@tangle-network/agent-runtime 0.37.0 → 0.38.0

package/dist/agent.d.ts

@@ -1,9 +1,9 @@
 import * as _tangle_network_agent_eval from '@tangle-network/agent-eval';
 import { TraceAnalystKindSpec, AnalystFinding } from '@tangle-network/agent-eval';
 import { R as RuntimeStreamEvent } from './types-CsCCryln.js';
-import { A as AgentSurfaces } from './improvement-adapter-CaZxFxTd.js';
-export { C as CreateSurfaceImprovementAdapterOpts, D as DraftPatchInput, a as DraftPatchOutput, R as ResolvedSurface, S as SurfaceImprovementEdit, b as SurfaceValidationIssue, c as createSurfaceImprovementAdapter, r as renderSurfaceIssues, d as resolveSubjectPath, v as validateSurfaces } from './improvement-adapter-CaZxFxTd.js';
-import { K as KnowledgeAdapter, a as RunAnalystLoopResult } from './types-D_MXrmJP.js';
+import { A as AgentSurfaces } from './improvement-adapter-BC4HhuAR.js';
+export { C as CreateSurfaceImprovementAdapterOpts, D as DraftPatchInput, a as DraftPatchOutput, R as ResolvedSurface, S as SurfaceImprovementEdit, b as SurfaceValidationIssue, c as createSurfaceImprovementAdapter, r as renderSurfaceIssues, d as resolveSubjectPath, v as validateSurfaces } from './improvement-adapter-BC4HhuAR.js';
+import { K as KnowledgeAdapter, a as RunAnalystLoopResult } from './types-p8dWBIXL.js';
 
 /**
  * The full agent manifest. Each agent ships ONE of these.

package/dist/analyst-loop.d.ts

@@ -1,5 +1,5 @@
-import { R as RunAnalystLoopOpts, a as RunAnalystLoopResult } from './types-D_MXrmJP.js';
-export { A as AnalystLoopEvent, b as AnalystRegistryLike, c as AnalystRegistryStreamingLike, d as AutoApplyPolicy, F as FindingsStoreLike, I as ImprovementAdapter, e as ImprovementEditBatch, f as ImprovementReport, K as KnowledgeAdapter, g as KnowledgeProposalBatch, h as KnowledgeReport } from './types-D_MXrmJP.js';
+import { R as RunAnalystLoopOpts, a as RunAnalystLoopResult } from './types-p8dWBIXL.js';
+export { A as AnalystLoopEvent, b as AnalystRegistryLike, c as AnalystRegistryStreamingLike, d as AutoApplyPolicy, F as FindingsStoreLike, I as ImprovementAdapter, e as ImprovementEditBatch, f as ImprovementReport, K as KnowledgeAdapter, g as KnowledgeProposalBatch, h as KnowledgeReport } from './types-p8dWBIXL.js';
 import '@tangle-network/agent-eval';
 
 /**

package/dist/analyst-loop.js

@@ -1,261 +1,7 @@
+import {
+  runAnalystLoop
+} from "./chunk-XBUG326M.js";
 import "./chunk-DGUM43GV.js";
-
-// src/analyst-loop/run-analyst-loop.ts
-import { diffFindings } from "@tangle-network/agent-eval";
-async function runAnalystLoop(opts) {
-  const log = opts.log ?? defaultLog;
-  const strategy = opts.priorFindingsStrategy ?? "per-kind";
-  const emit = makeEmitter(opts.onEvent);
-  const startedAt = Date.now();
-  const baselineRunId = resolveBaselineRunId(opts);
-  const priorAll = baselineRunId ? opts.findingsStore?.loadRun(baselineRunId) ?? [] : [];
-  log("baseline resolved", { baselineRunId, prior_findings: priorAll.length });
-  await emit({
-    type: "baseline-resolved",
-    runId: opts.runId,
-    baselineRunId,
-    priorFindingCount: priorAll.length
-  });
-  const priorFindings = buildPriorFindingsInput(priorAll, strategy, opts.registry.list());
-  const analystResult = await runRegistry(opts, priorFindings, emit);
-  log("analyst run complete", {
-    findings: analystResult.findings.length,
-    cost_usd: analystResult.total_cost_usd,
-    per_analyst: analystResult.per_analyst.map((s) => ({
-      id: s.analyst_id,
-      status: s.status,
-      n: s.findings_count
-    }))
-  });
-  if (opts.findingsStore && analystResult.findings.length > 0) {
-    await opts.findingsStore.append(opts.runId, analystResult.findings);
-    await emit({
-      type: "findings-persisted",
-      runId: opts.runId,
-      count: analystResult.findings.length
-    });
-  }
-  let diff = null;
-  if (baselineRunId && analystResult.findings.length > 0) {
-    diff = diffFindings(
-      priorAll.map((f) => ({ ...f })),
-      analystResult.findings.map((f) => ({ ...f, run_id: opts.runId }))
-    );
-    log("diff vs baseline", {
-      appeared: diff.appeared.length,
-      disappeared: diff.disappeared.length,
-      persisted: diff.persisted.length,
-      changed: diff.changed.length
-    });
-    await emit({
-      type: "diff-computed",
-      runId: opts.runId,
-      baselineRunId,
-      appeared: diff.appeared.length,
-      disappeared: diff.disappeared.length,
-      persisted: diff.persisted.length,
-      changed: diff.changed.length
-    });
-  }
-  let knowledge = null;
-  if (opts.knowledgeAdapter) {
-    knowledge = await runKnowledgeAdapter(opts, analystResult.findings, log, emit);
-  }
-  let improvement = null;
-  if (opts.improvementAdapter) {
-    improvement = await runImprovementAdapter(opts, analystResult.findings, log, emit);
-  }
-  await emit({
-    type: "loop-completed",
-    runId: opts.runId,
-    durationMs: Date.now() - startedAt
-  });
-  return {
-    runId: opts.runId,
-    baselineRunId,
-    analystResult,
-    diff,
-    knowledge,
-    improvement
-  };
-}
-function makeEmitter(onEvent) {
-  if (!onEvent) return async () => {
-  };
-  return async (event) => {
-    await onEvent(event);
-  };
-}
-async function runRegistry(opts, priorFindings, emit) {
-  const reg = opts.registry;
-  if (typeof reg.runStream === "function" && opts.onEvent) {
-    let final = null;
-    for await (const ev of reg.runStream(opts.runId, opts.inputs, { priorFindings })) {
-      await emit({ type: "analyst", runId: opts.runId, event: ev });
-      if (ev.type === "run-completed") final = ev.result;
-    }
-    if (!final) {
-      throw new Error("runAnalystLoop: registry.runStream ended without run-completed event");
-    }
-    return final;
-  }
-  return opts.registry.run(opts.runId, opts.inputs, { priorFindings });
-}
-function resolveBaselineRunId(opts) {
-  if (opts.baselineRunId === null) return null;
-  if (typeof opts.baselineRunId === "string") return opts.baselineRunId;
-  if (!opts.findingsStore) return null;
-  const all = opts.findingsStore.loadAll();
-  let last = null;
-  for (const row of all) {
-    if (row.run_id === opts.runId) continue;
-    last = row.run_id;
-  }
-  return last;
-}
-function buildPriorFindingsInput(prior, strategy, registry) {
-  if (strategy === "none" || prior.length === 0) return void 0;
-  const stripped = prior.map(({ run_id: _run_id, ...rest }) => rest);
-  if (strategy === "wildcard") {
-    return { "*": stripped };
-  }
-  void registry;
-  return stripped;
-}
-async function runKnowledgeAdapter(opts, findings, log, emit) {
-  const adapter = opts.knowledgeAdapter;
-  const batch = await adapter.proposeFromFindings(findings);
-  log("knowledge.proposeFromFindings", {
-    proposals: batch.proposals.length,
-    skipped: batch.skipped,
-    errors: batch.errors.length
-  });
-  await emit({
-    type: "knowledge-proposed",
-    runId: opts.runId,
-    proposalCount: batch.proposals.length,
-    skipped: batch.skipped,
-    errors: batch.errors.length
-  });
-  const auto = opts.autoApply?.knowledge ?? false;
-  const threshold = opts.autoApply?.knowledgeConfidenceThreshold ?? 0.85;
-  if (!auto || !adapter.apply) {
-    await emit({
-      type: "knowledge-applied",
-      runId: opts.runId,
-      writtenCount: 0,
-      withheldForReview: batch.proposals.length
-    });
-    return {
-      proposals: batch.proposals,
-      applied: [],
-      skipped: batch.skipped,
-      errors: batch.errors,
-      withheld_for_review: batch.proposals.length
-    };
-  }
-  const findingsById = new Map(findings.map((f) => [f.finding_id, f]));
-  const safe = [];
-  let withheld = 0;
-  for (const p of batch.proposals) {
-    const src = p.sourceFindingId ? findingsById.get(p.sourceFindingId) : void 0;
-    if (!src) {
-      withheld += 1;
-      continue;
-    }
-    if (src.confidence < threshold) {
-      withheld += 1;
-      continue;
-    }
-    safe.push(p);
-  }
-  const result = await adapter.apply(safe);
-  log("knowledge.apply", {
-    applied: result.written.length,
-    withheld_for_review: withheld,
-    warnings: result.warnings.length
-  });
-  await emit({
-    type: "knowledge-applied",
-    runId: opts.runId,
-    writtenCount: result.written.length,
-    withheldForReview: withheld
-  });
-  return {
-    proposals: batch.proposals,
-    applied: result.written,
-    skipped: batch.skipped,
-    errors: batch.errors,
-    withheld_for_review: withheld
-  };
-}
-async function runImprovementAdapter(opts, findings, log, emit) {
-  const adapter = opts.improvementAdapter;
-  const batch = await adapter.proposeFromFindings(findings);
-  log("improvement.proposeFromFindings", {
-    edits: batch.edits.length,
-    skipped: batch.skipped,
-    errors: batch.errors.length
-  });
-  await emit({
-    type: "improvement-proposed",
-    runId: opts.runId,
-    editCount: batch.edits.length,
-    skipped: batch.skipped,
-    errors: batch.errors.length
-  });
-  const auto = opts.autoApply?.improvement ?? false;
-  const threshold = opts.autoApply?.improvementConfidenceThreshold ?? 0.9;
-  if (!auto || !adapter.apply) {
-    await emit({
-      type: "improvement-applied",
-      runId: opts.runId,
-      appliedCount: 0,
-      withheldForReview: batch.edits.length
-    });
-    return {
-      edits: batch.edits,
-      applied: [],
-      skipped: batch.skipped,
-      errors: batch.errors,
-      withheld_for_review: batch.edits.length
-    };
-  }
-  const findingsById = new Map(findings.map((f) => [f.finding_id, f]));
-  const safe = [];
-  let withheld = 0;
-  for (const e of batch.edits) {
-    const src = e.sourceFindingId ? findingsById.get(e.sourceFindingId) : void 0;
-    if (!src || src.confidence < threshold) {
-      withheld += 1;
-      continue;
-    }
-    safe.push(e);
-  }
-  const result = await adapter.apply(safe);
-  log("improvement.apply", {
-    applied: result.applied.length,
-    withheld_for_review: withheld,
-    warnings: result.warnings.length
-  });
-  await emit({
-    type: "improvement-applied",
-    runId: opts.runId,
-    appliedCount: result.applied.length,
-    withheldForReview: withheld
-  });
-  return {
-    edits: batch.edits,
-    applied: result.applied,
-    skipped: batch.skipped,
-    errors: batch.errors,
-    withheld_for_review: withheld
-  };
-}
-function defaultLog(msg, fields) {
-  if (fields) console.log(`[analyst-loop] ${msg}`, fields);
-  else console.log(`[analyst-loop] ${msg}`);
-}
 export {
   runAnalystLoop
 };

package/dist/analyst-loop.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/analyst-loop/run-analyst-loop.ts"],"sourcesContent":["/**\n * `runAnalystLoop` — the one call agent apps reach for to close the\n * recursive-self-improvement loop.\n *\n *   1. Load baseline findings (last run, or the slice the caller specifies)\n *   2. Run the analyst registry with priorFindings injected\n *   3. Persist the new run's findings to the ledger\n *   4. Diff the new run against the baseline\n *   5. Hand the findings to the knowledge adapter → proposals (and\n *      optionally apply them) → wiki edits\n *   6. Hand the findings to the improvement adapter → prompt / tool /\n *      scaffolding edits (review-only by default)\n *   7. Return a single report the consumer renders / persists / acts on.\n *\n * Adapters are optional: the loop works as a \"run + diff + report\"\n * primitive when no adapters are wired; it closes end-to-end when\n * both adapters are wired.\n */\n\nimport type { AnalystFinding, AnalystRunResult, FindingsDiff } from '@tangle-network/agent-eval'\nimport { diffFindings } from '@tangle-network/agent-eval'\n\nimport type {\n  AnalystLoopEvent,\n  AnalystRegistryStreamingLike,\n  ImprovementReport,\n  KnowledgeReport,\n  RunAnalystLoopOpts,\n  RunAnalystLoopResult,\n} from './types'\n\nexport async function runAnalystLoop<TProposal = unknown, TEdit = unknown>(\n  opts: RunAnalystLoopOpts,\n): Promise<RunAnalystLoopResult<TProposal, TEdit>> {\n  const log = opts.log ?? defaultLog\n  const strategy = opts.priorFindingsStrategy ?? 'per-kind'\n  const emit = makeEmitter(opts.onEvent)\n  const startedAt = Date.now()\n\n  // 1. Resolve baseline + load prior findings.\n  const baselineRunId = resolveBaselineRunId(opts)\n  const priorAll: ReadonlyArray<AnalystFinding & { run_id: string }> = baselineRunId\n    ? (opts.findingsStore?.loadRun(baselineRunId) ?? [])\n    : []\n  log('baseline resolved', { baselineRunId, prior_findings: priorAll.length })\n  await emit({\n    type: 'baseline-resolved',\n    runId: opts.runId,\n    baselineRunId,\n    priorFindingCount: priorAll.length,\n  })\n\n  // 2. Run the registry. Strategy controls how analysts see priors.\n  //    When the registry exposes runStream, forward each event verbatim\n  //    so subscribers see per-analyst progress in real time.\n  const priorFindings = buildPriorFindingsInput(priorAll, strategy, opts.registry.list())\n  const analystResult = await runRegistry(opts, priorFindings, emit)\n  log('analyst run complete', {\n    findings: analystResult.findings.length,\n    cost_usd: analystResult.total_cost_usd,\n    per_analyst: analystResult.per_analyst.map((s) => ({\n      id: s.analyst_id,\n      status: s.status,\n      n: s.findings_count,\n    })),\n  })\n\n  // 3. Persist the new run before any side-effecting adapter runs so\n  //    the ledger is the source of truth even if an adapter throws.\n  if (opts.findingsStore && analystResult.findings.length > 0) {\n    await opts.findingsStore.append(opts.runId, analystResult.findings)\n    await emit({\n      type: 'findings-persisted',\n      runId: opts.runId,\n      count: analystResult.findings.length,\n    })\n  }\n\n  // 4. Diff vs baseline.\n  let diff: FindingsDiff | null = null\n  if (baselineRunId && analystResult.findings.length > 0) {\n    diff = diffFindings(\n      priorAll.map((f) => ({ ...f })),\n      analystResult.findings.map((f) => ({ ...f, run_id: opts.runId })),\n    )\n    log('diff vs baseline', {\n      appeared: diff.appeared.length,\n      disappeared: diff.disappeared.length,\n      persisted: diff.persisted.length,\n      changed: diff.changed.length,\n    })\n    await emit({\n      type: 'diff-computed',\n      runId: opts.runId,\n      baselineRunId,\n      appeared: diff.appeared.length,\n      disappeared: diff.disappeared.length,\n      persisted: diff.persisted.length,\n      changed: diff.changed.length,\n    })\n  }\n\n  // 5. Knowledge adapter — proposals + optional auto-apply.\n  let knowledge: KnowledgeReport<TProposal> | null = null\n  if (opts.knowledgeAdapter) {\n    knowledge = await runKnowledgeAdapter(opts, analystResult.findings, log, emit)\n  }\n\n  // 6. Improvement adapter — prompt / tool / scaffolding edits.\n  let improvement: ImprovementReport<TEdit> | null = null\n  if (opts.improvementAdapter) {\n    improvement = await runImprovementAdapter(opts, analystResult.findings, log, emit)\n  }\n\n  await emit({\n    type: 'loop-completed',\n    runId: opts.runId,\n    durationMs: Date.now() - startedAt,\n  })\n\n  return {\n    runId: opts.runId,\n    baselineRunId,\n    analystResult,\n    diff,\n    knowledge,\n    improvement,\n  }\n}\n\ntype Emitter = (event: AnalystLoopEvent) => Promise<void>\n\nfunction makeEmitter(onEvent: RunAnalystLoopOpts['onEvent']): Emitter {\n  if (!onEvent) return async () => {}\n  return async (event) => {\n    await onEvent(event)\n  }\n}\n\nasync function runRegistry(\n  opts: RunAnalystLoopOpts,\n  priorFindings: ReturnType<typeof buildPriorFindingsInput>,\n  emit: Emitter,\n): Promise<AnalystRunResult> {\n  const reg = opts.registry as AnalystRegistryStreamingLike\n  if (typeof reg.runStream === 'function' && opts.onEvent) {\n    let final: AnalystRunResult | null = null\n    for await (const ev of reg.runStream(opts.runId, opts.inputs, { priorFindings })) {\n      await emit({ type: 'analyst', runId: opts.runId, event: ev })\n      if (ev.type === 'run-completed') final = ev.result\n    }\n    if (!final) {\n      throw new Error('runAnalystLoop: registry.runStream ended without run-completed event')\n    }\n    return final\n  }\n  return opts.registry.run(opts.runId, opts.inputs, { priorFindings })\n}\n\nfunction resolveBaselineRunId(opts: RunAnalystLoopOpts): string | null {\n  if (opts.baselineRunId === null) return null\n  if (typeof opts.baselineRunId === 'string') return opts.baselineRunId\n  if (!opts.findingsStore) return null\n  const all = opts.findingsStore.loadAll()\n  let last: string | null = null\n  for (const row of all) {\n    if (row.run_id === opts.runId) continue\n    last = row.run_id\n  }\n  return last\n}\n\nfunction buildPriorFindingsInput(\n  prior: ReadonlyArray<AnalystFinding & { run_id: string }>,\n  strategy: 'per-kind' | 'wildcard' | 'none',\n  registry: ReadonlyArray<{ id: string }>,\n): ReadonlyArray<AnalystFinding> | Record<string, ReadonlyArray<AnalystFinding>> | undefined {\n  if (strategy === 'none' || prior.length === 0) return undefined\n  const stripped = prior.map(({ run_id: _run_id, ...rest }) => rest as AnalystFinding)\n  if (strategy === 'wildcard') {\n    return { '*': stripped }\n  }\n  void registry\n  return stripped\n}\n\nasync function runKnowledgeAdapter<TProposal>(\n  opts: RunAnalystLoopOpts,\n  findings: ReadonlyArray<AnalystFinding>,\n  log: NonNullable<RunAnalystLoopOpts['log']>,\n  emit: Emitter,\n): Promise<KnowledgeReport<TProposal>> {\n  const adapter = opts.knowledgeAdapter!\n  const batch = await adapter.proposeFromFindings(findings)\n  log('knowledge.proposeFromFindings', {\n    proposals: batch.proposals.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n  await emit({\n    type: 'knowledge-proposed',\n    runId: opts.runId,\n    proposalCount: batch.proposals.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n\n  const auto = opts.autoApply?.knowledge ?? false\n  const threshold = opts.autoApply?.knowledgeConfidenceThreshold ?? 0.85\n\n  if (!auto || !adapter.apply) {\n    await emit({\n      type: 'knowledge-applied',\n      runId: opts.runId,\n      writtenCount: 0,\n      withheldForReview: batch.proposals.length,\n    })\n    return {\n      proposals: batch.proposals as TProposal[],\n      applied: [],\n      skipped: batch.skipped,\n      errors: batch.errors,\n      withheld_for_review: batch.proposals.length,\n    }\n  }\n\n  const findingsById = new Map(findings.map((f) => [f.finding_id, f]))\n  const safe: TProposal[] = []\n  let withheld = 0\n  for (const p of batch.proposals as Array<TProposal & { sourceFindingId?: string }>) {\n    const src = p.sourceFindingId ? findingsById.get(p.sourceFindingId) : undefined\n    if (!src) {\n      withheld += 1\n      continue\n    }\n    if (src.confidence < threshold) {\n      withheld += 1\n      continue\n    }\n    safe.push(p)\n  }\n  const result = await adapter.apply(safe)\n  log('knowledge.apply', {\n    applied: result.written.length,\n    withheld_for_review: withheld,\n    warnings: result.warnings.length,\n  })\n  await emit({\n    type: 'knowledge-applied',\n    runId: opts.runId,\n    writtenCount: result.written.length,\n    withheldForReview: withheld,\n  })\n  return {\n    proposals: batch.proposals as TProposal[],\n    applied: result.written,\n    skipped: batch.skipped,\n    errors: batch.errors,\n    withheld_for_review: withheld,\n  }\n}\n\nasync function runImprovementAdapter<TEdit>(\n  opts: RunAnalystLoopOpts,\n  findings: ReadonlyArray<AnalystFinding>,\n  log: NonNullable<RunAnalystLoopOpts['log']>,\n  emit: Emitter,\n): Promise<ImprovementReport<TEdit>> {\n  const adapter = opts.improvementAdapter!\n  const batch = await adapter.proposeFromFindings(findings)\n  log('improvement.proposeFromFindings', {\n    edits: batch.edits.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n  await emit({\n    type: 'improvement-proposed',\n    runId: opts.runId,\n    editCount: batch.edits.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n\n  const auto = opts.autoApply?.improvement ?? false\n  const threshold = opts.autoApply?.improvementConfidenceThreshold ?? 0.9\n\n  if (!auto || !adapter.apply) {\n    await emit({\n      type: 'improvement-applied',\n      runId: opts.runId,\n      appliedCount: 0,\n      withheldForReview: batch.edits.length,\n    })\n    return {\n      edits: batch.edits as TEdit[],\n      applied: [],\n      skipped: batch.skipped,\n      errors: batch.errors,\n      withheld_for_review: batch.edits.length,\n    }\n  }\n\n  const findingsById = new Map(findings.map((f) => [f.finding_id, f]))\n  const safe: TEdit[] = []\n  let withheld = 0\n  for (const e of batch.edits as Array<TEdit & { sourceFindingId?: string }>) {\n    const src = e.sourceFindingId ? findingsById.get(e.sourceFindingId) : undefined\n    if (!src || src.confidence < threshold) {\n      withheld += 1\n      continue\n    }\n    safe.push(e)\n  }\n  const result = await adapter.apply(safe)\n  log('improvement.apply', {\n    applied: result.applied.length,\n    withheld_for_review: withheld,\n    warnings: result.warnings.length,\n  })\n  await emit({\n    type: 'improvement-applied',\n    runId: opts.runId,\n    appliedCount: result.applied.length,\n    withheldForReview: withheld,\n  })\n  return {\n    edits: batch.edits as TEdit[],\n    applied: result.applied,\n    skipped: batch.skipped,\n    errors: batch.errors,\n    withheld_for_review: withheld,\n  }\n}\n\nfunction defaultLog(msg: string, fields?: Record<string, unknown>): void {\n  if (fields) console.log(`[analyst-loop] ${msg}`, fields)\n  else console.log(`[analyst-loop] ${msg}`)\n}\n"],"mappings":";;;AAoBA,SAAS,oBAAoB;AAW7B,eAAsB,eACpB,MACiD;AACjD,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,WAAW,KAAK,yBAAyB;AAC/C,QAAM,OAAO,YAAY,KAAK,OAAO;AACrC,QAAM,YAAY,KAAK,IAAI;AAG3B,QAAM,gBAAgB,qBAAqB,IAAI;AAC/C,QAAM,WAA+D,gBAChE,KAAK,eAAe,QAAQ,aAAa,KAAK,CAAC,IAChD,CAAC;AACL,MAAI,qBAAqB,EAAE,eAAe,gBAAgB,SAAS,OAAO,CAAC;AAC3E,QAAM,KAAK;AAAA,IACT,MAAM;AAAA,IACN,OAAO,KAAK;AAAA,IACZ;AAAA,IACA,mBAAmB,SAAS;AAAA,EAC9B,CAAC;AAKD,QAAM,gBAAgB,wBAAwB,UAAU,UAAU,KAAK,SAAS,KAAK,CAAC;AACtF,QAAM,gBAAgB,MAAM,YAAY,MAAM,eAAe,IAAI;AACjE,MAAI,wBAAwB;AAAA,IAC1B,UAAU,cAAc,SAAS;AAAA,IACjC,UAAU,cAAc;AAAA,IACxB,aAAa,cAAc,YAAY,IAAI,CAAC,OAAO;AAAA,MACjD,IAAI,EAAE;AAAA,MACN,QAAQ,EAAE;AAAA,MACV,GAAG,EAAE;AAAA,IACP,EAAE;AAAA,EACJ,CAAC;AAID,MAAI,KAAK,iBAAiB,cAAc,SAAS,SAAS,GAAG;AAC3D,UAAM,KAAK,cAAc,OAAO,KAAK,OAAO,cAAc,QAAQ;AAClE,UAAM,KAAK;AAAA,MACT,MAAM;AAAA,MACN,OAAO,KAAK;AAAA,MACZ,OAAO,cAAc,SAAS;AAAA,IAChC,CAAC;AAAA,EACH;AAGA,MAAI,OAA4B;AAChC,MAAI,iBAAiB,cAAc,SAAS,SAAS,GAAG;AACtD,WAAO;AAAA,MACL,SAAS,IAAI,CAAC,OAAO,EAAE,GAAG,EAAE,EAAE;AAAA,MAC9B,cAAc,SAAS,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,QAAQ,KAAK,MAAM,EAAE;AAAA,IAClE;AACA,QAAI,oBAAoB;AAAA,MACtB,UAAU,KAAK,SAAS;AAAA,MACxB,aAAa,KAAK,YAAY;AAAA,MAC9B,WAAW,KAAK,UAAU;AAAA,MAC1B,SAAS,KAAK,QAAQ;AAAA,IACxB,CAAC;AACD,UAAM,KAAK;AAAA,MACT,MAAM;AAAA,MACN,OAAO,KAAK;AAAA,MACZ;AAAA,MACA,UAAU,KAAK,SAAS;AAAA,MACxB,aAAa,KAAK,YAAY;AAAA,MAC9B,WAAW,KAAK,UAAU;AAAA,MAC1B,SAAS,KAAK,QAAQ;AAAA,IACxB,CAAC;AAAA,EACH;AAGA,MAAI,YAA+C;AACnD,MAAI,KAAK,kBAAkB;AACzB,gBAAY,MAAM,oBAAoB,MAAM,cAAc,UAAU,KAAK,IAAI;AAAA,EAC/E;AAGA,MAAI,cAA+C;AACnD,MAAI,KAAK,oBAAoB;AAC3B,kBAAc,MAAM,sBAAsB,MAAM,cAAc,UAAU,KAAK,IAAI;AAAA,EACnF;AAEA,QAAM,KAAK;AAAA,IACT,MAAM;AAAA,IACN,OAAO,KAAK;AAAA,IACZ,YAAY,KAAK,IAAI,IAAI;AAAA,EAC3B,CAAC;AAED,SAAO;AAAA,IACL,OAAO,KAAK;AAAA,IACZ;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAIA,SAAS,YAAY,SAAiD;AACpE,MAAI,CAAC,QAAS,QAAO,YAAY;AAAA,EAAC;AAClC,SAAO,OAAO,UAAU;AACtB,UAAM,QAAQ,KAAK;AAAA,EACrB;AACF;AAEA,eAAe,YACb,MACA,eACA,MAC2B;AAC3B,QAAM,MAAM,KAAK;AACjB,MAAI,OAAO,IAAI,cAAc,cAAc,KAAK,SAAS;AACvD,QAAI,QAAiC;AACrC,qBAAiB,MAAM,IAAI,UAAU,KAAK,OAAO,KAAK,QAAQ,EAAE,cAAc,CAAC,GAAG;AAChF,YAAM,KAAK,EAAE,MAAM,WAAW,OAAO,KAAK,OAAO,OAAO,GAAG,CAAC;AAC5D,UAAI,GAAG,SAAS,gBAAiB,SAAQ,GAAG;AAAA,IAC9C;AACA,QAAI,CAAC,OAAO;AACV,YAAM,IAAI,MAAM,sEAAsE;AAAA,IACxF;AACA,WAAO;AAAA,EACT;AACA,SAAO,KAAK,SAAS,IAAI,KAAK,OAAO,KAAK,QAAQ,EAAE,cAAc,CAAC;AACrE;AAEA,SAAS,qBAAqB,MAAyC;AACrE,MAAI,KAAK,kBAAkB,KAAM,QAAO;AACxC,MAAI,OAAO,KAAK,kBAAkB,SAAU,QAAO,KAAK;AACxD,MAAI,CAAC,KAAK,cAAe,QAAO;AAChC,QAAM,MAAM,KAAK,cAAc,QAAQ;AACvC,MAAI,OAAsB;AAC1B,aAAW,OAAO,KAAK;AACrB,QAAI,IAAI,WAAW,KAAK,MAAO;AAC/B,WAAO,IAAI;AAAA,EACb;AACA,SAAO;AACT;AAEA,SAAS,wBACP,OACA,UACA,UAC2F;AAC3F,MAAI,aAAa,UAAU,MAAM,WAAW,EAAG,QAAO;AACtD,QAAM,WAAW,MAAM,IAAI,CAAC,EAAE,QAAQ,SAAS,GAAG,KAAK,MAAM,IAAsB;AACnF,MAAI,aAAa,YAAY;AAC3B,WAAO,EAAE,KAAK,SAAS;AAAA,EACzB;AACA,OAAK;AACL,SAAO;AACT;AAEA,eAAe,oBACb,MACA,UACA,KACA,MACqC;AACrC,QAAM,UAAU,KAAK;AACrB,QAAM,QAAQ,MAAM,QAAQ,oBAAoB,QAAQ;AACxD,MAAI,iCAAiC;AAAA,IACnC,WAAW,MAAM,UAAU;AAAA,IAC3B,SAAS,MAAM;AAAA,IACf,QAAQ,MAAM,OAAO;AAAA,EACvB,CAAC;AACD,QAAM,KAAK;AAAA,IACT,MAAM;AAAA,IACN,OAAO,KAAK;AAAA,IACZ,eAAe,MAAM,UAAU;AAAA,IAC/B,SAAS,MAAM;AAAA,IACf,QAAQ,MAAM,OAAO;AAAA,EACvB,CAAC;AAED,QAAM,OAAO,KAAK,WAAW,aAAa;AAC1C,QAAM,YAAY,KAAK,WAAW,gCAAgC;AAElE,MAAI,CAAC,QAAQ,CAAC,QAAQ,OAAO;AAC3B,UAAM,KAAK;AAAA,MACT,MAAM;AAAA,MACN,OAAO,KAAK;AAAA,MACZ,cAAc;AAAA,MACd,mBAAmB,MAAM,UAAU;AAAA,IACrC,CAAC;AACD,WAAO;AAAA,MACL,WAAW,MAAM;AAAA,MACjB,SAAS,CAAC;AAAA,MACV,SAAS,MAAM;AAAA,MACf,QAAQ,MAAM;AAAA,MACd,qBAAqB,MAAM,UAAU;AAAA,IACvC;AAAA,EACF;AAEA,QAAM,eAAe,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;AACnE,QAAM,OAAoB,CAAC;AAC3B,MAAI,WAAW;AACf,aAAW,KAAK,MAAM,WAA8D;AAClF,UAAM,MAAM,EAAE,kBAAkB,aAAa,IAAI,EAAE,eAAe,IAAI;AACtE,QAAI,CAAC,KAAK;AACR,kBAAY;AACZ;AAAA,IACF;AACA,QAAI,IAAI,aAAa,WAAW;AAC9B,kBAAY;AACZ;AAAA,IACF;AACA,SAAK,KAAK,CAAC;AAAA,EACb;AACA,QAAM,SAAS,MAAM,QAAQ,MAAM,IAAI;AACvC,MAAI,mBAAmB;AAAA,IACrB,SAAS,OAAO,QAAQ;AAAA,IACxB,qBAAqB;AAAA,IACrB,UAAU,OAAO,SAAS;AAAA,EAC5B,CAAC;AACD,QAAM,KAAK;AAAA,IACT,MAAM;AAAA,IACN,OAAO,KAAK;AAAA,IACZ,cAAc,OAAO,QAAQ;AAAA,IAC7B,mBAAmB;AAAA,EACrB,CAAC;AACD,SAAO;AAAA,IACL,WAAW,MAAM;AAAA,IACjB,SAAS,OAAO;AAAA,IAChB,SAAS,MAAM;AAAA,IACf,QAAQ,MAAM;AAAA,IACd,qBAAqB;AAAA,EACvB;AACF;AAEA,eAAe,sBACb,MACA,UACA,KACA,MACmC;AACnC,QAAM,UAAU,KAAK;AACrB,QAAM,QAAQ,MAAM,QAAQ,oBAAoB,QAAQ;AACxD,MAAI,mCAAmC;AAAA,IACrC,OAAO,MAAM,MAAM;AAAA,IACnB,SAAS,MAAM;AAAA,IACf,QAAQ,MAAM,OAAO;AAAA,EACvB,CAAC;AACD,QAAM,KAAK;AAAA,IACT,MAAM;AAAA,IACN,OAAO,KAAK;AAAA,IACZ,WAAW,MAAM,MAAM;AAAA,IACvB,SAAS,MAAM;AAAA,IACf,QAAQ,MAAM,OAAO;AAAA,EACvB,CAAC;AAED,QAAM,OAAO,KAAK,WAAW,eAAe;AAC5C,QAAM,YAAY,KAAK,WAAW,kCAAkC;AAEpE,MAAI,CAAC,QAAQ,CAAC,QAAQ,OAAO;AAC3B,UAAM,KAAK;AAAA,MACT,MAAM;AAAA,MACN,OAAO,KAAK;AAAA,MACZ,cAAc;AAAA,MACd,mBAAmB,MAAM,MAAM;AAAA,IACjC,CAAC;AACD,WAAO;AAAA,MACL,OAAO,MAAM;AAAA,MACb,SAAS,CAAC;AAAA,MACV,SAAS,MAAM;AAAA,MACf,QAAQ,MAAM;AAAA,MACd,qBAAqB,MAAM,MAAM;AAAA,IACnC;AAAA,EACF;AAEA,QAAM,eAAe,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;AACnE,QAAM,OAAgB,CAAC;AACvB,MAAI,WAAW;AACf,aAAW,KAAK,MAAM,OAAsD;AAC1E,UAAM,MAAM,EAAE,kBAAkB,aAAa,IAAI,EAAE,eAAe,IAAI;AACtE,QAAI,CAAC,OAAO,IAAI,aAAa,WAAW;AACtC,kBAAY;AACZ;AAAA,IACF;AACA,SAAK,KAAK,CAAC;AAAA,EACb;AACA,QAAM,SAAS,MAAM,QAAQ,MAAM,IAAI;AACvC,MAAI,qBAAqB;AAAA,IACvB,SAAS,OAAO,QAAQ;AAAA,IACxB,qBAAqB;AAAA,IACrB,UAAU,OAAO,SAAS;AAAA,EAC5B,CAAC;AACD,QAAM,KAAK;AAAA,IACT,MAAM;AAAA,IACN,OAAO,KAAK;AAAA,IACZ,cAAc,OAAO,QAAQ;AAAA,IAC7B,mBAAmB;AAAA,EACrB,CAAC;AACD,SAAO;AAAA,IACL,OAAO,MAAM;AAAA,IACb,SAAS,OAAO;AAAA,IAChB,SAAS,MAAM;AAAA,IACf,QAAQ,MAAM;AAAA,IACd,qBAAqB;AAAA,EACvB;AACF;AAEA,SAAS,WAAW,KAAa,QAAwC;AACvE,MAAI,OAAQ,SAAQ,IAAI,kBAAkB,GAAG,IAAI,MAAM;AAAA,MAClD,SAAQ,IAAI,kBAAkB,GAAG,EAAE;AAC1C;","names":[]}
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/chunk-VOX6Z3II.js

@@ -0,0 +1,90 @@
+import {
+  ConfigError
+} from "./chunk-SQSCRJ7U.js";
+
+// src/improvement/optimize-prompt.ts
+import { gepaDriver, heldOutGate, runImprovementLoop } from "@tangle-network/agent-eval/campaign";
+async function optimizePrompt(opts) {
+  if (!opts.driver && !opts.reflection) {
+    throw new ConfigError(
+      "optimizePrompt: pass `reflection` (builds the default gepaDriver) or a custom `driver`"
+    );
+  }
+  if (opts.scenarios.length === 0) {
+    throw new ConfigError("optimizePrompt: `scenarios` must be non-empty");
+  }
+  if (opts.holdoutScenarios.length === 0) {
+    throw new ConfigError(
+      "optimizePrompt: `holdoutScenarios` must be non-empty (the gate needs it)"
+    );
+  }
+  const driver = opts.driver ?? gepaDriver({
+    llm: opts.reflection.llm,
+    model: opts.reflection.model,
+    target: opts.reflection.target ?? "system prompt",
+    mutationPrimitives: opts.reflection.mutationPrimitives,
+    constraints: opts.reflection.preserveSections || opts.reflection.maxSentenceEdits !== void 0 ? {
+      preserveSections: opts.reflection.preserveSections,
+      maxSentenceEdits: opts.reflection.maxSentenceEdits
+    } : void 0
+  });
+  const gate = opts.gate ?? heldOutGate({
+    scenarios: opts.holdoutScenarios,
+    ...opts.deltaThreshold !== void 0 ? { deltaThreshold: opts.deltaThreshold } : {}
+  });
+  const result = await runImprovementLoop({
+    baselineSurface: opts.baselinePrompt,
+    dispatchWithSurface: (surface, scenario, ctx) => {
+      if (typeof surface !== "string") {
+        throw new ConfigError(
+          "optimizePrompt: received a CodeSurface \u2014 this entry point optimizes string prompts only"
+        );
+      }
+      return opts.runWithPrompt(surface, scenario, ctx);
+    },
+    driver,
+    populationSize: opts.populationSize ?? 4,
+    maxGenerations: opts.maxGenerations ?? 3,
+    ...opts.promoteTopK !== void 0 ? { promoteTopK: opts.promoteTopK } : {},
+    scenarios: opts.scenarios,
+    holdoutScenarios: opts.holdoutScenarios,
+    judges: opts.judges,
+    gate,
+    autoOnPromote: opts.autoOnPromote ?? "none",
+    ...opts.ghOwner !== void 0 ? { ghOwner: opts.ghOwner } : {},
+    ...opts.ghRepo !== void 0 ? { ghRepo: opts.ghRepo } : {},
+    runDir: opts.runDir,
+    ...opts.storage !== void 0 ? { storage: opts.storage } : {},
+    ...opts.seed !== void 0 ? { seed: opts.seed } : {},
+    ...opts.reps !== void 0 ? { reps: opts.reps } : {},
+    ...opts.maxConcurrency !== void 0 ? { maxConcurrency: opts.maxConcurrency } : {},
+    ...opts.now !== void 0 ? { now: opts.now } : {}
+  });
+  const improved = result.gateResult.decision === "ship";
+  const winnerSurface = typeof result.winnerSurface === "string" ? result.winnerSurface : opts.baselinePrompt;
+  const baselineComposite = meanComposite(result.baselineOnHoldout);
+  const winnerComposite = meanComposite(result.winnerOnHoldout);
+  return {
+    prompt: improved ? winnerSurface : opts.baselinePrompt,
+    improved,
+    decision: result.gateResult.decision,
+    reasons: result.gateResult.reasons,
+    baselineComposite,
+    winnerComposite,
+    delta: result.gateResult.delta ?? winnerComposite - baselineComposite,
+    ...improved && result.winnerRationale ? { rationale: result.winnerRationale } : {},
+    diff: result.promotedDiff,
+    raw: result
+  };
+}
+function meanComposite(campaign) {
+  const scenarios = Object.values(campaign.aggregates.byScenario);
+  if (scenarios.length === 0) return 0;
+  const sum = scenarios.reduce((acc, s) => acc + s.meanComposite, 0);
+  return sum / scenarios.length;
+}
+
+export {
+  optimizePrompt
+};
+//# sourceMappingURL=chunk-VOX6Z3II.js.map

package/dist/chunk-VOX6Z3II.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/improvement/optimize-prompt.ts"],"sourcesContent":["/**\n * @experimental\n *\n * `optimizePrompt` — identity-gated optimization for any TEXT prompt surface\n * (system prompt, planner prompt, judge rubric, skill doc).\n *\n * The text-surface sibling to this module's `improvementDriver` (the\n * CODE-surface / worktree path). Both feed agent-eval's `runImprovementLoop`;\n * this one defaults the driver to agent-eval's `gepaDriver` (reflective text\n * mutator) and the gate to `heldOutGate`.\n *\n * IDENTITY-GATED BY CONSTRUCTION — the whole point. The loop runs evals,\n * collects per-scenario signal, proposes candidates, and the gate compares\n * candidate-vs-baseline ON THE HELDOUT. `result.prompt` is the baseline\n * (identity) UNLESS the gate decided `'ship'`. So wiring a surface up is safe:\n * a surface with no beneficial mutation simply keeps its baseline. You never\n * regress by registering a prompt — you only ever improve when the held-out\n * data earns it.\n *\n * Generic over the runtime: `runWithPrompt` is the only domain seam — given a\n * candidate prompt + scenario, run it however the surface runs (sandbox\n * `streamPrompt`, a `runLoop`, a direct model call) and return the artifact the\n * judges score. The optimizer never assumes how a prompt is executed.\n */\n\nimport type { LlmClientOptions } from '@tangle-network/agent-eval'\nimport type {\n  CampaignResult,\n  CampaignStorage,\n  DispatchContext,\n  Gate,\n  GateResult,\n  ImprovementDriver,\n  JudgeConfig,\n  RunImprovementLoopResult,\n  Scenario,\n} from '@tangle-network/agent-eval/campaign'\nimport { gepaDriver, heldOutGate, runImprovementLoop } from '@tangle-network/agent-eval/campaign'\nimport { ConfigError } from '../errors'\n\n/** Reflection config for the default `gepaDriver`. Omit when passing a custom\n *  `driver`. */\nexport interface OptimizePromptReflection {\n  /** Router transport for the reflection model. */\n  llm: LlmClientOptions\n  /** Model that performs the reflective rewrite. */\n  model: string\n  /** What is being optimized — orients the reflection prompt. Default\n   *  `'system prompt'`. */\n  target?: string\n  /** Surface-specific mutation levers offered to the reflector. */\n  mutationPrimitives?: string[]\n  /** H2 (`## Foo`) headings that MUST survive every candidate. gepaDriver's\n   *  only structural guard — load-bearing sections of the prompt should be\n   *  `##` headings so a rewrite cannot drop them. */\n  preserveSections?: string[]\n  /** Max sentence-level edits per candidate vs the parent (a textual learning\n   *  rate). Caps a rewrite from wiping prior rules in one generation. */\n  maxSentenceEdits?: number\n}\n\n/** @experimental */\nexport interface OptimizePromptOptions<TScenario extends Scenario, TArtifact> {\n  /** The prompt being optimized — the identity baseline the gate protects. */\n  baselinePrompt: string\n  /** Domain seam: run a candidate prompt against a scenario → artifact the\n   *  judges score. The optimizer is agnostic to HOW the prompt runs. */\n  runWithPrompt: (prompt: string, scenario: TScenario, ctx: DispatchContext) => Promise<TArtifact>\n  /** Training pool — scored each generation to rank candidates. */\n  scenarios: TScenario[]\n  /** Held out of training — scored ONLY for the gate's baseline-vs-winner\n   *  delta. Disjoint from `scenarios`; this is what makes promotion measure\n   *  generalization, not memorization. */\n  holdoutScenarios: TScenario[]\n  /** Scorers — deterministic checks or LLM judges. */\n  judges: JudgeConfig<TArtifact, TScenario>[]\n  /** Where artifacts + traces land (opaque key under in-memory storage). */\n  runDir: string\n  /** Default driver = `gepaDriver` built from this. Required UNLESS `driver`\n   *  is supplied. */\n  reflection?: OptimizePromptReflection\n  /** Override the improvement strategy (custom driver / deterministic tests). */\n  driver?: ImprovementDriver\n  /** Override the promotion gate. Default `heldOutGate` over `holdoutScenarios`\n   *  — zero extra LLM. Wrap `defaultProductionGate` for red-team/reward-hacking\n   *  hardening on production wiring. */\n  gate?: Gate<TArtifact, TScenario>\n  /** Minimum held-out composite lift to ship, forwarded to the default\n   *  `heldOutGate`. When omitted the gate uses its own default. */\n  deltaThreshold?: number\n  /** Candidates proposed per generation. Default 4. */\n  populationSize?: number\n  /** Generations to run. Default 3. */\n  maxGenerations?: number\n  /** Candidates carried to the next generation. Default 2. */\n  promoteTopK?: number\n  /** Storage backend. Pass `inMemoryCampaignStorage()` for filesystem-less /\n   *  test runs. Default: Node filesystem. */\n  storage?: CampaignStorage\n  /** Reproducibility seed. Default 42. */\n  seed?: number\n  /** Per-scenario replicates for CI bands. Default 1. */\n  reps?: number\n  /** Max concurrent cells. Default 2. */\n  maxConcurrency?: number\n  /** Test seam — override the wall clock. */\n  now?: () => Date\n  /** On a shipped gate: `'pr'` opens a PR, `'none'` just reports. Default\n   *  `'none'`. */\n  autoOnPromote?: 'pr' | 'none'\n  ghOwner?: string\n  ghRepo?: string\n}\n\n/** @experimental */\nexport interface OptimizePromptResult<TArtifact, TScenario extends Scenario> {\n  /** The prompt to USE. Identity (the baseline) unless the gate shipped a\n   *  winner — so a caller can always assign `result.prompt` unconditionally. */\n  prompt: string\n  /** True only when the gate promoted a candidate over baseline on holdout. */\n  improved: boolean\n  /** The gate's verdict (`'ship' | 'hold' | 'need_more_work' | ...`). */\n  decision: GateResult['decision']\n  /** Human-readable reasons the gate gave. */\n  reasons: string[]\n  /** Mean held-out composite of the baseline. */\n  baselineComposite: number\n  /** Mean held-out composite of the winner candidate. */\n  winnerComposite: number\n  /** Held-out lift (winner − baseline); the gate's `delta` when it reported one. */\n  delta: number\n  /** Why the winner was proposed — present when a shipped winner carried a\n   *  driver rationale. */\n  rationale?: string\n  /** Unified baseline→winner diff (empty when the winner is the baseline). */\n  diff: string\n  /** The full loop result for callers that need generations / campaigns. */\n  raw: RunImprovementLoopResult<TArtifact, TScenario>\n}\n\n/** @experimental */\nexport async function optimizePrompt<TScenario extends Scenario, TArtifact>(\n  opts: OptimizePromptOptions<TScenario, TArtifact>,\n): Promise<OptimizePromptResult<TArtifact, TScenario>> {\n  if (!opts.driver && !opts.reflection) {\n    throw new ConfigError(\n      'optimizePrompt: pass `reflection` (builds the default gepaDriver) or a custom `driver`',\n    )\n  }\n  if (opts.scenarios.length === 0) {\n    throw new ConfigError('optimizePrompt: `scenarios` must be non-empty')\n  }\n  if (opts.holdoutScenarios.length === 0) {\n    throw new ConfigError(\n      'optimizePrompt: `holdoutScenarios` must be non-empty (the gate needs it)',\n    )\n  }\n\n  const driver =\n    opts.driver ??\n    gepaDriver({\n      llm: opts.reflection!.llm,\n      model: opts.reflection!.model,\n      target: opts.reflection!.target ?? 'system prompt',\n      mutationPrimitives: opts.reflection!.mutationPrimitives,\n      constraints:\n        opts.reflection!.preserveSections || opts.reflection!.maxSentenceEdits !== undefined\n          ? {\n              preserveSections: opts.reflection!.preserveSections,\n              maxSentenceEdits: opts.reflection!.maxSentenceEdits,\n            }\n          : undefined,\n    })\n\n  const gate =\n    opts.gate ??\n    heldOutGate<TArtifact, TScenario>({\n      scenarios: opts.holdoutScenarios,\n      ...(opts.deltaThreshold !== undefined ? { deltaThreshold: opts.deltaThreshold } : {}),\n    })\n\n  const result = await runImprovementLoop<TScenario, TArtifact>({\n    baselineSurface: opts.baselinePrompt,\n    dispatchWithSurface: (surface, scenario, ctx) => {\n      if (typeof surface !== 'string') {\n        // optimizePrompt is the TEXT-surface entry point; a CodeSurface means\n        // the caller wired the wrong driver. Fail loud — don't silently run the\n        // baseline and report a phantom score.\n        throw new ConfigError(\n          'optimizePrompt: received a CodeSurface — this entry point optimizes string prompts only',\n        )\n      }\n      return opts.runWithPrompt(surface, scenario, ctx)\n    },\n    driver,\n    populationSize: opts.populationSize ?? 4,\n    maxGenerations: opts.maxGenerations ?? 3,\n    ...(opts.promoteTopK !== undefined ? { promoteTopK: opts.promoteTopK } : {}),\n    scenarios: opts.scenarios,\n    holdoutScenarios: opts.holdoutScenarios,\n    judges: opts.judges,\n    gate,\n    autoOnPromote: opts.autoOnPromote ?? 'none',\n    ...(opts.ghOwner !== undefined ? { ghOwner: opts.ghOwner } : {}),\n    ...(opts.ghRepo !== undefined ? { ghRepo: opts.ghRepo } : {}),\n    runDir: opts.runDir,\n    ...(opts.storage !== undefined ? { storage: opts.storage } : {}),\n    ...(opts.seed !== undefined ? { seed: opts.seed } : {}),\n    ...(opts.reps !== undefined ? { reps: opts.reps } : {}),\n    ...(opts.maxConcurrency !== undefined ? { maxConcurrency: opts.maxConcurrency } : {}),\n    ...(opts.now !== undefined ? { now: opts.now } : {}),\n  })\n\n  const improved = result.gateResult.decision === 'ship'\n  const winnerSurface =\n    typeof result.winnerSurface === 'string' ? result.winnerSurface : opts.baselinePrompt\n  const baselineComposite = meanComposite(result.baselineOnHoldout)\n  const winnerComposite = meanComposite(result.winnerOnHoldout)\n\n  return {\n    prompt: improved ? winnerSurface : opts.baselinePrompt,\n    improved,\n    decision: result.gateResult.decision,\n    reasons: result.gateResult.reasons,\n    baselineComposite,\n    winnerComposite,\n    delta: result.gateResult.delta ?? winnerComposite - baselineComposite,\n    ...(improved && result.winnerRationale ? { rationale: result.winnerRationale } : {}),\n    diff: result.promotedDiff,\n    raw: result,\n  }\n}\n\n/** Mean composite over a campaign's per-scenario aggregates. The held-out\n *  campaigns score one surface across `holdoutScenarios`; averaging the\n *  per-scenario means gives the single number the gate's delta is built from. */\nfunction meanComposite(campaign: CampaignResult<unknown, Scenario>): number {\n  const scenarios = Object.values(campaign.aggregates.byScenario)\n  if (scenarios.length === 0) return 0\n  const sum = scenarios.reduce((acc, s) => acc + s.meanComposite, 0)\n  return sum / scenarios.length\n}\n"],"mappings":";;;;;AAqCA,SAAS,YAAY,aAAa,0BAA0B;AAwG5D,eAAsB,eACpB,MACqD;AACrD,MAAI,CAAC,KAAK,UAAU,CAAC,KAAK,YAAY;AACpC,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,MAAI,KAAK,UAAU,WAAW,GAAG;AAC/B,UAAM,IAAI,YAAY,+CAA+C;AAAA,EACvE;AACA,MAAI,KAAK,iBAAiB,WAAW,GAAG;AACtC,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AAEA,QAAM,SACJ,KAAK,UACL,WAAW;AAAA,IACT,KAAK,KAAK,WAAY;AAAA,IACtB,OAAO,KAAK,WAAY;AAAA,IACxB,QAAQ,KAAK,WAAY,UAAU;AAAA,IACnC,oBAAoB,KAAK,WAAY;AAAA,IACrC,aACE,KAAK,WAAY,oBAAoB,KAAK,WAAY,qBAAqB,SACvE;AAAA,MACE,kBAAkB,KAAK,WAAY;AAAA,MACnC,kBAAkB,KAAK,WAAY;AAAA,IACrC,IACA;AAAA,EACR,CAAC;AAEH,QAAM,OACJ,KAAK,QACL,YAAkC;AAAA,IAChC,WAAW,KAAK;AAAA,IAChB,GAAI,KAAK,mBAAmB,SAAY,EAAE,gBAAgB,KAAK,eAAe,IAAI,CAAC;AAAA,EACrF,CAAC;AAEH,QAAM,SAAS,MAAM,mBAAyC;AAAA,IAC5D,iBAAiB,KAAK;AAAA,IACtB,qBAAqB,CAAC,SAAS,UAAU,QAAQ;AAC/C,UAAI,OAAO,YAAY,UAAU;AAI/B,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AACA,aAAO,KAAK,cAAc,SAAS,UAAU,GAAG;AAAA,IAClD;AAAA,IACA;AAAA,IACA,gBAAgB,KAAK,kBAAkB;AAAA,IACvC,gBAAgB,KAAK,kBAAkB;AAAA,IACvC,GAAI,KAAK,gBAAgB,SAAY,EAAE,aAAa,KAAK,YAAY,IAAI,CAAC;AAAA,IAC1E,WAAW,KAAK;AAAA,IAChB,kBAAkB,KAAK;AAAA,IACvB,QAAQ,KAAK;AAAA,IACb;AAAA,IACA,eAAe,KAAK,iBAAiB;AAAA,IACrC,GAAI,KAAK,YAAY,SAAY,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;AAAA,IAC9D,GAAI,KAAK,WAAW,SAAY,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;AAAA,IAC3D,QAAQ,KAAK;AAAA,IACb,GAAI,KAAK,YAAY,SAAY,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;AAAA,IAC9D,GAAI,KAAK,SAAS,SAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;AAAA,IACrD,GAAI,KAAK,SAAS,SAAY,EAAE,MAAM,KAAK,KAAK,IAAI,CAAC;AAAA,IACrD,GAAI,KAAK,mBAAmB,SAAY,EAAE,gBAAgB,KAAK,eAAe,IAAI,CAAC;AAAA,IACnF,GAAI,KAAK,QAAQ,SAAY,EAAE,KAAK,KAAK,IAAI,IAAI,CAAC;AAAA,EACpD,CAAC;AAED,QAAM,WAAW,OAAO,WAAW,aAAa;AAChD,QAAM,gBACJ,OAAO,OAAO,kBAAkB,WAAW,OAAO,gBAAgB,KAAK;AACzE,QAAM,oBAAoB,cAAc,OAAO,iBAAiB;AAChE,QAAM,kBAAkB,cAAc,OAAO,eAAe;AAE5D,SAAO;AAAA,IACL,QAAQ,WAAW,gBAAgB,KAAK;AAAA,IACxC;AAAA,IACA,UAAU,OAAO,WAAW;AAAA,IAC5B,SAAS,OAAO,WAAW;AAAA,IAC3B;AAAA,IACA;AAAA,IACA,OAAO,OAAO,WAAW,SAAS,kBAAkB;AAAA,IACpD,GAAI,YAAY,OAAO,kBAAkB,EAAE,WAAW,OAAO,gBAAgB,IAAI,CAAC;AAAA,IAClF,MAAM,OAAO;AAAA,IACb,KAAK;AAAA,EACP;AACF;AAKA,SAAS,cAAc,UAAqD;AAC1E,QAAM,YAAY,OAAO,OAAO,SAAS,WAAW,UAAU;AAC9D,MAAI,UAAU,WAAW,EAAG,QAAO;AACnC,QAAM,MAAM,UAAU,OAAO,CAAC,KAAK,MAAM,MAAM,EAAE,eAAe,CAAC;AACjE,SAAO,MAAM,UAAU;AACzB;","names":[]}