInvalid = options.onInvalid ?? "throw";
-	const include = options.include ?? defaultInclude;
-	const filePaths = (await stat(path)).isDirectory() ? await collectFiles(path, include, options.recursive ?? false) : [path];
-	const runs = [];
-	const rejected = [];
-	for (const file of filePaths) {
-		const raw = await parseRecordFile(file);
-		for (const { index, value } of raw) {
-			const parsed = parseRunRecordSafe(value);
-			if (parsed.ok) {
-				runs.push(parsed.value);
-				continue;
-			}
-			const rejection = {
-				file,
-				index,
-				reason: parsed.error.message
-			};
-			if (onInvalid === "throw") throw new Error(`fromRunRecordDir: invalid RunRecord in '${file}' at index ${index}: ${parsed.error.message}`);
-			rejected.push(rejection);
-		}
-	}
-	return {
-		runs,
-		rejected,
-		files: filePaths
-	};
-}
-/** Read a single `.json` / `.jsonl` file into `{ index, value }` pairs. A
-*  malformed JSONL line throws with its line number rather than being skipped —
-*  silent line-dropping is how corpora quietly shrink. */
-async function parseRecordFile(file) {
-	const trimmed = (await readFile(file, "utf8")).trim();
-	if (trimmed.length === 0) return [];
-	if (trimmed.startsWith("[")) {
-		const parsed = JSON.parse(trimmed);
-		if (!Array.isArray(parsed)) throw new Error(`fromRunRecordDir: file '${file}' did not parse to an array`);
-		return parsed.map((value, index) => ({
-			index,
-			value
-		}));
-	}
-	const out = [];
-	const lines = trimmed.split("\n");
-	for (let i = 0; i < lines.length; i++) {
-		const line = lines[i].trim();
-		if (line.length === 0) continue;
-		try {
-			out.push({
-				index: i,
-				value: JSON.parse(line)
-			});
-		} catch (err) {
-			throw new Error(`fromRunRecordDir: file '${file}' line ${i + 1} is not valid JSON: ${err instanceof Error ? err.message : String(err)}`);
-		}
-	}
-	return out;
-}
-/** Sorted file list under a directory, filtered by `include`. Sorted so the
-*  resulting `RunRecord` order — and any downstream fingerprint — is stable
-*  across filesystems. */
-async function collectFiles(dir, include, recursive) {
-	const entries = await readdir(dir, { withFileTypes: true });
-	const files = [];
-	const subdirs = [];
-	for (const entry of entries) {
-		if (entry.isDirectory()) {
-			if (recursive) subdirs.push(join(dir, entry.name));
-			continue;
-		}
-		if (include(entry.name)) files.push(join(dir, entry.name));
-	}
-	files.sort();
-	subdirs.sort();
-	for (const sub of subdirs) files.push(...await collectFiles(sub, include, recursive));
-	return files;
-}
-//#endregion
-//#region src/contract/eval-reporting-suite.ts
-/**
-* # `evalReportingSuite` — one call from runs (or a run dir) to `analysis.json`.
-*
-* A thin wrapper over the analysis primitive (`analyzeRuns`) and the on-disk
-* intake adapter (`fromRunRecordDir`). It does NOT reimplement any statistics,
-* distributions, or clustering — it resolves the input into validated
-* `RunRecord[]`, calls `analyzeRuns` with the options you'd pass it directly,
-* wraps the result in a small provenance envelope, and (optionally) writes a
-* single `analysis.json` artifact.
-*
-* ```ts
-* // From a directory of run files, write ./runs/analysis.json:
-* const suite = await evalReportingSuite('./runs', { write: true })
-* // From records already in memory, no write:
-* const suite = await evalReportingSuite(records, { analyze: { decisionThreshold: 0.03 } })
-* suite.report // the InsightReport — distributions, paired lift, findings rollup
-* ```
-*/
-const ANALYSIS_ARTIFACT = "analysis.json";
-/**
-* Resolve runs (or a run dir/file), run `analyzeRuns`, and optionally persist a
-* single `analysis.json`. The only analysis logic lives in `analyzeRuns`; this
-* function is composition + I/O.
-*/
-async function evalReportingSuite(input, options = {}) {
-	const fromPath = typeof input === "string";
-	let runs;
-	let files = [];
-	let rejected = [];
-	if (fromPath) {
-		const loaded = await fromRunRecordDir(input, options.load);
-		runs = loaded.runs;
-		files = loaded.files;
-		rejected = loaded.rejected;
-	} else runs = input;
-	if (runs.length === 0) throw new Error(fromPath ? `evalReportingSuite: no RunRecords found at '${input}'` : "evalReportingSuite: no RunRecords to analyze");
-	const result = {
-		report: await analyzeRuns({
-			...options.analyze,
-			runs
-		}),
-		provenance: {
-			generatedAt: (/* @__PURE__ */ new Date()).toISOString(),
-			runCount: runs.length,
-			sourcePath: fromPath ? input : null,
-			files,
-			rejected
-		},
-		writtenTo: null
-	};
-	const target = resolveWriteTarget(options.write, fromPath ? input : null);
-	if (target) {
-		await mkdir(dirname(target), { recursive: true });
-		await writeFile(target, `${JSON.stringify(result, null, 2)}\n`, "utf8");
-		result.writtenTo = target;
-	}
-	return result;
-}
-/** Resolve where (if anywhere) to write `analysis.json`. Returns null when
-*  writing is disabled. Throws on `write: true` with in-memory input — there is
-*  no directory to anchor the artifact to, and silently inventing `cwd` would
-*  scatter files. */
-function resolveWriteTarget(write, sourcePath) {
-	if (!write) return null;
-	if (typeof write === "string") return write.endsWith("/") || !write.endsWith(".json") && !write.endsWith(".jsonl") ? join(write, ANALYSIS_ARTIFACT) : write;
-	if (sourcePath === null) throw new Error("evalReportingSuite: write:true needs a source path to anchor analysis.json — pass an explicit output path when analyzing in-memory records");
-	return sourcePath.endsWith(".json") || sourcePath.endsWith(".jsonl") ? join(dirname(sourcePath), ANALYSIS_ARTIFACT) : join(sourcePath, ANALYSIS_ARTIFACT);
-}
-//#endregion
 //#region src/contract/intake/agent-trace.ts
 function rangeLines(r) {
 	return Math.max(0, r.end_line - r.start_line + 1);
@@ -1385,9 +1290,6 @@ function partitionRunsByAuthoringModel(runs, index) {
 }
 //#endregion
 //#region src/contract/intake/code-agent-fields.ts
-function isRecord(value) {
-	return value !== null && typeof value === "object" && !Array.isArray(value);
-}
 function record(value) {
 	return isRecord(value) ? value : null;
 }
@@ -2200,7 +2102,8 @@ function emptyMetrics(entries) {
 }
 function penalizeErrors(base, metrics) {
 	const operations = Math.max(1, metrics.toolCalls + metrics.patchAttempts);
-	return clamp01(base * (1 - .5 * Math.min(1, (metrics.toolErrors + metrics.patchFailures) / operations)));
+	const errorRate = Math.min(1, (metrics.toolErrors + metrics.patchFailures) / operations);
+	return clamp01(base * (1 - .5 * errorRate));
 }
 function emptyTokenTotals() {
 	return {
@@ -2291,10 +2194,6 @@ function resolveSessionCost(explicit, metrics, model) {
 		}
 	};
 }
-function clamp01(value) {
-	if (!Number.isFinite(value)) return 0;
-	return Math.max(0, Math.min(1, value));
-}
 //#endregion
 //#region src/contract/intake/code-agent-codex-metrics.ts
 function codexMetrics(entries) {
@@ -2740,10 +2639,6 @@ function fromKimiCodeSession(options) {
 function fromPiSession(options) {
 	return fromCodeAgentSession("pi", options);
 }
-/** The graph IR a Pi graph run exports (`nodes[].ir`), which is not a Pi session. */
-function fromPiGraphSession(options) {
-	return fromCodeAgentSession("pi-graph", options);
-}
 function fromCodeAgentSession(source, options) {
 	const entries = options.entries.filter(isRecord);
 	if (entries.length === 0) return {
@@ -2974,197 +2869,6 @@ function finiteScore(value) {
 	return value !== void 0 && Number.isFinite(value) ? clamp01(value) : void 0;
 }
 //#endregion
-//#region src/contract/intake/code-agent-store.ts
-const SESSION_SUFFIX = ".jsonl";
-const REJECTION_SAMPLE_LIMIT = 8;
-/**
-* Read every session this run captured under `scope.root` and project them.
-*
-* A store directory that does not exist, or that holds no session the scope attributes, returns
-* `status: 'unavailable'`. That is not an error: a run whose harness store was never synced back
-* has no evidence, and saying so is the correct answer.
-*/
-async function observeCodeAgentStore(scope) {
-	if (!isAbsolute(scope.root)) throw new TypeError(`observeCodeAgentStore: scope.root must be an absolute path to this run's own store, received ${JSON.stringify(scope.root)}`);
-	const workspaceRoot = scope.workspaceRoot === void 0 ? void 0 : resolve(scope.workspaceRoot);
-	const paths = await listSessionFiles(resolve(scope.root));
-	const sessions = [];
-	const byReason = /* @__PURE__ */ new Map();
-	const sample = [];
-	const reject = (path, reason) => {
-		byReason.set(reason, (byReason.get(reason) ?? 0) + 1);
-		if (sample.length < REJECTION_SAMPLE_LIMIT) sample.push({
-			path,
-			reason
-		});
-	};
-	for (const path of paths) {
-		let entries;
-		try {
-			entries = (await parseCodeAgentJsonlFile(path)).entries;
-		} catch {
-			reject(path, "unreadable");
-			continue;
-		}
-		if (entries.length === 0) {
-			reject(path, "no-entries");
-			continue;
-		}
-		const workspace = sessionWorkspace(entries);
-		if (workspaceRoot !== void 0) {
-			if (workspace === void 0) {
-				reject(path, "no-workspace-recorded");
-				continue;
-			}
-			if (!isSameOrBelow(workspaceRoot, workspace)) {
-				reject(path, "outside-workspace");
-				continue;
-			}
-		}
-		const startedAtMs = sessionStartedAtMs(entries);
-		if (scope.window !== void 0) {
-			if (startedAtMs === void 0) {
-				reject(path, "no-start-recorded");
-				continue;
-			}
-			if (startedAtMs < scope.window.startedAtMs) {
-				reject(path, "before-run");
-				continue;
-			}
-			if (scope.window.finishedAtMs !== void 0 && startedAtMs > scope.window.finishedAtMs) {
-				reject(path, "after-run");
-				continue;
-			}
-		}
-		const observation = observeCodeAgentSession({
-			source: scope.source,
-			entries,
-			sourcePath: path
-		});
-		sessions.push({
-			path,
-			sessionId: observation.sessionId,
-			...workspace === void 0 ? {} : { workspace },
-			...startedAtMs === void 0 ? {} : { startedAtMs },
-			observation
-		});
-	}
-	const rejections = {
-		total: [...byReason.values()].reduce((sum, count) => sum + count, 0),
-		byReason: Object.fromEntries(byReason),
-		sample
-	};
-	if (sessions.length === 0) return {
-		status: "unavailable",
-		scope,
-		filesScanned: paths.length,
-		reason: unavailableReason(paths.length, rejections),
-		rejections
-	};
-	const bySurface = /* @__PURE__ */ new Map();
-	let actions = 0;
-	let subagentActions = 0;
-	let sessionsWithSubagentActions = 0;
-	for (const session of sessions) {
-		let sessionSubagents = 0;
-		for (const action of session.observation.actions) {
-			actions += 1;
-			bySurface.set(action.surface, (bySurface.get(action.surface) ?? 0) + 1);
-			if (action.surface === "subagent") sessionSubagents += 1;
-		}
-		subagentActions += sessionSubagents;
-		if (sessionSubagents > 0) sessionsWithSubagentActions += 1;
-	}
-	return {
-		status: "observed",
-		scope,
-		filesScanned: paths.length,
-		sessions,
-		actions,
-		subagentActions,
-		sessionsWithSubagentActions,
-		bySurface: Object.fromEntries(bySurface),
-		rejections
-	};
-}
-function unavailableReason(filesScanned, rejections) {
-	if (filesScanned === 0) return "the run captured no session files in this store";
-	return `no session in this store belongs to the run: ${filesScanned} file(s) scanned, all rejected (${Object.entries(rejections.byReason).map(([reason, count]) => `${count} ${reason}`).sort().join(", ")})`;
-}
-async function listSessionFiles(root) {
-	const found = [];
-	const walk = async (dir) => {
-		let entries;
-		try {
-			entries = await readdir(dir, { withFileTypes: true });
-		} catch {
-			return;
-		}
-		for (const entry of entries) {
-			const path = join(dir, entry.name);
-			if (entry.isDirectory()) await walk(path);
-			else if (entry.isFile() && entry.name.endsWith(SESSION_SUFFIX)) found.push(path);
-		}
-	};
-	await walk(root);
-	if (found.length === 0 && await isSessionFile(root)) found.push(root);
-	return found.sort();
-}
-async function isSessionFile(path) {
-	if (!path.endsWith(SESSION_SUFFIX)) return false;
-	try {
-		return (await stat(path)).isFile();
-	} catch {
-		return false;
-	}
-}
-/**
-* The working directory the session RECORDED, from the first entry that states one.
-*
-* codex writes it on `session_meta.payload.cwd`; claude-code writes `cwd` on every line. Neither is
-* inferred from the file's own location: where a captured file was copied to says nothing about
-* where the agent ran.
-*/
-function sessionWorkspace(entries) {
-	for (const entry of entries) {
-		const record = plainRecord(entry);
-		if (record === void 0) continue;
-		const direct = stringOf(record.cwd);
-		if (direct !== void 0) return direct;
-		const nested = stringOf(plainRecord(record.payload)?.cwd);
-		if (nested !== void 0) return nested;
-	}
-}
-/** When the session itself says it started, from the first entry carrying a parseable timestamp. */
-function sessionStartedAtMs(entries) {
-	for (const entry of entries) {
-		const record = plainRecord(entry);
-		if (record === void 0) continue;
-		const payload = plainRecord(record.payload);
-		for (const candidate of [payload?.timestamp, record.timestamp]) {
-			const text = stringOf(candidate);
-			if (text === void 0) continue;
-			const parsed = Date.parse(text);
-			if (Number.isFinite(parsed)) return parsed;
-		}
-		const epochMs = numberOf(record.timestampMs) ?? numberOf(payload?.timestampMs);
-		if (epochMs !== void 0) return epochMs;
-	}
-}
-function isSameOrBelow(root, candidate) {
-	const rel = relative(root, resolve(candidate));
-	return rel === "" || rel !== ".." && !rel.startsWith(`..${sep}`) && !isAbsolute(rel);
-}
-function plainRecord(value) {
-	return value !== null && typeof value === "object" && !Array.isArray(value) ? value : void 0;
-}
-function stringOf(value) {
-	return typeof value === "string" && value.length > 0 ? value : void 0;
-}
-function numberOf(value) {
-	return typeof value === "number" && Number.isFinite(value) ? value : void 0;
-}
-//#endregion
 //#region src/contract/intake/feedback-table.ts
 function fromFeedbackTable(opts) {
 	const { ratings, meta = [], scale, emitRaterScores = true } = opts;
@@ -3244,6 +2948,6 @@ function fromFeedbackTable(opts) {
 	};
 }
 //#endregion
-export { FileSystemOutcomeStore, InMemoryOutcomeStore, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, SelfImproveRunError, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiGraphSession, fromPiSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, observeCodeAgentStore, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };
+export { FileSystemOutcomeStore, InMemoryOutcomeStore, SelfImproveRunError, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, defaultProductionGate, defineAgentEval, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, makeProposalFinding, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, parseCodeAgentJsonlFile, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, streamCodeAgentJsonlFile, summarizeExecution, transientDispatchFailure, verifyAgentProfileImprovementExperimentComparison, verifyCandidateExperiment, verifyCandidateExperimentComparison };
 
 //# sourceMappingURL=index.js.map