@tangle-network/agent-eval 0.140.1 → 0.141.0

package/dist/index.js

@@ -6,18 +6,18 @@ import { a as estimateTokens, i as estimateCost, n as MetricsCollector, o as isM
 import { a as CostLedgerPersistenceError, c as costForTokenPricing, i as CostLedger, l as costForUsage, n as CostCallConflictError, o as CostReceiptCaptureError, r as CostCeilingReachedError, s as CostReservationExceededError, t as CostAccountingIncompleteError, u as modelPriceKey } from "./cost-ledger-DMFxsLKr.js";
 import { a as providerFromBaseUrl, i as defaultProviderRedactor, n as InMemoryRawProviderSink, r as NoopRawProviderSink, t as FileSystemRawProviderSink } from "./raw-provider-sink-BQd7mzyT.js";
 import { a as assertLlmRoute, c as callLlmJson, d as isTransientLlmError, f as maximumChargeForLlmRequest, i as LlmRouteAssertionError, l as costReceiptFromLlm, m as stripFencedJson, n as LlmClient, o as backoffMs, p as probeLlm, r as LlmResponseError, s as callLlm, t as LlmCallError, u as costReceiptFromLlmError } from "./llm-client-B3WXSH5Y.js";
-import { C as createChatClient, S as computeTraceMetrics, _ as CONTROL_INTEGRITY_ANALYST, a as analystFindingDigest, c as completedAnalystReviewQuality, d as validateAnalystReviewDecisions, f as DEFAULT_TRACE_ANALYST_KINDS, g as FAILURE_MODE_KIND_SPEC, h as IMPROVEMENT_KIND_SPEC, i as assertExactRegistryRunOpts, l as readAnalystReview, m as KNOWLEDGE_GAP_KIND_SPEC, n as AnalystRegistry, o as analystRunDigest, p as KNOWLEDGE_POISONING_KIND_SPEC, r as ExactAnalystRunExecutionError, s as assertUniqueFindingIds, t as buildDefaultAnalystRegistry, u as snapshotAnalystRun, v as ControlIntegrityAnalyst, y as emitControlIntegrityFindings } from "./default-registry-BOiBXNFh.js";
+import { C as createChatClient, S as computeTraceMetrics, _ as CONTROL_INTEGRITY_ANALYST, a as analystFindingDigest, c as completedAnalystReviewQuality, d as validateAnalystReviewDecisions, f as DEFAULT_TRACE_ANALYST_KINDS, g as FAILURE_MODE_KIND_SPEC, h as IMPROVEMENT_KIND_SPEC, i as assertExactRegistryRunOpts, l as readAnalystReview, m as KNOWLEDGE_GAP_KIND_SPEC, n as AnalystRegistry, o as analystRunDigest, p as KNOWLEDGE_POISONING_KIND_SPEC, r as ExactAnalystRunExecutionError, s as assertUniqueFindingIds, t as buildDefaultAnalystRegistry, u as snapshotAnalystRun, v as ControlIntegrityAnalyst, y as emitControlIntegrityFindings } from "./default-registry-BupFSXSN.js";
 import { INPUT_VALUE, LLM_CACHED_TOKENS, LLM_CACHED_TOKEN_ATTR_KEYS, LLM_CACHE_WRITE_TOKENS, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, LLM_CONTEXT_TOKENS, LLM_COST_ATTR_KEYS, LLM_COST_USD, LLM_INPUT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_MODEL_ATTR_KEYS, LLM_MODEL_NAME, LLM_OUTPUT_TOKENS, LLM_OUTPUT_TOKEN_ATTR_KEYS, LLM_REASONING_TOKENS, LLM_REASONING_TOKEN_ATTR_KEYS, OPENINFERENCE_SPAN_KIND, OUTPUT_VALUE, RUN_COST_ATTR_KEYS, SPAN_KIND_ATTR_KEYS, TOOL_ARGS_CAPTURED, TOOL_LATENCY_MS, TOOL_NAME, TOOL_NAME_ATTR_KEYS, applyLlmSpanOtlpAttributes, asNumber, contextInputTokens, firstNumberAttr } from "./trace-attributes.js";
-import { C as TraceNotFoundError, G as resolveTraceAnalystLimits, J as extractOtlpAttributes, K as asString, Q as readOtlpStatus, S as TraceFileTooLargeError, W as DEFAULT_TRACE_ANALYST_LIMITS, X as inferOtlpKind, Y as firstStringAttr, Z as projectOtlpFlatLine, _ as TraceAnalysisLimitError, b as TraceFileMalformedError, c as traceAnalystFunctionGroup, et as stringField, g as SpanNotFoundError, h as TRACE_ANALYSIS_LIMITS, i as runTraceAnalyst, it as traceSpanKindToOpenInferenceKind, l as createBoundedTraceAnalysisStore, m as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, n as renderPriorFindings, nt as classifyOtlpSpanRole, o as TRACE_ANALYST_TOOL_NAMESPACE, p as DEFAULT_TRACE_ANALYST_BUDGETS, r as renderUpstreamFindings, rt as isOtlpModelCall, s as buildTraceAnalysisToolDescriptors, t as createTraceAnalyst, tt as applyToolSpanOtlpAttributes, v as TraceAnalysisStoreContractError, x as TraceFileMissingError, y as TraceAnalysisValidationError } from "./kind-factory-DB7nIs35.js";
+import { C as TraceNotFoundError, G as resolveTraceAnalystLimits, J as extractOtlpAttributes, K as asString, Q as readOtlpStatus, S as TraceFileTooLargeError, W as DEFAULT_TRACE_ANALYST_LIMITS, X as inferOtlpKind, Y as firstStringAttr, Z as projectOtlpFlatLine, _ as TraceAnalysisLimitError, b as TraceFileMalformedError, c as traceAnalystFunctionGroup, et as stringField, g as SpanNotFoundError, h as TRACE_ANALYSIS_LIMITS, i as runTraceAnalyst, it as traceSpanKindToOpenInferenceKind, l as createBoundedTraceAnalysisStore, m as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, n as renderPriorFindings, nt as classifyOtlpSpanRole, o as TRACE_ANALYST_TOOL_NAMESPACE, p as DEFAULT_TRACE_ANALYST_BUDGETS, r as renderUpstreamFindings, rt as isOtlpModelCall, s as buildTraceAnalysisToolDescriptors, t as createTraceAnalyst, tt as applyToolSpanOtlpAttributes, v as TraceAnalysisStoreContractError, x as TraceFileMissingError, y as TraceAnalysisValidationError } from "./kind-factory-Bvwe3pup.js";
 import { a as computeFindingId, o as makeFinding, s as makeProposalFinding } from "./usage-receipt-CgxMEBZq.js";
 import { c as SUPERVISOR_RUN_INTEGRITY_SCHEMA, n as supervisorRunRolloutLines, t as analyzeSupervisorRunIntegrity } from "./integrity-hruKe6oe.js";
 import { c as assertMintedLines, f as isRolloutLine, i as ROLLOUT_SCHEMA, m as validateRolloutLine, p as isTrainableSplit, s as assertMinted, u as assertRolloutLine } from "./schema-C6DW4ZHR.js";
 import { a as scoreOrigin, n as observedScore, o as trainingReward, r as observedSplitScore, s as trainingScore, t as isRealnessGated } from "./reward-nw2xZGZG.js";
 import { a as clamp01, i as aggregateRunScore, r as DEFAULT_RUN_SCORE_WEIGHTS } from "./proposal-findings-2GIUo1et.js";
 import { a as RunCritic, d as defaultIsMaterial, f as diffFindings, h as defineTraceAnalyst, i as runSemanticConceptJudge, n as SEMANTIC_CONCEPT_JUDGE_VERSION, o as SKILL_USAGE_ANALYST, p as LockedJsonlAppender, r as createSemanticConceptJudge, s as SkillUsageAnalyst, t as DEFAULT_COMPLEXITY_WEIGHTS, u as FindingsStore } from "./semantic-concept-judge-DJQtFr95.js";
-import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-Bkak4nzo.js";
+import { t as createDspyRlmTraceEngine } from "./dspy-rlm-engine-Bw8Zz9CX.js";
 import { f as Mutex, p as mapConcurrent } from "./ledger-core-Dxz0Rkwa.js";
-import { At as pairedDeltaTest, Ct as paretoFrontier, D as runCanaries, Dt as decidePairedPromotion, Ft as summarizeBackendIntegrity, G as buildReflectionPrompt, It as JudgeParseError, K as parseReflectionResponse, Mt as assertRealAgentReceipts, Nt as assertRealBackend, Ot as pairedDecisionShape, Pt as summarizeAgentReceiptIntegrity, St as dominates, Tt as scalarScore, W as DEFAULT_MUTATION_PRIMITIVES, _t as canonicalJson, at as createReferenceEquivalenceJudge, bt as inMemoryVerdictCache, ct as redTeamDataset, dt as toolNamesForRun, ft as Dataset, gt as cachedJudge, ht as llmJudge, it as REFERENCE_EQUIVALENCE_JUDGE_VERSION, jt as BackendIntegrityError, kt as minimumPairsForPairedDeltaTest, lt as redTeamReport, mt as hashScenarios, ot as runReferenceEquivalenceJudge, pt as HoldoutLockedError, rt as REFERENCE_EQUIVALENCE_INPUT_LIMITS, st as DEFAULT_RED_TEAM_CORPUS, ut as scoreRedTeamOutput, vt as contentHash, wt as paretoFrontierWithCrowding, xt as crowdingDistance, yt as fileVerdictCache, z as surfaceContentHash } from "./skillopt-optimization-method-CSBQ8Qma.js";
+import { At as pairedDeltaTest, Ct as paretoFrontier, D as runCanaries, Dt as decidePairedPromotion, Ft as summarizeBackendIntegrity, G as buildReflectionPrompt, It as JudgeParseError, K as parseReflectionResponse, Mt as assertRealAgentReceipts, Nt as assertRealBackend, Ot as pairedDecisionShape, Pt as summarizeAgentReceiptIntegrity, St as dominates, Tt as scalarScore, W as DEFAULT_MUTATION_PRIMITIVES, _t as canonicalJson, at as createReferenceEquivalenceJudge, bt as inMemoryVerdictCache, ct as redTeamDataset, dt as toolNamesForRun, ft as Dataset, gt as cachedJudge, ht as llmJudge, it as REFERENCE_EQUIVALENCE_JUDGE_VERSION, jt as BackendIntegrityError, kt as minimumPairsForPairedDeltaTest, lt as redTeamReport, mt as hashScenarios, ot as runReferenceEquivalenceJudge, pt as HoldoutLockedError, rt as REFERENCE_EQUIVALENCE_INPUT_LIMITS, st as DEFAULT_RED_TEAM_CORPUS, ut as scoreRedTeamOutput, vt as contentHash, wt as paretoFrontierWithCrowding, xt as crowdingDistance, yt as fileVerdictCache, z as surfaceContentHash } from "./skillopt-optimization-method-DV6c4ooy.js";
 import { $ as selfPreference, A as pairedRiskDifference, B as requiredSampleSize, C as mulberry32, D as pairedCohensDz, E as pairedBootstrap, F as partialCredit, G as wilson, H as weightedComposite, I as passAtK, J as normalCdf, K as studentTCdf, L as pearsonR, M as pairedRiskDifferenceScore, N as pairedSignTest, O as pairedDeltaTieFraction, P as pairedTTest, Q as positionalBias, R as ranks, S as mcnemarRequiredN, T as pairedBinaryScale, U as weightedMean, V as spearmanR, W as wilcoxonSignedRank, X as calibrateJudgeContinuous, Y as calibrateJudge, Z as continuousAgreement, _ as interpretCliffs, a as MANN_WHITNEY_EXACT_MAX_WORK, b as mcnemar, c as bonferroni, d as confidenceInterval, et as verbosityBias, f as corpusInterRaterAgreement, g as interRaterReliability, h as holm, i as MANN_WHITNEY_EXACT_MAX_STATES, j as pairedRiskDifferenceExact, k as pairedMde, l as cliffsDelta, m as eProcess, n as DECISION_PAIRED_DELTA_STATISTIC, o as WILCOXON_EXACT_MAX_N, p as corpusInterRaterAgreementFromJudgeScores, r as DEFAULT_PERMUTATIONS, s as benjaminiHochberg, t as BOOTSTRAP_GATE_MIN_N, u as cohensD, v as isBinaryOutcomeVector, w as normalizeScores, x as mcnemarPower, y as mannWhitneyU, z as requiredPairedSampleSize } from "./statistics-ByxzSiOM.js";
 import { n as pairArms, r as pairRunRecords, t as comparePairedArms } from "./paired-arms-iZ08VFMN.js";
 import { n as llmSpanFromProvider, t as TraceEmitter } from "./emitter-CPBAhxum.js";
@@ -29,10 +29,10 @@ import { a as toHarborTrajectories, i as relabelImportedSplit, n as HARBOR_IMPOR
 import { t as buildTrajectory } from "./trajectory-D_7rLrvE.js";
 import { n as unmintableReasons, t as mintRolloutRows } from "./mint-Ctwk079K.js";
 import { C as SUPERVISOR_RUN_SCHEMA, T as showMeasured, _ as readClaudeCodeSupervisorRun, b as rollupSupervisorRuns, c as writeSupervisorRunReport, d as renderSupervisorRunHeadline, f as renderSupervisorRunMarkdown, g as claudeCodeSupervisorRunReader, t as analyzeSupervisorRun, v as analyzeSupervisorRunSources, w as isUnavailable } from "./supervisor-run-BEns49Nx.js";
-import { A as TRACE_ANALYST_ACTOR_DESCRIPTION, C as domainEvidencePattern, D as tokenizeDomainWords, E as scoreTraceInsightReadiness, O as traceAnalystOnRunComplete, S as describeTraceInsightScope, T as planTraceInsightQuestions, _ as otlpToTraceRunRecords, a as convertTraceStoresToOtlp, b as buildTraceInsightPrompt, c as otelRunCompleteHook, d as captureFetchToRawSink, f as ToolTraceMissingError, g as otlpToRunRecords, h as otlpRowsToTraceRunRecords, i as iterateRawCalls, j as TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, k as analyzeTraces, l as OTEL_AGENT_EVAL_SCOPE, m as otlpRowsToRunRecords, n as ReplayCacheMissError, o as createOtelExporter, p as toolSpansToTraceAnalysisStore, r as createReplayFetch, s as createOtelTracingStore, t as ReplayCache, u as exportRunAsOtlp, v as flattenOtlpExportToNdjson, w as inferDomainKeywords, x as defaultTraceInsightPanel, y as buildTraceInsightContext } from "./replay-k2MsOmv5.js";
-import { i as otlpTextToTraceAnalysisStore, n as OtlpFileTraceStore } from "./store-otlp-vRByAR6h.js";
-import { n as extractUsageFromResponse, r as extractUsageFromSse, t as extractUsage } from "./extract-usage-C5vMw-0R.js";
-import { B as agentProfileModelId, G as createLlmCorrectnessChecker, H as harnessAxisOf, I as CODING_HARNESSES, J as verifyCompletion, K as createTokenRecallChecker, L as HARNESS_NATIVE_MODEL, R as agentProfileHash, U as extractProducedState, V as expandProfileAxes, W as completionVerdict, q as parseCorrectnessResponse, z as agentProfileId } from "./campaign-BzjYNYVZ.js";
+import { A as TRACE_ANALYST_ACTOR_DESCRIPTION, C as domainEvidencePattern, D as tokenizeDomainWords, E as scoreTraceInsightReadiness, O as traceAnalystOnRunComplete, S as describeTraceInsightScope, T as planTraceInsightQuestions, _ as otlpToTraceRunRecords, a as convertTraceStoresToOtlp, b as buildTraceInsightPrompt, c as otelRunCompleteHook, d as captureFetchToRawSink, f as ToolTraceMissingError, g as otlpToRunRecords, h as otlpRowsToTraceRunRecords, i as iterateRawCalls, j as TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, k as analyzeTraces, l as OTEL_AGENT_EVAL_SCOPE, m as otlpRowsToRunRecords, n as ReplayCacheMissError, o as createOtelExporter, p as toolSpansToTraceAnalysisStore, r as createReplayFetch, s as createOtelTracingStore, t as ReplayCache, u as exportRunAsOtlp, v as flattenOtlpExportToNdjson, w as inferDomainKeywords, x as defaultTraceInsightPanel, y as buildTraceInsightContext } from "./replay-DjUfTrHD.js";
+import { i as otlpTextToTraceAnalysisStore, n as OtlpFileTraceStore } from "./store-otlp-D4I90_vR.js";
+import { n as extractUsageFromResponse, r as extractUsageFromSse, t as extractUsage } from "./extract-usage-7l1Xq5ti.js";
+import { B as agentProfileModelId, G as createLlmCorrectnessChecker, H as harnessAxisOf, I as CODING_HARNESSES, J as verifyCompletion, K as createTokenRecallChecker, L as HARNESS_NATIVE_MODEL, R as agentProfileHash, U as extractProducedState, V as expandProfileAxes, W as completionVerdict, q as parseCorrectnessResponse, z as agentProfileId } from "./campaign-COFTVy3Q.js";
 import { n as iqr, r as welchsTTest, t as compareToBaseline } from "./baseline-C-GocmIW.js";
 import { a as judgeSpans, c as runsForScenario, i as hasCapturedToolArgs, l as toolSpans, n as argHash, o as llmSpans, r as groupBy, s as runFailureClass, t as aggregateLlm } from "./query-Di7eEQ79.js";
 import { a as checkBehavioralCanary, i as canaryLeakView, o as checkCanaries, r as HoldoutAuditor, s as runBehavioralCanaries, t as analyzeRuns } from "./analyze-runs-BScZvqMV.js";
@@ -42,7 +42,7 @@ import { a as InMemoryTraceStore, i as FileSystemTraceStore, n as assertRunCaptu
 import { i as redactValue, n as REDACTION_VERSION, r as redactString, t as DEFAULT_REDACTION_RULES } from "./redact-7Aq1ukl-.js";
 import { n as DEFAULT_RULES, r as classifyFailure, t as computeToolUseMetrics } from "./tool-use-metrics-DEGMKycK.js";
 import { t as analyzeSeries } from "./series-convergence-CjO2QdRW.js";
-import { _ as deterministicSplit, g as BENCHMARK_SPLIT_SEED, t as benchmarks_exports } from "./benchmarks-HwoBE32G.js";
+import { _ as deterministicSplit, g as BENCHMARK_SPLIT_SEED, t as benchmarks_exports } from "./benchmarks-Ce8GdJbc.js";
 import { t as runEvalCampaign } from "./eval-campaign-YdkpWWoT.js";
 import { n as pairedEvalueSequence, t as evaluateInterimReleaseConfidence } from "./sequential-Br0mAPHA.js";
 import { accessSync, appendFileSync, constants, cpSync, existsSync, mkdirSync, promises, readFileSync, readdirSync, statSync, writeFileSync } from "node:fs";
@@ -831,19 +831,42 @@ function assertCapabilityHeadroom(result, opts = {}) {
 	}
 }
 //#endregion
+//#region src/deprecation.ts
+/**
+* One-shot deprecation warnings. A deprecated surface warns the FIRST time it
+* is used in a process — loud enough to be seen, quiet enough not to flood a
+* matrix run that constructs it per cell.
+*/
+const warned = /* @__PURE__ */ new Set();
+/** Emit `message` on console.warn once per process per `key`. */
+function warnDeprecatedOnce(key, message) {
+	if (warned.has(key)) return;
+	warned.add(key);
+	console.warn(`[agent-eval deprecation] ${message}`);
+}
+/** Test seam: forget prior warnings so once-semantics can be asserted. */
+function resetDeprecationWarnings() {
+	warned.clear();
+}
+//#endregion
 //#region src/client.ts
+const DEFAULT_REQUEST_TIMEOUT_MS = 3e4;
 /**
 * ProductClient — configurable HTTP client for exercising any agent's APIs.
 *
 * Routes are config, not hardcoded. Each agent provides its own RouteMap.
+*
+* @deprecated A one-product REST client does not belong in the generic
+* substrate. Moves to the product repo in the next major
+* (tangle-network/agent-runtime#694); vendor it if you construct it directly.
 */
-const DEFAULT_REQUEST_TIMEOUT_MS = 3e4;
 var ProductClient = class {
 	baseUrl;
 	routes;
 	cookies = "";
 	timeoutMs;
 	constructor(config) {
+		warnDeprecatedOnce("ProductClient", "ProductClient is deprecated and moves to the product repo in the next major (tangle-network/agent-runtime#694).");
 		this.baseUrl = config.baseUrl.replace(/\/+$/, "");
 		this.routes = config.routes;
 		this.timeoutMs = config.timeoutMs ?? DEFAULT_REQUEST_TIMEOUT_MS;
@@ -1373,6 +1396,11 @@ const RIGOR_STANCE = {
 * Uses a driver LLM (Claude/GPT-4o) to decide what to say each turn.
 * Not scripted — the driver gets the current product state and decides
 * the next realistic user message.
+*
+* @deprecated A one-product driver welded to `ProductClient` does not belong
+* in the generic substrate. Moves to the product repo in the next major
+* (tangle-network/agent-runtime#694). A driver is an `AgentProfile` on a
+* graph edge; use the agent-graph path once it lands, or vendor this class.
 */
 var AgentDriver = class {
 	chat;
@@ -1381,6 +1409,7 @@ var AgentDriver = class {
 	productContext;
 	costLedger;
 	constructor(chat, config) {
+		warnDeprecatedOnce("AgentDriver", "AgentDriver is deprecated and moves to the product repo in the next major (tangle-network/agent-runtime#694).");
 		this.chat = chat;
 		this.client = config.client;
 		this.driverModel = config.driverModel ?? "claude-sonnet-4-6";
@@ -1499,8 +1528,13 @@ function describeCompletion(persona, state) {
 * signs off (DONE) only when a real practitioner would act on the work
 * unmodified. Pure function of persona, product state, and product context
 * — exported so harness authors can inspect and regression-test it.
+*
+* @deprecated A role expressed as a code function can never be optimized.
+* Removal is a major; until then, treat this output as SEED data for a
+* registry-backed directive prompt (tangle-network/agent-runtime#694).
 */
 function buildDriverSystemPrompt(persona, state, productContext = "") {
+	warnDeprecatedOnce("buildDriverSystemPrompt", "buildDriverSystemPrompt is deprecated: roles belong in registry-backed prompt data, not code (tangle-network/agent-runtime#694).");
 	const rigor = persona.rigor ?? "demanding";
 	const expertise = persona.expertise ? ` You are ${persona.expertise}.` : "";
 	const pressure = persona.pressurePoints && persona.pressurePoints.length > 0 ? `\nA competent ${persona.role} here MUST get the agent to address each of:\n${persona.pressurePoints.map((p) => `  - ${p}`).join("\n")}\nDo NOT hand these to the agent. Probe whether it surfaces them itself. If it misses one, press on exactly that gap until it delivers or demonstrably fails.\n` : "";
@@ -1529,48 +1563,18 @@ Sign-off: respond with exactly "DONE" only when a ${persona.role} would act on t
 Output ONLY your next message to the agent — in character, first person, no meta-commentary, no stage directions.`;
 }
 /**
-* Build the WORKER-DRIVER system prompt — the harness-aware sibling of
-* `buildDriverSystemPrompt`. Where that one role-plays a demanding *user* of a
-* product, this one is a meta-agent that DRIVES a capable coding *worker* to
-* complete a goal: it writes rich, high-signal instructions that direct the
-* worker to exploit its harness's full power (parallelize, sub-agents, run-to-
-* completion, tools/MCP), matched to what THAT harness can actually do. The
-* load-bearing contract: the driver never writes a thin steer — every message
-* is the dense, specific directive a world-class engineering lead would write.
-* Pure function; exported so harness authors can inspect and regression-test it.
-*/
-function buildWorkerDriverSystemPrompt(ctx) {
-	const harness = ctx.harness ?? "a coding agent";
-	const brief = ctx.harnessBrief ? `\nThis worker's harness (${harness}) can:\n${ctx.harnessBrief}\nDrive it to USE these — and never ask for a capability it lacks.\n` : `\nThe worker runs in ${harness}. Drive it to exploit whatever its harness offers — parallel tool calls, sub-agents, run-to-completion — and never ask for a capability it lacks.\n`;
-	const progress = ctx.progress ? `\nWhat the worker has done so far:\n${ctx.progress}\n` : "\nThe worker has not started yet — your first instruction sets the whole plan.\n";
-	const context = ctx.context ? `\nContext:\n${ctx.context}\n` : "";
-	return `You are a DRIVER: a meta-agent whose entire job is to drive a capable coding worker to ACHIEVE A GOAL by writing it precise, high-signal instructions. You do not do the work yourself — you direct the worker to do it, brilliantly, and you hold it to a standard it would not hold itself to.
-
-The goal: ${ctx.goal}
-${context}${brief}${progress}
-THE BAR — non-negotiable. Every instruction you write is dense, specific, and complete. You write the message a world-class engineering lead writes to a strong report: the exact next objective, the concrete sub-steps, what to do in parallel vs in sequence, what to verify and how, what "done" looks like, and the failure modes to avoid. A thin steer — "try again", "fix the issues", a single vague sentence — is a failure on YOUR part, not the worker's. Out-drive a human power-user.
-
-HOW to drive (each turn):
-1. Read what the worker actually did (its trace/state above), not what it claims. Find the real gap between here and the goal.
-2. Decide the next objective — the largest correct step the worker can take now.
-3. Decompose it: name the sub-tasks that can run IN PARALLEL (independent files/checks/searches) and tell the worker to fan them out (sub-agents / parallel tool calls where its harness allows); name what must run in SEQUENCE and why.
-4. Make it exploit the harness: drive it to run to completion under its own autonomy, spawn sub-agents for separable work, use its tools/MCP, and not stop at the first plausible stopping point.
-5. Specify verification: the exact command / test / check that proves the step landed, and tell it to run that and report the result — never accept "done" without the check.
-6. Name the traps: the specific failure modes for THIS step (editing the wrong file, a check that passes vacuously, scope creep) and forbid them.
-7. As the task gets harder, decompose MORE, not less — more parallel branches, deeper sub-agent trees, tighter verification.
-
-COMPLETION: drive toward the goal's real acceptance check. Do not declare done — the deliverable's checker does. If the worker claims it is done, drive it to PROVE it with the check; if the check fails, drive the fix.
-
-Output ONLY your next instruction to the worker — direct, detailed, actionable, in the first person as the driver. No meta-commentary, no preamble.`;
-}
-/**
 * Decide the simulated user's next turn — the reactive, adversarial
 * turn-generation core of `AgentDriver`, exposed standalone so an in-process
 * eval harness can drive multi-shot conversations without the `ProductClient`
 * workspace machinery. Returns the next user message, or the literal "DONE"
 * when the simulated professional would sign off.
+*
+* @deprecated The persona-driver loop becomes a 2-node agent graph (driver
+* profile + delegates edge); removal is a major
+* (tangle-network/agent-runtime#694).
 */
 async function decideNextUserTurn(chat, opts) {
+	warnDeprecatedOnce("decideNextUserTurn", "decideNextUserTurn is deprecated: the persona-driver loop becomes a 2-node agent graph (tangle-network/agent-runtime#694).");
 	const { persona, state, history, productContext = "", model = "claude-sonnet-4-6" } = opts;
 	const lastResponse = history.length > 0 ? history[history.length - 1].content.slice(0, 2e3) : "(no conversation yet — this is the first message)";
 	const recentHistory = history.slice(-6).map((h) => `${h.role}: ${h.content.slice(0, 500)}`).join("\n\n");
@@ -3139,6 +3143,78 @@ function classifyTreatment(_record, gate) {
 	return gate.gated ? "treatment-not-applied" : "measurable";
 }
 //#endregion
+//#region src/worker-driver-seed.ts
+/**
+* Worker-driver seed DATA — the knowledge that used to be hardcoded inside
+* `buildWorkerDriverSystemPrompt` (deleted; zero callers), re-expressed as
+* plain data so a driver PROFILE can seed its directive prompt from it and an
+* optimizer (GEPA/DSPy) can then improve on it. A role expressed as a code
+* function can never improve; a role expressed as versionable prompt data can
+* (tangle-network/agent-runtime#694).
+*
+* Two exports:
+*   - `WORKER_DRIVER_DOCTRINE` — the "never write a thin steer" driving
+*     contract: the bar for every instruction a driver sends a worker.
+*   - `HARNESS_BRIEFS` — per-harness capability + caveat briefs the driver
+*     weaves into its instructions so it drives the worker to exploit what
+*     THAT harness can actually do, and never asks for a capability it lacks.
+*
+* Both are seeds, not law: a consumer copies them into its directive prompt
+* registry (e.g. `prompts/<surface>/v0.md`) and optimizes from there.
+*/
+/**
+* The driving contract a worker-driver directive seeds from. Every message the
+* driver writes is held to this bar; a thin steer is the driver's failure, not
+* the worker's.
+*/
+const WORKER_DRIVER_DOCTRINE = `You are a DRIVER: a meta-agent whose entire job is to drive a capable coding worker to ACHIEVE A GOAL by writing it precise, high-signal instructions. You do not do the work yourself — you direct the worker to do it, brilliantly, and you hold it to a standard it would not hold itself to.
+
+THE BAR — non-negotiable. Every instruction you write is dense, specific, and complete. You write the message a world-class engineering lead writes to a strong report: the exact next objective, the concrete sub-steps, what to do in parallel vs in sequence, what to verify and how, what "done" looks like, and the failure modes to avoid. A thin steer — "try again", "fix the issues", a single vague sentence — is a failure on YOUR part, not the worker's. Out-drive a human power-user.
+
+HOW to drive (each turn):
+1. Read what the worker actually did (its trace/state), not what it claims. Find the real gap between here and the goal.
+2. Decide the next objective — the largest correct step the worker can take now.
+3. Decompose it: name the sub-tasks that can run IN PARALLEL (independent files/checks/searches) and tell the worker to fan them out (sub-agents / parallel tool calls where its harness allows); name what must run in SEQUENCE and why.
+4. Make it exploit the harness: drive it to run to completion under its own autonomy, spawn sub-agents for separable work, use its tools/MCP, and not stop at the first plausible stopping point.
+5. Specify verification: the exact command / test / check that proves the step landed, and tell it to run that and report the result — never accept "done" without the check.
+6. Name the traps: the specific failure modes for THIS step (editing the wrong file, a check that passes vacuously, scope creep) and forbid them.
+7. As the task gets harder, decompose MORE, not less — more parallel branches, deeper sub-agent trees, tighter verification.
+
+COMPLETION: drive toward the goal's real acceptance check. Do not declare done — the deliverable's checker does. If the worker claims it is done, drive it to PROVE it with the check; if the check fails, drive the fix.
+
+Output ONLY your next instruction to the worker — direct, detailed, actionable, in the first person as the driver. No meta-commentary, no preamble.`;
+/**
+* Capability + caveat briefs per harness, keyed by harness id. Free text so
+* the substrate stays decoupled from any runtime harness type — the keys are
+* the conventional harness ids, but the record accepts any string so a
+* consumer can extend it with its own.
+*/
+const HARNESS_BRIEFS = {
+	"claude-code": [
+		"- parallel tool calls in a single turn (batch independent reads/greps/commands)",
+		"- parallel Task sub-agents (~10 concurrent) for separable investigation or bulk work; sub-agents cannot spawn sub-agents",
+		"- native WebSearch/WebFetch plus MCP servers when configured",
+		"- skills, hooks, and slash commands when installed in the workspace",
+		"- runs to completion under its own autonomy; drive it NOT to stop at the first plausible stopping point"
+	].join("\n"),
+	codex: [
+		"- long-horizon autonomous runs (strong run-to-completion; suited to hours-long missions)",
+		"- sandboxed exec with full shell access; MCP servers when configured",
+		"- no native sub-agent fan-out: express parallelism as explicit backgrounded commands or sequential decomposition",
+		"- structured output via --output-schema when a machine-parsed result is needed"
+	].join("\n"),
+	opencode: [
+		"- router-backed: any model the router serves; model choice is a live lever",
+		"- tool calls and MCP servers when configured",
+		"- no native sub-agent tree; decompose into sequential objectives with explicit verification between them"
+	].join("\n"),
+	"router-tools": [
+		"- a plain tool loop over router chat completions: no autonomy, no sub-agents, no workspace",
+		"- every capability must be driven explicitly, one tool call at a time",
+		"- keep objectives small and verification immediate; never assume it will continue on its own"
+	].join("\n")
+};
+//#endregion
 //#region src/anti-slop.ts
 const DEFAULT_HEDGES = [
 	/\bi\s+could\s+be\s+wrong\b/i,
@@ -12387,6 +12463,6 @@ function assertProductBenchmarkRun(runDir) {
 	return report;
 }
 //#endregion
-export { AGENT_PROFILE_KINDS, ATIF_SCHEMA_VERSION, ATTESTATION_ALGORITHM, AgentDriver, AgentEvalError, AgentProfileCellValidationError, AnalystRegistry, BENCHMARK_SPLIT_SEED, BOOTSTRAP_GATE_MIN_N, BackendIntegrityError, BenchmarkRunner, BudgetBreachError, BudgetGuard, CODING_HARNESSES, CONTROL_INTEGRITY_ANALYST, CallbackResearcher, CaptureIntegrityError, ConfigError, ControlIntegrityAnalyst, ConvergenceTracker, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, CostLedger, CostLedgerPersistenceError, CostReceiptCaptureError, CostReservationExceededError, CostTracker, CrossFamilyError, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_AGENT_SLOS, DEFAULT_COMPLEXITY_WEIGHTS, DEFAULT_RULES as DEFAULT_FAILURE_RULES, DEFAULT_FINDERS, DEFAULT_HARNESS_OBJECTIVES, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_PERMUTATIONS, DEFAULT_PR_REVIEW_SCORE_WEIGHTS, DEFAULT_REDACTION_RULES, DEFAULT_RED_TEAM_CORPUS, DEFAULT_RUN_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, DEFAULT_TRACE_ANALYST_LIMITS, Dataset, DescriptionLengthGate, DockerSandboxDriver, DualAgentBench, ERROR_COUNT_PATTERNS, EvalTraceStore, ExactAnalystRunExecutionError, ExperimentTracker, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, FindingsStore, HARBOR_IMPORT_GAP, HARNESS_NATIVE_MODEL, HeldOutGate, HoldoutAuditor, HoldoutLockedError, IMPROVEMENT_KIND_SPEC, INPUT_VALUE, INTENT_MATCH_JUDGE_VERSION, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, InMemoryWorkspaceInspector, JudgeError, JudgeParseError, JudgeRunner, KNOWLEDGE_GAP_KIND_SPEC, KNOWLEDGE_POISONING_KIND_SPEC, LLM_CACHED_TOKENS, LLM_CACHED_TOKEN_ATTR_KEYS, LLM_CACHE_WRITE_TOKENS, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, LLM_CONTEXT_TOKENS, LLM_COST_ATTR_KEYS, LLM_COST_USD, LLM_INPUT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_MODEL_ATTR_KEYS, LLM_MODEL_NAME, LLM_OUTPUT_TOKENS, LLM_OUTPUT_TOKEN_ATTR_KEYS, LLM_REASONING_TOKENS, LLM_REASONING_TOKEN_ATTR_KEYS, LimitExceededError, LlmCallError, LlmClient, LlmResponseError, LlmRouteAssertionError, LockedJsonlAppender, MANN_WHITNEY_EXACT_MAX_STATES, MANN_WHITNEY_EXACT_MAX_WORK, MODEL_PRICING, MetricsCollector, ModelsUnreachableError, MultiLayerVerifier, Mutex, NoopRawProviderSink, NoopResearcher, NotFoundError, OPENINFERENCE_SPAN_KIND, OTEL_AGENT_EVAL_SCOPE, OUTPUT_VALUE, OtlpFileTraceStore, PairwiseSteeringOptimizer, ProductClient, PromptRegistry, REDACTION_VERSION, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, RESEARCH_REPORT_HARD_PAIR_FLOOR, ROLLOUT_SCHEMA, RUN_COST_ATTR_KEYS, ReplayCache, ReplayCacheMissError, ReplayError, RunCritic, RunIntegrityError, RunRecordValidationError, SEMANTIC_CONCEPT_JUDGE_VERSION, SKILL_USAGE_ANALYST, SPAN_KIND_ATTR_KEYS, SUPERVISOR_RUN_INTEGRITY_SCHEMA, SUPERVISOR_RUN_SCHEMA, SandboxHarness, ScenarioRegistry, SeatUnsetError, SingleBackendError, SkillUsageAnalyst, SpanNotFoundError, SubprocessSandboxDriver, TOOL_ARGS_CAPTURED, TOOL_LATENCY_MS, TOOL_NAME, TOOL_NAME_ATTR_KEYS, TRACE_ANALYSIS_LIMITS, TRACE_ANALYST_ACTOR_DESCRIPTION, TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, TRACE_ANALYST_TOOL_NAMESPACE, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, TRACE_SCHEMA_VERSION, TokenCounter, ToolTraceMissingError, TraceAnalysisLimitError, TraceAnalysisStoreContractError, TraceAnalysisValidationError, TraceContractBuilder, TraceEmitter, TraceFileMalformedError, TraceFileMissingError, TraceFileTooLargeError, TraceNotFoundError, UI_FINDING_SEVERITIES, UI_LENSES, UNIVERSAL_FINDERS, ValidationError, VerificationError, WILCOXON_EXACT_MAX_N, acquisitionPlansForKnowledgeGaps, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, agentProfileModelId, agentVisibleFields, aggregateJudgeVerdicts, aggregateLlm, aggregatePrReviewScore, aggregateRunScore, allCriticalPassed, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeAntiSlop, analyzeSeries, analyzeSupervisorRun, analyzeSupervisorRunIntegrity, analyzeSupervisorRunSources, analyzeTraces, appendScorecard, applyLlmSpanOtlpAttributes, applyToolSpanOtlpAttributes, argHash, asNumber, asString, assertCapabilityHeadroom, assertCrossFamily, assertExactRegistryRunOpts, assertLlmRoute, assertMinted, assertMintedLines, assertModelsServed, assertNoHiddenLeak, assertProductBenchmarkRun, assertRealAgentReceipts, assertRealBackend, assertReleaseConfidence, assertRolloutLine, assertRunAgentProfileCell, assertRunCaptured, assertSingleBackend, assignFeedbackSplit, assignHeldOutTag, attachCostToReport, attest, attributeCounterfactuals, backoffMs, deterministicSplit as benchmarkDeterministicSplit, benchmarks_exports as benchmarks, benjaminiHochberg, bisect, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, buildAgentInterfaceProfileCell, buildAgentProfileCell, buildDefaultAnalystRegistry, buildDriverSystemPrompt, buildProductBenchmarkManifest, buildReflectionPrompt, buildReviewerPrompt, buildTraceAnalysisToolDescriptors, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, buildWorkerDriverSystemPrompt, byteLengthRange, cachedJudge, calibrateJudge, calibrateJudgeContinuous, callLlm, callLlmJson, canaryLeakView, canonicalJson, canonicalize, capabilityHeadroom, captureFetchToRawSink, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, checkTraceContracts, clamp01, classifyFailure, classifyOtlpSpanRole, classifyTreatment, claudeCodeSupervisorRunReader, cliffsDelta, clusteredPairedBinary, cohensD, collectionPreserved, commentsForSource, commitBisect, comparePairedArms, compareReferenceReplay, compareToBaseline, compilerJudge, completionVerdict, composeParsers, composeValidators, computeExperimentStats, computeFindingId, computeToolUseMetrics, computeTraceMetrics, confidenceInterval, containsAll, contentHash, contextInputTokens, continuousAgreement, contractJudge, controlFailureClassFromVerification, controlRunToFeedbackTrajectory, controlRunToRunRecord, convertTraceStoresToOtlp, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, costReport, createAntiSlopJudge, createBoundedTraceAnalysisStore, createChatClient, createDefaultReviewer, createDspyRlmTraceEngine, createFeedbackTrajectory, createIntentMatchJudge, createLlmCorrectnessChecker, createLlmReviewer, createOtelExporter, createOtelTracingStore, createReferenceEquivalenceJudge, createReplayFetch, createSandboxPool, createSemanticConceptJudge, createTokenRecallChecker, createTraceAnalyst, crossTraceDiff, crowdingDistance, dataDescriptionBits, decideNextUserTurn, decidePairedPromotion, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultBlendWeights, defaultIsMaterial, defaultProviderRedactor, defaultReferenceReplayMatcher, defaultTraceInsightPanel, defineTraceAnalyst, deployGateLayer, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, distillPlaybook, domainEvidencePattern, dominates, eProcess, emitControlIntegrityFindings, ensembleJudge, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateContract, evaluateHypothesis, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, evaluateTraceContract, executeScenario, expandProfileAxes, expectAgent, exportProductBenchmark, exportProductBenchmarkRuns, exportRewardModel, exportRunAsOtlp, extractAssetUrls, extractErrorCount, extractOtlpAttributes, extractProducedState, extractUsage, extractUsageFromResponse, extractUsageFromSse, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToDatasetScenario, feedbackTrajectoryToOptimizerRow, fileContains, fileExists, fileExperimentStore, fileVerdictCache, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findProductBenchmarkArtifacts, findSkipCountsAsPass, firstNumberAttr, firstStringAttr, flattenOtlpExportToNdjson, flowLayer, fnv1a32, formatBenchmarkReport, formatDriverReport, formatFindings, formatScorecardDiff, fromHarborTrajectory, gainHistogram, gateTreatmentApplied, gateTreatmentFromMetrics, gateTreatmentFromSpans, gateTreatmentFromToolSpans, ghCliClient, gitProvenanceReader, precision as goldenPrecision, gradeOnHidden, gradeSemanticStatus, groupBy, groupRunsByAgentProfileCell, harnessAxisOf, hasCapturedToolArgs, hashContent, hashJson, hashScenarios, hashToUnit, hiddenGrade, holm, htmlContainsElement, httpGithubClient, improvementVerdict, inMemoryExperimentStore, inMemoryReferenceReplayStore, inMemoryReviewStore, inMemoryRunRecordBackend, inMemoryVerdictCache, inferDomainKeywords, inferOtlpKind, interRaterReliability, interpretCliffs, iqr, isBinaryOutcomeVector, isHiddenDestination, isJudgeSpan, isLlmSpan, isModelPriced, isOtelConfigured, isOtlpModelCall, isRealnessGated, isRetrievalSpan, isRolloutLine, isRunRecord, isSandboxSpan, isToolSpan, isTrainableSplit, isTransientLlmError, isUnavailable, iterateRawCalls, jestTestParser, jsonHasKeys, jsonShape, jsonlReferenceReplayStore, jsonlReviewStore, jsonlRunRecordBackend, judgeFamily, judgeReplayGate, judgeSpans, keyPreserved, knowledgeReadinessTracePayload, leaderboard, linterJudge, llmJudge, llmSpanFromProvider, llmSpans, loadScorecard, loadScorerFromGrader, localCommandRunner, lowercaseMutator, makeEvalTools, makeFinding, makeProposalFinding, mannWhitneyU, mapConcurrent, matchGoldens, matchSpan, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, mergeLayerResults, mergeSteeringBundle, minimumPairsForPairedDeltaTest, mintRolloutRows, modelDescriptionBits, modelHasSnapshot, modelPriceKey, mulberry32, multiToolchainLayer, noProgressDetector, normalCdf, normalizeScores, notBlocked, objectiveEval, observeAll, observedScore, observedSplitScore, otelRunCompleteHook, otlpRowsToRunRecords, otlpRowsToTraceRunRecords, otlpTextToTraceAnalysisStore, otlpToRunRecords, otlpToTraceRunRecords, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedCohensDz, pairedDecisionShape, pairedDeltaTest, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paraphraseRobustness, paraphraseRobustnessScenarios, paretoChart, paretoFrontier, paretoFrontierWithCrowding, parseCorrectnessResponse, parseFeedbackTrajectoriesJsonl, parseReflectionResponse, parseRunRecordSafe, parseRuntimeTrajectoryHookEvent, partialCredit, partitionHeldOut, passAtK, passOrthogonality, pearsonR, pixelDeltaRatio, planTraceInsightQuestions, politenessPrefixMutator, positionalBias, preflightModels, printDriverSummary, probeLlm, productBenchmarkIntegrityFailures, productBenchmarkMutableSurfaces, productBenchmarkRepoIdentity, productBenchmarkSplits, profile_exports as profile, projectOtlpFlatLine, projectRuntimeTrajectoryEvidence, promptBisect, proposeSynthesisTargets, providerFromBaseUrl, pytestTestParser, ranks, readClaudeCodeSupervisorRun, readOtlpStatus, readProductBenchmarkManifest, readProductBenchmarkRecords, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redactString, redactValue, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatch, regexMatches, relabelImportedSplit, renderMarkdownReport, renderPlaybookMarkdown, renderPreferenceMemoryMarkdown, renderPriorFindings, renderReleaseReport, renderSteeringText, renderSupervisorRunHeadline, renderSupervisorRunMarkdown, renderUpstreamFindings, repeatedActionDetector, replayFeedbackTrajectories, replayFeedbackTrajectory, replayScorerOverCorpus, replayTraceThroughJudge, requireAgentProfileCell, requiredPairedSampleSize, requiredSampleSize, researchReport, resolveModelPricing, resolveSeat, resolveTraceAnalystLimits, rollupSupervisorRuns, roundTripRunRecord, routeFields, rowCount, rowWhere, runAgentControlLoop, runAssertions, runBehavioralCanaries, runCanaries, runCounterfactual, runE2EWorkflow, runEvalCampaign, runExpectations, runFailureClass, runHarnessExperiment, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runProposeReview, runProposeReviewAsControlLoop, runRecordToProductBenchmarkRecord, runReferenceEquivalenceJudge, runReferenceReplay, runScore, runSelfPlay, runSemanticConceptJudge, runTaskScore, runTestGradedScenario, runTraceAnalyst, runsForScenario, scalarScore, scanForMuffledGates, scoreContinuity, scoreFromEvals, scoreKnowledgeReadiness, scoreOrigin, scorePrReviewComments, scorePrReviewSource, scoreRedTeamOutput, scoreReferenceReplay, scoreTraceInsightReadiness, seatPresets, securityJudge, selectHarnessVariant, selfPreference, sentenceReorderMutator, serializeFeedbackTrajectoriesJsonl, showMeasured, signManifest, spearmanR, statusAdvanced, stopOnNoProgress, stopOnRepeatedAction, stringField, stripFencedJson, studentTCdf, subjectiveEval, summarizeAgentReceiptIntegrity, summarizeBackendIntegrity, summarizeHarnessResults, summarizePrReviewBenchmark, summarizePreferenceMemory, summaryTable, supervisorRunRolloutLines, testJudge, textInSnapshot, throwIfRunIncomplete, toAgentProfileJson, toHarborTrajectories, toHarborTrajectory, toJsonl, toLangfuseEnvelope, toOpenAiTool, toPrometheusText, toRewardRows, toSftRows, tokenizeDomainWords, toolNamesForRun, toolSpans, toolSpansToTraceAnalysisStore, traceAnalystFunctionGroup, traceAnalystOnRunComplete, traceContract, traceJudge, traceJudgeEnsemble, traceSpanKindToOpenInferenceKind, tracedAnalyzeTraces, trainingReward, trainingScore, typoMutator, unmintableReasons, urlContains, userQuestionsForKnowledgeGaps, validateAgentProfileCell, validateProductBenchmarkManifest, validateProductBenchmarkRecord, validateProductBenchmarkRun, validateRolloutLine, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, verifyManifest, visualDiff, viteDeployRunner, vitestTestParser, weightedComposite, weightedMean, weightedRecall, welchsTTest, whitespaceCollapseMutator, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withHeldoutBlend, withJudgeRetry, withOtelPipeline, wranglerDeployRunner, writeSupervisorRunReport };
+export { AGENT_PROFILE_KINDS, ATIF_SCHEMA_VERSION, ATTESTATION_ALGORITHM, AgentDriver, AgentEvalError, AgentProfileCellValidationError, AnalystRegistry, BENCHMARK_SPLIT_SEED, BOOTSTRAP_GATE_MIN_N, BackendIntegrityError, BenchmarkRunner, BudgetBreachError, BudgetGuard, CODING_HARNESSES, CONTROL_INTEGRITY_ANALYST, CallbackResearcher, CaptureIntegrityError, ConfigError, ControlIntegrityAnalyst, ConvergenceTracker, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, CostLedger, CostLedgerPersistenceError, CostReceiptCaptureError, CostReservationExceededError, CostTracker, CrossFamilyError, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_AGENT_SLOS, DEFAULT_COMPLEXITY_WEIGHTS, DEFAULT_RULES as DEFAULT_FAILURE_RULES, DEFAULT_FINDERS, DEFAULT_HARNESS_OBJECTIVES, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_PERMUTATIONS, DEFAULT_PR_REVIEW_SCORE_WEIGHTS, DEFAULT_REDACTION_RULES, DEFAULT_RED_TEAM_CORPUS, DEFAULT_RUN_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, DEFAULT_TRACE_ANALYST_LIMITS, Dataset, DescriptionLengthGate, DockerSandboxDriver, DualAgentBench, ERROR_COUNT_PATTERNS, EvalTraceStore, ExactAnalystRunExecutionError, ExperimentTracker, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, FindingsStore, HARBOR_IMPORT_GAP, HARNESS_BRIEFS, HARNESS_NATIVE_MODEL, HeldOutGate, HoldoutAuditor, HoldoutLockedError, IMPROVEMENT_KIND_SPEC, INPUT_VALUE, INTENT_MATCH_JUDGE_VERSION, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, InMemoryWorkspaceInspector, JudgeError, JudgeParseError, JudgeRunner, KNOWLEDGE_GAP_KIND_SPEC, KNOWLEDGE_POISONING_KIND_SPEC, LLM_CACHED_TOKENS, LLM_CACHED_TOKEN_ATTR_KEYS, LLM_CACHE_WRITE_TOKENS, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, LLM_CONTEXT_TOKENS, LLM_COST_ATTR_KEYS, LLM_COST_USD, LLM_INPUT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_MODEL_ATTR_KEYS, LLM_MODEL_NAME, LLM_OUTPUT_TOKENS, LLM_OUTPUT_TOKEN_ATTR_KEYS, LLM_REASONING_TOKENS, LLM_REASONING_TOKEN_ATTR_KEYS, LimitExceededError, LlmCallError, LlmClient, LlmResponseError, LlmRouteAssertionError, LockedJsonlAppender, MANN_WHITNEY_EXACT_MAX_STATES, MANN_WHITNEY_EXACT_MAX_WORK, MODEL_PRICING, MetricsCollector, ModelsUnreachableError, MultiLayerVerifier, Mutex, NoopRawProviderSink, NoopResearcher, NotFoundError, OPENINFERENCE_SPAN_KIND, OTEL_AGENT_EVAL_SCOPE, OUTPUT_VALUE, OtlpFileTraceStore, PairwiseSteeringOptimizer, ProductClient, PromptRegistry, REDACTION_VERSION, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, RESEARCH_REPORT_HARD_PAIR_FLOOR, ROLLOUT_SCHEMA, RUN_COST_ATTR_KEYS, ReplayCache, ReplayCacheMissError, ReplayError, RunCritic, RunIntegrityError, RunRecordValidationError, SEMANTIC_CONCEPT_JUDGE_VERSION, SKILL_USAGE_ANALYST, SPAN_KIND_ATTR_KEYS, SUPERVISOR_RUN_INTEGRITY_SCHEMA, SUPERVISOR_RUN_SCHEMA, SandboxHarness, ScenarioRegistry, SeatUnsetError, SingleBackendError, SkillUsageAnalyst, SpanNotFoundError, SubprocessSandboxDriver, TOOL_ARGS_CAPTURED, TOOL_LATENCY_MS, TOOL_NAME, TOOL_NAME_ATTR_KEYS, TRACE_ANALYSIS_LIMITS, TRACE_ANALYST_ACTOR_DESCRIPTION, TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, TRACE_ANALYST_TOOL_NAMESPACE, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, TRACE_SCHEMA_VERSION, TokenCounter, ToolTraceMissingError, TraceAnalysisLimitError, TraceAnalysisStoreContractError, TraceAnalysisValidationError, TraceContractBuilder, TraceEmitter, TraceFileMalformedError, TraceFileMissingError, TraceFileTooLargeError, TraceNotFoundError, UI_FINDING_SEVERITIES, UI_LENSES, UNIVERSAL_FINDERS, ValidationError, VerificationError, WILCOXON_EXACT_MAX_N, WORKER_DRIVER_DOCTRINE, acquisitionPlansForKnowledgeGaps, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, agentProfileModelId, agentVisibleFields, aggregateJudgeVerdicts, aggregateLlm, aggregatePrReviewScore, aggregateRunScore, allCriticalPassed, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeAntiSlop, analyzeSeries, analyzeSupervisorRun, analyzeSupervisorRunIntegrity, analyzeSupervisorRunSources, analyzeTraces, appendScorecard, applyLlmSpanOtlpAttributes, applyToolSpanOtlpAttributes, argHash, asNumber, asString, assertCapabilityHeadroom, assertCrossFamily, assertExactRegistryRunOpts, assertLlmRoute, assertMinted, assertMintedLines, assertModelsServed, assertNoHiddenLeak, assertProductBenchmarkRun, assertRealAgentReceipts, assertRealBackend, assertReleaseConfidence, assertRolloutLine, assertRunAgentProfileCell, assertRunCaptured, assertSingleBackend, assignFeedbackSplit, assignHeldOutTag, attachCostToReport, attest, attributeCounterfactuals, backoffMs, deterministicSplit as benchmarkDeterministicSplit, benchmarks_exports as benchmarks, benjaminiHochberg, bisect, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, buildAgentInterfaceProfileCell, buildAgentProfileCell, buildDefaultAnalystRegistry, buildDriverSystemPrompt, buildProductBenchmarkManifest, buildReflectionPrompt, buildReviewerPrompt, buildTraceAnalysisToolDescriptors, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, byteLengthRange, cachedJudge, calibrateJudge, calibrateJudgeContinuous, callLlm, callLlmJson, canaryLeakView, canonicalJson, canonicalize, capabilityHeadroom, captureFetchToRawSink, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, checkTraceContracts, clamp01, classifyFailure, classifyOtlpSpanRole, classifyTreatment, claudeCodeSupervisorRunReader, cliffsDelta, clusteredPairedBinary, cohensD, collectionPreserved, commentsForSource, commitBisect, comparePairedArms, compareReferenceReplay, compareToBaseline, compilerJudge, completionVerdict, composeParsers, composeValidators, computeExperimentStats, computeFindingId, computeToolUseMetrics, computeTraceMetrics, confidenceInterval, containsAll, contentHash, contextInputTokens, continuousAgreement, contractJudge, controlFailureClassFromVerification, controlRunToFeedbackTrajectory, controlRunToRunRecord, convertTraceStoresToOtlp, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, costReport, createAntiSlopJudge, createBoundedTraceAnalysisStore, createChatClient, createDefaultReviewer, createDspyRlmTraceEngine, createFeedbackTrajectory, createIntentMatchJudge, createLlmCorrectnessChecker, createLlmReviewer, createOtelExporter, createOtelTracingStore, createReferenceEquivalenceJudge, createReplayFetch, createSandboxPool, createSemanticConceptJudge, createTokenRecallChecker, createTraceAnalyst, crossTraceDiff, crowdingDistance, dataDescriptionBits, decideNextUserTurn, decidePairedPromotion, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultBlendWeights, defaultIsMaterial, defaultProviderRedactor, defaultReferenceReplayMatcher, defaultTraceInsightPanel, defineTraceAnalyst, deployGateLayer, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, distillPlaybook, domainEvidencePattern, dominates, eProcess, emitControlIntegrityFindings, ensembleJudge, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateContract, evaluateHypothesis, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, evaluateTraceContract, executeScenario, expandProfileAxes, expectAgent, exportProductBenchmark, exportProductBenchmarkRuns, exportRewardModel, exportRunAsOtlp, extractAssetUrls, extractErrorCount, extractOtlpAttributes, extractProducedState, extractUsage, extractUsageFromResponse, extractUsageFromSse, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToDatasetScenario, feedbackTrajectoryToOptimizerRow, fileContains, fileExists, fileExperimentStore, fileVerdictCache, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findProductBenchmarkArtifacts, findSkipCountsAsPass, firstNumberAttr, firstStringAttr, flattenOtlpExportToNdjson, flowLayer, fnv1a32, formatBenchmarkReport, formatDriverReport, formatFindings, formatScorecardDiff, fromHarborTrajectory, gainHistogram, gateTreatmentApplied, gateTreatmentFromMetrics, gateTreatmentFromSpans, gateTreatmentFromToolSpans, ghCliClient, gitProvenanceReader, precision as goldenPrecision, gradeOnHidden, gradeSemanticStatus, groupBy, groupRunsByAgentProfileCell, harnessAxisOf, hasCapturedToolArgs, hashContent, hashJson, hashScenarios, hashToUnit, hiddenGrade, holm, htmlContainsElement, httpGithubClient, improvementVerdict, inMemoryExperimentStore, inMemoryReferenceReplayStore, inMemoryReviewStore, inMemoryRunRecordBackend, inMemoryVerdictCache, inferDomainKeywords, inferOtlpKind, interRaterReliability, interpretCliffs, iqr, isBinaryOutcomeVector, isHiddenDestination, isJudgeSpan, isLlmSpan, isModelPriced, isOtelConfigured, isOtlpModelCall, isRealnessGated, isRetrievalSpan, isRolloutLine, isRunRecord, isSandboxSpan, isToolSpan, isTrainableSplit, isTransientLlmError, isUnavailable, iterateRawCalls, jestTestParser, jsonHasKeys, jsonShape, jsonlReferenceReplayStore, jsonlReviewStore, jsonlRunRecordBackend, judgeFamily, judgeReplayGate, judgeSpans, keyPreserved, knowledgeReadinessTracePayload, leaderboard, linterJudge, llmJudge, llmSpanFromProvider, llmSpans, loadScorecard, loadScorerFromGrader, localCommandRunner, lowercaseMutator, makeEvalTools, makeFinding, makeProposalFinding, mannWhitneyU, mapConcurrent, matchGoldens, matchSpan, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, mergeLayerResults, mergeSteeringBundle, minimumPairsForPairedDeltaTest, mintRolloutRows, modelDescriptionBits, modelHasSnapshot, modelPriceKey, mulberry32, multiToolchainLayer, noProgressDetector, normalCdf, normalizeScores, notBlocked, objectiveEval, observeAll, observedScore, observedSplitScore, otelRunCompleteHook, otlpRowsToRunRecords, otlpRowsToTraceRunRecords, otlpTextToTraceAnalysisStore, otlpToRunRecords, otlpToTraceRunRecords, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedCohensDz, pairedDecisionShape, pairedDeltaTest, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paraphraseRobustness, paraphraseRobustnessScenarios, paretoChart, paretoFrontier, paretoFrontierWithCrowding, parseCorrectnessResponse, parseFeedbackTrajectoriesJsonl, parseReflectionResponse, parseRunRecordSafe, parseRuntimeTrajectoryHookEvent, partialCredit, partitionHeldOut, passAtK, passOrthogonality, pearsonR, pixelDeltaRatio, planTraceInsightQuestions, politenessPrefixMutator, positionalBias, preflightModels, printDriverSummary, probeLlm, productBenchmarkIntegrityFailures, productBenchmarkMutableSurfaces, productBenchmarkRepoIdentity, productBenchmarkSplits, profile_exports as profile, projectOtlpFlatLine, projectRuntimeTrajectoryEvidence, promptBisect, proposeSynthesisTargets, providerFromBaseUrl, pytestTestParser, ranks, readClaudeCodeSupervisorRun, readOtlpStatus, readProductBenchmarkManifest, readProductBenchmarkRecords, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redactString, redactValue, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatch, regexMatches, relabelImportedSplit, renderMarkdownReport, renderPlaybookMarkdown, renderPreferenceMemoryMarkdown, renderPriorFindings, renderReleaseReport, renderSteeringText, renderSupervisorRunHeadline, renderSupervisorRunMarkdown, renderUpstreamFindings, repeatedActionDetector, replayFeedbackTrajectories, replayFeedbackTrajectory, replayScorerOverCorpus, replayTraceThroughJudge, requireAgentProfileCell, requiredPairedSampleSize, requiredSampleSize, researchReport, resetDeprecationWarnings, resolveModelPricing, resolveSeat, resolveTraceAnalystLimits, rollupSupervisorRuns, roundTripRunRecord, routeFields, rowCount, rowWhere, runAgentControlLoop, runAssertions, runBehavioralCanaries, runCanaries, runCounterfactual, runE2EWorkflow, runEvalCampaign, runExpectations, runFailureClass, runHarnessExperiment, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runProposeReview, runProposeReviewAsControlLoop, runRecordToProductBenchmarkRecord, runReferenceEquivalenceJudge, runReferenceReplay, runScore, runSelfPlay, runSemanticConceptJudge, runTaskScore, runTestGradedScenario, runTraceAnalyst, runsForScenario, scalarScore, scanForMuffledGates, scoreContinuity, scoreFromEvals, scoreKnowledgeReadiness, scoreOrigin, scorePrReviewComments, scorePrReviewSource, scoreRedTeamOutput, scoreReferenceReplay, scoreTraceInsightReadiness, seatPresets, securityJudge, selectHarnessVariant, selfPreference, sentenceReorderMutator, serializeFeedbackTrajectoriesJsonl, showMeasured, signManifest, spearmanR, statusAdvanced, stopOnNoProgress, stopOnRepeatedAction, stringField, stripFencedJson, studentTCdf, subjectiveEval, summarizeAgentReceiptIntegrity, summarizeBackendIntegrity, summarizeHarnessResults, summarizePrReviewBenchmark, summarizePreferenceMemory, summaryTable, supervisorRunRolloutLines, testJudge, textInSnapshot, throwIfRunIncomplete, toAgentProfileJson, toHarborTrajectories, toHarborTrajectory, toJsonl, toLangfuseEnvelope, toOpenAiTool, toPrometheusText, toRewardRows, toSftRows, tokenizeDomainWords, toolNamesForRun, toolSpans, toolSpansToTraceAnalysisStore, traceAnalystFunctionGroup, traceAnalystOnRunComplete, traceContract, traceJudge, traceJudgeEnsemble, traceSpanKindToOpenInferenceKind, tracedAnalyzeTraces, trainingReward, trainingScore, typoMutator, unmintableReasons, urlContains, userQuestionsForKnowledgeGaps, validateAgentProfileCell, validateProductBenchmarkManifest, validateProductBenchmarkRecord, validateProductBenchmarkRun, validateRolloutLine, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, verifyManifest, visualDiff, viteDeployRunner, vitestTestParser, warnDeprecatedOnce, weightedComposite, weightedMean, weightedRecall, welchsTTest, whitespaceCollapseMutator, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withHeldoutBlend, withJudgeRetry, withOtelPipeline, wranglerDeployRunner, writeSupervisorRunReport };
 
 //# sourceMappingURL=index.js.map