@tangle-network/agent-eval 0.29.1 → 0.31.0

package/dist/traces.js

@@ -17,13 +17,23 @@ import {
   TraceNotFoundError,
   analyzeTraces,
   buildTraceAnalystTools,
+  buildTraceInsightContext,
+  buildTraceInsightPrompt,
   createReplayFetch,
+  defaultTraceInsightPanel,
+  describeTraceInsightScope,
+  domainEvidencePattern,
   exportRunAsOtlp,
+  inferDomainKeywords,
   iterateRawCalls,
+  planTraceInsightQuestions,
   redactString,
   redactValue,
-  traceAnalystFunctionGroup
-} from "./chunk-UW4NOOZI.js";
+  scoreTraceInsightReadiness,
+  tokenizeDomainWords,
+  traceAnalystFunctionGroup,
+  traceAnalystOnRunComplete
+} from "./chunk-HIO4UIS5.js";
 import {
   aggregateLlm,
   argHash,
@@ -47,7 +57,7 @@ import {
   RunIntegrityError,
   assertRunCaptured,
   throwIfRunIncomplete
-} from "./chunk-KTGTIOFD.js";
+} from "./chunk-UBPIXOC4.js";
 import {
   FileSystemRawProviderSink,
   InMemoryRawProviderSink,
@@ -60,304 +70,8 @@ import {
   llmSpanFromProvider
 } from "./chunk-TVVP3ZZQ.js";
 import "./chunk-VSMTAMNK.js";
-import "./chunk-NG236HPC.js";
+import "./chunk-QYJT52YW.js";
 import "./chunk-PZ5AY32C.js";
-
-// src/trace-analyst/hook.ts
-var DEFAULT_QUESTION = "Summarise what happened in this run. Surface any failure modes, surprising findings, or evidence that the run's verdict is wrong.";
-function traceAnalystOnRunComplete(opts) {
-  return async (ctx) => {
-    if (opts.shouldRun && !opts.shouldRun(ctx)) return;
-    const source = opts.analyze.source;
-    if (source === void 0) {
-      await ctx.store.appendEvent({
-        eventId: `analyst-skip-${ctx.runId}`,
-        runId: ctx.runId,
-        kind: "log",
-        timestamp: Date.now(),
-        payload: { source: "trace_analyst_hook", reason: "no source configured" }
-      });
-      return;
-    }
-    const result = await analyzeTraces({ question: opts.question ?? DEFAULT_QUESTION }, {
-      ...opts.analyze,
-      source
-    });
-    if (opts.save) await opts.save(result, ctx);
-    if (opts.gateOn && !opts.gateOn(result, ctx)) {
-      await ctx.store.appendEvent({
-        eventId: `analyst-gate-${ctx.runId}`,
-        runId: ctx.runId,
-        kind: "log",
-        timestamp: Date.now(),
-        payload: {
-          source: "trace_analyst_hook",
-          reason: "analyst_gate_failed",
-          findings: result.findings
-        }
-      });
-    }
-  };
-}
-
-// src/trace-analyst/insights.ts
-var DOMAIN_STOP_WORDS = /* @__PURE__ */ new Set([
-  "and",
-  "advanced",
-  "app",
-  "build",
-  "create",
-  "easy",
-  "expert",
-  "extreme",
-  "for",
-  "from",
-  "hard",
-  "implementation",
-  "integrate",
-  "medium",
-  "project",
-  "task",
-  "the",
-  "this",
-  "with",
-  "workflow"
-]);
-function tokenizeDomainWords(value) {
-  return [...value.matchAll(/[A-Za-z][A-Za-z0-9.+#-]{2,}/g)].map((match) => match[0].toLowerCase()).filter((word) => !DOMAIN_STOP_WORDS.has(word));
-}
-function inferDomainKeywords(suite) {
-  const suiteWords = new Set(tokenizeDomainWords(`${suite.name} ${suite.collectionId ?? ""}`));
-  const source = [
-    suite.name,
-    suite.collectionId ?? "",
-    ...suite.tasks.flatMap((task) => [
-      task.id,
-      task.name,
-      task.prompt ?? "",
-      task.difficulty ?? "",
-      ...task.tags ?? [],
-      ...task.gaps ?? []
-    ])
-  ].join(" ");
-  const counts = /* @__PURE__ */ new Map();
-  for (const word of tokenizeDomainWords(source)) counts.set(word, (counts.get(word) ?? 0) + 1);
-  return [...counts.entries()].filter(([word, count]) => count >= 2 || suiteWords.has(word)).sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0])).map(([word]) => word).slice(0, 18);
-}
-function domainEvidencePattern(keywords) {
-  const escaped = keywords.filter((keyword) => keyword.length >= 3).map((keyword) => keyword.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"));
-  return escaped.length > 0 ? new RegExp(`(?<![A-Za-z0-9])(?:${escaped.join("|")})(?![A-Za-z0-9])`, "i") : /(?<![A-Za-z0-9])(?:sdk|api|css|dns|xml|provider|client|service|integration|webhook|transaction|auth|oauth|graphql|rest)(?![A-Za-z0-9])/i;
-}
-function describeTraceInsightScope(suite) {
-  const taskLabel = suite.tasks.length === 1 ? "1 implementation task" : `${suite.tasks.length} implementation tasks`;
-  const tags = /* @__PURE__ */ new Map();
-  for (const task of suite.tasks) {
-    for (const tag of task.tags ?? []) tags.set(tag, (tags.get(tag) ?? 0) + 1);
-  }
-  const topTags = [...tags.entries()].sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0])).slice(0, 8).map(([tag]) => tag);
-  if (topTags.length > 0) return `${taskLabel} across ${topTags.join(", ")}.`;
-  const difficulties = [
-    ...new Set(
-      suite.tasks.map((task) => task.difficulty).filter((value) => Boolean(value))
-    )
-  ].join(", ");
-  return `${taskLabel} across ${difficulties || "the selected benchmark scope"}.`;
-}
-function planTraceInsightQuestions(input) {
-  const hasFailures = input.suite.tasks.some((task) => task.outcome && task.outcome !== "satisfied");
-  const hasMultipleShots = input.suite.tasks.some(
-    (task) => (task.gaps ?? []).some((gap) => /shot|review|retry|continue/i.test(gap))
-  );
-  const questions = [
-    {
-      id: "execution-path",
-      question: "What did the worker actually do before the first meaningful implementation edit?",
-      why: "Separates grounded execution from polished but shallow output."
-    },
-    {
-      id: "research-grounding",
-      question: "Did the worker inspect docs, source, examples, or package references before committing to an implementation path?",
-      why: "Identifies whether failures came from weak retrieval, weak examples, or premature coding."
-    },
-    {
-      id: "domain-proof",
-      question: "Which tasks produced executable domain proof versus UI copy, placeholders, or inferred behavior?",
-      why: "Keeps product-quality claims tied to concrete evidence."
-    },
-    {
-      id: "root-cause",
-      question: "For each major failure cluster, is the likely root cause prompt/scaffold, docs/examples, SDK/API ergonomics, evaluator, runtime, or model behavior?",
-      why: "Turns trace observations into actionable ownership."
-    },
-    {
-      id: "evidence-quality",
-      question: "Which external-facing claims are directly supported by trace ids, span ids, verifier findings, reviewer notes, or generated code?",
-      why: "Prevents unsupported customer-report conclusions."
-    }
-  ];
-  if (hasMultipleShots) {
-    questions.push({
-      id: "reviewer-lift",
-      question: "Where did reviewer feedback improve score, stall, or regress across shots?",
-      why: "Shows whether the driver loop is learning or merely repeating work."
-    });
-  }
-  if (hasFailures) {
-    questions.push({
-      id: "optimization-targets",
-      question: "Which prompt, evaluator, scaffold, or workflow changes should feed the next GEPA/autoresearch optimization run?",
-      why: "Connects benchmark evidence to the optimization loop."
-    });
-  }
-  return questions;
-}
-function buildTraceInsightContext(input) {
-  return {
-    suite: input.suite,
-    scope: describeTraceInsightScope(input.suite),
-    keywords: inferDomainKeywords(input.suite),
-    questions: planTraceInsightQuestions(input),
-    panel: defaultTraceInsightPanel(),
-    findings: input.findings ?? [],
-    agent: input.agent ?? null,
-    totals: input.totals ?? null
-  };
-}
-function scoreTraceInsightReadiness(context) {
-  const failedTasks = context.suite.tasks.filter(
-    (task) => task.outcome && task.outcome !== "satisfied"
-  );
-  const findingTaskIds = new Set(context.findings.flatMap((finding) => finding.taskIds));
-  const failedTasksWithFindings = failedTasks.filter((task) => findingTaskIds.has(task.id));
-  const tasksWithGaps = context.suite.tasks.filter((task) => (task.gaps ?? []).length > 0);
-  const gates = [
-    {
-      id: "domain-context",
-      label: "Domain context inferred",
-      passed: context.keywords.length > 0,
-      severity: "high",
-      detail: context.keywords.length > 0 ? `${context.keywords.length} domain terms inferred: ${context.keywords.slice(0, 8).join(", ")}` : "No domain terms were inferred from suite, tasks, prompts, tags, or gaps."
-    },
-    {
-      id: "panel-coverage",
-      label: "Analyst panel planned",
-      passed: context.panel.length >= 4 && context.questions.length >= 5,
-      severity: "high",
-      detail: `${context.panel.length} panel roles and ${context.questions.length} investigation questions planned.`
-    },
-    {
-      id: "failure-coverage",
-      label: "Failures mapped to findings",
-      passed: failedTasks.length === 0 || failedTasksWithFindings.length / failedTasks.length >= 0.5,
-      severity: "critical",
-      detail: failedTasks.length === 0 ? "No failed tasks in suite." : `${failedTasksWithFindings.length}/${failedTasks.length} failed tasks appear in finding clusters.`
-    },
-    {
-      id: "gap-evidence",
-      label: "Task gaps captured",
-      passed: failedTasks.length === 0 || tasksWithGaps.length / failedTasks.length >= 0.5,
-      severity: "medium",
-      detail: `${tasksWithGaps.length} tasks include explicit evaluator or analyst gaps.`
-    }
-  ];
-  const penalty = gates.reduce((sum, gate) => {
-    if (gate.passed) return sum;
-    if (gate.severity === "critical") return sum + 35;
-    if (gate.severity === "high") return sum + 20;
-    if (gate.severity === "medium") return sum + 10;
-    return sum + 5;
-  }, 0);
-  const score = Math.max(0, Math.min(1, 1 - penalty / 100));
-  return {
-    score,
-    grade: score >= 0.9 ? "external-ready" : score >= 0.7 ? "internal-review" : "raw-analysis",
-    gates
-  };
-}
-function defaultTraceInsightPanel() {
-  return [
-    {
-      id: "trace-forensics",
-      name: "Trace Forensics",
-      responsibility: "Reconstruct what the worker did in order, including research, edits, reviewer interventions, verifier feedback, and stop reason."
-    },
-    {
-      id: "root-cause",
-      name: "Root Cause",
-      responsibility: "Map failures to prompt/scaffold, docs/examples, SDK/API/product ergonomics, evaluator, runtime, or model behavior."
-    },
-    {
-      id: "optimization",
-      name: "Optimization",
-      responsibility: "Identify prompt, reviewer, evaluator, scaffold, and GEPA/autoresearch changes that should be tested next."
-    },
-    {
-      id: "external-evidence",
-      name: "External Evidence",
-      responsibility: "Separate customer-safe claims from internal harness findings and reject conclusions without task, trace, span, code, reviewer, or verifier evidence."
-    }
-  ];
-}
-function buildTraceInsightPrompt(input) {
-  const context = buildTraceInsightContext(input);
-  const maxRepresentativeTraces = input.maxRepresentativeTraces ?? 6;
-  return `Analyze this benchmark run and produce evidence-backed trace intelligence.
-
-Audience:
-- internal AI/product leadership
-- possible customer-facing report for ${input.suite.name}
-
-Investigation plan:
-${context.questions.map((item, index) => `${index + 1}. ${item.question} (${item.why})`).join("\n")}
-
-Analyst panel:
-${context.panel.map((role) => `- ${role.name}: ${role.responsibility}`).join("\n")}
-
-If the task branches are independent, use subagents for the panel roles above and aggregate their findings. Do not run a panel role unless its answer will change the final report.
-
-Required output:
-1. Executive verdict: what this run proves and does not prove.
-2. The investigation questions you answered and the evidence used.
-3. Failure taxonomy: agent prompting, evaluator/harness, docs/examples, SDK/API/product integration, infra.
-4. Evidence-backed examples with trace ids/task ids and concrete verifier findings.
-5. Highest-ROI fixes for the benchmark harness, prompt/GEPA optimization, and customer-facing product/docs surface.
-6. What is safe for an external report versus what must stay internal.
-7. One rerun plan that would validate lift after optimization.
-
-Budget:
-- Inspect the dataset overview, the failure summary, and at most ${maxRepresentativeTraces} representative traces.
-- Prefer traces named in the failure summary over broad exploration.
-- Do not do exhaustive trace sweeps.
-- Return the final report as soon as the taxonomy and examples are supported.
-
-Run summary:
-${JSON.stringify(
-    {
-      suite: input.suite.name,
-      scope: context.scope,
-      inferredKeywords: context.keywords,
-      agent: context.agent,
-      totals: context.totals,
-      findings: context.findings.map((finding) => ({
-        kind: finding.kind,
-        severity: finding.severity,
-        taskCount: finding.taskIds.length,
-        proposedFixClass: finding.proposedFixClass
-      })),
-      failures: input.suite.tasks.filter((task) => task.outcome && task.outcome !== "satisfied").map((task) => ({
-        task: task.id,
-        difficulty: task.difficulty,
-        outcome: task.outcome,
-        score: task.score,
-        gaps: task.gaps ?? []
-      }))
-    },
-    null,
-    2
-  )}
-
-Use the trace tools. Do not invent facts. Cite task ids. Separate customer-facing claims from internal harness/model findings.`;
-}
 export {
   DEFAULT_REDACTION_RULES,
   DEFAULT_TRACE_ANALYST_BUDGETS,

package/dist/traces.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/trace-analyst/hook.ts","../src/trace-analyst/insights.ts"],"sourcesContent":["/**\n * Trace-analyst auto-execution hook.\n *\n * Wires `analyzeTraces` into a `TraceEmitter`'s `onRunComplete` so a\n * direct matrix run produces an analysis artifact without an out-of-band\n * step. Designed for the case where the consumer reports \"the analyst\n * never ran\" — the cause is almost always orchestration, not the analyst.\n *\n * Usage:\n *\n *   const emitter = new TraceEmitter(store, {\n *     onRunComplete: [traceAnalystOnRunComplete({ analyze: opts, save })],\n *   })\n *\n * Hooks are best-effort by default — they never crash the underlying run.\n * The caller decides whether to gate the run on the analysis result via\n * the `gateOn` callback.\n */\n\nimport type { RunCompleteHook, RunCompleteHookContext } from '../trace/emitter'\nimport { type AnalyzeTracesOptions, type AnalyzeTracesResult, analyzeTraces } from './analyst'\n\nexport interface TraceAnalystHookOptions {\n  /**\n   * Options forwarded to `analyzeTraces`. The hook supplies the question\n   * if you don't pass one — defaulting to a launch-grade prompt that asks\n   * for failure modes, surprising findings, and a recommendation.\n   */\n  analyze: Omit<AnalyzeTracesOptions, 'source'> & { source?: AnalyzeTracesOptions['source'] }\n  /**\n   * Override the question. The default is intentionally generic:\n   * \"Summarise what happened in this run, surface any failure modes,\n   *  surprising findings, or evidence the verdict is wrong.\"\n   */\n  question?: string\n  /**\n   * Persist the result. The hook calls this with the analysis output and\n   * the run context. Common implementations write to a TraceAnalysisStore\n   * or append to a per-run JSONL.\n   */\n  save?: (result: AnalyzeTracesResult, ctx: RunCompleteHookContext) => Promise<void>\n  /**\n   * Predicate gating execution per run. Default: every completed run.\n   * Use to skip aborted runs, debug runs, or runs without LLM activity.\n   */\n  shouldRun?: (ctx: RunCompleteHookContext) => boolean\n  /**\n   * Optional gate: if set and returns false, the hook records the failure\n   * as a log event on the run instead of staying quiet. The caller can\n   * then trigger downstream alerts off `analyst_gate_failed` log events.\n   */\n  gateOn?: (result: AnalyzeTracesResult, ctx: RunCompleteHookContext) => boolean\n}\n\nconst DEFAULT_QUESTION =\n  \"Summarise what happened in this run. Surface any failure modes, surprising findings, or evidence that the run's verdict is wrong.\"\n\nexport function traceAnalystOnRunComplete(opts: TraceAnalystHookOptions): RunCompleteHook {\n  return async (ctx: RunCompleteHookContext) => {\n    if (opts.shouldRun && !opts.shouldRun(ctx)) return\n    const source = opts.analyze.source\n    if (source === undefined) {\n      // The analyst needs a source. If the caller didn't supply one we don't\n      // run — but we do leave a breadcrumb so the absence is visible.\n      await ctx.store.appendEvent({\n        eventId: `analyst-skip-${ctx.runId}`,\n        runId: ctx.runId,\n        kind: 'log',\n        timestamp: Date.now(),\n        payload: { source: 'trace_analyst_hook', reason: 'no source configured' },\n      })\n      return\n    }\n    const result = await analyzeTraces({ question: opts.question ?? DEFAULT_QUESTION }, {\n      ...opts.analyze,\n      source,\n    } as AnalyzeTracesOptions)\n    if (opts.save) await opts.save(result, ctx)\n    if (opts.gateOn && !opts.gateOn(result, ctx)) {\n      await ctx.store.appendEvent({\n        eventId: `analyst-gate-${ctx.runId}`,\n        runId: ctx.runId,\n        kind: 'log',\n        timestamp: Date.now(),\n        payload: {\n          source: 'trace_analyst_hook',\n          reason: 'analyst_gate_failed',\n          findings: result.findings,\n        },\n      })\n    }\n  }\n}\n","export interface TraceInsightTask {\n  id: string\n  name: string\n  prompt?: string\n  difficulty?: string\n  tags?: string[]\n  outcome?: string\n  score?: number\n  gaps?: string[]\n}\n\nexport interface TraceInsightSuite {\n  name: string\n  collectionId?: string\n  tasks: TraceInsightTask[]\n}\n\nexport interface TraceInsightFinding {\n  kind: string\n  severity?: string\n  taskIds: string[]\n  evidence?: string\n  proposedFixClass?: string\n}\n\nexport interface TraceInsightQuestion {\n  id: string\n  question: string\n  why: string\n}\n\nexport interface TraceInsightPanelRole {\n  id: string\n  name: string\n  responsibility: string\n}\n\nexport interface TraceInsightPromptInput {\n  suite: TraceInsightSuite\n  findings?: TraceInsightFinding[]\n  agent?: Record<string, unknown>\n  totals?: Record<string, unknown>\n  maxRepresentativeTraces?: number\n}\n\nexport interface TraceInsightContext {\n  suite: TraceInsightSuite\n  scope: string\n  keywords: string[]\n  questions: TraceInsightQuestion[]\n  panel: TraceInsightPanelRole[]\n  findings: TraceInsightFinding[]\n  agent: Record<string, unknown> | null\n  totals: Record<string, unknown> | null\n}\n\nexport interface TraceInsightQualityGate {\n  id: string\n  label: string\n  passed: boolean\n  severity: 'critical' | 'high' | 'medium' | 'low'\n  detail: string\n}\n\nexport interface TraceInsightReadiness {\n  score: number\n  grade: 'external-ready' | 'internal-review' | 'raw-analysis'\n  gates: TraceInsightQualityGate[]\n}\n\nconst DOMAIN_STOP_WORDS = new Set([\n  'and',\n  'advanced',\n  'app',\n  'build',\n  'create',\n  'easy',\n  'expert',\n  'extreme',\n  'for',\n  'from',\n  'hard',\n  'implementation',\n  'integrate',\n  'medium',\n  'project',\n  'task',\n  'the',\n  'this',\n  'with',\n  'workflow',\n])\n\nexport function tokenizeDomainWords(value: string): string[] {\n  return [...value.matchAll(/[A-Za-z][A-Za-z0-9.+#-]{2,}/g)]\n    .map((match) => match[0].toLowerCase())\n    .filter((word) => !DOMAIN_STOP_WORDS.has(word))\n}\n\nexport function inferDomainKeywords(suite: TraceInsightSuite): string[] {\n  const suiteWords = new Set(tokenizeDomainWords(`${suite.name} ${suite.collectionId ?? ''}`))\n  const source = [\n    suite.name,\n    suite.collectionId ?? '',\n    ...suite.tasks.flatMap((task) => [\n      task.id,\n      task.name,\n      task.prompt ?? '',\n      task.difficulty ?? '',\n      ...(task.tags ?? []),\n      ...(task.gaps ?? []),\n    ]),\n  ].join(' ')\n  const counts = new Map<string, number>()\n  for (const word of tokenizeDomainWords(source)) counts.set(word, (counts.get(word) ?? 0) + 1)\n  return [...counts.entries()]\n    .filter(([word, count]) => count >= 2 || suiteWords.has(word))\n    .sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0]))\n    .map(([word]) => word)\n    .slice(0, 18)\n}\n\nexport function domainEvidencePattern(keywords: string[]): RegExp {\n  const escaped = keywords\n    .filter((keyword) => keyword.length >= 3)\n    .map((keyword) => keyword.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\\\$&'))\n  return escaped.length > 0\n    ? new RegExp(`(?<![A-Za-z0-9])(?:${escaped.join('|')})(?![A-Za-z0-9])`, 'i')\n    : /(?<![A-Za-z0-9])(?:sdk|api|css|dns|xml|provider|client|service|integration|webhook|transaction|auth|oauth|graphql|rest)(?![A-Za-z0-9])/i\n}\n\nexport function describeTraceInsightScope(suite: TraceInsightSuite): string {\n  const taskLabel =\n    suite.tasks.length === 1\n      ? '1 implementation task'\n      : `${suite.tasks.length} implementation tasks`\n  const tags = new Map<string, number>()\n  for (const task of suite.tasks) {\n    for (const tag of task.tags ?? []) tags.set(tag, (tags.get(tag) ?? 0) + 1)\n  }\n  const topTags = [...tags.entries()]\n    .sort((a, b) => b[1] - a[1] || a[0].localeCompare(b[0]))\n    .slice(0, 8)\n    .map(([tag]) => tag)\n  if (topTags.length > 0) return `${taskLabel} across ${topTags.join(', ')}.`\n  const difficulties = [\n    ...new Set(\n      suite.tasks.map((task) => task.difficulty).filter((value): value is string => Boolean(value)),\n    ),\n  ].join(', ')\n  return `${taskLabel} across ${difficulties || 'the selected benchmark scope'}.`\n}\n\nexport function planTraceInsightQuestions(input: TraceInsightPromptInput): TraceInsightQuestion[] {\n  const hasFailures = input.suite.tasks.some((task) => task.outcome && task.outcome !== 'satisfied')\n  const hasMultipleShots = input.suite.tasks.some((task) =>\n    (task.gaps ?? []).some((gap) => /shot|review|retry|continue/i.test(gap)),\n  )\n  const questions: TraceInsightQuestion[] = [\n    {\n      id: 'execution-path',\n      question: 'What did the worker actually do before the first meaningful implementation edit?',\n      why: 'Separates grounded execution from polished but shallow output.',\n    },\n    {\n      id: 'research-grounding',\n      question:\n        'Did the worker inspect docs, source, examples, or package references before committing to an implementation path?',\n      why: 'Identifies whether failures came from weak retrieval, weak examples, or premature coding.',\n    },\n    {\n      id: 'domain-proof',\n      question:\n        'Which tasks produced executable domain proof versus UI copy, placeholders, or inferred behavior?',\n      why: 'Keeps product-quality claims tied to concrete evidence.',\n    },\n    {\n      id: 'root-cause',\n      question:\n        'For each major failure cluster, is the likely root cause prompt/scaffold, docs/examples, SDK/API ergonomics, evaluator, runtime, or model behavior?',\n      why: 'Turns trace observations into actionable ownership.',\n    },\n    {\n      id: 'evidence-quality',\n      question:\n        'Which external-facing claims are directly supported by trace ids, span ids, verifier findings, reviewer notes, or generated code?',\n      why: 'Prevents unsupported customer-report conclusions.',\n    },\n  ]\n  if (hasMultipleShots) {\n    questions.push({\n      id: 'reviewer-lift',\n      question: 'Where did reviewer feedback improve score, stall, or regress across shots?',\n      why: 'Shows whether the driver loop is learning or merely repeating work.',\n    })\n  }\n  if (hasFailures) {\n    questions.push({\n      id: 'optimization-targets',\n      question:\n        'Which prompt, evaluator, scaffold, or workflow changes should feed the next GEPA/autoresearch optimization run?',\n      why: 'Connects benchmark evidence to the optimization loop.',\n    })\n  }\n  return questions\n}\n\nexport function buildTraceInsightContext(input: TraceInsightPromptInput): TraceInsightContext {\n  return {\n    suite: input.suite,\n    scope: describeTraceInsightScope(input.suite),\n    keywords: inferDomainKeywords(input.suite),\n    questions: planTraceInsightQuestions(input),\n    panel: defaultTraceInsightPanel(),\n    findings: input.findings ?? [],\n    agent: input.agent ?? null,\n    totals: input.totals ?? null,\n  }\n}\n\nexport function scoreTraceInsightReadiness(context: TraceInsightContext): TraceInsightReadiness {\n  const failedTasks = context.suite.tasks.filter(\n    (task) => task.outcome && task.outcome !== 'satisfied',\n  )\n  const findingTaskIds = new Set(context.findings.flatMap((finding) => finding.taskIds))\n  const failedTasksWithFindings = failedTasks.filter((task) => findingTaskIds.has(task.id))\n  const tasksWithGaps = context.suite.tasks.filter((task) => (task.gaps ?? []).length > 0)\n  const gates: TraceInsightQualityGate[] = [\n    {\n      id: 'domain-context',\n      label: 'Domain context inferred',\n      passed: context.keywords.length > 0,\n      severity: 'high',\n      detail:\n        context.keywords.length > 0\n          ? `${context.keywords.length} domain terms inferred: ${context.keywords.slice(0, 8).join(', ')}`\n          : 'No domain terms were inferred from suite, tasks, prompts, tags, or gaps.',\n    },\n    {\n      id: 'panel-coverage',\n      label: 'Analyst panel planned',\n      passed: context.panel.length >= 4 && context.questions.length >= 5,\n      severity: 'high',\n      detail: `${context.panel.length} panel roles and ${context.questions.length} investigation questions planned.`,\n    },\n    {\n      id: 'failure-coverage',\n      label: 'Failures mapped to findings',\n      passed:\n        failedTasks.length === 0 || failedTasksWithFindings.length / failedTasks.length >= 0.5,\n      severity: 'critical',\n      detail:\n        failedTasks.length === 0\n          ? 'No failed tasks in suite.'\n          : `${failedTasksWithFindings.length}/${failedTasks.length} failed tasks appear in finding clusters.`,\n    },\n    {\n      id: 'gap-evidence',\n      label: 'Task gaps captured',\n      passed: failedTasks.length === 0 || tasksWithGaps.length / failedTasks.length >= 0.5,\n      severity: 'medium',\n      detail: `${tasksWithGaps.length} tasks include explicit evaluator or analyst gaps.`,\n    },\n  ]\n  const penalty = gates.reduce((sum, gate) => {\n    if (gate.passed) return sum\n    if (gate.severity === 'critical') return sum + 35\n    if (gate.severity === 'high') return sum + 20\n    if (gate.severity === 'medium') return sum + 10\n    return sum + 5\n  }, 0)\n  const score = Math.max(0, Math.min(1, 1 - penalty / 100))\n  return {\n    score,\n    grade: score >= 0.9 ? 'external-ready' : score >= 0.7 ? 'internal-review' : 'raw-analysis',\n    gates,\n  }\n}\n\nexport function defaultTraceInsightPanel(): TraceInsightPanelRole[] {\n  return [\n    {\n      id: 'trace-forensics',\n      name: 'Trace Forensics',\n      responsibility:\n        'Reconstruct what the worker did in order, including research, edits, reviewer interventions, verifier feedback, and stop reason.',\n    },\n    {\n      id: 'root-cause',\n      name: 'Root Cause',\n      responsibility:\n        'Map failures to prompt/scaffold, docs/examples, SDK/API/product ergonomics, evaluator, runtime, or model behavior.',\n    },\n    {\n      id: 'optimization',\n      name: 'Optimization',\n      responsibility:\n        'Identify prompt, reviewer, evaluator, scaffold, and GEPA/autoresearch changes that should be tested next.',\n    },\n    {\n      id: 'external-evidence',\n      name: 'External Evidence',\n      responsibility:\n        'Separate customer-safe claims from internal harness findings and reject conclusions without task, trace, span, code, reviewer, or verifier evidence.',\n    },\n  ]\n}\n\nexport function buildTraceInsightPrompt(input: TraceInsightPromptInput): string {\n  const context = buildTraceInsightContext(input)\n  const maxRepresentativeTraces = input.maxRepresentativeTraces ?? 6\n  return `Analyze this benchmark run and produce evidence-backed trace intelligence.\n\nAudience:\n- internal AI/product leadership\n- possible customer-facing report for ${input.suite.name}\n\nInvestigation plan:\n${context.questions.map((item, index) => `${index + 1}. ${item.question} (${item.why})`).join('\\n')}\n\nAnalyst panel:\n${context.panel.map((role) => `- ${role.name}: ${role.responsibility}`).join('\\n')}\n\nIf the task branches are independent, use subagents for the panel roles above and aggregate their findings. Do not run a panel role unless its answer will change the final report.\n\nRequired output:\n1. Executive verdict: what this run proves and does not prove.\n2. The investigation questions you answered and the evidence used.\n3. Failure taxonomy: agent prompting, evaluator/harness, docs/examples, SDK/API/product integration, infra.\n4. Evidence-backed examples with trace ids/task ids and concrete verifier findings.\n5. Highest-ROI fixes for the benchmark harness, prompt/GEPA optimization, and customer-facing product/docs surface.\n6. What is safe for an external report versus what must stay internal.\n7. One rerun plan that would validate lift after optimization.\n\nBudget:\n- Inspect the dataset overview, the failure summary, and at most ${maxRepresentativeTraces} representative traces.\n- Prefer traces named in the failure summary over broad exploration.\n- Do not do exhaustive trace sweeps.\n- Return the final report as soon as the taxonomy and examples are supported.\n\nRun summary:\n${JSON.stringify(\n  {\n    suite: input.suite.name,\n    scope: context.scope,\n    inferredKeywords: context.keywords,\n    agent: context.agent,\n    totals: context.totals,\n    findings: context.findings.map((finding) => ({\n      kind: finding.kind,\n      severity: finding.severity,\n      taskCount: finding.taskIds.length,\n      proposedFixClass: finding.proposedFixClass,\n    })),\n    failures: input.suite.tasks\n      .filter((task) => task.outcome && task.outcome !== 'satisfied')\n      .map((task) => ({\n        task: task.id,\n        difficulty: task.difficulty,\n        outcome: task.outcome,\n        score: task.score,\n        gaps: task.gaps ?? [],\n      })),\n  },\n  null,\n  2,\n)}\n\nUse the trace tools. Do not invent facts. Cite task ids. Separate customer-facing claims from internal harness/model findings.`\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAsDA,IAAM,mBACJ;AAEK,SAAS,0BAA0B,MAAgD;AACxF,SAAO,OAAO,QAAgC;AAC5C,QAAI,KAAK,aAAa,CAAC,KAAK,UAAU,GAAG,EAAG;AAC5C,UAAM,SAAS,KAAK,QAAQ;AAC5B,QAAI,WAAW,QAAW;AAGxB,YAAM,IAAI,MAAM,YAAY;AAAA,QAC1B,SAAS,gBAAgB,IAAI,KAAK;AAAA,QAClC,OAAO,IAAI;AAAA,QACX,MAAM;AAAA,QACN,WAAW,KAAK,IAAI;AAAA,QACpB,SAAS,EAAE,QAAQ,sBAAsB,QAAQ,uBAAuB;AAAA,MAC1E,CAAC;AACD;AAAA,IACF;AACA,UAAM,SAAS,MAAM,cAAc,EAAE,UAAU,KAAK,YAAY,iBAAiB,GAAG;AAAA,MAClF,GAAG,KAAK;AAAA,MACR;AAAA,IACF,CAAyB;AACzB,QAAI,KAAK,KAAM,OAAM,KAAK,KAAK,QAAQ,GAAG;AAC1C,QAAI,KAAK,UAAU,CAAC,KAAK,OAAO,QAAQ,GAAG,GAAG;AAC5C,YAAM,IAAI,MAAM,YAAY;AAAA,QAC1B,SAAS,gBAAgB,IAAI,KAAK;AAAA,QAClC,OAAO,IAAI;AAAA,QACX,MAAM;AAAA,QACN,WAAW,KAAK,IAAI;AAAA,QACpB,SAAS;AAAA,UACP,QAAQ;AAAA,UACR,QAAQ;AAAA,UACR,UAAU,OAAO;AAAA,QACnB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACF;;;ACtBA,IAAM,oBAAoB,oBAAI,IAAI;AAAA,EAChC;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;AAAA,EACA;AACF,CAAC;AAEM,SAAS,oBAAoB,OAAyB;AAC3D,SAAO,CAAC,GAAG,MAAM,SAAS,8BAA8B,CAAC,EACtD,IAAI,CAAC,UAAU,MAAM,CAAC,EAAE,YAAY,CAAC,EACrC,OAAO,CAAC,SAAS,CAAC,kBAAkB,IAAI,IAAI,CAAC;AAClD;AAEO,SAAS,oBAAoB,OAAoC;AACtE,QAAM,aAAa,IAAI,IAAI,oBAAoB,GAAG,MAAM,IAAI,IAAI,MAAM,gBAAgB,EAAE,EAAE,CAAC;AAC3F,QAAM,SAAS;AAAA,IACb,MAAM;AAAA,IACN,MAAM,gBAAgB;AAAA,IACtB,GAAG,MAAM,MAAM,QAAQ,CAAC,SAAS;AAAA,MAC/B,KAAK;AAAA,MACL,KAAK;AAAA,MACL,KAAK,UAAU;AAAA,MACf,KAAK,cAAc;AAAA,MACnB,GAAI,KAAK,QAAQ,CAAC;AAAA,MAClB,GAAI,KAAK,QAAQ,CAAC;AAAA,IACpB,CAAC;AAAA,EACH,EAAE,KAAK,GAAG;AACV,QAAM,SAAS,oBAAI,IAAoB;AACvC,aAAW,QAAQ,oBAAoB,MAAM,EAAG,QAAO,IAAI,OAAO,OAAO,IAAI,IAAI,KAAK,KAAK,CAAC;AAC5F,SAAO,CAAC,GAAG,OAAO,QAAQ,CAAC,EACxB,OAAO,CAAC,CAAC,MAAM,KAAK,MAAM,SAAS,KAAK,WAAW,IAAI,IAAI,CAAC,EAC5D,KAAK,CAAC,GAAG,MAAM,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,EAAE,CAAC,EAAE,cAAc,EAAE,CAAC,CAAC,CAAC,EACtD,IAAI,CAAC,CAAC,IAAI,MAAM,IAAI,EACpB,MAAM,GAAG,EAAE;AAChB;AAEO,SAAS,sBAAsB,UAA4B;AAChE,QAAM,UAAU,SACb,OAAO,CAAC,YAAY,QAAQ,UAAU,CAAC,EACvC,IAAI,CAAC,YAAY,QAAQ,QAAQ,uBAAuB,MAAM,CAAC;AAClE,SAAO,QAAQ,SAAS,IACpB,IAAI,OAAO,sBAAsB,QAAQ,KAAK,GAAG,CAAC,oBAAoB,GAAG,IACzE;AACN;AAEO,SAAS,0BAA0B,OAAkC;AAC1E,QAAM,YACJ,MAAM,MAAM,WAAW,IACnB,0BACA,GAAG,MAAM,MAAM,MAAM;AAC3B,QAAM,OAAO,oBAAI,IAAoB;AACrC,aAAW,QAAQ,MAAM,OAAO;AAC9B,eAAW,OAAO,KAAK,QAAQ,CAAC,EAAG,MAAK,IAAI,MAAM,KAAK,IAAI,GAAG,KAAK,KAAK,CAAC;AAAA,EAC3E;AACA,QAAM,UAAU,CAAC,GAAG,KAAK,QAAQ,CAAC,EAC/B,KAAK,CAAC,GAAG,MAAM,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,EAAE,CAAC,EAAE,cAAc,EAAE,CAAC,CAAC,CAAC,EACtD,MAAM,GAAG,CAAC,EACV,IAAI,CAAC,CAAC,GAAG,MAAM,GAAG;AACrB,MAAI,QAAQ,SAAS,EAAG,QAAO,GAAG,SAAS,WAAW,QAAQ,KAAK,IAAI,CAAC;AACxE,QAAM,eAAe;AAAA,IACnB,GAAG,IAAI;AAAA,MACL,MAAM,MAAM,IAAI,CAAC,SAAS,KAAK,UAAU,EAAE,OAAO,CAAC,UAA2B,QAAQ,KAAK,CAAC;AAAA,IAC9F;AAAA,EACF,EAAE,KAAK,IAAI;AACX,SAAO,GAAG,SAAS,WAAW,gBAAgB,8BAA8B;AAC9E;AAEO,SAAS,0BAA0B,OAAwD;AAChG,QAAM,cAAc,MAAM,MAAM,MAAM,KAAK,CAAC,SAAS,KAAK,WAAW,KAAK,YAAY,WAAW;AACjG,QAAM,mBAAmB,MAAM,MAAM,MAAM;AAAA,IAAK,CAAC,UAC9C,KAAK,QAAQ,CAAC,GAAG,KAAK,CAAC,QAAQ,8BAA8B,KAAK,GAAG,CAAC;AAAA,EACzE;AACA,QAAM,YAAoC;AAAA,IACxC;AAAA,MACE,IAAI;AAAA,MACJ,UAAU;AAAA,MACV,KAAK;AAAA,IACP;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,UACE;AAAA,MACF,KAAK;AAAA,IACP;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,UACE;AAAA,MACF,KAAK;AAAA,IACP;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,UACE;AAAA,MACF,KAAK;AAAA,IACP;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,UACE;AAAA,MACF,KAAK;AAAA,IACP;AAAA,EACF;AACA,MAAI,kBAAkB;AACpB,cAAU,KAAK;AAAA,MACb,IAAI;AAAA,MACJ,UAAU;AAAA,MACV,KAAK;AAAA,IACP,CAAC;AAAA,EACH;AACA,MAAI,aAAa;AACf,cAAU,KAAK;AAAA,MACb,IAAI;AAAA,MACJ,UACE;AAAA,MACF,KAAK;AAAA,IACP,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEO,SAAS,yBAAyB,OAAqD;AAC5F,SAAO;AAAA,IACL,OAAO,MAAM;AAAA,IACb,OAAO,0BAA0B,MAAM,KAAK;AAAA,IAC5C,UAAU,oBAAoB,MAAM,KAAK;AAAA,IACzC,WAAW,0BAA0B,KAAK;AAAA,IAC1C,OAAO,yBAAyB;AAAA,IAChC,UAAU,MAAM,YAAY,CAAC;AAAA,IAC7B,OAAO,MAAM,SAAS;AAAA,IACtB,QAAQ,MAAM,UAAU;AAAA,EAC1B;AACF;AAEO,SAAS,2BAA2B,SAAqD;AAC9F,QAAM,cAAc,QAAQ,MAAM,MAAM;AAAA,IACtC,CAAC,SAAS,KAAK,WAAW,KAAK,YAAY;AAAA,EAC7C;AACA,QAAM,iBAAiB,IAAI,IAAI,QAAQ,SAAS,QAAQ,CAAC,YAAY,QAAQ,OAAO,CAAC;AACrF,QAAM,0BAA0B,YAAY,OAAO,CAAC,SAAS,eAAe,IAAI,KAAK,EAAE,CAAC;AACxF,QAAM,gBAAgB,QAAQ,MAAM,MAAM,OAAO,CAAC,UAAU,KAAK,QAAQ,CAAC,GAAG,SAAS,CAAC;AACvF,QAAM,QAAmC;AAAA,IACvC;AAAA,MACE,IAAI;AAAA,MACJ,OAAO;AAAA,MACP,QAAQ,QAAQ,SAAS,SAAS;AAAA,MAClC,UAAU;AAAA,MACV,QACE,QAAQ,SAAS,SAAS,IACtB,GAAG,QAAQ,SAAS,MAAM,2BAA2B,QAAQ,SAAS,MAAM,GAAG,CAAC,EAAE,KAAK,IAAI,CAAC,KAC5F;AAAA,IACR;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,OAAO;AAAA,MACP,QAAQ,QAAQ,MAAM,UAAU,KAAK,QAAQ,UAAU,UAAU;AAAA,MACjE,UAAU;AAAA,MACV,QAAQ,GAAG,QAAQ,MAAM,MAAM,oBAAoB,QAAQ,UAAU,MAAM;AAAA,IAC7E;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,OAAO;AAAA,MACP,QACE,YAAY,WAAW,KAAK,wBAAwB,SAAS,YAAY,UAAU;AAAA,MACrF,UAAU;AAAA,MACV,QACE,YAAY,WAAW,IACnB,8BACA,GAAG,wBAAwB,MAAM,IAAI,YAAY,MAAM;AAAA,IAC/D;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,OAAO;AAAA,MACP,QAAQ,YAAY,WAAW,KAAK,cAAc,SAAS,YAAY,UAAU;AAAA,MACjF,UAAU;AAAA,MACV,QAAQ,GAAG,cAAc,MAAM;AAAA,IACjC;AAAA,EACF;AACA,QAAM,UAAU,MAAM,OAAO,CAAC,KAAK,SAAS;AAC1C,QAAI,KAAK,OAAQ,QAAO;AACxB,QAAI,KAAK,aAAa,WAAY,QAAO,MAAM;AAC/C,QAAI,KAAK,aAAa,OAAQ,QAAO,MAAM;AAC3C,QAAI,KAAK,aAAa,SAAU,QAAO,MAAM;AAC7C,WAAO,MAAM;AAAA,EACf,GAAG,CAAC;AACJ,QAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,IAAI,UAAU,GAAG,CAAC;AACxD,SAAO;AAAA,IACL;AAAA,IACA,OAAO,SAAS,MAAM,mBAAmB,SAAS,MAAM,oBAAoB;AAAA,IAC5E;AAAA,EACF;AACF;AAEO,SAAS,2BAAoD;AAClE,SAAO;AAAA,IACL;AAAA,MACE,IAAI;AAAA,MACJ,MAAM;AAAA,MACN,gBACE;AAAA,IACJ;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,MAAM;AAAA,MACN,gBACE;AAAA,IACJ;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,MAAM;AAAA,MACN,gBACE;AAAA,IACJ;AAAA,IACA;AAAA,MACE,IAAI;AAAA,MACJ,MAAM;AAAA,MACN,gBACE;AAAA,IACJ;AAAA,EACF;AACF;AAEO,SAAS,wBAAwB,OAAwC;AAC9E,QAAM,UAAU,yBAAyB,KAAK;AAC9C,QAAM,0BAA0B,MAAM,2BAA2B;AACjE,SAAO;AAAA;AAAA;AAAA;AAAA,wCAI+B,MAAM,MAAM,IAAI;AAAA;AAAA;AAAA,EAGtD,QAAQ,UAAU,IAAI,CAAC,MAAM,UAAU,GAAG,QAAQ,CAAC,KAAK,KAAK,QAAQ,KAAK,KAAK,GAAG,GAAG,EAAE,KAAK,IAAI,CAAC;AAAA;AAAA;AAAA,EAGjG,QAAQ,MAAM,IAAI,CAAC,SAAS,KAAK,KAAK,IAAI,KAAK,KAAK,cAAc,EAAE,EAAE,KAAK,IAAI,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,mEAcf,uBAAuB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMxF,KAAK;AAAA,IACL;AAAA,MACE,OAAO,MAAM,MAAM;AAAA,MACnB,OAAO,QAAQ;AAAA,MACf,kBAAkB,QAAQ;AAAA,MAC1B,OAAO,QAAQ;AAAA,MACf,QAAQ,QAAQ;AAAA,MAChB,UAAU,QAAQ,SAAS,IAAI,CAAC,aAAa;AAAA,QAC3C,MAAM,QAAQ;AAAA,QACd,UAAU,QAAQ;AAAA,QAClB,WAAW,QAAQ,QAAQ;AAAA,QAC3B,kBAAkB,QAAQ;AAAA,MAC5B,EAAE;AAAA,MACF,UAAU,MAAM,MAAM,MACnB,OAAO,CAAC,SAAS,KAAK,WAAW,KAAK,YAAY,WAAW,EAC7D,IAAI,CAAC,UAAU;AAAA,QACd,MAAM,KAAK;AAAA,QACX,YAAY,KAAK;AAAA,QACjB,SAAS,KAAK;AAAA,QACd,OAAO,KAAK;AAAA,QACZ,MAAM,KAAK,QAAQ,CAAC;AAAA,MACtB,EAAE;AAAA,IACN;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AAAA;AAAA;AAGD;","names":[]}
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/{trajectory-BFmveYZt.d.ts → trajectory-CnoBo-JY.d.ts}

@@ -1,4 +1,4 @@
-import { S as Span, b as TraceEvent, T as TraceStore } from './store-BP5be6s7.js';
+import { S as Span, b as TraceEvent, T as TraceStore } from './store-Db2Bv8Cf.js';
 
 /**
  * Trajectory — ordered, structured view over a run's spans.

package/dist/wire/index.d.ts

@@ -1,14 +1,14 @@
-import { n as FeedbackTrajectoryStore } from '../feedback-trajectory-j0nJFgC6.js';
-import { T as TraceStore } from '../store-BP5be6s7.js';
+import { n as FeedbackTrajectoryStore } from '../feedback-trajectory-iATEAHmc.js';
+import { T as TraceStore } from '../store-Db2Bv8Cf.js';
 import { z } from 'zod';
 import { OpenAPIObject } from 'openapi3-ts/oas31';
 import * as hono_types from 'hono/types';
 import { ServerType } from '@hono/node-server';
 import { Hono } from 'hono';
-import '../control-runtime-BRdQ0wrx.js';
-import '../emitter-BqjeOvJh.js';
-import '../dataset-CiK_3LDr.js';
-import '../errors-BZ9sTdz7.js';
+import '../control-runtime-BZ_lVLYW.js';
+import '../emitter-DP_cSSiw.js';
+import '../dataset-ueRVTUoY.js';
+import '../errors-mje_cKOs.js';
 
 declare const RubricDimensionSchema: z.ZodObject<{
     id: z.ZodString;

package/dist/wire/index.js

@@ -33,10 +33,10 @@ import {
   runRpcBatch,
   runRpcOnce,
   startServer
-} from "../chunk-XFZCM5Z3.js";
-import "../chunk-4S4BM3QQ.js";
+} from "../chunk-SMSGXM74.js";
+import "../chunk-M6RZ5LJN.js";
 import "../chunk-PC4UYEBM.js";
-import "../chunk-NG236HPC.js";
+import "../chunk-QYJT52YW.js";
 import "../chunk-PZ5AY32C.js";
 export {
   BUILTIN_RUBRICS,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.29.1",
+  "version": "0.31.0",
   "description": "Substrate for self-improving agents: traces, verifiable rewards, preferences, GEPA / reflective mutation, auto-research, replay, sequential anytime-valid stats, and release gates.",
   "homepage": "https://github.com/tangle-network/agent-eval#readme",
   "repository": {

package/dist/chunk-NG236HPC.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/errors.ts"],"sourcesContent":["/**\n * Error taxonomy for `@tangle-network/agent-eval`.\n *\n * Every error this package throws as part of its *public contract* extends\n * `AgentEvalError`. Consumers can pattern-match by `instanceof <Subclass>` or\n * by the stable string `code` carried on the base class.\n *\n * The codes are stable across minor versions; new codes can be added, but\n * existing codes never change meaning. New subclasses are non-breaking.\n *\n * Internal invariant guards (`throw new Error('this should never happen')`)\n * remain plain `Error`s on purpose — they're programmer-mistake assertions,\n * not consumer-catchable contract failures.\n */\n\nexport type AgentEvalErrorCode =\n  | 'validation'\n  | 'not_found'\n  | 'config'\n  | 'capture_integrity'\n  | 'judge'\n  | 'verification'\n  | 'replay'\n\nexport class AgentEvalError extends Error {\n  /** Stable string code. Survives minification; safe to switch on. */\n  readonly code: AgentEvalErrorCode\n\n  constructor(code: AgentEvalErrorCode, message: string, options?: { cause?: unknown }) {\n    super(message, options)\n    this.name = this.constructor.name\n    this.code = code\n  }\n}\n\n/** Caller passed invalid arguments (out of range, mutually-exclusive options, bad shape). */\nexport class ValidationError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('validation', message, options)\n  }\n}\n\n/** A named resource (run, span, rubric, scenario, dataset row, route) does not exist. */\nexport class NotFoundError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('not_found', message, options)\n  }\n}\n\n/** Configuration missing or malformed (`HOME` unset, required image not supplied, env var absent). */\nexport class ConfigError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('config', message, options)\n  }\n}\n\n/**\n * A run is missing the artifacts a launch-grade check requires:\n * raw HTTP capture absent, no LLM spans, route assertion failed, run-end\n * assertion tripped. Block ship on this; do not catch and move on.\n */\nexport class CaptureIntegrityError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('capture_integrity', message, options)\n  }\n}\n\n/** A judge call failed in a way that's not retryable: schema parse failure, bad rubric, conflicting dimensions. */\nexport class JudgeError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('judge', message, options)\n  }\n}\n\n/** A verifier signalled a hard failure (compile, test, schema) — distinct from a low judge score. */\nexport class VerificationError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('verification', message, options)\n  }\n}\n\n/** Replay cache cannot satisfy a request: miss with no fallback, sink lacks list(), unsupported URL. */\nexport class ReplayError extends AgentEvalError {\n  constructor(message: string, options?: { cause?: unknown }) {\n    super('replay', message, options)\n  }\n}\n"],"mappings":";AAwBO,IAAM,iBAAN,cAA6B,MAAM;AAAA;AAAA,EAE/B;AAAA,EAET,YAAY,MAA0B,SAAiB,SAA+B;AACpF,UAAM,SAAS,OAAO;AACtB,SAAK,OAAO,KAAK,YAAY;AAC7B,SAAK,OAAO;AAAA,EACd;AACF;AAGO,IAAM,kBAAN,cAA8B,eAAe;AAAA,EAClD,YAAY,SAAiB,SAA+B;AAC1D,UAAM,cAAc,SAAS,OAAO;AAAA,EACtC;AACF;AAGO,IAAM,gBAAN,cAA4B,eAAe;AAAA,EAChD,YAAY,SAAiB,SAA+B;AAC1D,UAAM,aAAa,SAAS,OAAO;AAAA,EACrC;AACF;AAGO,IAAM,cAAN,cAA0B,eAAe;AAAA,EAC9C,YAAY,SAAiB,SAA+B;AAC1D,UAAM,UAAU,SAAS,OAAO;AAAA,EAClC;AACF;AAOO,IAAM,wBAAN,cAAoC,eAAe;AAAA,EACxD,YAAY,SAAiB,SAA+B;AAC1D,UAAM,qBAAqB,SAAS,OAAO;AAAA,EAC7C;AACF;AAGO,IAAM,aAAN,cAAyB,eAAe;AAAA,EAC7C,YAAY,SAAiB,SAA+B;AAC1D,UAAM,SAAS,SAAS,OAAO;AAAA,EACjC;AACF;AAGO,IAAM,oBAAN,cAAgC,eAAe;AAAA,EACpD,YAAY,SAAiB,SAA+B;AAC1D,UAAM,gBAAgB,SAAS,OAAO;AAAA,EACxC;AACF;AAGO,IAAM,cAAN,cAA0B,eAAe;AAAA,EAC9C,YAAY,SAAiB,SAA+B;AAC1D,UAAM,UAAU,SAAS,OAAO;AAAA,EAClC;AACF;","names":[]}