@tangle-network/agent-eval 0.91.0 → 0.93.0

package/dist/campaign/index.d.ts

@@ -7,7 +7,7 @@ export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, k
 import { L as LlmClientOptions } from '../llm-client-BeEcAokY.js';
 import { T as TraceAnalystKindSpec } from '../kind-factory-5b7xXXOr.js';
 import { c as AnalystFinding } from '../types-2VVIL04s.js';
-import { S as SignedManifest, A as AgentProfile, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-DELOEJ8v.js';
+import { S as SignedManifest, A as AgentProfile, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-mAnCugl9.js';
 import { E as EProcessState, a as PairedBootstrapResult } from '../statistics-C7PozGrZ.js';
 import { A as AgentEvalError } from '../errors-CzMUYo7b.js';
 import { a as RunSplitTag, R as RunRecord } from '../run-record-e7vj1uZQ.js';

package/dist/campaign/index.js

@@ -15,7 +15,7 @@ import {
   agentProfileHash,
   extractProducedState,
   verifyCompletion
-} from "../chunk-FIUKOSWI.js";
+} from "../chunk-S6OZEZQK.js";
 import {
   buildLoopProvenanceRecord,
   campaignBreakdown,

package/dist/{chunk-RTWFUK6A.js → chunk-E4GH6USR.js}

@@ -1,7 +1,7 @@
 import {
   objectiveEval,
   runAgentControlLoop
-} from "./chunk-NCRFYPS3.js";
+} from "./chunk-YEHAEDUD.js";
 import {
   validateRunRecord
 } from "./chunk-KWRRMR3J.js";
@@ -610,4 +610,4 @@ export {
   runProposeReviewAsControlLoop,
   controlFailureClassFromVerification
 };
-//# sourceMappingURL=chunk-RTWFUK6A.js.map
+//# sourceMappingURL=chunk-E4GH6USR.js.map

package/dist/{chunk-3CKU6VGU.js → chunk-Q2JRAWRI.js}

@@ -1,6 +1,6 @@
 import {
   objectiveEval
-} from "./chunk-NCRFYPS3.js";
+} from "./chunk-YEHAEDUD.js";
 
 // src/knowledge/readiness.ts
 function scoreKnowledgeReadiness(options) {
@@ -193,4 +193,4 @@ export {
   userQuestionsForKnowledgeGaps,
   acquisitionPlansForKnowledgeGaps
 };
-//# sourceMappingURL=chunk-3CKU6VGU.js.map
+//# sourceMappingURL=chunk-Q2JRAWRI.js.map

package/dist/{chunk-FIUKOSWI.js → chunk-S6OZEZQK.js}

@@ -38,15 +38,36 @@ var STOPWORDS = /* @__PURE__ */ new Set([
   "with",
   "by"
 ]);
+var REQUIREMENT_FORM_STOPWORDS = /* @__PURE__ */ new Set([
+  "generated",
+  "generate",
+  "view",
+  "render",
+  "rendered",
+  "persisted",
+  "persist",
+  "artifact",
+  "file",
+  "document",
+  "note",
+  "proposal",
+  "deliverable",
+  "output",
+  "created",
+  "create",
+  "produce",
+  "produced",
+  "flag"
+]);
 var MATCH_THRESHOLD = 0.5;
 var MIN_CONTENT_CHARS = 50;
-function tokens(s) {
+function tokens(s, extraStop) {
   return new Set(
-    s.toLowerCase().split(/[^a-z0-9]+/).filter((t) => t.length > 1 && !STOPWORDS.has(t))
+    s.toLowerCase().split(/[^a-z0-9]+/).filter((t) => t.length > 1 && !STOPWORDS.has(t) && !extraStop?.has(t))
   );
 }
 function tokenRecall(requirementText, candidateText) {
-  const req = tokens(requirementText);
+  const req = tokens(requirementText, REQUIREMENT_FORM_STOPWORDS);
   if (req.size === 0) return 0;
   const cand = tokens(candidateText);
   let hit = 0;
@@ -58,7 +79,10 @@ function artifactCandidates(req, reqIndex, artifacts) {
   const out = [];
   artifacts.forEach((a, i) => {
     if ((a.content ?? "").trim().length < MIN_CONTENT_CHARS) return;
-    let score = tokenRecall(reqText, `${a.path ?? ""} ${a.kind}`);
+    let score = tokenRecall(
+      reqText,
+      `${a.path ?? ""} ${a.kind} ${(a.content ?? "").slice(0, 4e3)}`
+    );
     if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {
       score = Math.max(score, 1);
     }
@@ -78,15 +102,16 @@ function proposalCandidates(req, reqIndex, proposals) {
   const out = [];
   for (const p of proposals) {
     if (p.status !== "approved") continue;
-    const score = tokenRecall(reqText, p.title);
+    const body = (p.content ?? "").trim();
+    if (body.length < MIN_CONTENT_CHARS) continue;
+    const score = tokenRecall(reqText, `${p.title} ${body}`);
     if (score < MATCH_THRESHOLD) continue;
-    const body = p.content ?? "";
     out.push({
       reqIndex,
       itemKey: `proposal:${p.id}`,
       score,
       evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,
-      content: body.trim().length >= MIN_CONTENT_CHARS ? body : null
+      content: body
     });
   }
   return out;
@@ -278,7 +303,12 @@ function extractProducedState(events) {
       });
     } else if (ev.type === "proposal_created") {
       const p = ev;
-      proposals.push({ id: p.proposalId, title: p.title, status: p.status ?? "pending" });
+      proposals.push({
+        id: p.proposalId,
+        title: p.title,
+        status: p.status ?? "pending",
+        ...p.content !== void 0 ? { content: p.content } : {}
+      });
     }
   }
   return { artifacts, proposals, toolCalls };
@@ -309,4 +339,4 @@ export {
   extractProducedState,
   agentProfileHash
 };
-//# sourceMappingURL=chunk-FIUKOSWI.js.map
+//# sourceMappingURL=chunk-S6OZEZQK.js.map

package/dist/chunk-S6OZEZQK.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/completion-verifier.ts","../src/produced-state.ts","../src/agent-profile.ts"],"sourcesContent":["/**\n * Completion verifier — the task-completion oracle.\n *\n * Answers the only eval question that is not a proxy: did the agent actually\n * COMPLETE the task — produce every required deliverable, persisted and\n * correct — rather than describe what should be done. A fluent transcript\n * that never produces the artifact scores zero here.\n *\n * Per requirement, a two-stage check:\n *   1. Structural — a produced item (vault artifact / approved proposal /\n *      tool call) of the right kind is matched against the requirement and\n *      carries non-empty content. Deterministic; no LLM.\n *   2. Correctness — only if structurally present AND the matched item\n *      carries content, one targeted check decides whether that item\n *      actually fulfils the requirement. A hallucinated artifact fails here;\n *      an absent one already failed stage 1.\n *\n * `completionRate` is satisfied / total. Quality dimensions are meaningless\n * on an incomplete task — callers gate on `fullyComplete` / `completionRate`\n * before scoring quality.\n */\n\nimport type { TCloud } from '@tangle-network/tcloud'\nimport type { Artifact } from './artifact-validator'\nimport type { DefaultVerdict } from './verdict'\n\n/** What kind of produced state can satisfy a requirement structurally. */\nexport type SatisfiedBy = 'artifact' | 'proposal' | 'tool-call' | 'any'\n\nexport interface CompletionRequirement {\n  /** Stable id from the task gold (e.g. a persona's `expected_requirements[].req_id`). */\n  reqId: string\n  /** Human-readable description of the required deliverable. */\n  title: string\n  /** Optional kind/category hint, matched against a produced item's kind. */\n  category?: string\n  /** What produced state satisfies this requirement. Defaults to 'any'. */\n  satisfiedBy?: SatisfiedBy\n}\n\nexport interface TaskGold {\n  taskId: string\n  requirements: CompletionRequirement[]\n}\n\nexport interface ProducedProposal {\n  id: string\n  title: string\n  status: 'pending' | 'approved' | 'rejected'\n  /** Optional persisted body — when present, enables a correctness check. */\n  content?: string\n}\n\n/** Everything observable about what a run actually produced. */\nexport interface ProducedState {\n  /** Persisted vault artifacts. Reuses the shared `Artifact` shape. */\n  artifacts: Artifact[]\n  /** Proposals / filings the agent created. */\n  proposals: ProducedProposal[]\n  /** Names of tools the agent invoked. */\n  toolCalls: string[]\n}\n\nexport interface RequirementCheck {\n  reqId: string\n  title: string\n  /** A produced item of the right kind matched the requirement, non-empty. */\n  structurallyPresent: boolean\n  /**\n   * Whether the matched item actually fulfils the requirement. `null` when\n   * not structurally present, or when the matched item carries no content\n   * to assess.\n   */\n  correct: boolean | null\n  /** structurallyPresent && correct !== false. */\n  satisfied: boolean\n  /** Human-readable evidence for the verdict. */\n  evidence: string[]\n}\n\n/** Extends the substrate verdict spine: `valid` = `fullyComplete` and\n *  `score` = `completionRate` — derived in `completionVerdict()`, the one\n *  place those equalities hold by construction. */\nexport interface CompletionVerdict extends DefaultVerdict {\n  taskId: string\n  requirements: RequirementCheck[]\n  /** satisfied / total requirements. */\n  completionRate: number\n  /** Every requirement satisfied. */\n  fullyComplete: boolean\n}\n\n/**\n * Construct a `CompletionVerdict` from the per-requirement checks, deriving\n * `completionRate` / `fullyComplete` and the spine fields (`valid` =\n * `fullyComplete`, `score` = `completionRate`) in one place. Throws on zero\n * requirements — a verdict over nothing is a misconfiguration, mirroring\n * `verifyCompletion`'s gold-spec guard.\n */\nexport function completionVerdict(input: {\n  taskId: string\n  requirements: RequirementCheck[]\n}): CompletionVerdict {\n  if (input.requirements.length === 0) {\n    throw new Error(\n      `completionVerdict: task '${input.taskId}' has no requirement checks — nothing to derive a verdict from`,\n    )\n  }\n  const satisfiedCount = input.requirements.filter((r) => r.satisfied).length\n  const completionRate = satisfiedCount / input.requirements.length\n  const fullyComplete = satisfiedCount === input.requirements.length\n  return {\n    taskId: input.taskId,\n    requirements: input.requirements,\n    completionRate,\n    fullyComplete,\n    valid: fullyComplete,\n    score: completionRate,\n  }\n}\n\n/**\n * Decides whether a produced item's content actually fulfils a requirement.\n * Injected so the structural verifier stays pure and unit-testable; the\n * production implementation is `createLlmCorrectnessChecker`.\n */\nexport type CorrectnessChecker = (\n  requirement: CompletionRequirement,\n  content: string,\n) => Promise<{ correct: boolean; reason: string }>\n\nconst STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'of',\n  'for',\n  'and',\n  'or',\n  'to',\n  'in',\n  'on',\n  'with',\n  'by',\n])\n\n// Deliverable-FORM vocabulary — words that name the SHAPE of an output, not its\n// domain content. A correct \"swap-comparison view persisted as a ui/** artifact\"\n// is an OpenUI JSON whose body says nothing about \"artifact\" / \"persisted\" /\n// \"view\"; the discriminative tokens are the domain nouns (swap, comparison).\n// Stripped from the REQUIREMENT side of structural recall so a deliverable is\n// matched on what it IS about, not on the boilerplate describing its form. The\n// correctness checker strips the same class via TITLE_STOPWORDS. Anti-game holds:\n// the distinctive domain tokens remain, so an off-topic item still fails.\nconst REQUIREMENT_FORM_STOPWORDS = new Set([\n  'generated',\n  'generate',\n  'view',\n  'render',\n  'rendered',\n  'persisted',\n  'persist',\n  'artifact',\n  'file',\n  'document',\n  'note',\n  'proposal',\n  'deliverable',\n  'output',\n  'created',\n  'create',\n  'produce',\n  'produced',\n  'flag',\n])\n\nconst MATCH_THRESHOLD = 0.5\nconst MIN_CONTENT_CHARS = 50\n\nfunction tokens(s: string, extraStop?: Set<string>): Set<string> {\n  return new Set(\n    s\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 1 && !STOPWORDS.has(t) && !extraStop?.has(t)),\n  )\n}\n\n/**\n * Recall of the requirement's tokens within a candidate's identifying text.\n * Recall, not Jaccard — a candidate's path/id legitimately carries extra\n * tokens the requirement does not name. The requirement side drops\n * deliverable-FORM vocabulary so recall keys on the distinctive domain tokens.\n */\nfunction tokenRecall(requirementText: string, candidateText: string): number {\n  const req = tokens(requirementText, REQUIREMENT_FORM_STOPWORDS)\n  if (req.size === 0) return 0\n  const cand = tokens(candidateText)\n  let hit = 0\n  for (const t of req) if (cand.has(t)) hit++\n  return hit / req.size\n}\n\ninterface Candidate {\n  reqIndex: number\n  /** Unique key for a produced item — each item satisfies at most one requirement. */\n  itemKey: string\n  score: number\n  evidence: string\n  /** Content to correctness-check, or null when the matched item has none. */\n  content: string | null\n}\n\nfunction artifactCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  artifacts: Artifact[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  artifacts.forEach((a, i) => {\n    if ((a.content ?? '').trim().length < MIN_CONTENT_CHARS) return\n    // Match against the artifact CONTENT too, not just its path + kind — a\n    // generated view / note whose path is generic still satisfies a requirement\n    // when its body covers it (e.g. an OpenUI comparison grounded in the on-file\n    // figures). Bounded slice keeps the recall text cheap; MATCH_THRESHOLD holds.\n    let score = tokenRecall(\n      reqText,\n      `${a.path ?? ''} ${a.kind} ${(a.content ?? '').slice(0, 4000)}`,\n    )\n    if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {\n      score = Math.max(score, 1)\n    }\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `artifact:${i}`,\n      score,\n      evidence: `artifact '${a.path ?? a.kind}' matched (token recall ${score.toFixed(2)})`,\n      content: a.content ?? null,\n    })\n  })\n  return out\n}\n\nfunction proposalCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  proposals: ProducedProposal[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  for (const p of proposals) {\n    // Pending or rejected work is not a completed deliverable.\n    if (p.status !== 'approved') continue\n    // A proposal needs an assessable BODY to be a deliverable. A bare title is\n    // not completion: correctness cannot be judged on it, so a title-only match\n    // would auto-pass the oracle (structurallyPresent && correct===null →\n    // satisfied) with no verifiable content. Tool calls are the only\n    // legitimately content-less deliverable (`toolCallCandidates`).\n    const body = (p.content ?? '').trim()\n    if (body.length < MIN_CONTENT_CHARS) continue\n    // Match against the body as well as the (often short) title — a refusal /\n    // flag / analysis proposal whose title is a label still satisfies a\n    // descriptively-worded requirement when its content covers it. MATCH_THRESHOLD\n    // + the requirement's distinctive tokens keep an off-topic proposal out;\n    // correctness (a SEMANTIC checker, NOT this lexical pass) then judges\n    // polarity/fulfilment, so a negation that merely contains the tokens fails.\n    // Structural and correctness must use different evidence or the two-stage\n    // check collapses to one lexical gate.\n    const score = tokenRecall(reqText, `${p.title} ${body}`)\n    if (score < MATCH_THRESHOLD) continue\n    out.push({\n      reqIndex,\n      itemKey: `proposal:${p.id}`,\n      score,\n      evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,\n      content: body,\n    })\n  }\n  return out\n}\n\nfunction toolCallCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  toolCalls: string[],\n): Candidate[] {\n  const out: Candidate[] = []\n  toolCalls.forEach((name, i) => {\n    const score = tokenRecall(req.title, name)\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `tool:${i}`,\n      score,\n      evidence: `tool call '${name}' matched (token recall ${score.toFixed(2)})`,\n      content: null,\n    })\n  })\n  return out\n}\n\n/**\n * Verify whether a run completed the task. `checkCorrectness` is injected —\n * `createLlmCorrectnessChecker` for production, a deterministic stub in tests.\n *\n * Throws on a gold spec with no requirements: an eval task that requires\n * nothing is a misconfiguration, not a vacuously-complete task.\n */\nexport async function verifyCompletion(\n  gold: TaskGold,\n  state: ProducedState,\n  checkCorrectness: CorrectnessChecker,\n): Promise<CompletionVerdict> {\n  if (gold.requirements.length === 0) {\n    throw new Error(\n      `verifyCompletion: task '${gold.taskId}' has no requirements — malformed gold spec`,\n    )\n  }\n\n  // Collect every above-threshold (requirement, produced-item) candidate, then\n  // assign greedily by descending score: each requirement and each produced\n  // item is used at most once. One deliverable fulfils one requirement.\n  const candidates: Candidate[] = []\n  gold.requirements.forEach((req, i) => {\n    const by = req.satisfiedBy ?? 'any'\n    if (by === 'artifact' || by === 'any') {\n      candidates.push(...artifactCandidates(req, i, state.artifacts))\n    }\n    if (by === 'proposal' || by === 'any') {\n      candidates.push(...proposalCandidates(req, i, state.proposals))\n    }\n    if (by === 'tool-call' || by === 'any') {\n      candidates.push(...toolCallCandidates(req, i, state.toolCalls))\n    }\n  })\n  candidates.sort((a, b) => b.score - a.score)\n\n  const assigned = new Map<number, Candidate>()\n  const itemTaken = new Set<string>()\n  for (const c of candidates) {\n    if (assigned.has(c.reqIndex) || itemTaken.has(c.itemKey)) continue\n    assigned.set(c.reqIndex, c)\n    itemTaken.add(c.itemKey)\n  }\n\n  const requirements: RequirementCheck[] = []\n  for (let i = 0; i < gold.requirements.length; i++) {\n    const req = gold.requirements[i]!\n    const match = assigned.get(i)\n    const evidence: string[] = []\n    let correct: boolean | null = null\n\n    if (match) {\n      evidence.push(match.evidence)\n      if (match.content !== null) {\n        const r = await checkCorrectness(req, match.content)\n        correct = r.correct\n        evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\n      } else {\n        evidence.push('correctness: not assessed — matched item carries no content')\n      }\n    } else {\n      const by = req.satisfiedBy ?? 'any'\n      const kind = by === 'any' ? 'artifact/proposal/tool-call' : by\n      evidence.push(`no produced ${kind} matched this requirement`)\n    }\n\n    const structurallyPresent = match !== undefined\n    const satisfied = structurallyPresent && correct !== false\n    requirements.push({\n      reqId: req.reqId,\n      title: req.title,\n      structurallyPresent,\n      correct,\n      satisfied,\n      evidence,\n    })\n  }\n\n  return completionVerdict({ taskId: gold.taskId, requirements })\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n  model?: string\n  /** Max chars of artifact content sent to the checker. */\n  maxContentChars?: number\n}\n\n/** Parse the correctness checker's model response. Fails loud on a bad shape. */\nexport function parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n  const match = raw.match(/\\{[\\s\\S]*\\}/)\n  if (!match) {\n    throw new Error(`correctness checker: no JSON object in model response: ${raw.slice(0, 200)}`)\n  }\n  const parsed = JSON.parse(match[0]) as { correct?: unknown; reason?: unknown }\n  if (typeof parsed.correct !== 'boolean') {\n    throw new Error(`correctness checker: 'correct' is not a boolean in: ${match[0].slice(0, 200)}`)\n  }\n  return { correct: parsed.correct, reason: typeof parsed.reason === 'string' ? parsed.reason : '' }\n}\n\n/**\n * Production `CorrectnessChecker` — one LLM call per matched artifact,\n * deterministic (temperature 0), structured JSON out. Judges fulfilment\n * only: a plan, a gesture, or a description of what should be done does not\n * fulfil a requirement — the artifact must BE the deliverable.\n */\nexport function createLlmCorrectnessChecker(\n  tc: TCloud,\n  opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n  const model = opts.model ?? 'claude-sonnet-4-6'\n  const maxContentChars = opts.maxContentChars ?? 8000\n  return async (requirement, content) => {\n    const resp = await tc.chat({\n      model,\n      messages: [\n        {\n          role: 'system',\n          content:\n            'You verify whether a produced work artifact actually fulfils a stated requirement. Judge fulfilment only — is the deliverable substantively present and on-point — not polish. A plan to do it later, a vague gesture, or a description of what should be done does NOT fulfil a requirement; the artifact must BE the deliverable. Respond with a single JSON object: {\"correct\": boolean, \"reason\": string (<= 30 words)}.',\n        },\n        {\n          role: 'user',\n          content: `Requirement: ${requirement.title}\\n${\n            requirement.category ? `Category: ${requirement.category}\\n` : ''\n          }\\nProduced artifact:\\n${content.slice(0, maxContentChars)}`,\n        },\n      ],\n      temperature: 0,\n      maxTokens: 200,\n    })\n    const raw =\n      (resp as { choices?: { message?: { content?: string } }[] }).choices?.[0]?.message?.content ??\n      ''\n    return parseCorrectnessResponse(raw)\n  }\n}\n\n/** Stopwords for requirement-title tokenization — drops the imperative verbs\n *  ('review', 'update', …) common to deliverable titles so recall keys on the\n *  substantive nouns, not the boilerplate ask. */\nconst TITLE_STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'and',\n  'or',\n  'for',\n  'to',\n  'of',\n  'in',\n  'on',\n  'with',\n  'review',\n  'update',\n  'new',\n  'proposed',\n])\n\n/**\n * Deterministic `CorrectnessChecker` — the no-LLM counterpart to\n * `createLlmCorrectnessChecker`. A produced item fulfils a requirement when its\n * content is substantive (≥ `minContentLength` chars) AND recalls ≥ `minRecall`\n * of the requirement title's significant tokens. No network.\n *\n * Polarity-blind: token recall credits a negation that contains the\n * requirement's tokens (\"I will NOT produce the comparison\" recalls every token\n * of \"produce the comparison\"). The structural match stage is ALSO lexical, so\n * pairing the two collapses to a single gameable gate. Use this only as an\n * opt-in structural pre-filter or for tasks whose requirements have no polarity\n * to invert; for produced-state grading the correctness checker MUST be semantic\n * (`createLlmCorrectnessChecker`). See the anti-game fixtures in the test suite.\n */\nexport function createTokenRecallChecker(\n  opts: { minRecall?: number; minContentLength?: number } = {},\n): CorrectnessChecker {\n  const minRecall = opts.minRecall ?? 0.5\n  const minLen = opts.minContentLength ?? 120\n  return async (requirement, content) => {\n    const body = content.trim()\n    if (body.length < minLen)\n      return {\n        correct: false,\n        reason: `content too thin (${body.length} chars) to be the deliverable`,\n      }\n    const titleTokens = requirement.title\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 2 && !TITLE_STOPWORDS.has(t))\n    if (titleTokens.length === 0)\n      return {\n        correct: true,\n        reason: 'requirement title has no significant tokens — structural match accepted',\n      }\n    const lower = body.toLowerCase()\n    const hits = titleTokens.filter((t) => lower.includes(t)).length\n    const recall = hits / titleTokens.length\n    return recall >= minRecall\n      ? {\n          correct: true,\n          reason: `content recalls ${hits}/${titleTokens.length} requirement tokens`,\n        }\n      : {\n          correct: false,\n          reason: `content recalls only ${hits}/${titleTokens.length} requirement tokens`,\n        }\n  }\n}\n","/**\n * Produced-state extraction — normalize a run's runtime event stream into the\n * typed `ProducedState` the completion oracle consumes.\n *\n * `ProducedState` answers \"what did the agent actually produce\" — vault\n * artifacts, proposals, tool calls. The runtime emits these as a stream of\n * events; this module is the single normalization point from that stream to\n * the shape `verifyCompletion` expects.\n *\n * Input is structurally typed (`RuntimeEventLike`) so this module does not\n * depend on agent-runtime — agent-runtime's `RuntimeStreamEvent` satisfies it\n * structurally. The `content` on `ArtifactEventLike` and the whole\n * `proposal_created` variant are the runtime-side enrichments this contract\n * requires; the runtime emits them, this module consumes them.\n */\n\nimport type { Artifact } from './artifact-validator'\nimport type { ProducedProposal, ProducedState } from './completion-verifier'\n\n/** A tool the agent invoked. */\nexport interface ToolCallEventLike {\n  type: 'tool_call'\n  toolName: string\n}\n\n/**\n * An artifact the agent produced. `content` is the enriched field — the\n * runtime's base `artifact` event carries only metadata; the completion\n * oracle needs the body to verify the deliverable, so the runtime emits it.\n */\nexport interface ArtifactEventLike {\n  type: 'artifact'\n  artifactId: string\n  name?: string\n  mimeType?: string\n  uri?: string\n  content?: string\n}\n\n/** A proposal / filing the agent created. */\nexport interface ProposalEventLike {\n  type: 'proposal_created'\n  proposalId: string\n  title: string\n  status?: 'pending' | 'approved' | 'rejected'\n  // body of the proposal (e.g. a submit_proposal `description`). When present,\n  // the completion oracle correctness-checks it like artifact content; absent,\n  // the proposal is graded presence-only.\n  content?: string\n}\n\n/**\n * The subset of runtime stream events `extractProducedState` consumes.\n * agent-runtime's full `RuntimeStreamEvent` union satisfies this structurally;\n * the `{ type: string }` catch-all keeps the input permissive so callers can\n * pass the whole unfiltered telemetry stream — unrecognized events are skipped.\n */\nexport type RuntimeEventLike =\n  | ToolCallEventLike\n  | ArtifactEventLike\n  | ProposalEventLike\n  | { type: string }\n\nfunction artifactKind(mimeType: string | undefined): string {\n  if (!mimeType) return 'file'\n  if (mimeType.includes('json')) return 'json'\n  if (mimeType.startsWith('text/')) return 'text'\n  return 'file'\n}\n\n/**\n * Normalize a run's runtime event stream into `ProducedState`.\n *\n * Pure and total — unrecognized event types are skipped. `toolCalls` is\n * deduplicated by name in first-seen order (completion cares about a tool's\n * presence, not its call count). An artifact with neither a name nor a uri\n * still yields an entry keyed by its `artifactId` so it is never silently\n * dropped; an artifact with no `content` yields empty content, which the\n * completion oracle's structural check then rejects on its own.\n */\nexport function extractProducedState(events: readonly RuntimeEventLike[]): ProducedState {\n  const artifacts: Artifact[] = []\n  const proposals: ProducedProposal[] = []\n  const toolCalls: string[] = []\n  const seenTools = new Set<string>()\n\n  for (const ev of events) {\n    if (ev.type === 'tool_call') {\n      const name = (ev as ToolCallEventLike).toolName\n      if (name && !seenTools.has(name)) {\n        seenTools.add(name)\n        toolCalls.push(name)\n      }\n    } else if (ev.type === 'artifact') {\n      const a = ev as ArtifactEventLike\n      artifacts.push({\n        kind: artifactKind(a.mimeType),\n        path: a.name ?? a.uri ?? a.artifactId,\n        content: a.content ?? '',\n      })\n    } else if (ev.type === 'proposal_created') {\n      const p = ev as ProposalEventLike\n      proposals.push({\n        id: p.proposalId,\n        title: p.title,\n        status: p.status ?? 'pending',\n        ...(p.content !== undefined ? { content: p.content } : {}),\n      })\n    }\n  }\n\n  return { artifacts, proposals, toolCalls }\n}\n","/**\n * @stable\n *\n * AgentProfile — the eval harness's unit of variation.\n *\n * A profile pins everything that changes agent behaviour for a benchmark\n * cell: the model, the active skills, the prompt version, the available\n * tools. Vary the profile — swap a model, add a skill — and re-run the suite\n * to benchmark the change. The scorecard keys a cell on\n * `(scenarioId, profileHash)`, so the model is not a separate axis: it lives\n * inside the profile, and two profiles with the same model but different\n * skills are different cells.\n *\n * `agentProfileHash` is the profile's behaviour identity. Two profiles that\n * produce the same agent behaviour share a hash (and a scorecard cell);\n * reordering `skills` or `tools` does not change it; the human-facing `id`\n * label does not affect it.\n */\n\nimport { createHash } from 'node:crypto'\nimport { ValidationError } from './errors'\nimport { canonicalize } from './pre-registration'\n\nexport interface AgentProfile {\n  /** Human-facing label, e.g. `sonnet-legal-skills-v3`. Not part of the hash. */\n  id: string\n  /** Model snapshot id this profile pins, e.g. `claude-sonnet-4-6@2025-04-15`. */\n  model: string\n  /** Skill ids/versions active in this profile — the primary behaviour lever. */\n  skills?: string[]\n  /** Prompt version identifier. */\n  promptVersion?: string\n  /** Tool ids available to the agent. */\n  tools?: string[]\n  /** Any other behaviour-bearing knobs that should fingerprint into the hash. */\n  metadata?: Record<string, string | number | boolean>\n}\n\n/**\n * Deterministic behaviour identity of a profile — a sha256 over the\n * behaviour-bearing fields. `skills` and `tools` are order-insensitive; the\n * `id` label is excluded. Throws on a profile with no `model` — an unkeyable\n * profile must fail loud rather than collapse into a blank-model cell.\n */\nexport function agentProfileHash(profile: AgentProfile): string {\n  if (typeof profile.model !== 'string' || profile.model.trim().length === 0) {\n    throw new ValidationError(`AgentProfile \"${profile.id}\" has no model — cannot hash`)\n  }\n  const behaviour = {\n    model: profile.model.trim(),\n    skills: [...(profile.skills ?? [])].sort(),\n    promptVersion: profile.promptVersion ?? null,\n    tools: [...(profile.tools ?? [])].sort(),\n    metadata: profile.metadata ?? {},\n  }\n  return createHash('sha256')\n    .update(JSON.stringify(canonicalize(behaviour)))\n    .digest('hex')\n}\n"],"mappings":";;;;;;;;AAmGO,SAAS,kBAAkB,OAGZ;AACpB,MAAI,MAAM,aAAa,WAAW,GAAG;AACnC,UAAM,IAAI;AAAA,MACR,4BAA4B,MAAM,MAAM;AAAA,IAC1C;AAAA,EACF;AACA,QAAM,iBAAiB,MAAM,aAAa,OAAO,CAAC,MAAM,EAAE,SAAS,EAAE;AACrE,QAAM,iBAAiB,iBAAiB,MAAM,aAAa;AAC3D,QAAM,gBAAgB,mBAAmB,MAAM,aAAa;AAC5D,SAAO;AAAA,IACL,QAAQ,MAAM;AAAA,IACd,cAAc,MAAM;AAAA,IACpB;AAAA,IACA;AAAA,IACA,OAAO;AAAA,IACP,OAAO;AAAA,EACT;AACF;AAYA,IAAM,YAAY,oBAAI,IAAI;AAAA,EACxB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUD,IAAM,6BAA6B,oBAAI,IAAI;AAAA,EACzC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,IAAM,kBAAkB;AACxB,IAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAW,WAAsC;AAC/D,SAAO,IAAI;AAAA,IACT,EACG,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,KAAK,CAAC,WAAW,IAAI,CAAC,CAAC;AAAA,EAC1E;AACF;AAQA,SAAS,YAAY,iBAAyB,eAA+B;AAC3E,QAAM,MAAM,OAAO,iBAAiB,0BAA0B;AAC9D,MAAI,IAAI,SAAS,EAAG,QAAO;AAC3B,QAAM,OAAO,OAAO,aAAa;AACjC,MAAI,MAAM;AACV,aAAW,KAAK,IAAK,KAAI,KAAK,IAAI,CAAC,EAAG;AACtC,SAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,GAAG,MAAM;AAC1B,SAAK,EAAE,WAAW,IAAI,KAAK,EAAE,SAAS,kBAAmB;AAKzD,QAAI,QAAQ;AAAA,MACV;AAAA,MACA,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,IAAI,KAAK,EAAE,WAAW,IAAI,MAAM,GAAG,GAAI,CAAC;AAAA,IAC/D;AACA,QAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAAG;AACjF,cAAQ,KAAK,IAAI,OAAO,CAAC;AAAA,IAC3B;AACA,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,CAAC;AAAA,MACtB;AAAA,MACA,UAAU,aAAa,EAAE,QAAQ,EAAE,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS,EAAE,WAAW;AAAA,IACxB,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,aAAW,KAAK,WAAW;AAEzB,QAAI,EAAE,WAAW,WAAY;AAM7B,UAAM,QAAQ,EAAE,WAAW,IAAI,KAAK;AACpC,QAAI,KAAK,SAAS,kBAAmB;AASrC,UAAM,QAAQ,YAAY,SAAS,GAAG,EAAE,KAAK,IAAI,IAAI,EAAE;AACvD,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,EAAE,EAAE;AAAA,MACzB;AAAA,MACA,UAAU,sBAAsB,EAAE,KAAK,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS;AAAA,IACX,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,MAAM,MAAM;AAC7B,UAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;AACzC,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,QAAQ,CAAC;AAAA,MAClB;AAAA,MACA,UAAU,cAAc,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MACvE,SAAS;AAAA,IACX,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;AAC5B,MAAI,KAAK,aAAa,WAAW,GAAG;AAClC,UAAM,IAAI;AAAA,MACR,2BAA2B,KAAK,MAAM;AAAA,IACxC;AAAA,EACF;AAKA,QAAM,aAA0B,CAAC;AACjC,OAAK,aAAa,QAAQ,CAAC,KAAK,MAAM;AACpC,UAAM,KAAK,IAAI,eAAe;AAC9B,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,eAAe,OAAO,OAAO;AACtC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AAAA,EACF,CAAC;AACD,aAAW,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAE3C,QAAM,WAAW,oBAAI,IAAuB;AAC5C,QAAM,YAAY,oBAAI,IAAY;AAClC,aAAW,KAAK,YAAY;AAC1B,QAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,EAAG;AAC1D,aAAS,IAAI,EAAE,UAAU,CAAC;AAC1B,cAAU,IAAI,EAAE,OAAO;AAAA,EACzB;AAEA,QAAM,eAAmC,CAAC;AAC1C,WAAS,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;AACjD,UAAM,MAAM,KAAK,aAAa,CAAC;AAC/B,UAAM,QAAQ,SAAS,IAAI,CAAC;AAC5B,UAAM,WAAqB,CAAC;AAC5B,QAAI,UAA0B;AAE9B,QAAI,OAAO;AACT,eAAS,KAAK,MAAM,QAAQ;AAC5B,UAAI,MAAM,YAAY,MAAM;AAC1B,cAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;AACnD,kBAAU,EAAE;AACZ,iBAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,MAAM,WAAM,EAAE,MAAM,EAAE;AAAA,MAC3E,OAAO;AACL,iBAAS,KAAK,kEAA6D;AAAA,MAC7E;AAAA,IACF,OAAO;AACL,YAAM,KAAK,IAAI,eAAe;AAC9B,YAAM,OAAO,OAAO,QAAQ,gCAAgC;AAC5D,eAAS,KAAK,eAAe,IAAI,2BAA2B;AAAA,IAC9D;AAEA,UAAM,sBAAsB,UAAU;AACtC,UAAM,YAAY,uBAAuB,YAAY;AACrD,iBAAa,KAAK;AAAA,MAChB,OAAO,IAAI;AAAA,MACX,OAAO,IAAI;AAAA,MACX;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AAEA,SAAO,kBAAkB,EAAE,QAAQ,KAAK,QAAQ,aAAa,CAAC;AAChE;AASO,SAAS,yBAAyB,KAAmD;AAC1F,QAAM,QAAQ,IAAI,MAAM,aAAa;AACrC,MAAI,CAAC,OAAO;AACV,UAAM,IAAI,MAAM,0DAA0D,IAAI,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EAC/F;AACA,QAAM,SAAS,KAAK,MAAM,MAAM,CAAC,CAAC;AAClC,MAAI,OAAO,OAAO,YAAY,WAAW;AACvC,UAAM,IAAI,MAAM,uDAAuD,MAAM,CAAC,EAAE,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EACjG;AACA,SAAO,EAAE,SAAS,OAAO,SAAS,QAAQ,OAAO,OAAO,WAAW,WAAW,OAAO,SAAS,GAAG;AACnG;AAQO,SAAS,4BACd,IACA,OAAkC,CAAC,GACf;AACpB,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,kBAAkB,KAAK,mBAAmB;AAChD,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,OAAO,MAAM,GAAG,KAAK;AAAA,MACzB;AAAA,MACA,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SACE;AAAA,QACJ;AAAA,QACA;AAAA,UACE,MAAM;AAAA,UACN,SAAS,gBAAgB,YAAY,KAAK;AAAA,EACxC,YAAY,WAAW,aAAa,YAAY,QAAQ;AAAA,IAAO,EACjE;AAAA;AAAA,EAAyB,QAAQ,MAAM,GAAG,eAAe,CAAC;AAAA,QAC5D;AAAA,MACF;AAAA,MACA,aAAa;AAAA,MACb,WAAW;AAAA,IACb,CAAC;AACD,UAAM,MACH,KAA4D,UAAU,CAAC,GAAG,SAAS,WACpF;AACF,WAAO,yBAAyB,GAAG;AAAA,EACrC;AACF;AAKA,IAAM,kBAAkB,oBAAI,IAAI;AAAA,EAC9B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAgBM,SAAS,yBACd,OAA0D,CAAC,GACvC;AACpB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,SAAS,KAAK,oBAAoB;AACxC,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,OAAO,QAAQ,KAAK;AAC1B,QAAI,KAAK,SAAS;AAChB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ,qBAAqB,KAAK,MAAM;AAAA,MAC1C;AACF,UAAM,cAAc,YAAY,MAC7B,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,gBAAgB,IAAI,CAAC,CAAC;AACxD,QAAI,YAAY,WAAW;AACzB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ;AAAA,MACV;AACF,UAAM,QAAQ,KAAK,YAAY;AAC/B,UAAM,OAAO,YAAY,OAAO,CAAC,MAAM,MAAM,SAAS,CAAC,CAAC,EAAE;AAC1D,UAAM,SAAS,OAAO,YAAY;AAClC,WAAO,UAAU,YACb;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,mBAAmB,IAAI,IAAI,YAAY,MAAM;AAAA,IACvD,IACA;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,wBAAwB,IAAI,IAAI,YAAY,MAAM;AAAA,IAC5D;AAAA,EACN;AACF;;;AC/bA,SAAS,aAAa,UAAsC;AAC1D,MAAI,CAAC,SAAU,QAAO;AACtB,MAAI,SAAS,SAAS,MAAM,EAAG,QAAO;AACtC,MAAI,SAAS,WAAW,OAAO,EAAG,QAAO;AACzC,SAAO;AACT;AAYO,SAAS,qBAAqB,QAAoD;AACvF,QAAM,YAAwB,CAAC;AAC/B,QAAM,YAAgC,CAAC;AACvC,QAAM,YAAsB,CAAC;AAC7B,QAAM,YAAY,oBAAI,IAAY;AAElC,aAAW,MAAM,QAAQ;AACvB,QAAI,GAAG,SAAS,aAAa;AAC3B,YAAM,OAAQ,GAAyB;AACvC,UAAI,QAAQ,CAAC,UAAU,IAAI,IAAI,GAAG;AAChC,kBAAU,IAAI,IAAI;AAClB,kBAAU,KAAK,IAAI;AAAA,MACrB;AAAA,IACF,WAAW,GAAG,SAAS,YAAY;AACjC,YAAM,IAAI;AACV,gBAAU,KAAK;AAAA,QACb,MAAM,aAAa,EAAE,QAAQ;AAAA,QAC7B,MAAM,EAAE,QAAQ,EAAE,OAAO,EAAE;AAAA,QAC3B,SAAS,EAAE,WAAW;AAAA,MACxB,CAAC;AAAA,IACH,WAAW,GAAG,SAAS,oBAAoB;AACzC,YAAM,IAAI;AACV,gBAAU,KAAK;AAAA,QACb,IAAI,EAAE;AAAA,QACN,OAAO,EAAE;AAAA,QACT,QAAQ,EAAE,UAAU;AAAA,QACpB,GAAI,EAAE,YAAY,SAAY,EAAE,SAAS,EAAE,QAAQ,IAAI,CAAC;AAAA,MAC1D,CAAC;AAAA,IACH;AAAA,EACF;AAEA,SAAO,EAAE,WAAW,WAAW,UAAU;AAC3C;;;AC7FA,SAAS,kBAAkB;AAyBpB,SAAS,iBAAiB,SAA+B;AAC9D,MAAI,OAAO,QAAQ,UAAU,YAAY,QAAQ,MAAM,KAAK,EAAE,WAAW,GAAG;AAC1E,UAAM,IAAI,gBAAgB,iBAAiB,QAAQ,EAAE,mCAA8B;AAAA,EACrF;AACA,QAAM,YAAY;AAAA,IAChB,OAAO,QAAQ,MAAM,KAAK;AAAA,IAC1B,QAAQ,CAAC,GAAI,QAAQ,UAAU,CAAC,CAAE,EAAE,KAAK;AAAA,IACzC,eAAe,QAAQ,iBAAiB;AAAA,IACxC,OAAO,CAAC,GAAI,QAAQ,SAAS,CAAC,CAAE,EAAE,KAAK;AAAA,IACvC,UAAU,QAAQ,YAAY,CAAC;AAAA,EACjC;AACA,SAAO,WAAW,QAAQ,EACvB,OAAO,KAAK,UAAU,aAAa,SAAS,CAAC,CAAC,EAC9C,OAAO,KAAK;AACjB;","names":[]}

package/dist/{chunk-NCRFYPS3.js → chunk-YEHAEDUD.js}

@@ -2,6 +2,112 @@ import {
   TraceEmitter
 } from "./chunk-TVVP3ZZQ.js";
 
+// src/detectors/index.ts
+function repeatedActionDetector(opts = {}) {
+  const max = opts.maxRepeated ?? 3;
+  const severity = opts.severity ?? "warn";
+  const failureClass = opts.failureClass ?? "tool_recovery_failure";
+  let last;
+  let streak = 0;
+  return {
+    id: "repeated-action",
+    get streak() {
+      return streak;
+    },
+    observe(event) {
+      const fp = event.actionFingerprint;
+      if (fp === void 0) return null;
+      streak = fp === last ? streak + 1 : 1;
+      last = fp;
+      if (max <= 0 || streak < max) return null;
+      return {
+        detector: "repeated-action",
+        severity,
+        failureClass,
+        reason: `stuck: repeated same action for ${streak} step(s)`,
+        streak,
+        ...event.label ? { evidence: { action: event.label } } : {}
+      };
+    },
+    reset() {
+      last = void 0;
+      streak = 0;
+    }
+  };
+}
+function noProgressDetector(opts = {}) {
+  const max = opts.maxNoProgress ?? 3;
+  const minScoreDelta = opts.minScoreDelta ?? 1e-3;
+  const severity = opts.severity ?? "warn";
+  const failureClass = opts.failureClass ?? "tool_recovery_failure";
+  let lastState;
+  let lastScore;
+  let streak = 0;
+  return {
+    id: "no-progress",
+    get streak() {
+      return streak;
+    },
+    observe(event) {
+      const stateUnchanged = lastState !== void 0 && lastState === event.stateFingerprint;
+      const scoreFlat = Math.abs((event.score ?? 0) - (lastScore ?? 0)) < minScoreDelta;
+      streak = stateUnchanged && scoreFlat ? streak + 1 : 0;
+      lastState = event.stateFingerprint;
+      lastScore = event.score;
+      if (max <= 0 || streak < max) return null;
+      return {
+        detector: "no-progress",
+        severity,
+        failureClass,
+        reason: `stuck: no state/score progress for ${streak} step(s)`,
+        streak,
+        ...event.label ? { evidence: { state: event.label } } : {}
+      };
+    },
+    reset() {
+      lastState = void 0;
+      lastScore = void 0;
+      streak = 0;
+    }
+  };
+}
+function errorStreakDetector(opts = {}) {
+  const max = opts.maxErrors ?? 3;
+  const severity = opts.severity ?? "warn";
+  const failureClass = opts.failureClass ?? "tool_recovery_failure";
+  let streak = 0;
+  return {
+    id: "error-streak",
+    get streak() {
+      return streak;
+    },
+    observe(event) {
+      if (event.status === void 0) return null;
+      streak = event.status === "error" ? streak + 1 : 0;
+      if (max <= 0 || streak < max) return null;
+      return {
+        detector: "error-streak",
+        severity,
+        failureClass,
+        reason: `stuck: ${streak} consecutive errored step(s)`,
+        streak,
+        ...event.label ? { evidence: { lastError: event.label } } : {}
+      };
+    },
+    reset() {
+      streak = 0;
+    }
+  };
+}
+function observeAll(detectors, event) {
+  const signals = [];
+  for (const d of detectors) {
+    const s = d.observe(event);
+    if (s) signals.push(s);
+  }
+  return signals;
+}
+
 // src/control-runtime.ts
 var DEFAULT_BUDGET = {
   maxSteps: 8,
@@ -25,10 +131,13 @@ async function runAgentControlLoop(config) {
   const emitter = config.store ? new TraceEmitter(config.store) : void 0;
   let spentCostUsd = 0;
   const runtimeErrors = [];
-  let lastStateFingerprint;
-  let lastActionFingerprint;
-  let noProgressStreak = 0;
-  let repeatedActionStreak = 0;
+  const repeatedDetector = repeatedActionDetector({
+    maxRepeated: config.stopPolicies?.maxRepeatedActions ?? 0
+  });
+  const progressDetector = noProgressDetector({
+    maxNoProgress: config.stopPolicies?.maxNoProgressSteps ?? 0,
+    minScoreDelta: config.stopPolicies?.minScoreDelta ?? 1e-3
+  });
   try {
     if (emitter) {
       await runTrace(
@@ -97,7 +206,10 @@ async function runAgentControlLoop(config) {
         stoppedBy: "runtime-error"
       });
     }
-    lastStateFingerprint = fingerprintState(state, config.stopPolicies);
+    progressDetector.observe({
+      stateFingerprint: fingerprintState(state, config.stopPolicies),
+      score: averageScore(evals)
+    });
     for (let stepIndex = 0; stepIndex < budget.maxSteps; stepIndex++) {
       if (controller.signal.aborted) {
         return finish(emitter, {
@@ -229,18 +341,14 @@ async function runAgentControlLoop(config) {
         });
       }
       const actionFingerprint = fingerprintAction(decision.action, config.stopPolicies);
-      repeatedActionStreak = actionFingerprint === lastActionFingerprint ? repeatedActionStreak + 1 : 1;
-      lastActionFingerprint = actionFingerprint;
-      const repeatedActionStop = repeatedActionStopDecision(
-        config.stopPolicies,
-        repeatedActionStreak
-      );
-      if (repeatedActionStop.stop) {
+      const repeatedActionSignal = repeatedDetector.observe({ actionFingerprint });
+      const repeatedActionStreak = repeatedDetector.streak;
+      if (repeatedActionSignal) {
         return finish(emitter, {
           intent: config.intent,
           pass: false,
           completed: true,
-          reason: repeatedActionStop.reason,
+          reason: repeatedActionSignal.reason,
           score: averageScore(evals),
           steps: history,
           finalState: state,
@@ -435,16 +543,8 @@ async function runAgentControlLoop(config) {
       }
       const scoreAfter = averageScore(evals);
       const stateFingerprint = fingerprintState(state, config.stopPolicies);
-      const noProgressStop = noProgressStopDecision({
-        policies: config.stopPolicies,
-        lastStateFingerprint,
-        stateFingerprint,
-        scoreBefore,
-        scoreAfter,
-        currentStreak: noProgressStreak
-      });
-      noProgressStreak = noProgressStop.streak;
-      lastStateFingerprint = stateFingerprint;
+      const noProgressSignal = progressDetector.observe({ stateFingerprint, score: scoreAfter });
+      const noProgressStreak = progressDetector.streak;
       const step = {
         index: stepIndex,
         decision,
@@ -484,12 +584,12 @@ async function runAgentControlLoop(config) {
         );
       }
       await runOnStep(config.onStep, step, runtimeErrors);
-      if (noProgressStop.stop) {
+      if (noProgressSignal) {
         return finish(emitter, {
           intent: config.intent,
           pass: false,
           completed: true,
-          reason: noProgressStop.reason,
+          reason: noProgressSignal.reason,
           score: scoreAfter,
           steps: history,
           finalState: state,
@@ -759,24 +859,6 @@ async function runTrace(runtimeErrors, stepIndex, write) {
     return void 0;
   }
 }
-function noProgressStopDecision(args) {
-  const max = args.policies?.maxNoProgressSteps;
-  if (!max || max <= 0) return { stop: false, reason: "", streak: 0 };
-  const minScoreDelta = args.policies?.minScoreDelta ?? 1e-3;
-  const scoreDelta = Math.abs((args.scoreAfter ?? 0) - (args.scoreBefore ?? 0));
-  const stateUnchanged = args.lastStateFingerprint !== void 0 && args.lastStateFingerprint === args.stateFingerprint;
-  const scoreFlat = scoreDelta < minScoreDelta;
-  const streak = stateUnchanged && scoreFlat ? args.currentStreak + 1 : 0;
-  return streak >= max ? { stop: true, reason: `stuck: no state/score progress for ${streak} step(s)`, streak } : { stop: false, reason: "", streak };
-}
-function repeatedActionStopDecision(policies, streak) {
-  const max = policies?.maxRepeatedActions;
-  if (!max || max <= 0 || streak < max) return { stop: false, reason: "" };
-  return {
-    stop: true,
-    reason: `stuck: repeated same action for ${streak} step(s)`
-  };
-}
 function fingerprintState(state, policies) {
   if (policies?.stateFingerprint) return policies.stateFingerprint(state);
   return stableFingerprint(state);
@@ -828,6 +910,10 @@ async function finish(emitter, result) {
 }
 
 export {
+  repeatedActionDetector,
+  noProgressDetector,
+  errorStreakDetector,
+  observeAll,
   runAgentControlLoop,
   stopOnNoProgress,
   stopOnRepeatedAction,
@@ -835,4 +921,4 @@ export {
   subjectiveEval,
   allCriticalPassed
 };
-//# sourceMappingURL=chunk-NCRFYPS3.js.map
+//# sourceMappingURL=chunk-YEHAEDUD.js.map

package/dist/chunk-YEHAEDUD.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/detectors/index.ts","../src/control-runtime.ts"],"sourcesContent":["/**\n * Streaming detectors — the canonical, online failure-mode kernel.\n *\n * A `StreamingDetector` is a pure incremental reducer: fold one normalized step, get a signal the\n * moment a threshold trips (else null). The SAME kernel runs in two places, so the logic lives once:\n *   - ONLINE, over a live agent pipe (a worker's tool-call stream) → raise a finding mid-run.\n *   - INSIDE the control loop (`control-runtime`) → its `stopOn*` policies fold these detectors.\n *\n * Detectors are stateful but self-contained; `streak` is exposed for telemetry and `reset` clears it.\n * Fingerprinting is the CALLER's job — the control loop hashes state/action with `stableFingerprint`,\n * a tool-call pipe hashes args with `argHash` — so a detector stays agnostic to what it's watching.\n */\n\nimport type { FailureClass } from '../trace/schema'\n\nexport type DetectorSeverity = 'info' | 'warn' | 'critical'\n\n/** A normalized step a detector folds over. Fields are optional; each detector reads only what it\n *  needs (repeated-action → `actionFingerprint`; no-progress → `stateFingerprint` + `score`;\n *  error-streak → `status`). */\nexport interface DetectorEvent {\n  /** Fingerprint of the action/tool-call this step took (caller pre-hashes). */\n  readonly actionFingerprint?: string\n  /** Fingerprint of observable state AFTER this step. */\n  readonly stateFingerprint?: string\n  /** Score after this step (the score-flat half of no-progress). */\n  readonly score?: number\n  /** Whether this step errored. */\n  readonly status?: 'ok' | 'error'\n  /** Free-form label carried into the signal evidence (e.g. the tool name). */\n  readonly label?: string\n}\n\nexport interface DetectorSignal {\n  readonly detector: string\n  readonly severity: DetectorSeverity\n  readonly failureClass?: FailureClass\n  readonly reason: string\n  /** Consecutive matching steps at the moment the signal fired. */\n  readonly streak: number\n  readonly evidence?: Record<string, unknown>\n}\n\nexport interface StreamingDetector {\n  readonly id: string\n  /** Current streak (consecutive matching steps) — for telemetry/introspection between observes. */\n  readonly streak: number\n  /** Fold one event; return a signal when the threshold trips, else null. */\n  observe(event: DetectorEvent): DetectorSignal | null\n  /** Clear all state. */\n  reset(): void\n}\n\nexport interface RepeatedActionOptions {\n  /** Signal once the same action fingerprint repeats this many CONSECUTIVE steps (default 3).\n   *  `<= 0` disables the signal (the streak is still tracked, for telemetry). */\n  readonly maxRepeated?: number\n  readonly severity?: DetectorSeverity\n  /** Failure class to stamp on the signal (default `tool_recovery_failure`). */\n  readonly failureClass?: FailureClass\n}\n\n/** Same action fingerprint N consecutive steps = a stuck loop (the #1 long-horizon failure mode). */\nexport function repeatedActionDetector(opts: RepeatedActionOptions = {}): StreamingDetector {\n  const max = opts.maxRepeated ?? 3\n  const severity = opts.severity ?? 'warn'\n  const failureClass: FailureClass = opts.failureClass ?? 'tool_recovery_failure'\n  let last: string | undefined\n  let streak = 0\n  return {\n    id: 'repeated-action',\n    get streak() {\n      return streak\n    },\n    observe(event) {\n      const fp = event.actionFingerprint\n      if (fp === undefined) return null\n      streak = fp === last ? streak + 1 : 1\n      last = fp\n      if (max <= 0 || streak < max) return null\n      return {\n        detector: 'repeated-action',\n        severity,\n        failureClass,\n        reason: `stuck: repeated same action for ${streak} step(s)`,\n        streak,\n        ...(event.label ? { evidence: { action: event.label } } : {}),\n      }\n    },\n    reset() {\n      last = undefined\n      streak = 0\n    },\n  }\n}\n\nexport interface NoProgressOptions {\n  /** Signal once state is unchanged AND score is flat for this many CONSECUTIVE steps (default 3).\n   *  `<= 0` disables the signal. */\n  readonly maxNoProgress?: number\n  /** Minimum |score change| that counts as progress (default 0.001). */\n  readonly minScoreDelta?: number\n  readonly severity?: DetectorSeverity\n  readonly failureClass?: FailureClass\n}\n\n/** State + score unchanged across N steps = spinning wheels. Compares each step to the previous one,\n *  so prime with the initial state (observe it once before the first real step) to detect on step 1. */\nexport function noProgressDetector(opts: NoProgressOptions = {}): StreamingDetector {\n  const max = opts.maxNoProgress ?? 3\n  const minScoreDelta = opts.minScoreDelta ?? 0.001\n  const severity = opts.severity ?? 'warn'\n  const failureClass: FailureClass = opts.failureClass ?? 'tool_recovery_failure'\n  let lastState: string | undefined\n  let lastScore: number | undefined\n  let streak = 0\n  return {\n    id: 'no-progress',\n    get streak() {\n      return streak\n    },\n    observe(event) {\n      const stateUnchanged = lastState !== undefined && lastState === event.stateFingerprint\n      const scoreFlat = Math.abs((event.score ?? 0) - (lastScore ?? 0)) < minScoreDelta\n      streak = stateUnchanged && scoreFlat ? streak + 1 : 0\n      lastState = event.stateFingerprint\n      lastScore = event.score\n      if (max <= 0 || streak < max) return null\n      return {\n        detector: 'no-progress',\n        severity,\n        failureClass,\n        reason: `stuck: no state/score progress for ${streak} step(s)`,\n        streak,\n        ...(event.label ? { evidence: { state: event.label } } : {}),\n      }\n    },\n    reset() {\n      lastState = undefined\n      lastScore = undefined\n      streak = 0\n    },\n  }\n}\n\nexport interface ErrorStreakOptions {\n  /** Signal once this many CONSECUTIVE steps errored (default 3). `<= 0` disables. */\n  readonly maxErrors?: number\n  readonly severity?: DetectorSeverity\n  readonly failureClass?: FailureClass\n}\n\n/** N consecutive tool errors = the worker is hammering a broken approach. */\nexport function errorStreakDetector(opts: ErrorStreakOptions = {}): StreamingDetector {\n  const max = opts.maxErrors ?? 3\n  const severity = opts.severity ?? 'warn'\n  const failureClass: FailureClass = opts.failureClass ?? 'tool_recovery_failure'\n  let streak = 0\n  return {\n    id: 'error-streak',\n    get streak() {\n      return streak\n    },\n    observe(event) {\n      if (event.status === undefined) return null\n      streak = event.status === 'error' ? streak + 1 : 0\n      if (max <= 0 || streak < max) return null\n      return {\n        detector: 'error-streak',\n        severity,\n        failureClass,\n        reason: `stuck: ${streak} consecutive errored step(s)`,\n        streak,\n        ...(event.label ? { evidence: { lastError: event.label } } : {}),\n      }\n    },\n    reset() {\n      streak = 0\n    },\n  }\n}\n\n/** Fold one event through many detectors at once; returns every signal that fired this step. The\n *  natural shape for an online pipe watching with a whole panel of detectors. */\nexport function observeAll(\n  detectors: ReadonlyArray<StreamingDetector>,\n  event: DetectorEvent,\n): DetectorSignal[] {\n  const signals: DetectorSignal[] = []\n  for (const d of detectors) {\n    const s = d.observe(event)\n    if (s) signals.push(s)\n  }\n  return signals\n}\n","/**\n * Policy-based agent control runtime.\n *\n * This is the minimal reusable loop behind driver-agent patterns:\n *\n *   observe state -> validate -> decide next action -> act -> observe -> ...\n *\n * It deliberately does not model named \"topologies\". Direct execution,\n * critic/revise, driver intervention, specialist calls, and human escalation\n * are all just actions chosen by the control policy.\n */\n\nimport { noProgressDetector, repeatedActionDetector } from './detectors'\nimport { type SpanHandle, TraceEmitter } from './trace/emitter'\nimport type { FailureClass } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport type ControlSeverity = 'info' | 'warning' | 'error' | 'critical'\nexport type ControlActionFailureMode = 'continue' | 'stop'\n\nexport interface ControlEvalResult {\n  /** Stable validator or judge id. */\n  id: string\n  /** Whether this check passed. */\n  passed: boolean\n  /** Optional normalized score. 1 = best, 0 = worst. */\n  score?: number\n  /** Objective validators should usually be \"error\" or \"critical\" when failed. */\n  severity?: ControlSeverity\n  /** Human-readable result. */\n  detail?: string\n  /** Small evidence string or pointer. Avoid large payloads. */\n  evidence?: string\n  /** True when the result came from deterministic state, not LLM judgment. */\n  objective?: boolean\n  /** Structured details for downstream control policies and reports. */\n  metadata?: Record<string, unknown>\n}\n\nexport interface ControlBudget {\n  maxSteps: number\n  maxWallMs?: number\n  maxCostUsd?: number\n}\n\nexport interface ControlStopPolicies<TState, TAction> {\n  /**\n   * Stop after N consecutive steps with no state fingerprint change and\n   * less than `minScoreDelta` score movement. Disabled when omitted.\n   */\n  maxNoProgressSteps?: number\n  /**\n   * Stop after the same action fingerprint is selected N consecutive\n   * times. Disabled when omitted.\n   */\n  maxRepeatedActions?: number\n  /** Minimum score movement that counts as progress. Default 0.001. */\n  minScoreDelta?: number\n  /** Override the default JSON/string fingerprint for state comparisons. */\n  stateFingerprint?: (state: TState) => string\n  /** Override the default JSON/string fingerprint for repeated-action checks. */\n  actionFingerprint?: (action: TAction) => string\n}\n\nexport interface ControlContext<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  intent: string\n  state: TState\n  evals: TEval[]\n  history: ControlStep<TState, TAction, TActionResult, TEval>[]\n  budget: ControlBudget\n  stepIndex: number\n  wallMs: number\n  spentCostUsd: number\n  remainingCostUsd?: number\n  abortSignal: AbortSignal\n  emitter?: TraceEmitter\n}\n\nexport type ControlDecision<TAction> =\n  | {\n      type: 'continue'\n      action: TAction\n      reason?: string\n    }\n  | {\n      type: 'stop'\n      reason: string\n      pass?: boolean\n      score?: number\n    }\n\nexport interface StopDecision {\n  stop: boolean\n  pass: boolean\n  reason: string\n  score?: number\n  failureClass?: FailureClass\n}\n\nexport interface ControlActionOutcome<TActionResult> {\n  ok: boolean\n  result?: TActionResult\n  error?: string\n  costUsd?: number\n  durationMs: number\n}\n\nexport interface ControlRuntimeError {\n  phase: 'observe' | 'validate' | 'decide' | 'act' | 'stop-policy' | 'on-step' | 'trace'\n  stepIndex: number\n  message: string\n}\n\nexport interface ControlStep<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  index: number\n  decision: ControlDecision<TAction>\n  beforeState: TState\n  afterState: TState\n  evalsBefore: TEval[]\n  evalsAfter: TEval[]\n  actionOutcome?: ControlActionOutcome<TActionResult>\n  startedAt: string\n  endedAt: string\n}\n\nexport interface ControlRunResult<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  intent: string\n  pass: boolean\n  completed: boolean\n  reason: string\n  score?: number\n  steps: ControlStep<TState, TAction, TActionResult, TEval>[]\n  finalState: TState | undefined\n  finalEvals: TEval[]\n  wallMs: number\n  spentCostUsd: number\n  /** null when the run executed without a TraceEmitter wired (no run record was persisted). */\n  runId: string | null\n  failureClass?: FailureClass\n  runtimeErrors: ControlRuntimeError[]\n  stoppedBy: 'policy' | 'stop-policy' | 'budget' | 'abort' | 'runtime-error'\n}\n\nexport interface ControlRuntimeConfig<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n> {\n  intent: string\n  budget?: Partial<ControlBudget>\n  signal?: AbortSignal\n  /** Defaults to `continue`: action failures are recorded, then the policy gets another chance. */\n  actionFailure?: ControlActionFailureMode\n  /**\n   * Extract cost from an action result. Used for `maxCostUsd` budget\n   * enforcement and trace budget ledger emission.\n   */\n  getActionCostUsd?: (ctx: {\n    action: TAction\n    result: TActionResult\n    state: TState\n    evals: TEval[]\n    history: ControlStep<TState, TAction, TActionResult, TEval>[]\n  }) => number | undefined\n\n  /** Read typed task/product state. Prefer structured state over transcript-only context. */\n  observe: (ctx: {\n    history: ControlStep<TState, TAction, TActionResult, TEval>[]\n    abortSignal: AbortSignal\n  }) => Promise<TState> | TState\n\n  /** Objective validators first, subjective judges only where objective state is insufficient. */\n  validate: (ctx: {\n    intent: string\n    state: TState\n    history: ControlStep<TState, TAction, TActionResult, TEval>[]\n    abortSignal: AbortSignal\n  }) => Promise<TEval[]> | TEval[]\n\n  /** Choose the next control action. Can call a worker, ask user, run critic, inspect state, or stop. */\n  decide: (\n    ctx: ControlContext<TState, TAction, TActionResult, TEval>,\n  ) => Promise<ControlDecision<TAction>> | ControlDecision<TAction>\n\n  /** Execute the action selected by the policy. */\n  act: (\n    action: TAction,\n    ctx: ControlContext<TState, TAction, TActionResult, TEval>,\n  ) => Promise<TActionResult> | TActionResult\n\n  /** Final stopping policy. Called before decide and after each action. */\n  shouldStop?: (\n    ctx: ControlContext<TState, TAction, TActionResult, TEval>,\n  ) => Promise<StopDecision> | StopDecision\n\n  /** Optional hook for tracing or live progress updates. */\n  onStep?: (step: ControlStep<TState, TAction, TActionResult, TEval>) => Promise<void> | void\n\n  /** Optional generic stuck-loop policies. Custom `shouldStop` still runs first. */\n  stopPolicies?: ControlStopPolicies<TState, TAction>\n\n  /** Optional trace sink. Emits one run plus one span per control step. */\n  store?: TraceStore\n  scenarioId?: string\n  projectId?: string\n  variantId?: string\n}\n\nconst DEFAULT_BUDGET: ControlBudget = {\n  maxSteps: 8,\n  maxWallMs: 5 * 60 * 1000,\n}\n\nexport async function runAgentControlLoop<\n  TState,\n  TAction,\n  TActionResult,\n  TEval extends ControlEvalResult = ControlEvalResult,\n>(\n  config: ControlRuntimeConfig<TState, TAction, TActionResult, TEval>,\n): Promise<ControlRunResult<TState, TAction, TActionResult, TEval>> {\n  const budget = normalizeBudget(config.budget)\n  const actionFailure = config.actionFailure ?? 'continue'\n  const controller = new AbortController()\n  const upstreamAbort = () => controller.abort(config.signal?.reason)\n  if (config.signal) {\n    if (config.signal.aborted) controller.abort(config.signal.reason)\n    else config.signal.addEventListener('abort', upstreamAbort, { once: true })\n  }\n\n  const started = Date.now()\n  const wallTimer = budget.maxWallMs\n    ? setTimeout(\n        () => controller.abort(new Error('control runtime wall timeout')),\n        budget.maxWallMs,\n      )\n    : undefined\n  const history: ControlStep<TState, TAction, TActionResult, TEval>[] = []\n  const emitter = config.store ? new TraceEmitter(config.store) : undefined\n  let spentCostUsd = 0\n  const runtimeErrors: ControlRuntimeError[] = []\n  // The stuck-detection kernels (shared with the online detectors). Thresholds 0 = disabled; the\n  // streak is still tracked for telemetry. Both stops report `tool_recovery_failure`.\n  const repeatedDetector = repeatedActionDetector({\n    maxRepeated: config.stopPolicies?.maxRepeatedActions ?? 0,\n  })\n  const progressDetector = noProgressDetector({\n    maxNoProgress: config.stopPolicies?.maxNoProgressSteps ?? 0,\n    minScoreDelta: config.stopPolicies?.minScoreDelta ?? 0.001,\n  })\n\n  try {\n    if (emitter) {\n      await runTrace(runtimeErrors, 0, () =>\n        emitter.startRun({\n          scenarioId: config.scenarioId ?? 'agent-control-loop',\n          projectId: config.projectId,\n          variantId: config.variantId,\n          layer: 'meta',\n          tags: {\n            intent: config.intent.slice(0, 120),\n            maxSteps: String(budget.maxSteps),\n            ...(budget.maxCostUsd !== undefined ? { maxCostUsd: String(budget.maxCostUsd) } : {}),\n          },\n        }),\n      )\n    }\n\n    let state: TState\n    let evals: TEval[]\n    try {\n      state = await config.observe({ history, abortSignal: controller.signal })\n    } catch (err) {\n      const error = runtimeError('observe', 0, err)\n      runtimeErrors.push(error)\n      return finish(emitter, {\n        intent: config.intent,\n        pass: false,\n        completed: false,\n        reason: error.message,\n        steps: history,\n        finalState: undefined,\n        finalEvals: [],\n        wallMs: Date.now() - started,\n        spentCostUsd,\n        runId: emitter?.runId ?? null,\n        failureClass: 'unknown',\n        runtimeErrors,\n        stoppedBy: 'runtime-error',\n      })\n    }\n    try {\n      evals = await config.validate({\n        intent: config.intent,\n        state,\n        history,\n        abortSignal: controller.signal,\n      })\n      await recordEvalSpans(emitter, evals, 'initial', runtimeErrors, 0)\n    } catch (err) {\n      const error = runtimeError('validate', 0, err)\n      runtimeErrors.push(error)\n      return finish(emitter, {\n        intent: config.intent,\n        pass: false,\n        completed: false,\n        reason: error.message,\n        steps: history,\n        finalState: state,\n        finalEvals: [],\n        wallMs: Date.now() - started,\n        spentCostUsd,\n        runId: emitter?.runId ?? null,\n        failureClass: 'unknown',\n        runtimeErrors,\n        stoppedBy: 'runtime-error',\n      })\n    }\n    // Prime the no-progress detector with the initial state + score so step 0 can already detect a\n    // no-op (mirrors the pre-loop fingerprint baseline). The priming observe never signals (streak 0).\n    progressDetector.observe({\n      stateFingerprint: fingerprintState(state, config.stopPolicies),\n      score: averageScore(evals),\n    })\n\n    for (let stepIndex = 0; stepIndex < budget.maxSteps; stepIndex++) {\n      if (controller.signal.aborted) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: abortReason(controller.signal),\n          score: undefined,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'timeout',\n          runtimeErrors,\n          stoppedBy: 'abort',\n        })\n      }\n\n      const budgetStop = budgetStopDecision(budget, spentCostUsd)\n      if (budgetStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: budgetStop.reason,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'budget_exceeded',\n          runtimeErrors,\n          stoppedBy: 'budget',\n        })\n      }\n\n      const ctx = makeContext(\n        config.intent,\n        state,\n        evals,\n        history,\n        budget,\n        stepIndex,\n        started,\n        spentCostUsd,\n        controller.signal,\n        emitter,\n      )\n      let stop: StopDecision\n      try {\n        stop = config.shouldStop ? await config.shouldStop(ctx) : defaultStopDecision(evals)\n      } catch (err) {\n        runtimeErrors.push(runtimeError('stop-policy', stepIndex, err))\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      if (stop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: stop.pass,\n          completed: true,\n          reason: stop.reason,\n          score: stop.score,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: stop.failureClass,\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n\n      let decision: ControlDecision<TAction>\n      try {\n        decision = await config.decide(ctx)\n      } catch (err) {\n        runtimeErrors.push(runtimeError('decide', stepIndex, err))\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      if (decision.type === 'stop') {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: decision.pass ?? false,\n          completed: true,\n          reason: decision.reason,\n          score: decision.score,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: decision.pass === false ? 'unknown' : undefined,\n          runtimeErrors,\n          stoppedBy: 'policy',\n        })\n      }\n\n      const actionFingerprint = fingerprintAction(decision.action, config.stopPolicies)\n      const repeatedActionSignal = repeatedDetector.observe({ actionFingerprint })\n      const repeatedActionStreak = repeatedDetector.streak\n      if (repeatedActionSignal) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: true,\n          reason: repeatedActionSignal.reason,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'tool_recovery_failure',\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n\n      const beforeState = state\n      const evalsBefore = evals\n      const scoreBefore = averageScore(evals)\n      const actionStarted = Date.now()\n      const stepHandle = emitter\n        ? await runTrace(runtimeErrors, stepIndex, () =>\n            emitter.tool({\n              name: `control-step-${stepIndex}`,\n              toolName: 'agent-control-action',\n              args: decision.action,\n              attributes: {\n                decision: decision.reason ?? 'continue',\n                repeatedActionStreak,\n              },\n            }),\n          )\n        : undefined\n      let actionOutcome: ControlActionOutcome<TActionResult>\n      try {\n        const result = await config.act(decision.action, ctx)\n        const rawCostUsd = config.getActionCostUsd?.({\n          action: decision.action,\n          result,\n          state,\n          evals,\n          history,\n        })\n        const costUsd = normalizeActionCostUsd(rawCostUsd, runtimeErrors, stepIndex)\n        if (costUsd !== undefined && Number.isFinite(costUsd) && costUsd > 0) {\n          spentCostUsd += costUsd\n          await recordCostBudget(\n            emitter,\n            budget,\n            spentCostUsd,\n            stepHandle,\n            runtimeErrors,\n            stepIndex,\n          )\n        }\n        actionOutcome = {\n          ok: true,\n          result,\n          ...(costUsd !== undefined ? { costUsd } : {}),\n          durationMs: Date.now() - actionStarted,\n        }\n      } catch (err) {\n        runtimeErrors.push(runtimeError('act', stepIndex, err))\n        actionOutcome = {\n          ok: false,\n          error: runtimeErrors[runtimeErrors.length - 1]!.message,\n          durationMs: Date.now() - actionStarted,\n        }\n        if (actionFailure === 'stop') {\n          await runTrace(runtimeErrors, stepIndex, () =>\n            stepHandle?.fail(actionOutcome.error ?? 'action failed'),\n          )\n          const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n            index: stepIndex,\n            decision,\n            beforeState,\n            afterState: state,\n            evalsBefore,\n            evalsAfter: evals,\n            actionOutcome,\n            startedAt: new Date(actionStarted).toISOString(),\n            endedAt: new Date().toISOString(),\n          }\n          history.push(step)\n          await runOnStep(config.onStep, step, runtimeErrors)\n          return finish(emitter, {\n            intent: config.intent,\n            pass: false,\n            completed: false,\n            reason: actionOutcome.error ?? 'action failed',\n            score: averageScore(evals),\n            steps: history,\n            finalState: state,\n            finalEvals: evals,\n            wallMs: Date.now() - started,\n            spentCostUsd,\n            runId: emitter?.runId ?? null,\n            failureClass: 'unknown',\n            runtimeErrors,\n            stoppedBy: 'runtime-error',\n          })\n        }\n      }\n\n      try {\n        state = await config.observe({ history, abortSignal: controller.signal })\n      } catch (err) {\n        runtimeErrors.push(runtimeError('observe', stepIndex, err))\n        const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n          index: stepIndex,\n          decision,\n          beforeState,\n          afterState: beforeState,\n          evalsBefore,\n          evalsAfter: evals,\n          actionOutcome,\n          startedAt: new Date(actionStarted).toISOString(),\n          endedAt: new Date().toISOString(),\n        }\n        history.push(step)\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.fail(runtimeErrors[runtimeErrors.length - 1]!.message),\n        )\n        await runOnStep(config.onStep, step, runtimeErrors)\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: beforeState,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      try {\n        evals = await config.validate({\n          intent: config.intent,\n          state,\n          history,\n          abortSignal: controller.signal,\n        })\n        await recordEvalSpans(\n          emitter,\n          evals,\n          `step-${stepIndex}`,\n          runtimeErrors,\n          stepIndex,\n          stepHandle?.span.spanId,\n        )\n      } catch (err) {\n        runtimeErrors.push(runtimeError('validate', stepIndex, err))\n        const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n          index: stepIndex,\n          decision,\n          beforeState,\n          afterState: state,\n          evalsBefore,\n          evalsAfter: evals,\n          actionOutcome,\n          startedAt: new Date(actionStarted).toISOString(),\n          endedAt: new Date().toISOString(),\n        }\n        history.push(step)\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.fail(runtimeErrors[runtimeErrors.length - 1]!.message),\n        )\n        await runOnStep(config.onStep, step, runtimeErrors)\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      const scoreAfter = averageScore(evals)\n      const stateFingerprint = fingerprintState(state, config.stopPolicies)\n      const noProgressSignal = progressDetector.observe({ stateFingerprint, score: scoreAfter })\n      const noProgressStreak = progressDetector.streak\n\n      const step: ControlStep<TState, TAction, TActionResult, TEval> = {\n        index: stepIndex,\n        decision,\n        beforeState,\n        afterState: state,\n        evalsBefore,\n        evalsAfter: evals,\n        actionOutcome,\n        startedAt: new Date(actionStarted).toISOString(),\n        endedAt: new Date().toISOString(),\n      }\n      history.push(step)\n      if (actionOutcome.ok) {\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.end({\n            attributes: {\n              actionCostUsd: actionOutcome.costUsd ?? null,\n              spentCostUsd,\n              scoreBefore: scoreBefore ?? null,\n              scoreAfter: scoreAfter ?? null,\n              noProgressStreak,\n            },\n          }),\n        )\n      } else {\n        await runTrace(runtimeErrors, stepIndex, () =>\n          stepHandle?.fail(actionOutcome.error ?? 'action failed', {\n            attributes: {\n              spentCostUsd,\n              noProgressStreak,\n            },\n          }),\n        )\n      }\n      await runOnStep(config.onStep, step, runtimeErrors)\n\n      if (noProgressSignal) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: true,\n          reason: noProgressSignal.reason,\n          score: scoreAfter,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'tool_recovery_failure',\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n\n      const postStepBudgetStop = budgetStopDecision(budget, spentCostUsd)\n      if (postStepBudgetStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: postStepBudgetStop.reason,\n          score: scoreAfter,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'budget_exceeded',\n          runtimeErrors,\n          stoppedBy: 'budget',\n        })\n      }\n\n      const postStepCtx = makeContext(\n        config.intent,\n        state,\n        evals,\n        history,\n        budget,\n        stepIndex + 1,\n        started,\n        spentCostUsd,\n        controller.signal,\n        emitter,\n      )\n      let postStepStop: StopDecision\n      try {\n        postStepStop = config.shouldStop\n          ? await config.shouldStop(postStepCtx)\n          : defaultStopDecision(evals)\n      } catch (err) {\n        runtimeErrors.push(runtimeError('stop-policy', stepIndex + 1, err))\n        return finish(emitter, {\n          intent: config.intent,\n          pass: false,\n          completed: false,\n          reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n          score: averageScore(evals),\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: 'unknown',\n          runtimeErrors,\n          stoppedBy: 'runtime-error',\n        })\n      }\n      if (postStepStop.stop) {\n        return finish(emitter, {\n          intent: config.intent,\n          pass: postStepStop.pass,\n          completed: true,\n          reason: postStepStop.reason,\n          score: postStepStop.score,\n          steps: history,\n          finalState: state,\n          finalEvals: evals,\n          wallMs: Date.now() - started,\n          spentCostUsd,\n          runId: emitter?.runId ?? null,\n          failureClass: postStepStop.failureClass,\n          runtimeErrors,\n          stoppedBy: 'stop-policy',\n        })\n      }\n    }\n\n    return finish(emitter, {\n      intent: config.intent,\n      pass: false,\n      completed: false,\n      reason: `budget exhausted: maxSteps=${budget.maxSteps}`,\n      steps: history,\n      finalState: state,\n      finalEvals: evals,\n      wallMs: Date.now() - started,\n      spentCostUsd,\n      runId: emitter?.runId ?? null,\n      failureClass: 'budget_exceeded',\n      runtimeErrors,\n      stoppedBy: 'budget',\n    })\n  } catch (err) {\n    runtimeErrors.push(runtimeError('act', history.length, err))\n    return finish(emitter, {\n      intent: config.intent,\n      pass: false,\n      completed: false,\n      reason: runtimeErrors[runtimeErrors.length - 1]!.message,\n      steps: history,\n      finalState: undefined,\n      finalEvals: [],\n      wallMs: Date.now() - started,\n      spentCostUsd,\n      runId: emitter?.runId ?? null,\n      failureClass: 'unknown',\n      runtimeErrors,\n      stoppedBy: 'runtime-error',\n    })\n  } finally {\n    if (wallTimer) clearTimeout(wallTimer)\n    if (config.signal) config.signal.removeEventListener('abort', upstreamAbort)\n  }\n}\n\nexport function stopOnNoProgress<TState, TAction>(\n  maxNoProgressSteps: number,\n  options: Omit<ControlStopPolicies<TState, TAction>, 'maxNoProgressSteps'> = {},\n): ControlStopPolicies<TState, TAction> {\n  return { ...options, maxNoProgressSteps }\n}\n\nexport function stopOnRepeatedAction<TState, TAction>(\n  maxRepeatedActions: number,\n  options: Omit<ControlStopPolicies<TState, TAction>, 'maxRepeatedActions'> = {},\n): ControlStopPolicies<TState, TAction> {\n  return { ...options, maxRepeatedActions }\n}\n\nexport function objectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult {\n  return { ...input, objective: true }\n}\n\nexport function subjectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult {\n  return { ...input, objective: false }\n}\n\nfunction normalizeBudget(input: Partial<ControlBudget> | undefined): ControlBudget {\n  const raw = { ...DEFAULT_BUDGET, ...input } as Record<string, unknown>\n  if (!Number.isInteger(raw.maxSteps) || (raw.maxSteps as number) < 1) {\n    throw new RangeError(\n      `ControlRuntime budget.maxSteps must be an integer >= 1, got ${String(raw.maxSteps)}`,\n    )\n  }\n  const budget: ControlBudget = { maxSteps: raw.maxSteps as number }\n  if (raw.maxWallMs !== undefined) {\n    if (\n      typeof raw.maxWallMs !== 'number' ||\n      !Number.isFinite(raw.maxWallMs) ||\n      raw.maxWallMs <= 0\n    ) {\n      throw new RangeError(\n        `ControlRuntime budget.maxWallMs must be a positive finite number, got ${String(raw.maxWallMs)}`,\n      )\n    }\n    budget.maxWallMs = raw.maxWallMs\n  }\n  if (raw.maxCostUsd !== undefined) {\n    if (\n      typeof raw.maxCostUsd !== 'number' ||\n      !Number.isFinite(raw.maxCostUsd) ||\n      raw.maxCostUsd < 0\n    ) {\n      throw new RangeError(\n        `ControlRuntime budget.maxCostUsd must be a nonnegative finite number, got ${String(raw.maxCostUsd)}`,\n      )\n    }\n    budget.maxCostUsd = raw.maxCostUsd\n  }\n  return budget\n}\n\nfunction normalizeActionCostUsd(\n  costUsd: number | undefined,\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n): number | undefined {\n  if (costUsd === undefined) return undefined\n  if (!Number.isFinite(costUsd) || costUsd < 0) {\n    runtimeErrors.push(\n      runtimeError('act', stepIndex, new Error(`invalid action costUsd: ${String(costUsd)}`)),\n    )\n    return undefined\n  }\n  return costUsd\n}\n\nexport function allCriticalPassed(evals: ControlEvalResult[]): boolean {\n  return evals.every(\n    (result) => result.passed || (result.severity !== 'critical' && result.severity !== 'error'),\n  )\n}\n\nfunction makeContext<TState, TAction, TActionResult, TEval extends ControlEvalResult>(\n  intent: string,\n  state: TState,\n  evals: TEval[],\n  history: ControlStep<TState, TAction, TActionResult, TEval>[],\n  budget: ControlBudget,\n  stepIndex: number,\n  started: number,\n  spentCostUsd: number,\n  abortSignal: AbortSignal,\n  emitter?: TraceEmitter,\n): ControlContext<TState, TAction, TActionResult, TEval> {\n  return {\n    intent,\n    state,\n    evals,\n    history,\n    budget,\n    stepIndex,\n    wallMs: Date.now() - started,\n    spentCostUsd,\n    remainingCostUsd:\n      budget.maxCostUsd === undefined ? undefined : Math.max(0, budget.maxCostUsd - spentCostUsd),\n    abortSignal,\n    emitter,\n  }\n}\n\nfunction defaultStopDecision(evals: ControlEvalResult[]): StopDecision {\n  if (!evals.length) return { stop: false, pass: false, reason: 'no evals yet' }\n  const pass = allCriticalPassed(evals)\n  return pass\n    ? { stop: true, pass: true, reason: 'all critical evals passed', score: averageScore(evals) }\n    : {\n        stop: false,\n        pass: false,\n        reason: 'critical evals still failing',\n        score: averageScore(evals),\n      }\n}\n\nfunction averageScore(evals: ControlEvalResult[]): number | undefined {\n  const scored = evals\n    .map((result) => result.score)\n    .filter((score): score is number => typeof score === 'number')\n  if (!scored.length) return undefined\n  return Math.round((scored.reduce((sum, score) => sum + score, 0) / scored.length) * 1000) / 1000\n}\n\nfunction budgetStopDecision(\n  budget: ControlBudget,\n  spentCostUsd: number,\n): { stop: boolean; reason: string } {\n  if (budget.maxCostUsd !== undefined && spentCostUsd >= budget.maxCostUsd) {\n    return {\n      stop: true,\n      reason: `budget exhausted: maxCostUsd=${budget.maxCostUsd}`,\n    }\n  }\n  return { stop: false, reason: '' }\n}\n\nasync function recordCostBudget(\n  emitter: TraceEmitter | undefined,\n  budget: ControlBudget,\n  spentCostUsd: number,\n  handle: SpanHandle | undefined,\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n): Promise<void> {\n  if (!emitter || budget.maxCostUsd === undefined) return\n  const maxCostUsd = budget.maxCostUsd\n  await runTrace(runtimeErrors, stepIndex, () =>\n    emitter.recordBudget({\n      dimension: 'usd',\n      limit: maxCostUsd,\n      consumed: spentCostUsd,\n      remaining: Math.max(0, maxCostUsd - spentCostUsd),\n      breached: spentCostUsd >= maxCostUsd,\n      spanId: handle?.span.spanId,\n    }),\n  )\n}\n\nasync function recordEvalSpans(\n  emitter: TraceEmitter | undefined,\n  evals: ControlEvalResult[],\n  phase: string,\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n  targetSpanId?: string,\n): Promise<void> {\n  if (!emitter) return\n  for (const result of evals) {\n    await runTrace(runtimeErrors, stepIndex, () =>\n      emitter.recordJudge({\n        judgeId: result.objective ? 'objective-validator' : 'subjective-judge',\n        targetSpanId: targetSpanId ?? emitter.runId,\n        name: `control-eval/${result.id}`,\n        dimension: result.id,\n        score: typeof result.score === 'number' ? result.score : result.passed ? 1 : 0,\n        rationale: result.detail,\n        evidence: result.evidence,\n        attributes: {\n          phase,\n          passed: result.passed,\n          severity: result.severity,\n          objective: result.objective,\n        },\n      }),\n    )\n  }\n}\n\nasync function runOnStep<TState, TAction, TActionResult, TEval extends ControlEvalResult>(\n  onStep: ControlRuntimeConfig<TState, TAction, TActionResult, TEval>['onStep'] | undefined,\n  step: ControlStep<TState, TAction, TActionResult, TEval>,\n  runtimeErrors: ControlRuntimeError[],\n): Promise<void> {\n  if (!onStep) return\n  try {\n    await onStep(step)\n  } catch (err) {\n    runtimeErrors.push(runtimeError('on-step', step.index, err))\n  }\n}\n\nasync function runTrace<T>(\n  runtimeErrors: ControlRuntimeError[],\n  stepIndex: number,\n  write: () => Promise<T | undefined> | T | undefined,\n): Promise<T | undefined> {\n  try {\n    return await write()\n  } catch (err) {\n    runtimeErrors.push(runtimeError('trace', stepIndex, err))\n    return undefined\n  }\n}\n\nfunction fingerprintState<TState, TAction>(\n  state: TState,\n  policies?: ControlStopPolicies<TState, TAction>,\n): string {\n  if (policies?.stateFingerprint) return policies.stateFingerprint(state)\n  return stableFingerprint(state)\n}\n\nfunction fingerprintAction<TState, TAction>(\n  action: TAction,\n  policies?: ControlStopPolicies<TState, TAction>,\n): string {\n  if (policies?.actionFingerprint) return policies.actionFingerprint(action)\n  return stableFingerprint(action)\n}\n\nfunction stableFingerprint(value: unknown): string {\n  if (typeof value === 'string') return value\n  if (typeof value === 'number' || typeof value === 'boolean' || value == null) return String(value)\n  try {\n    return JSON.stringify(sortForFingerprint(value))\n  } catch {\n    return String(value)\n  }\n}\n\nfunction sortForFingerprint(value: unknown): unknown {\n  if (Array.isArray(value)) return value.map(sortForFingerprint)\n  if (!value || typeof value !== 'object') return value\n  const record = value as Record<string, unknown>\n  const sorted: Record<string, unknown> = {}\n  for (const key of Object.keys(record).sort()) {\n    sorted[key] = sortForFingerprint(record[key])\n  }\n  return sorted\n}\n\nfunction abortReason(signal: AbortSignal): string {\n  const reason = signal.reason\n  if (reason instanceof Error) return reason.message\n  return reason ? String(reason) : 'aborted'\n}\n\nfunction runtimeError(\n  phase: ControlRuntimeError['phase'],\n  stepIndex: number,\n  err: unknown,\n): ControlRuntimeError {\n  const message = err instanceof Error ? err.message : String(err)\n  return { phase, stepIndex, message }\n}\n\nasync function finish<TState, TAction, TActionResult, TEval extends ControlEvalResult>(\n  emitter: TraceEmitter | undefined,\n  result: ControlRunResult<TState, TAction, TActionResult, TEval>,\n): Promise<ControlRunResult<TState, TAction, TActionResult, TEval>> {\n  await runTrace(result.runtimeErrors, result.steps.length, () =>\n    emitter?.endRun({\n      pass: result.pass,\n      score: result.score ?? averageScore(result.finalEvals),\n      failureClass: result.failureClass,\n      notes: result.reason,\n    }),\n  )\n  return result\n}\n"],"mappings":";;;;;AA+DO,SAAS,uBAAuB,OAA8B,CAAC,GAAsB;AAC1F,QAAM,MAAM,KAAK,eAAe;AAChC,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,eAA6B,KAAK,gBAAgB;AACxD,MAAI;AACJ,MAAI,SAAS;AACb,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,IAAI,SAAS;AACX,aAAO;AAAA,IACT;AAAA,IACA,QAAQ,OAAO;AACb,YAAM,KAAK,MAAM;AACjB,UAAI,OAAO,OAAW,QAAO;AAC7B,eAAS,OAAO,OAAO,SAAS,IAAI;AACpC,aAAO;AACP,UAAI,OAAO,KAAK,SAAS,IAAK,QAAO;AACrC,aAAO;AAAA,QACL,UAAU;AAAA,QACV;AAAA,QACA;AAAA,QACA,QAAQ,mCAAmC,MAAM;AAAA,QACjD;AAAA,QACA,GAAI,MAAM,QAAQ,EAAE,UAAU,EAAE,QAAQ,MAAM,MAAM,EAAE,IAAI,CAAC;AAAA,MAC7D;AAAA,IACF;AAAA,IACA,QAAQ;AACN,aAAO;AACP,eAAS;AAAA,IACX;AAAA,EACF;AACF;AAcO,SAAS,mBAAmB,OAA0B,CAAC,GAAsB;AAClF,QAAM,MAAM,KAAK,iBAAiB;AAClC,QAAM,gBAAgB,KAAK,iBAAiB;AAC5C,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,eAA6B,KAAK,gBAAgB;AACxD,MAAI;AACJ,MAAI;AACJ,MAAI,SAAS;AACb,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,IAAI,SAAS;AACX,aAAO;AAAA,IACT;AAAA,IACA,QAAQ,OAAO;AACb,YAAM,iBAAiB,cAAc,UAAa,cAAc,MAAM;AACtE,YAAM,YAAY,KAAK,KAAK,MAAM,SAAS,MAAM,aAAa,EAAE,IAAI;AACpE,eAAS,kBAAkB,YAAY,SAAS,IAAI;AACpD,kBAAY,MAAM;AAClB,kBAAY,MAAM;AAClB,UAAI,OAAO,KAAK,SAAS,IAAK,QAAO;AACrC,aAAO;AAAA,QACL,UAAU;AAAA,QACV;AAAA,QACA;AAAA,QACA,QAAQ,sCAAsC,MAAM;AAAA,QACpD;AAAA,QACA,GAAI,MAAM,QAAQ,EAAE,UAAU,EAAE,OAAO,MAAM,MAAM,EAAE,IAAI,CAAC;AAAA,MAC5D;AAAA,IACF;AAAA,IACA,QAAQ;AACN,kBAAY;AACZ,kBAAY;AACZ,eAAS;AAAA,IACX;AAAA,EACF;AACF;AAUO,SAAS,oBAAoB,OAA2B,CAAC,GAAsB;AACpF,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,WAAW,KAAK,YAAY;AAClC,QAAM,eAA6B,KAAK,gBAAgB;AACxD,MAAI,SAAS;AACb,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,IAAI,SAAS;AACX,aAAO;AAAA,IACT;AAAA,IACA,QAAQ,OAAO;AACb,UAAI,MAAM,WAAW,OAAW,QAAO;AACvC,eAAS,MAAM,WAAW,UAAU,SAAS,IAAI;AACjD,UAAI,OAAO,KAAK,SAAS,IAAK,QAAO;AACrC,aAAO;AAAA,QACL,UAAU;AAAA,QACV;AAAA,QACA;AAAA,QACA,QAAQ,UAAU,MAAM;AAAA,QACxB;AAAA,QACA,GAAI,MAAM,QAAQ,EAAE,UAAU,EAAE,WAAW,MAAM,MAAM,EAAE,IAAI,CAAC;AAAA,MAChE;AAAA,IACF;AAAA,IACA,QAAQ;AACN,eAAS;AAAA,IACX;AAAA,EACF;AACF;AAIO,SAAS,WACd,WACA,OACkB;AAClB,QAAM,UAA4B,CAAC;AACnC,aAAW,KAAK,WAAW;AACzB,UAAM,IAAI,EAAE,QAAQ,KAAK;AACzB,QAAI,EAAG,SAAQ,KAAK,CAAC;AAAA,EACvB;AACA,SAAO;AACT;;;AC8BA,IAAM,iBAAgC;AAAA,EACpC,UAAU;AAAA,EACV,WAAW,IAAI,KAAK;AACtB;AAEA,eAAsB,oBAMpB,QACkE;AAClE,QAAM,SAAS,gBAAgB,OAAO,MAAM;AAC5C,QAAM,gBAAgB,OAAO,iBAAiB;AAC9C,QAAM,aAAa,IAAI,gBAAgB;AACvC,QAAM,gBAAgB,MAAM,WAAW,MAAM,OAAO,QAAQ,MAAM;AAClE,MAAI,OAAO,QAAQ;AACjB,QAAI,OAAO,OAAO,QAAS,YAAW,MAAM,OAAO,OAAO,MAAM;AAAA,QAC3D,QAAO,OAAO,iBAAiB,SAAS,eAAe,EAAE,MAAM,KAAK,CAAC;AAAA,EAC5E;AAEA,QAAM,UAAU,KAAK,IAAI;AACzB,QAAM,YAAY,OAAO,YACrB;AAAA,IACE,MAAM,WAAW,MAAM,IAAI,MAAM,8BAA8B,CAAC;AAAA,IAChE,OAAO;AAAA,EACT,IACA;AACJ,QAAM,UAAgE,CAAC;AACvE,QAAM,UAAU,OAAO,QAAQ,IAAI,aAAa,OAAO,KAAK,IAAI;AAChE,MAAI,eAAe;AACnB,QAAM,gBAAuC,CAAC;AAG9C,QAAM,mBAAmB,uBAAuB;AAAA,IAC9C,aAAa,OAAO,cAAc,sBAAsB;AAAA,EAC1D,CAAC;AACD,QAAM,mBAAmB,mBAAmB;AAAA,IAC1C,eAAe,OAAO,cAAc,sBAAsB;AAAA,IAC1D,eAAe,OAAO,cAAc,iBAAiB;AAAA,EACvD,CAAC;AAED,MAAI;AACF,QAAI,SAAS;AACX,YAAM;AAAA,QAAS;AAAA,QAAe;AAAA,QAAG,MAC/B,QAAQ,SAAS;AAAA,UACf,YAAY,OAAO,cAAc;AAAA,UACjC,WAAW,OAAO;AAAA,UAClB,WAAW,OAAO;AAAA,UAClB,OAAO;AAAA,UACP,MAAM;AAAA,YACJ,QAAQ,OAAO,OAAO,MAAM,GAAG,GAAG;AAAA,YAClC,UAAU,OAAO,OAAO,QAAQ;AAAA,YAChC,GAAI,OAAO,eAAe,SAAY,EAAE,YAAY,OAAO,OAAO,UAAU,EAAE,IAAI,CAAC;AAAA,UACrF;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF;AAEA,QAAI;AACJ,QAAI;AACJ,QAAI;AACF,cAAQ,MAAM,OAAO,QAAQ,EAAE,SAAS,aAAa,WAAW,OAAO,CAAC;AAAA,IAC1E,SAAS,KAAK;AACZ,YAAM,QAAQ,aAAa,WAAW,GAAG,GAAG;AAC5C,oBAAc,KAAK,KAAK;AACxB,aAAO,OAAO,SAAS;AAAA,QACrB,QAAQ,OAAO;AAAA,QACf,MAAM;AAAA,QACN,WAAW;AAAA,QACX,QAAQ,MAAM;AAAA,QACd,OAAO;AAAA,QACP,YAAY;AAAA,QACZ,YAAY,CAAC;AAAA,QACb,QAAQ,KAAK,IAAI,IAAI;AAAA,QACrB;AAAA,QACA,OAAO,SAAS,SAAS;AAAA,QACzB,cAAc;AAAA,QACd;AAAA,QACA,WAAW;AAAA,MACb,CAAC;AAAA,IACH;AACA,QAAI;AACF,cAAQ,MAAM,OAAO,SAAS;AAAA,QAC5B,QAAQ,OAAO;AAAA,QACf;AAAA,QACA;AAAA,QACA,aAAa,WAAW;AAAA,MAC1B,CAAC;AACD,YAAM,gBAAgB,SAAS,OAAO,WAAW,eAAe,CAAC;AAAA,IACnE,SAAS,KAAK;AACZ,YAAM,QAAQ,aAAa,YAAY,GAAG,GAAG;AAC7C,oBAAc,KAAK,KAAK;AACxB,aAAO,OAAO,SAAS;AAAA,QACrB,QAAQ,OAAO;AAAA,QACf,MAAM;AAAA,QACN,WAAW;AAAA,QACX,QAAQ,MAAM;AAAA,QACd,OAAO;AAAA,QACP,YAAY;AAAA,QACZ,YAAY,CAAC;AAAA,QACb,QAAQ,KAAK,IAAI,IAAI;AAAA,QACrB;AAAA,QACA,OAAO,SAAS,SAAS;AAAA,QACzB,cAAc;AAAA,QACd;AAAA,QACA,WAAW;AAAA,MACb,CAAC;AAAA,IACH;AAGA,qBAAiB,QAAQ;AAAA,MACvB,kBAAkB,iBAAiB,OAAO,OAAO,YAAY;AAAA,MAC7D,OAAO,aAAa,KAAK;AAAA,IAC3B,CAAC;AAED,aAAS,YAAY,GAAG,YAAY,OAAO,UAAU,aAAa;AAChE,UAAI,WAAW,OAAO,SAAS;AAC7B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,YAAY,WAAW,MAAM;AAAA,UACrC,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,aAAa,mBAAmB,QAAQ,YAAY;AAC1D,UAAI,WAAW,MAAM;AACnB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,WAAW;AAAA,UACnB,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,MAAM;AAAA,QACV,OAAO;AAAA,QACP;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,WAAW;AAAA,QACX;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,eAAO,OAAO,aAAa,MAAM,OAAO,WAAW,GAAG,IAAI,oBAAoB,KAAK;AAAA,MACrF,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,eAAe,WAAW,GAAG,CAAC;AAC9D,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI,KAAK,MAAM;AACb,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,QAAQ,KAAK;AAAA,UACb,OAAO,KAAK;AAAA,UACZ,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc,KAAK;AAAA,UACnB;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,UAAI;AACJ,UAAI;AACF,mBAAW,MAAM,OAAO,OAAO,GAAG;AAAA,MACpC,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,UAAU,WAAW,GAAG,CAAC;AACzD,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI,SAAS,SAAS,QAAQ;AAC5B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM,SAAS,QAAQ;AAAA,UACvB,WAAW;AAAA,UACX,QAAQ,SAAS;AAAA,UACjB,OAAO,SAAS;AAAA,UAChB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc,SAAS,SAAS,QAAQ,YAAY;AAAA,UACpD;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,oBAAoB,kBAAkB,SAAS,QAAQ,OAAO,YAAY;AAChF,YAAM,uBAAuB,iBAAiB,QAAQ,EAAE,kBAAkB,CAAC;AAC3E,YAAM,uBAAuB,iBAAiB;AAC9C,UAAI,sBAAsB;AACxB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,qBAAqB;AAAA,UAC7B,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,cAAc;AACpB,YAAM,cAAc;AACpB,YAAM,cAAc,aAAa,KAAK;AACtC,YAAM,gBAAgB,KAAK,IAAI;AAC/B,YAAM,aAAa,UACf,MAAM;AAAA,QAAS;AAAA,QAAe;AAAA,QAAW,MACvC,QAAQ,KAAK;AAAA,UACX,MAAM,gBAAgB,SAAS;AAAA,UAC/B,UAAU;AAAA,UACV,MAAM,SAAS;AAAA,UACf,YAAY;AAAA,YACV,UAAU,SAAS,UAAU;AAAA,YAC7B;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,IACA;AACJ,UAAI;AACJ,UAAI;AACF,cAAM,SAAS,MAAM,OAAO,IAAI,SAAS,QAAQ,GAAG;AACpD,cAAM,aAAa,OAAO,mBAAmB;AAAA,UAC3C,QAAQ,SAAS;AAAA,UACjB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,QACF,CAAC;AACD,cAAM,UAAU,uBAAuB,YAAY,eAAe,SAAS;AAC3E,YAAI,YAAY,UAAa,OAAO,SAAS,OAAO,KAAK,UAAU,GAAG;AACpE,0BAAgB;AAChB,gBAAM;AAAA,YACJ;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,YACA;AAAA,UACF;AAAA,QACF;AACA,wBAAgB;AAAA,UACd,IAAI;AAAA,UACJ;AAAA,UACA,GAAI,YAAY,SAAY,EAAE,QAAQ,IAAI,CAAC;AAAA,UAC3C,YAAY,KAAK,IAAI,IAAI;AAAA,QAC3B;AAAA,MACF,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,OAAO,WAAW,GAAG,CAAC;AACtD,wBAAgB;AAAA,UACd,IAAI;AAAA,UACJ,OAAO,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UAChD,YAAY,KAAK,IAAI,IAAI;AAAA,QAC3B;AACA,YAAI,kBAAkB,QAAQ;AAC5B,gBAAM;AAAA,YAAS;AAAA,YAAe;AAAA,YAAW,MACvC,YAAY,KAAK,cAAc,SAAS,eAAe;AAAA,UACzD;AACA,gBAAMA,QAA2D;AAAA,YAC/D,OAAO;AAAA,YACP;AAAA,YACA;AAAA,YACA,YAAY;AAAA,YACZ;AAAA,YACA,YAAY;AAAA,YACZ;AAAA,YACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,YAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,UAClC;AACA,kBAAQ,KAAKA,KAAI;AACjB,gBAAM,UAAU,OAAO,QAAQA,OAAM,aAAa;AAClD,iBAAO,OAAO,SAAS;AAAA,YACrB,QAAQ,OAAO;AAAA,YACf,MAAM;AAAA,YACN,WAAW;AAAA,YACX,QAAQ,cAAc,SAAS;AAAA,YAC/B,OAAO,aAAa,KAAK;AAAA,YACzB,OAAO;AAAA,YACP,YAAY;AAAA,YACZ,YAAY;AAAA,YACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,YACrB;AAAA,YACA,OAAO,SAAS,SAAS;AAAA,YACzB,cAAc;AAAA,YACd;AAAA,YACA,WAAW;AAAA,UACb,CAAC;AAAA,QACH;AAAA,MACF;AAEA,UAAI;AACF,gBAAQ,MAAM,OAAO,QAAQ,EAAE,SAAS,aAAa,WAAW,OAAO,CAAC;AAAA,MAC1E,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,WAAW,WAAW,GAAG,CAAC;AAC1D,cAAMA,QAA2D;AAAA,UAC/D,OAAO;AAAA,UACP;AAAA,UACA;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,UAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,QAClC;AACA,gBAAQ,KAAKA,KAAI;AACjB,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,KAAK,cAAc,cAAc,SAAS,CAAC,EAAG,OAAO;AAAA,QACnE;AACA,cAAM,UAAU,OAAO,QAAQA,OAAM,aAAa;AAClD,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI;AACF,gBAAQ,MAAM,OAAO,SAAS;AAAA,UAC5B,QAAQ,OAAO;AAAA,UACf;AAAA,UACA;AAAA,UACA,aAAa,WAAW;AAAA,QAC1B,CAAC;AACD,cAAM;AAAA,UACJ;AAAA,UACA;AAAA,UACA,QAAQ,SAAS;AAAA,UACjB;AAAA,UACA;AAAA,UACA,YAAY,KAAK;AAAA,QACnB;AAAA,MACF,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,YAAY,WAAW,GAAG,CAAC;AAC3D,cAAMA,QAA2D;AAAA,UAC/D,OAAO;AAAA,UACP;AAAA,UACA;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,YAAY;AAAA,UACZ;AAAA,UACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,UAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,QAClC;AACA,gBAAQ,KAAKA,KAAI;AACjB,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,KAAK,cAAc,cAAc,SAAS,CAAC,EAAG,OAAO;AAAA,QACnE;AACA,cAAM,UAAU,OAAO,QAAQA,OAAM,aAAa;AAClD,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,YAAM,aAAa,aAAa,KAAK;AACrC,YAAM,mBAAmB,iBAAiB,OAAO,OAAO,YAAY;AACpE,YAAM,mBAAmB,iBAAiB,QAAQ,EAAE,kBAAkB,OAAO,WAAW,CAAC;AACzF,YAAM,mBAAmB,iBAAiB;AAE1C,YAAM,OAA2D;AAAA,QAC/D,OAAO;AAAA,QACP;AAAA,QACA;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,QACA,WAAW,IAAI,KAAK,aAAa,EAAE,YAAY;AAAA,QAC/C,UAAS,oBAAI,KAAK,GAAE,YAAY;AAAA,MAClC;AACA,cAAQ,KAAK,IAAI;AACjB,UAAI,cAAc,IAAI;AACpB,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,IAAI;AAAA,YACd,YAAY;AAAA,cACV,eAAe,cAAc,WAAW;AAAA,cACxC;AAAA,cACA,aAAa,eAAe;AAAA,cAC5B,YAAY,cAAc;AAAA,cAC1B;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF,OAAO;AACL,cAAM;AAAA,UAAS;AAAA,UAAe;AAAA,UAAW,MACvC,YAAY,KAAK,cAAc,SAAS,iBAAiB;AAAA,YACvD,YAAY;AAAA,cACV;AAAA,cACA;AAAA,YACF;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,YAAM,UAAU,OAAO,QAAQ,MAAM,aAAa;AAElD,UAAI,kBAAkB;AACpB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,iBAAiB;AAAA,UACzB,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,qBAAqB,mBAAmB,QAAQ,YAAY;AAClE,UAAI,mBAAmB,MAAM;AAC3B,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,mBAAmB;AAAA,UAC3B,OAAO;AAAA,UACP,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAEA,YAAM,cAAc;AAAA,QAClB,OAAO;AAAA,QACP;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,QACA;AAAA,QACA,WAAW;AAAA,QACX;AAAA,MACF;AACA,UAAI;AACJ,UAAI;AACF,uBAAe,OAAO,aAClB,MAAM,OAAO,WAAW,WAAW,IACnC,oBAAoB,KAAK;AAAA,MAC/B,SAAS,KAAK;AACZ,sBAAc,KAAK,aAAa,eAAe,YAAY,GAAG,GAAG,CAAC;AAClE,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM;AAAA,UACN,WAAW;AAAA,UACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,UACjD,OAAO,aAAa,KAAK;AAAA,UACzB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc;AAAA,UACd;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AACA,UAAI,aAAa,MAAM;AACrB,eAAO,OAAO,SAAS;AAAA,UACrB,QAAQ,OAAO;AAAA,UACf,MAAM,aAAa;AAAA,UACnB,WAAW;AAAA,UACX,QAAQ,aAAa;AAAA,UACrB,OAAO,aAAa;AAAA,UACpB,OAAO;AAAA,UACP,YAAY;AAAA,UACZ,YAAY;AAAA,UACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,UACrB;AAAA,UACA,OAAO,SAAS,SAAS;AAAA,UACzB,cAAc,aAAa;AAAA,UAC3B;AAAA,UACA,WAAW;AAAA,QACb,CAAC;AAAA,MACH;AAAA,IACF;AAEA,WAAO,OAAO,SAAS;AAAA,MACrB,QAAQ,OAAO;AAAA,MACf,MAAM;AAAA,MACN,WAAW;AAAA,MACX,QAAQ,8BAA8B,OAAO,QAAQ;AAAA,MACrD,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,QAAQ,KAAK,IAAI,IAAI;AAAA,MACrB;AAAA,MACA,OAAO,SAAS,SAAS;AAAA,MACzB,cAAc;AAAA,MACd;AAAA,MACA,WAAW;AAAA,IACb,CAAC;AAAA,EACH,SAAS,KAAK;AACZ,kBAAc,KAAK,aAAa,OAAO,QAAQ,QAAQ,GAAG,CAAC;AAC3D,WAAO,OAAO,SAAS;AAAA,MACrB,QAAQ,OAAO;AAAA,MACf,MAAM;AAAA,MACN,WAAW;AAAA,MACX,QAAQ,cAAc,cAAc,SAAS,CAAC,EAAG;AAAA,MACjD,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY,CAAC;AAAA,MACb,QAAQ,KAAK,IAAI,IAAI;AAAA,MACrB;AAAA,MACA,OAAO,SAAS,SAAS;AAAA,MACzB,cAAc;AAAA,MACd;AAAA,MACA,WAAW;AAAA,IACb,CAAC;AAAA,EACH,UAAE;AACA,QAAI,UAAW,cAAa,SAAS;AACrC,QAAI,OAAO,OAAQ,QAAO,OAAO,oBAAoB,SAAS,aAAa;AAAA,EAC7E;AACF;AAEO,SAAS,iBACd,oBACA,UAA4E,CAAC,GACvC;AACtC,SAAO,EAAE,GAAG,SAAS,mBAAmB;AAC1C;AAEO,SAAS,qBACd,oBACA,UAA4E,CAAC,GACvC;AACtC,SAAO,EAAE,GAAG,SAAS,mBAAmB;AAC1C;AAEO,SAAS,cAAc,OAAgE;AAC5F,SAAO,EAAE,GAAG,OAAO,WAAW,KAAK;AACrC;AAEO,SAAS,eAAe,OAAgE;AAC7F,SAAO,EAAE,GAAG,OAAO,WAAW,MAAM;AACtC;AAEA,SAAS,gBAAgB,OAA0D;AACjF,QAAM,MAAM,EAAE,GAAG,gBAAgB,GAAG,MAAM;AAC1C,MAAI,CAAC,OAAO,UAAU,IAAI,QAAQ,KAAM,IAAI,WAAsB,GAAG;AACnE,UAAM,IAAI;AAAA,MACR,+DAA+D,OAAO,IAAI,QAAQ,CAAC;AAAA,IACrF;AAAA,EACF;AACA,QAAM,SAAwB,EAAE,UAAU,IAAI,SAAmB;AACjE,MAAI,IAAI,cAAc,QAAW;AAC/B,QACE,OAAO,IAAI,cAAc,YACzB,CAAC,OAAO,SAAS,IAAI,SAAS,KAC9B,IAAI,aAAa,GACjB;AACA,YAAM,IAAI;AAAA,QACR,yEAAyE,OAAO,IAAI,SAAS,CAAC;AAAA,MAChG;AAAA,IACF;AACA,WAAO,YAAY,IAAI;AAAA,EACzB;AACA,MAAI,IAAI,eAAe,QAAW;AAChC,QACE,OAAO,IAAI,eAAe,YAC1B,CAAC,OAAO,SAAS,IAAI,UAAU,KAC/B,IAAI,aAAa,GACjB;AACA,YAAM,IAAI;AAAA,QACR,6EAA6E,OAAO,IAAI,UAAU,CAAC;AAAA,MACrG;AAAA,IACF;AACA,WAAO,aAAa,IAAI;AAAA,EAC1B;AACA,SAAO;AACT;AAEA,SAAS,uBACP,SACA,eACA,WACoB;AACpB,MAAI,YAAY,OAAW,QAAO;AAClC,MAAI,CAAC,OAAO,SAAS,OAAO,KAAK,UAAU,GAAG;AAC5C,kBAAc;AAAA,MACZ,aAAa,OAAO,WAAW,IAAI,MAAM,2BAA2B,OAAO,OAAO,CAAC,EAAE,CAAC;AAAA,IACxF;AACA,WAAO;AAAA,EACT;AACA,SAAO;AACT;AAEO,SAAS,kBAAkB,OAAqC;AACrE,SAAO,MAAM;AAAA,IACX,CAAC,WAAW,OAAO,UAAW,OAAO,aAAa,cAAc,OAAO,aAAa;AAAA,EACtF;AACF;AAEA,SAAS,YACP,QACA,OACA,OACA,SACA,QACA,WACA,SACA,cACA,aACA,SACuD;AACvD,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,QAAQ,KAAK,IAAI,IAAI;AAAA,IACrB;AAAA,IACA,kBACE,OAAO,eAAe,SAAY,SAAY,KAAK,IAAI,GAAG,OAAO,aAAa,YAAY;AAAA,IAC5F;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,oBAAoB,OAA0C;AACrE,MAAI,CAAC,MAAM,OAAQ,QAAO,EAAE,MAAM,OAAO,MAAM,OAAO,QAAQ,eAAe;AAC7E,QAAM,OAAO,kBAAkB,KAAK;AACpC,SAAO,OACH,EAAE,MAAM,MAAM,MAAM,MAAM,QAAQ,6BAA6B,OAAO,aAAa,KAAK,EAAE,IAC1F;AAAA,IACE,MAAM;AAAA,IACN,MAAM;AAAA,IACN,QAAQ;AAAA,IACR,OAAO,aAAa,KAAK;AAAA,EAC3B;AACN;AAEA,SAAS,aAAa,OAAgD;AACpE,QAAM,SAAS,MACZ,IAAI,CAAC,WAAW,OAAO,KAAK,EAC5B,OAAO,CAAC,UAA2B,OAAO,UAAU,QAAQ;AAC/D,MAAI,CAAC,OAAO,OAAQ,QAAO;AAC3B,SAAO,KAAK,MAAO,OAAO,OAAO,CAAC,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO,SAAU,GAAI,IAAI;AAC9F;AAEA,SAAS,mBACP,QACA,cACmC;AACnC,MAAI,OAAO,eAAe,UAAa,gBAAgB,OAAO,YAAY;AACxE,WAAO;AAAA,MACL,MAAM;AAAA,MACN,QAAQ,gCAAgC,OAAO,UAAU;AAAA,IAC3D;AAAA,EACF;AACA,SAAO,EAAE,MAAM,OAAO,QAAQ,GAAG;AACnC;AAEA,eAAe,iBACb,SACA,QACA,cACA,QACA,eACA,WACe;AACf,MAAI,CAAC,WAAW,OAAO,eAAe,OAAW;AACjD,QAAM,aAAa,OAAO;AAC1B,QAAM;AAAA,IAAS;AAAA,IAAe;AAAA,IAAW,MACvC,QAAQ,aAAa;AAAA,MACnB,WAAW;AAAA,MACX,OAAO;AAAA,MACP,UAAU;AAAA,MACV,WAAW,KAAK,IAAI,GAAG,aAAa,YAAY;AAAA,MAChD,UAAU,gBAAgB;AAAA,MAC1B,QAAQ,QAAQ,KAAK;AAAA,IACvB,CAAC;AAAA,EACH;AACF;AAEA,eAAe,gBACb,SACA,OACA,OACA,eACA,WACA,cACe;AACf,MAAI,CAAC,QAAS;AACd,aAAW,UAAU,OAAO;AAC1B,UAAM;AAAA,MAAS;AAAA,MAAe;AAAA,MAAW,MACvC,QAAQ,YAAY;AAAA,QAClB,SAAS,OAAO,YAAY,wBAAwB;AAAA,QACpD,cAAc,gBAAgB,QAAQ;AAAA,QACtC,MAAM,gBAAgB,OAAO,EAAE;AAAA,QAC/B,WAAW,OAAO;AAAA,QAClB,OAAO,OAAO,OAAO,UAAU,WAAW,OAAO,QAAQ,OAAO,SAAS,IAAI;AAAA,QAC7E,WAAW,OAAO;AAAA,QAClB,UAAU,OAAO;AAAA,QACjB,YAAY;AAAA,UACV;AAAA,UACA,QAAQ,OAAO;AAAA,UACf,UAAU,OAAO;AAAA,UACjB,WAAW,OAAO;AAAA,QACpB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAEA,eAAe,UACb,QACA,MACA,eACe;AACf,MAAI,CAAC,OAAQ;AACb,MAAI;AACF,UAAM,OAAO,IAAI;AAAA,EACnB,SAAS,KAAK;AACZ,kBAAc,KAAK,aAAa,WAAW,KAAK,OAAO,GAAG,CAAC;AAAA,EAC7D;AACF;AAEA,eAAe,SACb,eACA,WACA,OACwB;AACxB,MAAI;AACF,WAAO,MAAM,MAAM;AAAA,EACrB,SAAS,KAAK;AACZ,kBAAc,KAAK,aAAa,SAAS,WAAW,GAAG,CAAC;AACxD,WAAO;AAAA,EACT;AACF;AAEA,SAAS,iBACP,OACA,UACQ;AACR,MAAI,UAAU,iBAAkB,QAAO,SAAS,iBAAiB,KAAK;AACtE,SAAO,kBAAkB,KAAK;AAChC;AAEA,SAAS,kBACP,QACA,UACQ;AACR,MAAI,UAAU,kBAAmB,QAAO,SAAS,kBAAkB,MAAM;AACzE,SAAO,kBAAkB,MAAM;AACjC;AAEA,SAAS,kBAAkB,OAAwB;AACjD,MAAI,OAAO,UAAU,SAAU,QAAO;AACtC,MAAI,OAAO,UAAU,YAAY,OAAO,UAAU,aAAa,SAAS,KAAM,QAAO,OAAO,KAAK;AACjG,MAAI;AACF,WAAO,KAAK,UAAU,mBAAmB,KAAK,CAAC;AAAA,EACjD,QAAQ;AACN,WAAO,OAAO,KAAK;AAAA,EACrB;AACF;AAEA,SAAS,mBAAmB,OAAyB;AACnD,MAAI,MAAM,QAAQ,KAAK,EAAG,QAAO,MAAM,IAAI,kBAAkB;AAC7D,MAAI,CAAC,SAAS,OAAO,UAAU,SAAU,QAAO;AAChD,QAAM,SAAS;AACf,QAAM,SAAkC,CAAC;AACzC,aAAW,OAAO,OAAO,KAAK,MAAM,EAAE,KAAK,GAAG;AAC5C,WAAO,GAAG,IAAI,mBAAmB,OAAO,GAAG,CAAC;AAAA,EAC9C;AACA,SAAO;AACT;AAEA,SAAS,YAAY,QAA6B;AAChD,QAAM,SAAS,OAAO;AACtB,MAAI,kBAAkB,MAAO,QAAO,OAAO;AAC3C,SAAO,SAAS,OAAO,MAAM,IAAI;AACnC;AAEA,SAAS,aACP,OACA,WACA,KACqB;AACrB,QAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,SAAO,EAAE,OAAO,WAAW,QAAQ;AACrC;AAEA,eAAe,OACb,SACA,QACkE;AAClE,QAAM;AAAA,IAAS,OAAO;AAAA,IAAe,OAAO,MAAM;AAAA,IAAQ,MACxD,SAAS,OAAO;AAAA,MACd,MAAM,OAAO;AAAA,MACb,OAAO,OAAO,SAAS,aAAa,OAAO,UAAU;AAAA,MACrD,cAAc,OAAO;AAAA,MACrB,OAAO,OAAO;AAAA,IAChB,CAAC;AAAA,EACH;AACA,SAAO;AACT;","names":["step"]}

package/dist/control.js

@@ -4,7 +4,7 @@ import {
   runProposeReview,
   runProposeReviewAsControlLoop,
   scoreFromEvals
-} from "./chunk-RTWFUK6A.js";
+} from "./chunk-E4GH6USR.js";
 import {
   allCriticalPassed,
   objectiveEval,
@@ -12,7 +12,7 @@ import {
   stopOnNoProgress,
   stopOnRepeatedAction,
   subjectiveEval
-} from "./chunk-NCRFYPS3.js";
+} from "./chunk-YEHAEDUD.js";
 import "./chunk-KWRRMR3J.js";
 import "./chunk-TVVP3ZZQ.js";
 import "./chunk-VSMTAMNK.js";