@tangle-network/agent-runtime 0.37.0 → 0.39.0

package/dist/chunk-XBUG326M.js

@@ -0,0 +1,261 @@
+// 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
+};
+//# sourceMappingURL=chunk-XBUG326M.js.map

package/dist/chunk-XBUG326M.js.map

@@ -0,0 +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":[]}

package/dist/dynamic-DcrwVGuV.d.ts

@@ -0,0 +1,106 @@
+import { I as Iteration, D as Driver } from './types-B9O7l-ij.js';
+
+/**
+ * @experimental
+ *
+ * Dynamic driver — the agent authors the loop topology at runtime.
+ *
+ * Where `refine` and `fanout-vote` encode a fixed shape as a pure function of
+ * history, this driver delegates the per-round shape to an injected
+ * `TopologyPlanner`. Each round the planner inspects the task + iteration
+ * history and emits one `TopologyMove`:
+ *   - `refine` → one task next round (optionally rewritten from the prior attempt)
+ *   - `fanout` → N tasks next round (the kernel round-robins `agentRuns`, so a
+ *     2-harness fanout dispatches branch 0 to harness A and branch 1 to harness B)
+ *   - `stop`   → terminate; the kernel selects the winner across all iterations
+ *
+ * The planner is the brain; this driver is the structure. It maps moves onto
+ * the kernel's `plan`/`decide` contract, enforces the iteration + fanout caps,
+ * and fails loud on a malformed move. The planner is injected exactly like
+ * `refine`'s `refineTask` and `fanout-vote`'s `selector` — so a test can drive
+ * a deterministic policy through the real kernel, and production can wire it to
+ * an LLM via `createSandboxPlanner`.
+ *
+ * Topology is orthogonal to harness: the planner never names a backend. Which
+ * harness runs a branch is decided by the `AgentRunSpec` the kernel round-robins
+ * to, so one dynamic driver works across claude-code, codex, opencode, pi —
+ * including fanning a single round across several at once.
+ */
+
+/** Terminal once `decide` returns `'done'` (a kernel terminal decision). */
+type DynamicDecision = 'continue' | 'done';
+/**
+ * One topology decision for the next round. `fanout` carries explicit tasks
+ * rather than a count so the planner can issue heterogeneous branches (a
+ * different sub-task per harness); pass N copies of one task for a homogeneous
+ * fanout that relies on `agentRuns` diversity instead.
+ *
+ * @experimental
+ */
+type TopologyMove<Task> = {
+    kind: 'refine';
+    task: Task;
+    rationale?: string;
+} | {
+    kind: 'fanout';
+    tasks: Task[];
+    rationale?: string;
+} | {
+    kind: 'stop';
+    rationale?: string;
+};
+/** @experimental */
+interface PlannerContext<Task, Output> {
+    /** The root task the loop was invoked with — stable across rounds. */
+    task: Task;
+    /** Every iteration so far, in dispatch order, with outputs + verdicts. */
+    history: ReadonlyArray<Iteration<Task, Output>>;
+    /** `history.length` — iterations already spent. */
+    iterationsSpent: number;
+    /** Iterations left before the driver's `maxIterations` cap forces a stop. */
+    iterationsRemaining: number;
+}
+/**
+ * Chooses the next topology move from the task + history. Sync or async; an
+ * async planner is where an LLM call goes (see `createSandboxPlanner`).
+ *
+ * @experimental
+ */
+type TopologyPlanner<Task, Output> = (ctx: PlannerContext<Task, Output>) => TopologyMove<Task> | Promise<TopologyMove<Task>>;
+/** @experimental */
+interface CreateDynamicDriverOptions<Task, Output> {
+    /** The agent-authored topology policy. Invoked once per round in `plan`. */
+    planner: TopologyPlanner<Task, Output>;
+    /**
+     * Hard safety cap on total iterations. When reached, the driver stops before
+     * consulting the planner. Default 8. Set the kernel's `runLoop`
+     * `maxIterations >= ` this so the driver's cap governs and the loop closes on
+     * a clean `'done'` rather than a truncated `'continue'`.
+     */
+    maxIterations?: number;
+    /** Max branches a single `fanout` move may dispatch. Default 4. */
+    maxFanout?: number;
+    /** Stable identifier surfaced in trace events. Default `'dynamic'`. */
+    name?: string;
+}
+/** @experimental */
+declare function createDynamicDriver<Task, Output>(options: CreateDynamicDriverOptions<Task, Output>): Driver<Task, Output, DynamicDecision>;
+/**
+ * Compact, planner-friendly view of iteration history — what an LLM planner
+ * needs to choose the next move without the raw event streams. Output is
+ * truncated so a long run's prompt stays bounded.
+ *
+ * @experimental
+ */
+declare function summarizeHistory<Task, Output>(history: ReadonlyArray<Iteration<Task, Output>>, opts?: {
+    maxOutputChars?: number;
+}): Array<{
+    index: number;
+    agentRunName: string;
+    valid?: boolean;
+    score?: number;
+    error?: string;
+    output?: string;
+}>;
+
+export { type CreateDynamicDriverOptions as C, type DynamicDecision as D, type PlannerContext as P, type TopologyPlanner as T, type TopologyMove as a, createDynamicDriver as c, summarizeHistory as s };

package/dist/{improvement-adapter-CaZxFxTd.d.ts → improvement-adapter-BC4HhuAR.d.ts}

@@ -1,5 +1,5 @@
 import { FindingSubject, AnalystFinding } from '@tangle-network/agent-eval';
-import { I as ImprovementAdapter } from './types-D_MXrmJP.js';
+import { I as ImprovementAdapter } from './types-p8dWBIXL.js';
 
 /**
  * `AgentSurfaces` — declarative map of the mutable file/directory paths

package/dist/improvement.d.ts

@@ -1,8 +1,9 @@
-import { AnalystFinding, LlmClientOptions } from '@tangle-network/agent-eval';
+import { AnalystFinding } from '@tangle-network/agent-eval';
 import { L as LocalHarness, r as runLocalHarness } from './local-harness-KrdFTY5R.js';
-import { LabeledScenarioStore, WorktreeAdapter, ImprovementDriver, Scenario, DispatchContext, JudgeConfig, Gate, CampaignStorage, GateResult, RunImprovementLoopResult } from '@tangle-network/agent-eval/campaign';
-import { S as SurfaceImprovementEdit } from './improvement-adapter-CaZxFxTd.js';
-import { I as ImprovementAdapter } from './types-D_MXrmJP.js';
+import { LabeledScenarioStore, WorktreeAdapter, ImprovementDriver } from '@tangle-network/agent-eval/campaign';
+export { O as OptimizePromptOptions, b as OptimizePromptReflection, a as OptimizePromptResult, o as optimizePrompt } from './optimize-prompt-D-urF2wW.js';
+import { S as SurfaceImprovementEdit } from './improvement-adapter-BC4HhuAR.js';
+import { I as ImprovementAdapter } from './types-p8dWBIXL.js';
 import 'node:child_process';
 
 /**
@@ -98,131 +99,6 @@ interface AgenticGeneratorOptions {
 }
 declare function agenticGenerator(opts?: AgenticGeneratorOptions): CandidateGenerator;
 
-/**
- * @experimental
- *
- * `optimizePrompt` — identity-gated optimization for any TEXT prompt surface
- * (system prompt, planner prompt, judge rubric, skill doc).
- *
- * The text-surface sibling to this module's `improvementDriver` (the
- * CODE-surface / worktree path). Both feed agent-eval's `runImprovementLoop`;
- * this one defaults the driver to agent-eval's `gepaDriver` (reflective text
- * mutator) and the gate to `heldOutGate`.
- *
- * IDENTITY-GATED BY CONSTRUCTION — the whole point. The loop runs evals,
- * collects per-scenario signal, proposes candidates, and the gate compares
- * candidate-vs-baseline ON THE HELDOUT. `result.prompt` is the baseline
- * (identity) UNLESS the gate decided `'ship'`. So wiring a surface up is safe:
- * a surface with no beneficial mutation simply keeps its baseline. You never
- * regress by registering a prompt — you only ever improve when the held-out
- * data earns it.
- *
- * Generic over the runtime: `runWithPrompt` is the only domain seam — given a
- * candidate prompt + scenario, run it however the surface runs (sandbox
- * `streamPrompt`, a `runLoop`, a direct model call) and return the artifact the
- * judges score. The optimizer never assumes how a prompt is executed.
- */
-
-/** Reflection config for the default `gepaDriver`. Omit when passing a custom
- *  `driver`. */
-interface OptimizePromptReflection {
-    /** Router transport for the reflection model. */
-    llm: LlmClientOptions;
-    /** Model that performs the reflective rewrite. */
-    model: string;
-    /** What is being optimized — orients the reflection prompt. Default
-     *  `'system prompt'`. */
-    target?: string;
-    /** Surface-specific mutation levers offered to the reflector. */
-    mutationPrimitives?: string[];
-    /** H2 (`## Foo`) headings that MUST survive every candidate. gepaDriver's
-     *  only structural guard — load-bearing sections of the prompt should be
-     *  `##` headings so a rewrite cannot drop them. */
-    preserveSections?: string[];
-    /** Max sentence-level edits per candidate vs the parent (a textual learning
-     *  rate). Caps a rewrite from wiping prior rules in one generation. */
-    maxSentenceEdits?: number;
-}
-/** @experimental */
-interface OptimizePromptOptions<TScenario extends Scenario, TArtifact> {
-    /** The prompt being optimized — the identity baseline the gate protects. */
-    baselinePrompt: string;
-    /** Domain seam: run a candidate prompt against a scenario → artifact the
-     *  judges score. The optimizer is agnostic to HOW the prompt runs. */
-    runWithPrompt: (prompt: string, scenario: TScenario, ctx: DispatchContext) => Promise<TArtifact>;
-    /** Training pool — scored each generation to rank candidates. */
-    scenarios: TScenario[];
-    /** Held out of training — scored ONLY for the gate's baseline-vs-winner
-     *  delta. Disjoint from `scenarios`; this is what makes promotion measure
-     *  generalization, not memorization. */
-    holdoutScenarios: TScenario[];
-    /** Scorers — deterministic checks or LLM judges. */
-    judges: JudgeConfig<TArtifact, TScenario>[];
-    /** Where artifacts + traces land (opaque key under in-memory storage). */
-    runDir: string;
-    /** Default driver = `gepaDriver` built from this. Required UNLESS `driver`
-     *  is supplied. */
-    reflection?: OptimizePromptReflection;
-    /** Override the improvement strategy (custom driver / deterministic tests). */
-    driver?: ImprovementDriver;
-    /** Override the promotion gate. Default `heldOutGate` over `holdoutScenarios`
-     *  — zero extra LLM. Wrap `defaultProductionGate` for red-team/reward-hacking
-     *  hardening on production wiring. */
-    gate?: Gate<TArtifact, TScenario>;
-    /** Minimum held-out composite lift to ship, forwarded to the default
-     *  `heldOutGate`. When omitted the gate uses its own default. */
-    deltaThreshold?: number;
-    /** Candidates proposed per generation. Default 4. */
-    populationSize?: number;
-    /** Generations to run. Default 3. */
-    maxGenerations?: number;
-    /** Candidates carried to the next generation. Default 2. */
-    promoteTopK?: number;
-    /** Storage backend. Pass `inMemoryCampaignStorage()` for filesystem-less /
-     *  test runs. Default: Node filesystem. */
-    storage?: CampaignStorage;
-    /** Reproducibility seed. Default 42. */
-    seed?: number;
-    /** Per-scenario replicates for CI bands. Default 1. */
-    reps?: number;
-    /** Max concurrent cells. Default 2. */
-    maxConcurrency?: number;
-    /** Test seam — override the wall clock. */
-    now?: () => Date;
-    /** On a shipped gate: `'pr'` opens a PR, `'none'` just reports. Default
-     *  `'none'`. */
-    autoOnPromote?: 'pr' | 'none';
-    ghOwner?: string;
-    ghRepo?: string;
-}
-/** @experimental */
-interface OptimizePromptResult<TArtifact, TScenario extends Scenario> {
-    /** The prompt to USE. Identity (the baseline) unless the gate shipped a
-     *  winner — so a caller can always assign `result.prompt` unconditionally. */
-    prompt: string;
-    /** True only when the gate promoted a candidate over baseline on holdout. */
-    improved: boolean;
-    /** The gate's verdict (`'ship' | 'hold' | 'need_more_work' | ...`). */
-    decision: GateResult['decision'];
-    /** Human-readable reasons the gate gave. */
-    reasons: string[];
-    /** Mean held-out composite of the baseline. */
-    baselineComposite: number;
-    /** Mean held-out composite of the winner candidate. */
-    winnerComposite: number;
-    /** Held-out lift (winner − baseline); the gate's `delta` when it reported one. */
-    delta: number;
-    /** Why the winner was proposed — present when a shipped winner carried a
-     *  driver rationale. */
-    rationale?: string;
-    /** Unified baseline→winner diff (empty when the winner is the baseline). */
-    diff: string;
-    /** The full loop result for callers that need generations / campaigns. */
-    raw: RunImprovementLoopResult<TArtifact, TScenario>;
-}
-/** @experimental */
-declare function optimizePrompt<TScenario extends Scenario, TArtifact>(opts: OptimizePromptOptions<TScenario, TArtifact>): Promise<OptimizePromptResult<TArtifact, TScenario>>;
-
 /**
  * @experimental
  *
@@ -242,4 +118,4 @@ interface ReflectiveGeneratorOptions {
 }
 declare function reflectiveGenerator(opts: ReflectiveGeneratorOptions): CandidateGenerator;
 
-export { type AgenticGeneratorOptions, type CandidateGenerator, type ImprovementDriverOptions, type OptimizePromptOptions, type OptimizePromptReflection, type OptimizePromptResult, type ReflectiveGeneratorOptions, agenticGenerator, improvementDriver, optimizePrompt, reflectiveGenerator };
+export { type AgenticGeneratorOptions, type CandidateGenerator, type ImprovementDriverOptions, type ReflectiveGeneratorOptions, agenticGenerator, improvementDriver, reflectiveGenerator };

package/dist/improvement.js

@@ -1,9 +1,10 @@
+import {
+  optimizePrompt
+} from "./chunk-VOX6Z3II.js";
 import {
   runLocalHarness
 } from "./chunk-GLR25NG7.js";
-import {
-  ConfigError
-} from "./chunk-SQSCRJ7U.js";
+import "./chunk-SQSCRJ7U.js";
 import "./chunk-DGUM43GV.js";
 
 // src/improvement/agentic-generator.ts
@@ -130,88 +131,6 @@ function resolveFindings(ctx) {
   return ctx.findings;
 }
 
-// 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;
-}
-
 // src/improvement/reflective-generator.ts
 import { spawnSync as spawnSync2 } from "child_process";
 function reflectiveGenerator(opts) {