@tangle-network/agent-eval 0.118.2 → 0.119.0

package/dist/{chunk-63KG3ASG.js → chunk-FIUFRXP3.js}

@@ -2,7 +2,7 @@ import {
   OtlpFileTraceStore,
   TraceFileMissingError,
   buildTraceAnalystTools
-} from "./chunk-BUBU3YHT.js";
+} from "./chunk-6QIM2EAP.js";
 
 // src/trace-analyst/prompts.ts
 var TRACE_ANALYST_ACTOR_DESCRIPTION = `You answer questions about an OTLP-shaped JSONL trace dataset using the trace tools provided in the \`traces\` namespace.
@@ -198,4 +198,4 @@ export {
   TRACE_ANALYST_SUBAGENT_DESCRIPTION,
   analyzeTraces
 };
-//# sourceMappingURL=chunk-63KG3ASG.js.map
+//# sourceMappingURL=chunk-FIUFRXP3.js.map

package/dist/{chunk-J23QK2U5.js → chunk-GERDAIAL.js}

@@ -6,19 +6,19 @@ import {
   DEFAULT_TRACE_ANALYST_KINDS,
   createTraceAnalystKind,
   makeFinding
-} from "./chunk-DZBCDDIE.js";
+} from "./chunk-7A4LIMMY.js";
 import {
   LlmClient
 } from "./chunk-NJC7U437.js";
 import {
-  LLM_INPUT_TOKEN_ATTR_KEYS,
-  LLM_OUTPUT_TOKEN_ATTR_KEYS,
-  TOOL_NAME_ATTR_KEYS,
   spanEpochMillis
-} from "./chunk-BUBU3YHT.js";
+} from "./chunk-6QIM2EAP.js";
 import {
-  LLM_CONTEXT_TOKENS
-} from "./chunk-NW4AQBXD.js";
+  LLM_CONTEXT_TOKENS,
+  LLM_INPUT_TOKEN_ATTR_KEYS,
+  LLM_OUTPUT_TOKEN_ATTR_KEYS,
+  TOOL_NAME_ATTR_KEYS
+} from "./chunk-K4DBDHLK.js";
 
 // src/analyst/chat-client.ts
 function createChatClient(opts) {
@@ -547,4 +547,4 @@ export {
   behavioralAnalyst,
   buildDefaultAnalystRegistry
 };
-//# sourceMappingURL=chunk-J23QK2U5.js.map
+//# sourceMappingURL=chunk-GERDAIAL.js.map

package/dist/{chunk-J23QK2U5.js.map → chunk-GERDAIAL.js.map}

@@ -1 +1 @@
-{"version":3,"sources":["../src/analyst/chat-client.ts","../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/default-registry.ts"],"sourcesContent":["/**\n * ChatClient — the single LLM abstraction analysts call.\n *\n * agent-eval already ships an `LlmClient` (OpenAI-compatible, retry,\n * graceful JSON-schema degrade) and judges that talk to `TCloud`. Two\n * mixed patterns force every analyst author to pick a transport, which\n * couples analyst code to runtime concerns (cli-bridge vs router vs\n * sandbox-sdk) it shouldn't know about.\n *\n * `ChatClient` is one interface every analyst takes via `AnalystContext.chat`.\n * The operator decides at the registry boundary which transport binds\n * to it. Analyst code stays transport-agnostic; swapping production\n * (sandbox-sdk) for local dev (cli-bridge) or tests (mock) is a one-\n * line factory call.\n *\n * Designed to coexist: existing `LlmClient` callers and existing\n * `TCloud`-based judges keep working untouched. New analyst code uses\n * `ChatClient`. When old call sites migrate, they pick up budgeting,\n * cancellation, and unified telemetry for free.\n */\n\nimport {\n  type LlmCallRequest,\n  type LlmCallResult,\n  LlmClient,\n  type LlmClientOptions,\n} from '../llm-client'\n\n/**\n * Unified chat interface. Mirrors LlmCallRequest/Result so the OpenAI-\n * compatible mental model stays. Two methods: a one-shot `chat()` and\n * an `streamChat()` for future agentic loops (not yet exposed).\n */\nexport interface ChatClient {\n  /** Display name of the bound transport — included in telemetry. */\n  readonly transport: ChatTransport\n  /** Default model when caller omits — operators bind this per environment. */\n  readonly defaultModel?: string\n  /** Total provider attempts this transport can make for one chat call. */\n  readonly maximumAttempts?: number\n\n  /** Implementations must enforce `req.maxTokens` when it is present. */\n  chat(req: ChatRequest, opts?: ChatCallOpts): Promise<ChatResponse>\n}\n\nexport type ChatTransport =\n  | 'router' // router.tangle.tools — production paid models\n  | 'sandbox-sdk' // box.streamPrompt() — chat completion via sandbox SDK\n  | 'cli-bridge' // local cli-bridge for dev / local-only runs\n  | 'direct-provider' // direct OpenAI / Anthropic / etc. — bypass router\n  | 'mock' // test-time injection\n\nexport interface ChatRequest extends Omit<LlmCallRequest, 'model'> {\n  /** Optional — falls back to ChatClient.defaultModel. */\n  model?: string\n}\n\nexport type ChatResponse = LlmCallResult\n\nexport interface ChatCallOpts {\n  /** Cancel the in-flight request. */\n  signal?: AbortSignal\n  /** Hard USD ceiling for this single call (informational; the underlying transport may not enforce). */\n  maxCostUsd?: number\n  /** Correlation tag carried into request headers when the transport allows. */\n  correlationId?: string\n  /** Stable provider idempotency key for retries/redrives of one paid call. */\n  idempotencyKey?: string\n}\n\n// ── Factory ─────────────────────────────────────────────────────────\n\nexport type CreateChatClientOpts =\n  | RouterTransportOpts\n  | CliBridgeTransportOpts\n  | DirectProviderTransportOpts\n  | SandboxSdkTransportOpts\n  | MockTransportOpts\n\ninterface BaseTransportOpts {\n  defaultModel?: string\n  /** Total provider attempts. Required for opaque transports used in capped runs. */\n  maximumAttempts?: number\n}\n\nexport interface RouterTransportOpts extends BaseTransportOpts {\n  transport: 'router'\n  baseUrl?: string\n  apiKey: string\n}\n\nexport interface CliBridgeTransportOpts extends BaseTransportOpts {\n  transport: 'cli-bridge'\n  baseUrl?: string\n  bearer?: string\n}\n\nexport interface DirectProviderTransportOpts extends BaseTransportOpts {\n  transport: 'direct-provider'\n  baseUrl: string\n  apiKey: string\n}\n\n/**\n * Sandbox-SDK transport. Provided as a thin pass-through: the caller\n * supplies a callable that mimics LlmClient.chat() against an already-\n * configured Sandbox handle. We don't import the SDK here to keep\n * agent-eval dep-free of @tangle-network/sandbox.\n */\nexport interface SandboxSdkTransportOpts extends BaseTransportOpts {\n  transport: 'sandbox-sdk'\n  chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Mock transport for tests. The handler receives the request and returns\n * whatever the test wants. No retries, no JSON-schema degrade.\n */\nexport interface MockTransportOpts extends BaseTransportOpts {\n  transport: 'mock'\n  handler: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Build a ChatClient bound to a specific transport. The returned client\n * is safe to share across analysts in a single registry run.\n */\nexport function createChatClient(opts: CreateChatClientOpts): ChatClient {\n  switch (opts.transport) {\n    case 'router':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl ?? 'https://router.tangle.tools/v1',\n          apiKey: opts.apiKey,\n          maxRetries: opts.maximumAttempts,\n        } as LlmClientOptions),\n      )\n    case 'cli-bridge':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl ?? 'http://127.0.0.1:3344/v1',\n          apiKey: opts.bearer ?? '',\n          maxRetries: opts.maximumAttempts,\n        } as LlmClientOptions),\n      )\n    case 'direct-provider':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl,\n          apiKey: opts.apiKey,\n          maxRetries: opts.maximumAttempts,\n        } as LlmClientOptions),\n      )\n    case 'sandbox-sdk':\n      return {\n        transport: 'sandbox-sdk',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: opts.maximumAttempts,\n        chat: async (req, callOpts) => opts.chat(resolveModel(req, opts.defaultModel), callOpts),\n      }\n    case 'mock':\n      return {\n        transport: 'mock',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: 1,\n        chat: async (req, callOpts) => opts.handler(resolveModel(req, opts.defaultModel), callOpts),\n      }\n  }\n}\n\nfunction wrapLlmClient(\n  transport: ChatTransport,\n  defaultModel: string | undefined,\n  inner: LlmClient,\n): ChatClient {\n  return {\n    transport,\n    defaultModel,\n    maximumAttempts: inner.maximumAttempts,\n    chat: (req, callOpts) => {\n      const resolved = resolveModel(req, defaultModel)\n      const request: LlmCallRequest = {\n        model: resolved.model!,\n        messages: req.messages,\n        jsonMode: req.jsonMode,\n        jsonSchema: req.jsonSchema,\n        temperature: req.temperature,\n        maxTokens: req.maxTokens,\n        timeoutMs: req.timeoutMs,\n      }\n      return inner.call(request, {\n        signal: callOpts?.signal,\n        idempotencyKey: callOpts?.idempotencyKey,\n      })\n    },\n  }\n}\n\nfunction resolveModel(req: ChatRequest, defaultModel: string | undefined): ChatRequest {\n  if (req.model) return req\n  if (!defaultModel) {\n    throw new Error(\n      'ChatClient.chat: no model on request and no defaultModel on the client. ' +\n        'Either pass req.model or bind defaultModel at createChatClient().',\n    )\n  }\n  return { ...req, model: defaultModel }\n}\n","/**\n * Deterministic behavioral metrics over OTLP spans — pure arithmetic, no LLM.\n *\n * These are the model-independent multiplier: the four trace-quality signals a\n * tolerant analyzer (e.g. HALO) re-derives per run inside the model — token\n * growth, output decay, tool monoculture, missing self-verification — computed\n * here once, in TypeScript, with zero model judgment. A finding that falls out\n * of arithmetic is trivially model-agnostic and cannot hallucinate the trend.\n *\n * General, not trace-specific: the detectors key off token trajectories and\n * tool usage present in any agentic OTLP trace, not any one benchmark.\n */\n\nimport { LLM_CONTEXT_TOKENS } from '../trace/attribute-vocabulary'\nimport { executionTrackByLane } from '../trace/execution-tracks'\nimport {\n  LLM_INPUT_TOKEN_ATTR_KEYS,\n  LLM_OUTPUT_TOKEN_ATTR_KEYS,\n  TOOL_NAME_ATTR_KEYS,\n} from '../trace/otlp-attributes'\nimport { spanEpochMillis } from './otlp-span'\nimport type { TraceAnalystSpan } from './types'\n\nexport type SuboptimalCode =\n  | 'monotonic-input-growth'\n  | 'output-length-decay'\n  | 'single-tool-dependency'\n  | 'no-self-verification'\n\nexport interface SuboptimalSignal {\n  code: SuboptimalCode\n  severity: 'high' | 'medium' | 'low'\n  /** Human-readable claim, with the backing numbers inlined. */\n  detail: string\n  /** The exact figures the detector fired on — auditable, no model in the loop. */\n  evidence: Record<string, number | string | boolean>\n}\n\nexport interface BehavioralMetrics {\n  /** The only trace represented by these metrics; null when spans are empty. */\n  traceId: string | null\n  llmCallCount: number\n  /** Causally serial LLM timelines. Parallel branches are never joined. */\n  tokenSequences: BehavioralTokenSequence[]\n  /** Token values from the longest serial timeline, retained for convenience. */\n  inputTokenTrajectory: number[]\n  outputTokenTrajectory: number[]\n  toolHistogram: Record<string, number>\n  totalToolCalls: number\n  distinctTools: number\n  /** distinct/total tool calls; 1.0 when there are no tool calls. */\n  toolDiversityRatio: number\n  hasSelfVerification: boolean\n  signals: SuboptimalSignal[]\n}\n\nexport interface BehavioralTokenSequence {\n  scopeId: string\n  spanIds: string[]\n  inputTokenTrajectory: Array<number | null>\n  outputTokenTrajectory: Array<number | null>\n}\n\ninterface TokenSample {\n  span: TraceAnalystSpan\n  input: number | null\n  output: number | null\n  step: number | null\n}\n\n/** ≥ this input-token growth ratio across a run, with no compression, fires. */\nconst INPUT_GROWTH_FACTOR = 3\n/** Tool-usage signals need at least this many calls to be meaningful. */\nconst MIN_TOOL_CALLS = 3\n/** Tool names that read or check state count as self-verification, not mutation.\n *  Covers the inspect verbs plus the read/search tools real harnesses use to\n *  verify (Claude Code Read/Grep/Glob, codex read_file/ls/cat, git status/diff,\n *  test/lint). A pure shell tool (Bash/exec_command) is intentionally NOT matched\n *  — its name can't tell a `pytest` from an `rm`. */\nconst VERIFY_RE =\n  /verif|eval|inspect|check|assert|validat|review|confirm|read|grep|glob|search|view|\\blist\\b|\\bls\\b|\\bcat\\b|\\bfind\\b|diff|status|\\btest|lint|typecheck/i\n\nfunction num(v: unknown): number | null {\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\nfunction numAttr(attrs: Record<string, unknown>, keys: readonly string[]): number | null {\n  for (const key of keys) {\n    const value = num(attrs[key])\n    if (value !== null) return value\n  }\n  return null\n}\nfunction inputTokensOf(s: TraceAnalystSpan): number | null {\n  const exactContext = num(s.attributes[LLM_CONTEXT_TOKENS])\n  if (exactContext !== null) return exactContext\n  // Foreign producers disagree on whether prompt includes cached tokens.\n  // Without an exact total, retain the reported prompt instead of guessing.\n  return (\n    numAttr(s.attributes, LLM_INPUT_TOKEN_ATTR_KEYS) ?? num(s.attributes['llm.usage.input_tokens'])\n  )\n}\nfunction outputTokensOf(s: TraceAnalystSpan): number | null {\n  return (\n    numAttr(s.attributes, LLM_OUTPUT_TOKEN_ATTR_KEYS) ??\n    num(s.attributes['llm.usage.output_tokens'])\n  )\n}\nfunction stepOf(s: TraceAnalystSpan): number | null {\n  return num(s.attributes.step)\n}\nfunction toolNameOf(s: TraceAnalystSpan): string | null {\n  if (s.tool_name) return s.tool_name\n  for (const key of TOOL_NAME_ATTR_KEYS) {\n    const t = s.attributes[key]\n    if (typeof t === 'string' && t.length > 0) return t\n  }\n  return null\n}\n\n/**\n * Reduce a span list to behavioral metrics + fired suboptimality signals.\n * Pure + deterministic: same spans → same output, on any machine, no model.\n */\nexport function computeTraceMetrics(spans: readonly TraceAnalystSpan[]): BehavioralMetrics {\n  const traceIds = new Set(spans.map((span) => span.trace_id))\n  if (traceIds.size > 1) {\n    throw new Error(\n      `computeTraceMetrics: expected spans from one trace, received ${traceIds.size} traces`,\n    )\n  }\n  const traceId = traceIds.values().next().value ?? null\n\n  const samples: TokenSample[] = spans.map((span) => ({\n    span,\n    input: inputTokensOf(span),\n    output: outputTokensOf(span),\n    step: stepOf(span),\n  }))\n  const llmSamples = samples.filter((sample) => sample.span.kind === 'LLM')\n  const tokenSamples =\n    llmSamples.length > 0\n      ? llmSamples\n      : samples.filter((sample) => sample.input !== null || sample.output !== null)\n  const tokenSequences = buildTokenSequences(tokenSamples, spans)\n  const primarySequence = tokenSequences[0]\n  const inputTokenTrajectory =\n    primarySequence?.inputTokenTrajectory.filter((value): value is number => value !== null) ?? []\n  const outputTokenTrajectory =\n    primarySequence?.outputTokenTrajectory.filter((value): value is number => value !== null) ?? []\n  const toolHistogram: Record<string, number> = {}\n  let hasSelfVerification = false\n\n  for (const s of spans) {\n    const tool = toolNameOf(s)\n    if (tool) {\n      toolHistogram[tool] = (toolHistogram[tool] ?? 0) + 1\n      if (VERIFY_RE.test(tool)) hasSelfVerification = true\n    }\n  }\n\n  const totalToolCalls = Object.values(toolHistogram).reduce((a, b) => a + b, 0)\n  const distinctTools = Object.keys(toolHistogram).length\n  const toolDiversityRatio = totalToolCalls === 0 ? 1 : distinctTools / totalToolCalls\n\n  const signals: SuboptimalSignal[] = []\n  const seenTokenSignals = new Set<SuboptimalCode>()\n  for (const sequence of tokenSequences) {\n    for (const signal of tokenSignals(sequence)) {\n      if (seenTokenSignals.has(signal.code)) continue\n      seenTokenSignals.add(signal.code)\n      signals.push(signal)\n    }\n  }\n\n  if (totalToolCalls >= MIN_TOOL_CALLS && distinctTools === 1) {\n    const only = Object.keys(toolHistogram)[0]!\n    signals.push({\n      code: 'single-tool-dependency',\n      severity: 'medium',\n      detail: `All ${totalToolCalls} observed tool calls are \\`${only}\\`; no alternate tool call was observed.`,\n      evidence: { tool: only, calls: totalToolCalls, distinct_tools: 1 },\n    })\n  }\n\n  if (totalToolCalls >= MIN_TOOL_CALLS && !hasSelfVerification) {\n    signals.push({\n      code: 'no-self-verification',\n      severity: 'medium',\n      detail: `${totalToolCalls} tool calls were observed without a verification-named tool call.`,\n      evidence: { tool_calls: totalToolCalls, verification_calls: 0 },\n    })\n  }\n\n  return {\n    traceId,\n    llmCallCount: tokenSamples.length,\n    tokenSequences,\n    inputTokenTrajectory,\n    outputTokenTrajectory,\n    toolHistogram,\n    totalToolCalls,\n    distinctTools,\n    toolDiversityRatio,\n    hasSelfVerification,\n    signals,\n  }\n}\n\nfunction buildTokenSequences(\n  samples: readonly TokenSample[],\n  spans: readonly TraceAnalystSpan[],\n): BehavioralTokenSequence[] {\n  const spansById = new Map(spans.map((span) => [span.span_id, span]))\n  const executionScopeFor = createTokenExecutionScopeResolver(spansById)\n  const scopedSamples = samples.map((sample) => ({\n    sample,\n    ...executionScopeFor(sample.span),\n  }))\n  const trackByLane = executionTrackByLane(scopedSamples)\n  const byTrack = new Map<string, { scopeId: string; samples: TokenSample[] }>()\n  for (const scoped of scopedSamples) {\n    const trackId = trackByLane.get(scoped.key)!\n    const track = byTrack.get(trackId) ?? { scopeId: scoped.scopeId, samples: [] }\n    track.samples.push(scoped.sample)\n    byTrack.set(trackId, track)\n  }\n\n  const sequences: BehavioralTokenSequence[] = []\n  for (const { scopeId, samples: tracked } of byTrack.values()) {\n    const runs = serialTokenRuns([...tracked].sort(compareTokenSamples))\n    runs.forEach((run, index) => {\n      sequences.push({\n        scopeId: runs.length === 1 ? scopeId : `${scopeId}#${index + 1}`,\n        spanIds: run.map((sample) => sample.span.span_id),\n        inputTokenTrajectory: run.map((sample) => sample.input),\n        outputTokenTrajectory: run.map((sample) => sample.output),\n      })\n    })\n  }\n\n  return sequences.sort(\n    (a, b) =>\n      b.spanIds.length - a.spanIds.length ||\n      a.scopeId.localeCompare(b.scopeId) ||\n      a.spanIds[0]!.localeCompare(b.spanIds[0]!),\n  )\n}\n\ninterface AncestorScope {\n  agentId: string | null\n  rootId: string | null\n  missingParentId: string | null\n  laneSpanId: string | null\n}\n\nfunction createTokenExecutionScopeResolver(spansById: ReadonlyMap<string, TraceAnalystSpan>): (\n  span: TraceAnalystSpan,\n) => {\n  key: string\n  scopeKey: string\n  scopeId: string\n  start: number | null\n  end: number | null\n} {\n  const cache = new Map<string, AncestorScope>()\n  return (span) => {\n    const ancestry = resolveAncestorScope(span.parent_span_id, spansById, cache)\n    const rootId = ancestry.rootId\n    const scopeId = ancestry.agentId\n      ? `span:${ancestry.agentId}`\n      : ancestry.missingParentId\n        ? `parent:${ancestry.missingParentId}`\n        : rootId\n          ? `root:${rootId}`\n          : span.agent_name\n            ? `agent:${span.agent_name}`\n            : `trace:${span.trace_id}`\n    const scopeSpanId = ancestry.agentId ?? ancestry.missingParentId ?? rootId\n    const laneSpan = ancestry.laneSpanId ? spansById.get(ancestry.laneSpanId) : undefined\n    const direct = scopeSpanId === null || ancestry.laneSpanId === scopeSpanId\n    const timedSpan = direct ? span : laneSpan\n    return {\n      key: JSON.stringify([scopeId, direct ? span.span_id : ancestry.laneSpanId]),\n      scopeKey: scopeId,\n      scopeId,\n      start: timedSpan ? spanEpochMillis(timedSpan.start_time) : null,\n      end: timedSpan ? spanEpochMillis(timedSpan.end_time) : null,\n    }\n  }\n}\n\nfunction resolveAncestorScope(\n  startId: string | null,\n  spansById: ReadonlyMap<string, TraceAnalystSpan>,\n  cache: Map<string, AncestorScope>,\n): AncestorScope {\n  const empty: AncestorScope = {\n    agentId: null,\n    rootId: null,\n    missingParentId: null,\n    laneSpanId: null,\n  }\n  if (!startId) return empty\n\n  const path: string[] = []\n  const pathIndex = new Map<string, number>()\n  let currentId: string | null = startId\n  let resolved = empty\n  while (currentId) {\n    const cached = cache.get(currentId)\n    if (cached) {\n      resolved = cached\n      break\n    }\n    const cycleStart = pathIndex.get(currentId)\n    if (cycleStart !== undefined) {\n      resolved = {\n        agentId: null,\n        rootId: [...path.slice(cycleStart)].sort()[0]!,\n        missingParentId: null,\n        laneSpanId: [...path.slice(cycleStart)].sort()[0]!,\n      }\n      break\n    }\n    const current = spansById.get(currentId)\n    if (!current) {\n      resolved = {\n        agentId: null,\n        rootId: null,\n        missingParentId: currentId,\n        laneSpanId: currentId,\n      }\n      break\n    }\n    if (current.kind === 'AGENT') {\n      resolved = {\n        agentId: current.span_id,\n        rootId: null,\n        missingParentId: null,\n        laneSpanId: current.span_id,\n      }\n      break\n    }\n    pathIndex.set(currentId, path.length)\n    path.push(currentId)\n    currentId = current.parent_span_id\n  }\n\n  for (let index = path.length - 1; index >= 0; index -= 1) {\n    if (resolved.agentId === null && resolved.rootId === null) {\n      resolved = { ...resolved, rootId: path[index]!, laneSpanId: path[index]! }\n    } else if (\n      resolved.laneSpanId === (resolved.agentId ?? resolved.missingParentId ?? resolved.rootId)\n    ) {\n      resolved = { ...resolved, laneSpanId: path[index]! }\n    }\n    cache.set(path[index]!, resolved)\n  }\n  return resolved\n}\n\nfunction compareTokenSamples(a: TokenSample, b: TokenSample): number {\n  const aStart = spanEpochMillis(a.span.start_time)\n  const bStart = spanEpochMillis(b.span.start_time)\n  if (aStart === null && bStart !== null) return 1\n  if (aStart !== null && bStart === null) return -1\n  if (aStart !== null && bStart !== null && aStart !== bStart) return aStart - bStart\n  if (a.step !== null && b.step !== null && a.step !== b.step) return a.step - b.step\n  return a.span.span_id.localeCompare(b.span.span_id)\n}\n\nfunction serialTokenRuns(ordered: readonly TokenSample[]): TokenSample[][] {\n  const runs: TokenSample[][] = []\n  let serial: TokenSample[] = []\n  let overlap: TokenSample[] = []\n  let overlapEnd = Number.NEGATIVE_INFINITY\n\n  const flushSerial = () => {\n    if (serial.length > 0) runs.push(serial)\n    serial = []\n  }\n  const flushOverlap = () => {\n    if (overlap.length === 1) {\n      serial.push(overlap[0]!)\n    } else if (overlap.length > 1) {\n      flushSerial()\n      for (const sample of overlap) runs.push([sample])\n    }\n    overlap = []\n    overlapEnd = Number.NEGATIVE_INFINITY\n  }\n\n  for (const sample of ordered) {\n    const start = spanEpochMillis(sample.span.start_time)\n    const end = spanEpochMillis(sample.span.end_time)\n    if (start === null || end === null || sample.span.duration_ms <= 0 || end < start) {\n      flushOverlap()\n      flushSerial()\n      runs.push([sample])\n      continue\n    }\n    if (overlap.length > 0 && start >= overlapEnd) flushOverlap()\n    overlap.push(sample)\n    overlapEnd = Math.max(overlapEnd, end)\n  }\n  flushOverlap()\n  flushSerial()\n  return runs\n}\n\nfunction tokenSignals(sequence: BehavioralTokenSequence): SuboptimalSignal[] {\n  const signals: SuboptimalSignal[] = []\n  const inputs = sequence.inputTokenTrajectory\n  const outputs = sequence.outputTokenTrajectory\n  if (inputs.length >= 3 && inputs.every((value): value is number => value !== null)) {\n    const first = inputs[0]!\n    const last = inputs[inputs.length - 1]!\n    const isMonotonic = everyAdjacent(inputs, (previous, current) => current >= previous)\n    const growthFromZero = first === 0 && last > 0\n    const growth = growthFromZero ? Infinity : first > 0 ? last / first : 0\n    if (isMonotonic && last > first && growth >= INPUT_GROWTH_FACTOR) {\n      const growthLabel = growthFromZero ? '0→nonzero (unbounded)' : `${growth.toFixed(1)}x`\n      signals.push({\n        code: 'monotonic-input-growth',\n        severity: 'high',\n        detail: `LLM input tokens grew ${growthLabel} (${first}→${last}) across ${inputs.length} serial calls without an intervening decrease.`,\n        evidence: {\n          first,\n          last,\n          growth_x: growthFromZero ? 'unbounded' : Number(growth.toFixed(2)),\n          calls: inputs.length,\n          scope: sequence.scopeId,\n          first_span_id: sequence.spanIds[0]!,\n          last_span_id: sequence.spanIds[sequence.spanIds.length - 1]!,\n        },\n      })\n    }\n  }\n\n  if (\n    inputs.length >= 3 &&\n    inputs.length === outputs.length &&\n    inputs.every((value): value is number => value !== null) &&\n    outputs.every((value): value is number => value !== null)\n  ) {\n    const first = outputs[0]!\n    const last = outputs[outputs.length - 1]!\n    const inputIsMonotonic = everyAdjacent(inputs, (previous, current) => current >= previous)\n    const outputIsMonotonic = everyAdjacent(outputs, (previous, current) => current <= previous)\n    const inputGrew = inputs[inputs.length - 1]! > inputs[0]!\n    if (inputIsMonotonic && inputGrew && outputIsMonotonic && last < first) {\n      signals.push({\n        code: 'output-length-decay',\n        severity: 'medium',\n        detail: `LLM output tokens shrank ${first}→${last} over ${outputs.length} serial calls while input tokens increased monotonically.`,\n        evidence: {\n          first,\n          last,\n          calls: outputs.length,\n          scope: sequence.scopeId,\n          first_span_id: sequence.spanIds[0]!,\n          last_span_id: sequence.spanIds[sequence.spanIds.length - 1]!,\n        },\n      })\n    }\n  }\n  return signals\n}\n\nfunction everyAdjacent(\n  values: readonly number[],\n  predicate: (previous: number, current: number) => boolean,\n): boolean {\n  return values.slice(1).every((current, index) => predicate(values[index]!, current))\n}\n","/**\n * `behavioralAnalyst` — a DETERMINISTIC analyst (cost.kind = 'deterministic',\n * never calls the LLM). It produces the efficiency/behavioral findings a\n * tolerant agentic analyzer (HALO) re-derives per run inside the model —\n * context bloat, output decay, tool monoculture, missing self-verification —\n * directly from arithmetic over spans (`computeTraceMetrics`).\n *\n * Why it matters: these findings are model-agnostic BY CONSTRUCTION (no model\n * in the loop), so they cannot return 0 on a weak model the way the Ax-RLM\n * does — and they are strictly more reliable than HALO, which spends tokens\n * re-deriving the same numbers and can hallucinate the trend. The agentic\n * RLM kinds remain for SEMANTIC findings that genuinely need a model; this\n * analyst owns the behavioral class.\n */\n\nimport {\n  type BehavioralMetrics,\n  computeTraceMetrics,\n  type SuboptimalCode,\n} from '../trace-analyst/behavioral-metrics'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport { type Analyst, type AnalystFinding, makeFinding } from './types'\n\nconst RECOMMENDED_ACTION: Record<SuboptimalCode, string> = {\n  'monotonic-input-growth':\n    'Inspect context assembly; if prior history is repeatedly included, summarize completed work before the next model call.',\n  'output-length-decay':\n    'Check late-step completeness; if shorter responses omit required work, add explicit completion criteria to the agent instructions.',\n  'single-tool-dependency':\n    'Test whether an inspect or verification tool improves outcomes after the repeated call fails or returns no progress.',\n  'no-self-verification':\n    'After state-changing actions, require an observable check before the agent proceeds.',\n}\n\nconst ANALYST_ID = 'efficiency-behavioral'\n\nconst AGGREGATE_CLAIM: Record<SuboptimalCode, (observed: number, analyzed: number) => string> = {\n  'monotonic-input-growth': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces showed input tokens grow from zero to nonzero or to at least 3x their initial value across at least 3 serial model calls without a decrease.`,\n  'output-length-decay': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces showed output tokens decrease while input tokens increased monotonically across at least 3 serial model calls.`,\n  'single-tool-dependency': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces used only one named tool across at least 3 tool calls.`,\n  'no-self-verification': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces had at least 3 tool calls without a verification-named tool call.`,\n}\n\n/**\n * Map computed signals → structured AnalystFindings. Pure: no LLM, no clock\n * dependence beyond `produced_at` (overridable for deterministic tests).\n */\nexport function deriveEfficiencyFindings(\n  metrics: BehavioralMetrics,\n  opts: { analystId?: string; producedAt?: string } = {},\n): AnalystFinding[] {\n  const analystId = opts.analystId ?? ANALYST_ID\n  const traceId = metrics.traceId\n  return metrics.signals.map((sig) =>\n    makeFinding({\n      analyst_id: analystId,\n      area: 'efficiency',\n      subject: sig.code,\n      claim: sig.detail,\n      severity: sig.severity,\n      // Deterministic arithmetic over spans, not a model judgment → certain.\n      confidence: 1,\n      evidence_refs: [\n        {\n          kind: 'metric',\n          uri: traceId\n            ? `metric://trace/${encodeURIComponent(traceId)}/efficiency/${sig.code}`\n            : `metric://efficiency/${sig.code}`,\n          excerpt: JSON.stringify(sig.evidence),\n        },\n      ],\n      recommended_action: RECOMMENDED_ACTION[sig.code],\n      metadata: {\n        deterministic: true,\n        evidence: sig.evidence,\n        ...(traceId ? { trace_id: traceId } : {}),\n      },\n      id_basis: sig.code,\n      ...(opts.producedAt ? { produced_at: opts.producedAt } : {}),\n    }),\n  )\n}\n\n/** The deterministic behavioral/efficiency analyst (no LLM, any-model). */\nexport function behavioralAnalyst(): Analyst<TraceAnalysisStore> {\n  return {\n    id: ANALYST_ID,\n    description:\n      'Deterministic behavioral/efficiency findings over OTLP spans — token-growth, output-decay, tool-monoculture, missing self-verification. Zero LLM; model-agnostic by construction.',\n    inputKind: 'trace-store',\n    cost: { kind: 'deterministic' },\n    version: '2.0.0',\n    async analyze(store) {\n      const overview = await store.getOverview()\n      const analyzedTraceIds = [...new Set(overview.sample_trace_ids)].sort()\n      const findingsById = new Map<\n        string,\n        { finding: AnalystFinding; traceIds: string[]; evidence: AnalystFinding['evidence_refs'] }\n      >()\n      for (const traceId of analyzedTraceIds) {\n        const viewed = await store.viewTrace({ trace_id: traceId })\n        if (viewed.trace_id !== traceId) {\n          throw new Error(\n            `behavioralAnalyst: requested trace '${traceId}', received '${viewed.trace_id}'`,\n          )\n        }\n        if (!viewed.spans) {\n          throw new Error(\n            `behavioralAnalyst: trace '${traceId}' is oversized; complete spans are required`,\n          )\n        }\n        const metrics = computeTraceMetrics(viewed.spans)\n        if (metrics.traceId !== null && metrics.traceId !== traceId) {\n          throw new Error(\n            `behavioralAnalyst: requested trace '${traceId}', received '${metrics.traceId}'`,\n          )\n        }\n        for (const finding of deriveEfficiencyFindings(metrics)) {\n          const current = findingsById.get(finding.finding_id)\n          if (!current) {\n            findingsById.set(finding.finding_id, {\n              finding,\n              traceIds: [traceId],\n              evidence: [...finding.evidence_refs],\n            })\n            continue\n          }\n          current.traceIds.push(traceId)\n          current.evidence.push(...finding.evidence_refs)\n        }\n      }\n      return [...findingsById.values()].map(({ finding, traceIds, evidence }) => ({\n        ...finding,\n        claim: AGGREGATE_CLAIM[finding.subject as SuboptimalCode](\n          traceIds.length,\n          analyzedTraceIds.length,\n        ),\n        rationale: `${traceIds.length}/${analyzedTraceIds.length} analyzed traces exhibited this pattern.`,\n        evidence_refs: evidence,\n        metadata: {\n          deterministic: true,\n          trace_ids: traceIds,\n          observed_trace_count: traceIds.length,\n          analyzed_trace_count: analyzedTraceIds.length,\n        },\n      }))\n    },\n  }\n}\n","/**\n * `buildDefaultAnalystRegistry` — the canonical analyst suite, so consumers\n * stop hand-wiring `new AnalystRegistry()` + per-kind `createTraceAnalystKind`.\n *\n * The deterministic `behavioralAnalyst` is ALWAYS registered (it needs no\n * model and is model-agnostic by construction). The agentic RLM kinds are\n * registered only when an `ai` service is supplied — so a caller with no LLM\n * still gets the full behavioral/efficiency diagnosis, and the substrate's\n * \"any model (including no model)\" guarantee holds at the suite level.\n */\n\nimport type { AxAIService } from '@ax-llm/ax'\nimport { behavioralAnalyst } from './behavioral-analyst'\nimport { createTraceAnalystKind, type TraceAnalystKindSpec } from './kind-factory'\nimport { DEFAULT_TRACE_ANALYST_KINDS } from './kinds'\nimport { AnalystRegistry, type AnalystRegistryOptions } from './registry'\n\nexport interface DefaultAnalystRegistryOptions {\n  /** Ax service for the agentic RLM kinds. Omit → only the deterministic analyst. */\n  ai?: AxAIService\n  /** Model for the agentic kinds (falls back to the ai service default). */\n  model?: string\n  /** Which agentic kinds to register when `ai` is present. Default = the shipped suite. */\n  kinds?: readonly TraceAnalystKindSpec[]\n  /** Set false to omit the deterministic behavioral analyst (default: include). */\n  includeBehavioral?: boolean\n  /** Forwarded to the AnalystRegistry constructor (signal, tags, priorFindings). */\n  registry?: AnalystRegistryOptions\n}\n\nexport function buildDefaultAnalystRegistry(\n  opts: DefaultAnalystRegistryOptions = {},\n): AnalystRegistry {\n  const registry = new AnalystRegistry(opts.registry)\n  if (opts.includeBehavioral !== false) {\n    registry.register(behavioralAnalyst())\n  }\n  if (opts.ai) {\n    const kinds = opts.kinds ?? DEFAULT_TRACE_ANALYST_KINDS\n    for (const spec of kinds) {\n      registry.register(createTraceAnalystKind(spec, { ai: opts.ai, model: opts.model }))\n    }\n  }\n  return registry\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;AA+HO,SAAS,iBAAiB,MAAwC;AACvE,UAAQ,KAAK,WAAW;AAAA,IACtB,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK,WAAW;AAAA,UACzB,QAAQ,KAAK;AAAA,UACb,YAAY,KAAK;AAAA,QACnB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK,WAAW;AAAA,UACzB,QAAQ,KAAK,UAAU;AAAA,UACvB,YAAY,KAAK;AAAA,QACnB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK;AAAA,UACd,QAAQ,KAAK;AAAA,UACb,YAAY,KAAK;AAAA,QACnB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,WAAW;AAAA,QACX,cAAc,KAAK;AAAA,QACnB,iBAAiB,KAAK;AAAA,QACtB,MAAM,OAAO,KAAK,aAAa,KAAK,KAAK,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;AAAA,MACzF;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,WAAW;AAAA,QACX,cAAc,KAAK;AAAA,QACnB,iBAAiB;AAAA,QACjB,MAAM,OAAO,KAAK,aAAa,KAAK,QAAQ,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;AAAA,MAC5F;AAAA,EACJ;AACF;AAEA,SAAS,cACP,WACA,cACA,OACY;AACZ,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,iBAAiB,MAAM;AAAA,IACvB,MAAM,CAAC,KAAK,aAAa;AACvB,YAAM,WAAW,aAAa,KAAK,YAAY;AAC/C,YAAM,UAA0B;AAAA,QAC9B,OAAO,SAAS;AAAA,QAChB,UAAU,IAAI;AAAA,QACd,UAAU,IAAI;AAAA,QACd,YAAY,IAAI;AAAA,QAChB,aAAa,IAAI;AAAA,QACjB,WAAW,IAAI;AAAA,QACf,WAAW,IAAI;AAAA,MACjB;AACA,aAAO,MAAM,KAAK,SAAS;AAAA,QACzB,QAAQ,UAAU;AAAA,QAClB,gBAAgB,UAAU;AAAA,MAC5B,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAEA,SAAS,aAAa,KAAkB,cAA+C;AACrF,MAAI,IAAI,MAAO,QAAO;AACtB,MAAI,CAAC,cAAc;AACjB,UAAM,IAAI;AAAA,MACR;AAAA,IAEF;AAAA,EACF;AACA,SAAO,EAAE,GAAG,KAAK,OAAO,aAAa;AACvC;;;AC9IA,IAAM,sBAAsB;AAE5B,IAAM,iBAAiB;AAMvB,IAAM,YACJ;AAEF,SAAS,IAAI,GAA2B;AACtC,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AACA,SAAS,QAAQ,OAAgC,MAAwC;AACvF,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,MAAM,GAAG,CAAC;AAC5B,QAAI,UAAU,KAAM,QAAO;AAAA,EAC7B;AACA,SAAO;AACT;AACA,SAAS,cAAc,GAAoC;AACzD,QAAM,eAAe,IAAI,EAAE,WAAW,kBAAkB,CAAC;AACzD,MAAI,iBAAiB,KAAM,QAAO;AAGlC,SACE,QAAQ,EAAE,YAAY,yBAAyB,KAAK,IAAI,EAAE,WAAW,wBAAwB,CAAC;AAElG;AACA,SAAS,eAAe,GAAoC;AAC1D,SACE,QAAQ,EAAE,YAAY,0BAA0B,KAChD,IAAI,EAAE,WAAW,yBAAyB,CAAC;AAE/C;AACA,SAAS,OAAO,GAAoC;AAClD,SAAO,IAAI,EAAE,WAAW,IAAI;AAC9B;AACA,SAAS,WAAW,GAAoC;AACtD,MAAI,EAAE,UAAW,QAAO,EAAE;AAC1B,aAAW,OAAO,qBAAqB;AACrC,UAAM,IAAI,EAAE,WAAW,GAAG;AAC1B,QAAI,OAAO,MAAM,YAAY,EAAE,SAAS,EAAG,QAAO;AAAA,EACpD;AACA,SAAO;AACT;AAMO,SAAS,oBAAoB,OAAuD;AACzF,QAAM,WAAW,IAAI,IAAI,MAAM,IAAI,CAAC,SAAS,KAAK,QAAQ,CAAC;AAC3D,MAAI,SAAS,OAAO,GAAG;AACrB,UAAM,IAAI;AAAA,MACR,gEAAgE,SAAS,IAAI;AAAA,IAC/E;AAAA,EACF;AACA,QAAM,UAAU,SAAS,OAAO,EAAE,KAAK,EAAE,SAAS;AAElD,QAAM,UAAyB,MAAM,IAAI,CAAC,UAAU;AAAA,IAClD;AAAA,IACA,OAAO,cAAc,IAAI;AAAA,IACzB,QAAQ,eAAe,IAAI;AAAA,IAC3B,MAAM,OAAO,IAAI;AAAA,EACnB,EAAE;AACF,QAAM,aAAa,QAAQ,OAAO,CAAC,WAAW,OAAO,KAAK,SAAS,KAAK;AACxE,QAAM,eACJ,WAAW,SAAS,IAChB,aACA,QAAQ,OAAO,CAAC,WAAW,OAAO,UAAU,QAAQ,OAAO,WAAW,IAAI;AAChF,QAAM,iBAAiB,oBAAoB,cAAc,KAAK;AAC9D,QAAM,kBAAkB,eAAe,CAAC;AACxC,QAAM,uBACJ,iBAAiB,qBAAqB,OAAO,CAAC,UAA2B,UAAU,IAAI,KAAK,CAAC;AAC/F,QAAM,wBACJ,iBAAiB,sBAAsB,OAAO,CAAC,UAA2B,UAAU,IAAI,KAAK,CAAC;AAChG,QAAM,gBAAwC,CAAC;AAC/C,MAAI,sBAAsB;AAE1B,aAAW,KAAK,OAAO;AACrB,UAAM,OAAO,WAAW,CAAC;AACzB,QAAI,MAAM;AACR,oBAAc,IAAI,KAAK,cAAc,IAAI,KAAK,KAAK;AACnD,UAAI,UAAU,KAAK,IAAI,EAAG,uBAAsB;AAAA,IAClD;AAAA,EACF;AAEA,QAAM,iBAAiB,OAAO,OAAO,aAAa,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC7E,QAAM,gBAAgB,OAAO,KAAK,aAAa,EAAE;AACjD,QAAM,qBAAqB,mBAAmB,IAAI,IAAI,gBAAgB;AAEtE,QAAM,UAA8B,CAAC;AACrC,QAAM,mBAAmB,oBAAI,IAAoB;AACjD,aAAW,YAAY,gBAAgB;AACrC,eAAW,UAAU,aAAa,QAAQ,GAAG;AAC3C,UAAI,iBAAiB,IAAI,OAAO,IAAI,EAAG;AACvC,uBAAiB,IAAI,OAAO,IAAI;AAChC,cAAQ,KAAK,MAAM;AAAA,IACrB;AAAA,EACF;AAEA,MAAI,kBAAkB,kBAAkB,kBAAkB,GAAG;AAC3D,UAAM,OAAO,OAAO,KAAK,aAAa,EAAE,CAAC;AACzC,YAAQ,KAAK;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,OAAO,cAAc,8BAA8B,IAAI;AAAA,MAC/D,UAAU,EAAE,MAAM,MAAM,OAAO,gBAAgB,gBAAgB,EAAE;AAAA,IACnE,CAAC;AAAA,EACH;AAEA,MAAI,kBAAkB,kBAAkB,CAAC,qBAAqB;AAC5D,YAAQ,KAAK;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,GAAG,cAAc;AAAA,MACzB,UAAU,EAAE,YAAY,gBAAgB,oBAAoB,EAAE;AAAA,IAChE,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL;AAAA,IACA,cAAc,aAAa;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,oBACP,SACA,OAC2B;AAC3B,QAAM,YAAY,IAAI,IAAI,MAAM,IAAI,CAAC,SAAS,CAAC,KAAK,SAAS,IAAI,CAAC,CAAC;AACnE,QAAM,oBAAoB,kCAAkC,SAAS;AACrE,QAAM,gBAAgB,QAAQ,IAAI,CAAC,YAAY;AAAA,IAC7C;AAAA,IACA,GAAG,kBAAkB,OAAO,IAAI;AAAA,EAClC,EAAE;AACF,QAAM,cAAc,qBAAqB,aAAa;AACtD,QAAM,UAAU,oBAAI,IAAyD;AAC7E,aAAW,UAAU,eAAe;AAClC,UAAM,UAAU,YAAY,IAAI,OAAO,GAAG;AAC1C,UAAM,QAAQ,QAAQ,IAAI,OAAO,KAAK,EAAE,SAAS,OAAO,SAAS,SAAS,CAAC,EAAE;AAC7E,UAAM,QAAQ,KAAK,OAAO,MAAM;AAChC,YAAQ,IAAI,SAAS,KAAK;AAAA,EAC5B;AAEA,QAAM,YAAuC,CAAC;AAC9C,aAAW,EAAE,SAAS,SAAS,QAAQ,KAAK,QAAQ,OAAO,GAAG;AAC5D,UAAM,OAAO,gBAAgB,CAAC,GAAG,OAAO,EAAE,KAAK,mBAAmB,CAAC;AACnE,SAAK,QAAQ,CAAC,KAAK,UAAU;AAC3B,gBAAU,KAAK;AAAA,QACb,SAAS,KAAK,WAAW,IAAI,UAAU,GAAG,OAAO,IAAI,QAAQ,CAAC;AAAA,QAC9D,SAAS,IAAI,IAAI,CAAC,WAAW,OAAO,KAAK,OAAO;AAAA,QAChD,sBAAsB,IAAI,IAAI,CAAC,WAAW,OAAO,KAAK;AAAA,QACtD,uBAAuB,IAAI,IAAI,CAAC,WAAW,OAAO,MAAM;AAAA,MAC1D,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AAEA,SAAO,UAAU;AAAA,IACf,CAAC,GAAG,MACF,EAAE,QAAQ,SAAS,EAAE,QAAQ,UAC7B,EAAE,QAAQ,cAAc,EAAE,OAAO,KACjC,EAAE,QAAQ,CAAC,EAAG,cAAc,EAAE,QAAQ,CAAC,CAAE;AAAA,EAC7C;AACF;AASA,SAAS,kCAAkC,WAQzC;AACA,QAAM,QAAQ,oBAAI,IAA2B;AAC7C,SAAO,CAAC,SAAS;AACf,UAAM,WAAW,qBAAqB,KAAK,gBAAgB,WAAW,KAAK;AAC3E,UAAM,SAAS,SAAS;AACxB,UAAM,UAAU,SAAS,UACrB,QAAQ,SAAS,OAAO,KACxB,SAAS,kBACP,UAAU,SAAS,eAAe,KAClC,SACE,QAAQ,MAAM,KACd,KAAK,aACH,SAAS,KAAK,UAAU,KACxB,SAAS,KAAK,QAAQ;AAChC,UAAM,cAAc,SAAS,WAAW,SAAS,mBAAmB;AACpE,UAAM,WAAW,SAAS,aAAa,UAAU,IAAI,SAAS,UAAU,IAAI;AAC5E,UAAM,SAAS,gBAAgB,QAAQ,SAAS,eAAe;AAC/D,UAAM,YAAY,SAAS,OAAO;AAClC,WAAO;AAAA,MACL,KAAK,KAAK,UAAU,CAAC,SAAS,SAAS,KAAK,UAAU,SAAS,UAAU,CAAC;AAAA,MAC1E,UAAU;AAAA,MACV;AAAA,MACA,OAAO,YAAY,gBAAgB,UAAU,UAAU,IAAI;AAAA,MAC3D,KAAK,YAAY,gBAAgB,UAAU,QAAQ,IAAI;AAAA,IACzD;AAAA,EACF;AACF;AAEA,SAAS,qBACP,SACA,WACA,OACe;AACf,QAAM,QAAuB;AAAA,IAC3B,SAAS;AAAA,IACT,QAAQ;AAAA,IACR,iBAAiB;AAAA,IACjB,YAAY;AAAA,EACd;AACA,MAAI,CAAC,QAAS,QAAO;AAErB,QAAM,OAAiB,CAAC;AACxB,QAAM,YAAY,oBAAI,IAAoB;AAC1C,MAAI,YAA2B;AAC/B,MAAI,WAAW;AACf,SAAO,WAAW;AAChB,UAAM,SAAS,MAAM,IAAI,SAAS;AAClC,QAAI,QAAQ;AACV,iBAAW;AACX;AAAA,IACF;AACA,UAAM,aAAa,UAAU,IAAI,SAAS;AAC1C,QAAI,eAAe,QAAW;AAC5B,iBAAW;AAAA,QACT,SAAS;AAAA,QACT,QAAQ,CAAC,GAAG,KAAK,MAAM,UAAU,CAAC,EAAE,KAAK,EAAE,CAAC;AAAA,QAC5C,iBAAiB;AAAA,QACjB,YAAY,CAAC,GAAG,KAAK,MAAM,UAAU,CAAC,EAAE,KAAK,EAAE,CAAC;AAAA,MAClD;AACA;AAAA,IACF;AACA,UAAM,UAAU,UAAU,IAAI,SAAS;AACvC,QAAI,CAAC,SAAS;AACZ,iBAAW;AAAA,QACT,SAAS;AAAA,QACT,QAAQ;AAAA,QACR,iBAAiB;AAAA,QACjB,YAAY;AAAA,MACd;AACA;AAAA,IACF;AACA,QAAI,QAAQ,SAAS,SAAS;AAC5B,iBAAW;AAAA,QACT,SAAS,QAAQ;AAAA,QACjB,QAAQ;AAAA,QACR,iBAAiB;AAAA,QACjB,YAAY,QAAQ;AAAA,MACtB;AACA;AAAA,IACF;AACA,cAAU,IAAI,WAAW,KAAK,MAAM;AACpC,SAAK,KAAK,SAAS;AACnB,gBAAY,QAAQ;AAAA,EACtB;AAEA,WAAS,QAAQ,KAAK,SAAS,GAAG,SAAS,GAAG,SAAS,GAAG;AACxD,QAAI,SAAS,YAAY,QAAQ,SAAS,WAAW,MAAM;AACzD,iBAAW,EAAE,GAAG,UAAU,QAAQ,KAAK,KAAK,GAAI,YAAY,KAAK,KAAK,EAAG;AAAA,IAC3E,WACE,SAAS,gBAAgB,SAAS,WAAW,SAAS,mBAAmB,SAAS,SAClF;AACA,iBAAW,EAAE,GAAG,UAAU,YAAY,KAAK,KAAK,EAAG;AAAA,IACrD;AACA,UAAM,IAAI,KAAK,KAAK,GAAI,QAAQ;AAAA,EAClC;AACA,SAAO;AACT;AAEA,SAAS,oBAAoB,GAAgB,GAAwB;AACnE,QAAM,SAAS,gBAAgB,EAAE,KAAK,UAAU;AAChD,QAAM,SAAS,gBAAgB,EAAE,KAAK,UAAU;AAChD,MAAI,WAAW,QAAQ,WAAW,KAAM,QAAO;AAC/C,MAAI,WAAW,QAAQ,WAAW,KAAM,QAAO;AAC/C,MAAI,WAAW,QAAQ,WAAW,QAAQ,WAAW,OAAQ,QAAO,SAAS;AAC7E,MAAI,EAAE,SAAS,QAAQ,EAAE,SAAS,QAAQ,EAAE,SAAS,EAAE,KAAM,QAAO,EAAE,OAAO,EAAE;AAC/E,SAAO,EAAE,KAAK,QAAQ,cAAc,EAAE,KAAK,OAAO;AACpD;AAEA,SAAS,gBAAgB,SAAkD;AACzE,QAAM,OAAwB,CAAC;AAC/B,MAAI,SAAwB,CAAC;AAC7B,MAAI,UAAyB,CAAC;AAC9B,MAAI,aAAa,OAAO;AAExB,QAAM,cAAc,MAAM;AACxB,QAAI,OAAO,SAAS,EAAG,MAAK,KAAK,MAAM;AACvC,aAAS,CAAC;AAAA,EACZ;AACA,QAAM,eAAe,MAAM;AACzB,QAAI,QAAQ,WAAW,GAAG;AACxB,aAAO,KAAK,QAAQ,CAAC,CAAE;AAAA,IACzB,WAAW,QAAQ,SAAS,GAAG;AAC7B,kBAAY;AACZ,iBAAW,UAAU,QAAS,MAAK,KAAK,CAAC,MAAM,CAAC;AAAA,IAClD;AACA,cAAU,CAAC;AACX,iBAAa,OAAO;AAAA,EACtB;AAEA,aAAW,UAAU,SAAS;AAC5B,UAAM,QAAQ,gBAAgB,OAAO,KAAK,UAAU;AACpD,UAAM,MAAM,gBAAgB,OAAO,KAAK,QAAQ;AAChD,QAAI,UAAU,QAAQ,QAAQ,QAAQ,OAAO,KAAK,eAAe,KAAK,MAAM,OAAO;AACjF,mBAAa;AACb,kBAAY;AACZ,WAAK,KAAK,CAAC,MAAM,CAAC;AAClB;AAAA,IACF;AACA,QAAI,QAAQ,SAAS,KAAK,SAAS,WAAY,cAAa;AAC5D,YAAQ,KAAK,MAAM;AACnB,iBAAa,KAAK,IAAI,YAAY,GAAG;AAAA,EACvC;AACA,eAAa;AACb,cAAY;AACZ,SAAO;AACT;AAEA,SAAS,aAAa,UAAuD;AAC3E,QAAM,UAA8B,CAAC;AACrC,QAAM,SAAS,SAAS;AACxB,QAAM,UAAU,SAAS;AACzB,MAAI,OAAO,UAAU,KAAK,OAAO,MAAM,CAAC,UAA2B,UAAU,IAAI,GAAG;AAClF,UAAM,QAAQ,OAAO,CAAC;AACtB,UAAM,OAAO,OAAO,OAAO,SAAS,CAAC;AACrC,UAAM,cAAc,cAAc,QAAQ,CAAC,UAAU,YAAY,WAAW,QAAQ;AACpF,UAAM,iBAAiB,UAAU,KAAK,OAAO;AAC7C,UAAM,SAAS,iBAAiB,WAAW,QAAQ,IAAI,OAAO,QAAQ;AACtE,QAAI,eAAe,OAAO,SAAS,UAAU,qBAAqB;AAChE,YAAM,cAAc,iBAAiB,+BAA0B,GAAG,OAAO,QAAQ,CAAC,CAAC;AACnF,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ,yBAAyB,WAAW,KAAK,KAAK,SAAI,IAAI,YAAY,OAAO,MAAM;AAAA,QACvF,UAAU;AAAA,UACR;AAAA,UACA;AAAA,UACA,UAAU,iBAAiB,cAAc,OAAO,OAAO,QAAQ,CAAC,CAAC;AAAA,UACjE,OAAO,OAAO;AAAA,UACd,OAAO,SAAS;AAAA,UAChB,eAAe,SAAS,QAAQ,CAAC;AAAA,UACjC,cAAc,SAAS,QAAQ,SAAS,QAAQ,SAAS,CAAC;AAAA,QAC5D;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,MACE,OAAO,UAAU,KACjB,OAAO,WAAW,QAAQ,UAC1B,OAAO,MAAM,CAAC,UAA2B,UAAU,IAAI,KACvD,QAAQ,MAAM,CAAC,UAA2B,UAAU,IAAI,GACxD;AACA,UAAM,QAAQ,QAAQ,CAAC;AACvB,UAAM,OAAO,QAAQ,QAAQ,SAAS,CAAC;AACvC,UAAM,mBAAmB,cAAc,QAAQ,CAAC,UAAU,YAAY,WAAW,QAAQ;AACzF,UAAM,oBAAoB,cAAc,SAAS,CAAC,UAAU,YAAY,WAAW,QAAQ;AAC3F,UAAM,YAAY,OAAO,OAAO,SAAS,CAAC,IAAK,OAAO,CAAC;AACvD,QAAI,oBAAoB,aAAa,qBAAqB,OAAO,OAAO;AACtE,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ,4BAA4B,KAAK,SAAI,IAAI,SAAS,QAAQ,MAAM;AAAA,QACxE,UAAU;AAAA,UACR;AAAA,UACA;AAAA,UACA,OAAO,QAAQ;AAAA,UACf,OAAO,SAAS;AAAA,UAChB,eAAe,SAAS,QAAQ,CAAC;AAAA,UACjC,cAAc,SAAS,QAAQ,SAAS,QAAQ,SAAS,CAAC;AAAA,QAC5D;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,cACP,QACA,WACS;AACT,SAAO,OAAO,MAAM,CAAC,EAAE,MAAM,CAAC,SAAS,UAAU,UAAU,OAAO,KAAK,GAAI,OAAO,CAAC;AACrF;;;ACncA,IAAM,qBAAqD;AAAA,EACzD,0BACE;AAAA,EACF,uBACE;AAAA,EACF,0BACE;AAAA,EACF,wBACE;AACJ;AAEA,IAAM,aAAa;AAEnB,IAAM,kBAA0F;AAAA,EAC9F,0BAA0B,CAAC,UAAU,aACnC,GAAG,QAAQ,IAAI,QAAQ;AAAA,EACzB,uBAAuB,CAAC,UAAU,aAChC,GAAG,QAAQ,IAAI,QAAQ;AAAA,EACzB,0BAA0B,CAAC,UAAU,aACnC,GAAG,QAAQ,IAAI,QAAQ;AAAA,EACzB,wBAAwB,CAAC,UAAU,aACjC,GAAG,QAAQ,IAAI,QAAQ;AAC3B;AAMO,SAAS,yBACd,SACA,OAAoD,CAAC,GACnC;AAClB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,UAAU,QAAQ;AACxB,SAAO,QAAQ,QAAQ;AAAA,IAAI,CAAC,QAC1B,YAAY;AAAA,MACV,YAAY;AAAA,MACZ,MAAM;AAAA,MACN,SAAS,IAAI;AAAA,MACb,OAAO,IAAI;AAAA,MACX,UAAU,IAAI;AAAA;AAAA,MAEd,YAAY;AAAA,MACZ,eAAe;AAAA,QACb;AAAA,UACE,MAAM;AAAA,UACN,KAAK,UACD,kBAAkB,mBAAmB,OAAO,CAAC,eAAe,IAAI,IAAI,KACpE,uBAAuB,IAAI,IAAI;AAAA,UACnC,SAAS,KAAK,UAAU,IAAI,QAAQ;AAAA,QACtC;AAAA,MACF;AAAA,MACA,oBAAoB,mBAAmB,IAAI,IAAI;AAAA,MAC/C,UAAU;AAAA,QACR,eAAe;AAAA,QACf,UAAU,IAAI;AAAA,QACd,GAAI,UAAU,EAAE,UAAU,QAAQ,IAAI,CAAC;AAAA,MACzC;AAAA,MACA,UAAU,IAAI;AAAA,MACd,GAAI,KAAK,aAAa,EAAE,aAAa,KAAK,WAAW,IAAI,CAAC;AAAA,IAC5D,CAAC;AAAA,EACH;AACF;AAGO,SAAS,oBAAiD;AAC/D,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS;AAAA,IACT,MAAM,QAAQ,OAAO;AACnB,YAAM,WAAW,MAAM,MAAM,YAAY;AACzC,YAAM,mBAAmB,CAAC,GAAG,IAAI,IAAI,SAAS,gBAAgB,CAAC,EAAE,KAAK;AACtE,YAAM,eAAe,oBAAI,IAGvB;AACF,iBAAW,WAAW,kBAAkB;AACtC,cAAM,SAAS,MAAM,MAAM,UAAU,EAAE,UAAU,QAAQ,CAAC;AAC1D,YAAI,OAAO,aAAa,SAAS;AAC/B,gBAAM,IAAI;AAAA,YACR,uCAAuC,OAAO,gBAAgB,OAAO,QAAQ;AAAA,UAC/E;AAAA,QACF;AACA,YAAI,CAAC,OAAO,OAAO;AACjB,gBAAM,IAAI;AAAA,YACR,6BAA6B,OAAO;AAAA,UACtC;AAAA,QACF;AACA,cAAM,UAAU,oBAAoB,OAAO,KAAK;AAChD,YAAI,QAAQ,YAAY,QAAQ,QAAQ,YAAY,SAAS;AAC3D,gBAAM,IAAI;AAAA,YACR,uCAAuC,OAAO,gBAAgB,QAAQ,OAAO;AAAA,UAC/E;AAAA,QACF;AACA,mBAAW,WAAW,yBAAyB,OAAO,GAAG;AACvD,gBAAM,UAAU,aAAa,IAAI,QAAQ,UAAU;AACnD,cAAI,CAAC,SAAS;AACZ,yBAAa,IAAI,QAAQ,YAAY;AAAA,cACnC;AAAA,cACA,UAAU,CAAC,OAAO;AAAA,cAClB,UAAU,CAAC,GAAG,QAAQ,aAAa;AAAA,YACrC,CAAC;AACD;AAAA,UACF;AACA,kBAAQ,SAAS,KAAK,OAAO;AAC7B,kBAAQ,SAAS,KAAK,GAAG,QAAQ,aAAa;AAAA,QAChD;AAAA,MACF;AACA,aAAO,CAAC,GAAG,aAAa,OAAO,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS,UAAU,SAAS,OAAO;AAAA,QAC1E,GAAG;AAAA,QACH,OAAO,gBAAgB,QAAQ,OAAyB;AAAA,UACtD,SAAS;AAAA,UACT,iBAAiB;AAAA,QACnB;AAAA,QACA,WAAW,GAAG,SAAS,MAAM,IAAI,iBAAiB,MAAM;AAAA,QACxD,eAAe;AAAA,QACf,UAAU;AAAA,UACR,eAAe;AAAA,UACf,WAAW;AAAA,UACX,sBAAsB,SAAS;AAAA,UAC/B,sBAAsB,iBAAiB;AAAA,QACzC;AAAA,MACF,EAAE;AAAA,IACJ;AAAA,EACF;AACF;;;AC1HO,SAAS,4BACd,OAAsC,CAAC,GACtB;AACjB,QAAM,WAAW,IAAI,gBAAgB,KAAK,QAAQ;AAClD,MAAI,KAAK,sBAAsB,OAAO;AACpC,aAAS,SAAS,kBAAkB,CAAC;AAAA,EACvC;AACA,MAAI,KAAK,IAAI;AACX,UAAM,QAAQ,KAAK,SAAS;AAC5B,eAAW,QAAQ,OAAO;AACxB,eAAS,SAAS,uBAAuB,MAAM,EAAE,IAAI,KAAK,IAAI,OAAO,KAAK,MAAM,CAAC,CAAC;AAAA,IACpF;AAAA,EACF;AACA,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../src/analyst/chat-client.ts","../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/default-registry.ts"],"sourcesContent":["/**\n * ChatClient — the single LLM abstraction analysts call.\n *\n * agent-eval already ships an `LlmClient` (OpenAI-compatible, retry,\n * graceful JSON-schema degrade) and judges that talk to `TCloud`. Two\n * mixed patterns force every analyst author to pick a transport, which\n * couples analyst code to runtime concerns (cli-bridge vs router vs\n * sandbox-sdk) it shouldn't know about.\n *\n * `ChatClient` is one interface every analyst takes via `AnalystContext.chat`.\n * The operator decides at the registry boundary which transport binds\n * to it. Analyst code stays transport-agnostic; swapping production\n * (sandbox-sdk) for local dev (cli-bridge) or tests (mock) is a one-\n * line factory call.\n *\n * Designed to coexist: existing `LlmClient` callers and existing\n * `TCloud`-based judges keep working untouched. New analyst code uses\n * `ChatClient`. When old call sites migrate, they pick up budgeting,\n * cancellation, and unified telemetry for free.\n */\n\nimport {\n  type LlmCallRequest,\n  type LlmCallResult,\n  LlmClient,\n  type LlmClientOptions,\n} from '../llm-client'\n\n/**\n * Unified chat interface. Mirrors LlmCallRequest/Result so the OpenAI-\n * compatible mental model stays. Two methods: a one-shot `chat()` and\n * an `streamChat()` for future agentic loops (not yet exposed).\n */\nexport interface ChatClient {\n  /** Display name of the bound transport — included in telemetry. */\n  readonly transport: ChatTransport\n  /** Default model when caller omits — operators bind this per environment. */\n  readonly defaultModel?: string\n  /** Total provider attempts this transport can make for one chat call. */\n  readonly maximumAttempts?: number\n\n  /** Implementations must enforce `req.maxTokens` when it is present. */\n  chat(req: ChatRequest, opts?: ChatCallOpts): Promise<ChatResponse>\n}\n\nexport type ChatTransport =\n  | 'router' // router.tangle.tools — production paid models\n  | 'sandbox-sdk' // box.streamPrompt() — chat completion via sandbox SDK\n  | 'cli-bridge' // local cli-bridge for dev / local-only runs\n  | 'direct-provider' // direct OpenAI / Anthropic / etc. — bypass router\n  | 'mock' // test-time injection\n\nexport interface ChatRequest extends Omit<LlmCallRequest, 'model'> {\n  /** Optional — falls back to ChatClient.defaultModel. */\n  model?: string\n}\n\nexport type ChatResponse = LlmCallResult\n\nexport interface ChatCallOpts {\n  /** Cancel the in-flight request. */\n  signal?: AbortSignal\n  /** Hard USD ceiling for this single call (informational; the underlying transport may not enforce). */\n  maxCostUsd?: number\n  /** Correlation tag carried into request headers when the transport allows. */\n  correlationId?: string\n  /** Stable provider idempotency key for retries/redrives of one paid call. */\n  idempotencyKey?: string\n}\n\n// ── Factory ─────────────────────────────────────────────────────────\n\nexport type CreateChatClientOpts =\n  | RouterTransportOpts\n  | CliBridgeTransportOpts\n  | DirectProviderTransportOpts\n  | SandboxSdkTransportOpts\n  | MockTransportOpts\n\ninterface BaseTransportOpts {\n  defaultModel?: string\n  /** Total provider attempts. Required for opaque transports used in capped runs. */\n  maximumAttempts?: number\n}\n\nexport interface RouterTransportOpts extends BaseTransportOpts {\n  transport: 'router'\n  baseUrl?: string\n  apiKey: string\n}\n\nexport interface CliBridgeTransportOpts extends BaseTransportOpts {\n  transport: 'cli-bridge'\n  baseUrl?: string\n  bearer?: string\n}\n\nexport interface DirectProviderTransportOpts extends BaseTransportOpts {\n  transport: 'direct-provider'\n  baseUrl: string\n  apiKey: string\n}\n\n/**\n * Sandbox-SDK transport. Provided as a thin pass-through: the caller\n * supplies a callable that mimics LlmClient.chat() against an already-\n * configured Sandbox handle. We don't import the SDK here to keep\n * agent-eval dep-free of @tangle-network/sandbox.\n */\nexport interface SandboxSdkTransportOpts extends BaseTransportOpts {\n  transport: 'sandbox-sdk'\n  chat: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Mock transport for tests. The handler receives the request and returns\n * whatever the test wants. No retries, no JSON-schema degrade.\n */\nexport interface MockTransportOpts extends BaseTransportOpts {\n  transport: 'mock'\n  handler: (req: ChatRequest, opts?: ChatCallOpts) => Promise<ChatResponse>\n}\n\n/**\n * Build a ChatClient bound to a specific transport. The returned client\n * is safe to share across analysts in a single registry run.\n */\nexport function createChatClient(opts: CreateChatClientOpts): ChatClient {\n  switch (opts.transport) {\n    case 'router':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl ?? 'https://router.tangle.tools/v1',\n          apiKey: opts.apiKey,\n          maxRetries: opts.maximumAttempts,\n        } as LlmClientOptions),\n      )\n    case 'cli-bridge':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl ?? 'http://127.0.0.1:3344/v1',\n          apiKey: opts.bearer ?? '',\n          maxRetries: opts.maximumAttempts,\n        } as LlmClientOptions),\n      )\n    case 'direct-provider':\n      return wrapLlmClient(\n        opts.transport,\n        opts.defaultModel,\n        new LlmClient({\n          baseUrl: opts.baseUrl,\n          apiKey: opts.apiKey,\n          maxRetries: opts.maximumAttempts,\n        } as LlmClientOptions),\n      )\n    case 'sandbox-sdk':\n      return {\n        transport: 'sandbox-sdk',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: opts.maximumAttempts,\n        chat: async (req, callOpts) => opts.chat(resolveModel(req, opts.defaultModel), callOpts),\n      }\n    case 'mock':\n      return {\n        transport: 'mock',\n        defaultModel: opts.defaultModel,\n        maximumAttempts: 1,\n        chat: async (req, callOpts) => opts.handler(resolveModel(req, opts.defaultModel), callOpts),\n      }\n  }\n}\n\nfunction wrapLlmClient(\n  transport: ChatTransport,\n  defaultModel: string | undefined,\n  inner: LlmClient,\n): ChatClient {\n  return {\n    transport,\n    defaultModel,\n    maximumAttempts: inner.maximumAttempts,\n    chat: (req, callOpts) => {\n      const resolved = resolveModel(req, defaultModel)\n      const request: LlmCallRequest = {\n        model: resolved.model!,\n        messages: req.messages,\n        jsonMode: req.jsonMode,\n        jsonSchema: req.jsonSchema,\n        temperature: req.temperature,\n        maxTokens: req.maxTokens,\n        timeoutMs: req.timeoutMs,\n      }\n      return inner.call(request, {\n        signal: callOpts?.signal,\n        idempotencyKey: callOpts?.idempotencyKey,\n      })\n    },\n  }\n}\n\nfunction resolveModel(req: ChatRequest, defaultModel: string | undefined): ChatRequest {\n  if (req.model) return req\n  if (!defaultModel) {\n    throw new Error(\n      'ChatClient.chat: no model on request and no defaultModel on the client. ' +\n        'Either pass req.model or bind defaultModel at createChatClient().',\n    )\n  }\n  return { ...req, model: defaultModel }\n}\n","/**\n * Deterministic behavioral metrics over OTLP spans — pure arithmetic, no LLM.\n *\n * These are the model-independent multiplier: the four trace-quality signals a\n * tolerant analyzer (e.g. HALO) re-derives per run inside the model — token\n * growth, output decay, tool monoculture, missing self-verification — computed\n * here once, in TypeScript, with zero model judgment. A finding that falls out\n * of arithmetic is trivially model-agnostic and cannot hallucinate the trend.\n *\n * General, not trace-specific: the detectors key off token trajectories and\n * tool usage present in any agentic OTLP trace, not any one benchmark.\n */\n\nimport { LLM_CONTEXT_TOKENS } from '../trace/attribute-vocabulary'\nimport { executionTrackByLane } from '../trace/execution-tracks'\nimport {\n  LLM_INPUT_TOKEN_ATTR_KEYS,\n  LLM_OUTPUT_TOKEN_ATTR_KEYS,\n  TOOL_NAME_ATTR_KEYS,\n} from '../trace/otlp-attributes'\nimport { spanEpochMillis } from './otlp-span'\nimport type { TraceAnalystSpan } from './types'\n\nexport type SuboptimalCode =\n  | 'monotonic-input-growth'\n  | 'output-length-decay'\n  | 'single-tool-dependency'\n  | 'no-self-verification'\n\nexport interface SuboptimalSignal {\n  code: SuboptimalCode\n  severity: 'high' | 'medium' | 'low'\n  /** Human-readable claim, with the backing numbers inlined. */\n  detail: string\n  /** The exact figures the detector fired on — auditable, no model in the loop. */\n  evidence: Record<string, number | string | boolean>\n}\n\nexport interface BehavioralMetrics {\n  /** The only trace represented by these metrics; null when spans are empty. */\n  traceId: string | null\n  llmCallCount: number\n  /** Causally serial LLM timelines. Parallel branches are never joined. */\n  tokenSequences: BehavioralTokenSequence[]\n  /** Token values from the longest serial timeline, retained for convenience. */\n  inputTokenTrajectory: number[]\n  outputTokenTrajectory: number[]\n  toolHistogram: Record<string, number>\n  totalToolCalls: number\n  distinctTools: number\n  /** distinct/total tool calls; 1.0 when there are no tool calls. */\n  toolDiversityRatio: number\n  hasSelfVerification: boolean\n  signals: SuboptimalSignal[]\n}\n\nexport interface BehavioralTokenSequence {\n  scopeId: string\n  spanIds: string[]\n  inputTokenTrajectory: Array<number | null>\n  outputTokenTrajectory: Array<number | null>\n}\n\ninterface TokenSample {\n  span: TraceAnalystSpan\n  input: number | null\n  output: number | null\n  step: number | null\n}\n\n/** ≥ this input-token growth ratio across a run, with no compression, fires. */\nconst INPUT_GROWTH_FACTOR = 3\n/** Tool-usage signals need at least this many calls to be meaningful. */\nconst MIN_TOOL_CALLS = 3\n/** Tool names that read or check state count as self-verification, not mutation.\n *  Covers the inspect verbs plus the read/search tools real harnesses use to\n *  verify (Claude Code Read/Grep/Glob, codex read_file/ls/cat, git status/diff,\n *  test/lint). A pure shell tool (Bash/exec_command) is intentionally NOT matched\n *  — its name can't tell a `pytest` from an `rm`. */\nconst VERIFY_RE =\n  /verif|eval|inspect|check|assert|validat|review|confirm|read|grep|glob|search|view|\\blist\\b|\\bls\\b|\\bcat\\b|\\bfind\\b|diff|status|\\btest|lint|typecheck/i\n\nfunction num(v: unknown): number | null {\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\nfunction numAttr(attrs: Record<string, unknown>, keys: readonly string[]): number | null {\n  for (const key of keys) {\n    const value = num(attrs[key])\n    if (value !== null) return value\n  }\n  return null\n}\nfunction inputTokensOf(s: TraceAnalystSpan): number | null {\n  const exactContext = num(s.attributes[LLM_CONTEXT_TOKENS])\n  if (exactContext !== null) return exactContext\n  // Foreign producers disagree on whether prompt includes cached tokens.\n  // Without an exact total, retain the reported prompt instead of guessing.\n  return (\n    numAttr(s.attributes, LLM_INPUT_TOKEN_ATTR_KEYS) ?? num(s.attributes['llm.usage.input_tokens'])\n  )\n}\nfunction outputTokensOf(s: TraceAnalystSpan): number | null {\n  return (\n    numAttr(s.attributes, LLM_OUTPUT_TOKEN_ATTR_KEYS) ??\n    num(s.attributes['llm.usage.output_tokens'])\n  )\n}\nfunction stepOf(s: TraceAnalystSpan): number | null {\n  return num(s.attributes.step)\n}\nfunction toolNameOf(s: TraceAnalystSpan): string | null {\n  if (s.tool_name) return s.tool_name\n  for (const key of TOOL_NAME_ATTR_KEYS) {\n    const t = s.attributes[key]\n    if (typeof t === 'string' && t.length > 0) return t\n  }\n  return null\n}\n\n/**\n * Reduce a span list to behavioral metrics + fired suboptimality signals.\n * Pure + deterministic: same spans → same output, on any machine, no model.\n */\nexport function computeTraceMetrics(spans: readonly TraceAnalystSpan[]): BehavioralMetrics {\n  const traceIds = new Set(spans.map((span) => span.trace_id))\n  if (traceIds.size > 1) {\n    throw new Error(\n      `computeTraceMetrics: expected spans from one trace, received ${traceIds.size} traces`,\n    )\n  }\n  const traceId = traceIds.values().next().value ?? null\n\n  const samples: TokenSample[] = spans.map((span) => ({\n    span,\n    input: inputTokensOf(span),\n    output: outputTokensOf(span),\n    step: stepOf(span),\n  }))\n  const llmSamples = samples.filter((sample) => sample.span.kind === 'LLM')\n  const tokenSamples =\n    llmSamples.length > 0\n      ? llmSamples\n      : samples.filter((sample) => sample.input !== null || sample.output !== null)\n  const tokenSequences = buildTokenSequences(tokenSamples, spans)\n  const primarySequence = tokenSequences[0]\n  const inputTokenTrajectory =\n    primarySequence?.inputTokenTrajectory.filter((value): value is number => value !== null) ?? []\n  const outputTokenTrajectory =\n    primarySequence?.outputTokenTrajectory.filter((value): value is number => value !== null) ?? []\n  const toolHistogram: Record<string, number> = {}\n  let hasSelfVerification = false\n\n  for (const s of spans) {\n    const tool = toolNameOf(s)\n    if (tool) {\n      toolHistogram[tool] = (toolHistogram[tool] ?? 0) + 1\n      if (VERIFY_RE.test(tool)) hasSelfVerification = true\n    }\n  }\n\n  const totalToolCalls = Object.values(toolHistogram).reduce((a, b) => a + b, 0)\n  const distinctTools = Object.keys(toolHistogram).length\n  const toolDiversityRatio = totalToolCalls === 0 ? 1 : distinctTools / totalToolCalls\n\n  const signals: SuboptimalSignal[] = []\n  const seenTokenSignals = new Set<SuboptimalCode>()\n  for (const sequence of tokenSequences) {\n    for (const signal of tokenSignals(sequence)) {\n      if (seenTokenSignals.has(signal.code)) continue\n      seenTokenSignals.add(signal.code)\n      signals.push(signal)\n    }\n  }\n\n  if (totalToolCalls >= MIN_TOOL_CALLS && distinctTools === 1) {\n    const only = Object.keys(toolHistogram)[0]!\n    signals.push({\n      code: 'single-tool-dependency',\n      severity: 'medium',\n      detail: `All ${totalToolCalls} observed tool calls are \\`${only}\\`; no alternate tool call was observed.`,\n      evidence: { tool: only, calls: totalToolCalls, distinct_tools: 1 },\n    })\n  }\n\n  if (totalToolCalls >= MIN_TOOL_CALLS && !hasSelfVerification) {\n    signals.push({\n      code: 'no-self-verification',\n      severity: 'medium',\n      detail: `${totalToolCalls} tool calls were observed without a verification-named tool call.`,\n      evidence: { tool_calls: totalToolCalls, verification_calls: 0 },\n    })\n  }\n\n  return {\n    traceId,\n    llmCallCount: tokenSamples.length,\n    tokenSequences,\n    inputTokenTrajectory,\n    outputTokenTrajectory,\n    toolHistogram,\n    totalToolCalls,\n    distinctTools,\n    toolDiversityRatio,\n    hasSelfVerification,\n    signals,\n  }\n}\n\nfunction buildTokenSequences(\n  samples: readonly TokenSample[],\n  spans: readonly TraceAnalystSpan[],\n): BehavioralTokenSequence[] {\n  const spansById = new Map(spans.map((span) => [span.span_id, span]))\n  const executionScopeFor = createTokenExecutionScopeResolver(spansById)\n  const scopedSamples = samples.map((sample) => ({\n    sample,\n    ...executionScopeFor(sample.span),\n  }))\n  const trackByLane = executionTrackByLane(scopedSamples)\n  const byTrack = new Map<string, { scopeId: string; samples: TokenSample[] }>()\n  for (const scoped of scopedSamples) {\n    const trackId = trackByLane.get(scoped.key)!\n    const track = byTrack.get(trackId) ?? { scopeId: scoped.scopeId, samples: [] }\n    track.samples.push(scoped.sample)\n    byTrack.set(trackId, track)\n  }\n\n  const sequences: BehavioralTokenSequence[] = []\n  for (const { scopeId, samples: tracked } of byTrack.values()) {\n    const runs = serialTokenRuns([...tracked].sort(compareTokenSamples))\n    runs.forEach((run, index) => {\n      sequences.push({\n        scopeId: runs.length === 1 ? scopeId : `${scopeId}#${index + 1}`,\n        spanIds: run.map((sample) => sample.span.span_id),\n        inputTokenTrajectory: run.map((sample) => sample.input),\n        outputTokenTrajectory: run.map((sample) => sample.output),\n      })\n    })\n  }\n\n  return sequences.sort(\n    (a, b) =>\n      b.spanIds.length - a.spanIds.length ||\n      a.scopeId.localeCompare(b.scopeId) ||\n      a.spanIds[0]!.localeCompare(b.spanIds[0]!),\n  )\n}\n\ninterface AncestorScope {\n  agentId: string | null\n  rootId: string | null\n  missingParentId: string | null\n  laneSpanId: string | null\n}\n\nfunction createTokenExecutionScopeResolver(spansById: ReadonlyMap<string, TraceAnalystSpan>): (\n  span: TraceAnalystSpan,\n) => {\n  key: string\n  scopeKey: string\n  scopeId: string\n  start: number | null\n  end: number | null\n} {\n  const cache = new Map<string, AncestorScope>()\n  return (span) => {\n    const ancestry = resolveAncestorScope(span.parent_span_id, spansById, cache)\n    const rootId = ancestry.rootId\n    const scopeId = ancestry.agentId\n      ? `span:${ancestry.agentId}`\n      : ancestry.missingParentId\n        ? `parent:${ancestry.missingParentId}`\n        : rootId\n          ? `root:${rootId}`\n          : span.agent_name\n            ? `agent:${span.agent_name}`\n            : `trace:${span.trace_id}`\n    const scopeSpanId = ancestry.agentId ?? ancestry.missingParentId ?? rootId\n    const laneSpan = ancestry.laneSpanId ? spansById.get(ancestry.laneSpanId) : undefined\n    const direct = scopeSpanId === null || ancestry.laneSpanId === scopeSpanId\n    const timedSpan = direct ? span : laneSpan\n    return {\n      key: JSON.stringify([scopeId, direct ? span.span_id : ancestry.laneSpanId]),\n      scopeKey: scopeId,\n      scopeId,\n      start: timedSpan ? spanEpochMillis(timedSpan.start_time) : null,\n      end: timedSpan ? spanEpochMillis(timedSpan.end_time) : null,\n    }\n  }\n}\n\nfunction resolveAncestorScope(\n  startId: string | null,\n  spansById: ReadonlyMap<string, TraceAnalystSpan>,\n  cache: Map<string, AncestorScope>,\n): AncestorScope {\n  const empty: AncestorScope = {\n    agentId: null,\n    rootId: null,\n    missingParentId: null,\n    laneSpanId: null,\n  }\n  if (!startId) return empty\n\n  const path: string[] = []\n  const pathIndex = new Map<string, number>()\n  let currentId: string | null = startId\n  let resolved = empty\n  while (currentId) {\n    const cached = cache.get(currentId)\n    if (cached) {\n      resolved = cached\n      break\n    }\n    const cycleStart = pathIndex.get(currentId)\n    if (cycleStart !== undefined) {\n      resolved = {\n        agentId: null,\n        rootId: [...path.slice(cycleStart)].sort()[0]!,\n        missingParentId: null,\n        laneSpanId: [...path.slice(cycleStart)].sort()[0]!,\n      }\n      break\n    }\n    const current = spansById.get(currentId)\n    if (!current) {\n      resolved = {\n        agentId: null,\n        rootId: null,\n        missingParentId: currentId,\n        laneSpanId: currentId,\n      }\n      break\n    }\n    if (current.kind === 'AGENT') {\n      resolved = {\n        agentId: current.span_id,\n        rootId: null,\n        missingParentId: null,\n        laneSpanId: current.span_id,\n      }\n      break\n    }\n    pathIndex.set(currentId, path.length)\n    path.push(currentId)\n    currentId = current.parent_span_id\n  }\n\n  for (let index = path.length - 1; index >= 0; index -= 1) {\n    if (resolved.agentId === null && resolved.rootId === null) {\n      resolved = { ...resolved, rootId: path[index]!, laneSpanId: path[index]! }\n    } else if (\n      resolved.laneSpanId === (resolved.agentId ?? resolved.missingParentId ?? resolved.rootId)\n    ) {\n      resolved = { ...resolved, laneSpanId: path[index]! }\n    }\n    cache.set(path[index]!, resolved)\n  }\n  return resolved\n}\n\nfunction compareTokenSamples(a: TokenSample, b: TokenSample): number {\n  const aStart = spanEpochMillis(a.span.start_time)\n  const bStart = spanEpochMillis(b.span.start_time)\n  if (aStart === null && bStart !== null) return 1\n  if (aStart !== null && bStart === null) return -1\n  if (aStart !== null && bStart !== null && aStart !== bStart) return aStart - bStart\n  if (a.step !== null && b.step !== null && a.step !== b.step) return a.step - b.step\n  return a.span.span_id.localeCompare(b.span.span_id)\n}\n\nfunction serialTokenRuns(ordered: readonly TokenSample[]): TokenSample[][] {\n  const runs: TokenSample[][] = []\n  let serial: TokenSample[] = []\n  let overlap: TokenSample[] = []\n  let overlapEnd = Number.NEGATIVE_INFINITY\n\n  const flushSerial = () => {\n    if (serial.length > 0) runs.push(serial)\n    serial = []\n  }\n  const flushOverlap = () => {\n    if (overlap.length === 1) {\n      serial.push(overlap[0]!)\n    } else if (overlap.length > 1) {\n      flushSerial()\n      for (const sample of overlap) runs.push([sample])\n    }\n    overlap = []\n    overlapEnd = Number.NEGATIVE_INFINITY\n  }\n\n  for (const sample of ordered) {\n    const start = spanEpochMillis(sample.span.start_time)\n    const end = spanEpochMillis(sample.span.end_time)\n    if (start === null || end === null || sample.span.duration_ms <= 0 || end < start) {\n      flushOverlap()\n      flushSerial()\n      runs.push([sample])\n      continue\n    }\n    if (overlap.length > 0 && start >= overlapEnd) flushOverlap()\n    overlap.push(sample)\n    overlapEnd = Math.max(overlapEnd, end)\n  }\n  flushOverlap()\n  flushSerial()\n  return runs\n}\n\nfunction tokenSignals(sequence: BehavioralTokenSequence): SuboptimalSignal[] {\n  const signals: SuboptimalSignal[] = []\n  const inputs = sequence.inputTokenTrajectory\n  const outputs = sequence.outputTokenTrajectory\n  if (inputs.length >= 3 && inputs.every((value): value is number => value !== null)) {\n    const first = inputs[0]!\n    const last = inputs[inputs.length - 1]!\n    const isMonotonic = everyAdjacent(inputs, (previous, current) => current >= previous)\n    const growthFromZero = first === 0 && last > 0\n    const growth = growthFromZero ? Infinity : first > 0 ? last / first : 0\n    if (isMonotonic && last > first && growth >= INPUT_GROWTH_FACTOR) {\n      const growthLabel = growthFromZero ? '0→nonzero (unbounded)' : `${growth.toFixed(1)}x`\n      signals.push({\n        code: 'monotonic-input-growth',\n        severity: 'high',\n        detail: `LLM input tokens grew ${growthLabel} (${first}→${last}) across ${inputs.length} serial calls without an intervening decrease.`,\n        evidence: {\n          first,\n          last,\n          growth_x: growthFromZero ? 'unbounded' : Number(growth.toFixed(2)),\n          calls: inputs.length,\n          scope: sequence.scopeId,\n          first_span_id: sequence.spanIds[0]!,\n          last_span_id: sequence.spanIds[sequence.spanIds.length - 1]!,\n        },\n      })\n    }\n  }\n\n  if (\n    inputs.length >= 3 &&\n    inputs.length === outputs.length &&\n    inputs.every((value): value is number => value !== null) &&\n    outputs.every((value): value is number => value !== null)\n  ) {\n    const first = outputs[0]!\n    const last = outputs[outputs.length - 1]!\n    const inputIsMonotonic = everyAdjacent(inputs, (previous, current) => current >= previous)\n    const outputIsMonotonic = everyAdjacent(outputs, (previous, current) => current <= previous)\n    const inputGrew = inputs[inputs.length - 1]! > inputs[0]!\n    if (inputIsMonotonic && inputGrew && outputIsMonotonic && last < first) {\n      signals.push({\n        code: 'output-length-decay',\n        severity: 'medium',\n        detail: `LLM output tokens shrank ${first}→${last} over ${outputs.length} serial calls while input tokens increased monotonically.`,\n        evidence: {\n          first,\n          last,\n          calls: outputs.length,\n          scope: sequence.scopeId,\n          first_span_id: sequence.spanIds[0]!,\n          last_span_id: sequence.spanIds[sequence.spanIds.length - 1]!,\n        },\n      })\n    }\n  }\n  return signals\n}\n\nfunction everyAdjacent(\n  values: readonly number[],\n  predicate: (previous: number, current: number) => boolean,\n): boolean {\n  return values.slice(1).every((current, index) => predicate(values[index]!, current))\n}\n","/**\n * `behavioralAnalyst` — a DETERMINISTIC analyst (cost.kind = 'deterministic',\n * never calls the LLM). It produces the efficiency/behavioral findings a\n * tolerant agentic analyzer (HALO) re-derives per run inside the model —\n * context bloat, output decay, tool monoculture, missing self-verification —\n * directly from arithmetic over spans (`computeTraceMetrics`).\n *\n * Why it matters: these findings are model-agnostic BY CONSTRUCTION (no model\n * in the loop), so they cannot return 0 on a weak model the way the Ax-RLM\n * does — and they are strictly more reliable than HALO, which spends tokens\n * re-deriving the same numbers and can hallucinate the trend. The agentic\n * RLM kinds remain for SEMANTIC findings that genuinely need a model; this\n * analyst owns the behavioral class.\n */\n\nimport {\n  type BehavioralMetrics,\n  computeTraceMetrics,\n  type SuboptimalCode,\n} from '../trace-analyst/behavioral-metrics'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport { type Analyst, type AnalystFinding, makeFinding } from './types'\n\nconst RECOMMENDED_ACTION: Record<SuboptimalCode, string> = {\n  'monotonic-input-growth':\n    'Inspect context assembly; if prior history is repeatedly included, summarize completed work before the next model call.',\n  'output-length-decay':\n    'Check late-step completeness; if shorter responses omit required work, add explicit completion criteria to the agent instructions.',\n  'single-tool-dependency':\n    'Test whether an inspect or verification tool improves outcomes after the repeated call fails or returns no progress.',\n  'no-self-verification':\n    'After state-changing actions, require an observable check before the agent proceeds.',\n}\n\nconst ANALYST_ID = 'efficiency-behavioral'\n\nconst AGGREGATE_CLAIM: Record<SuboptimalCode, (observed: number, analyzed: number) => string> = {\n  'monotonic-input-growth': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces showed input tokens grow from zero to nonzero or to at least 3x their initial value across at least 3 serial model calls without a decrease.`,\n  'output-length-decay': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces showed output tokens decrease while input tokens increased monotonically across at least 3 serial model calls.`,\n  'single-tool-dependency': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces used only one named tool across at least 3 tool calls.`,\n  'no-self-verification': (observed, analyzed) =>\n    `${observed}/${analyzed} analyzed traces had at least 3 tool calls without a verification-named tool call.`,\n}\n\n/**\n * Map computed signals → structured AnalystFindings. Pure: no LLM, no clock\n * dependence beyond `produced_at` (overridable for deterministic tests).\n */\nexport function deriveEfficiencyFindings(\n  metrics: BehavioralMetrics,\n  opts: { analystId?: string; producedAt?: string } = {},\n): AnalystFinding[] {\n  const analystId = opts.analystId ?? ANALYST_ID\n  const traceId = metrics.traceId\n  return metrics.signals.map((sig) =>\n    makeFinding({\n      analyst_id: analystId,\n      area: 'efficiency',\n      subject: sig.code,\n      claim: sig.detail,\n      severity: sig.severity,\n      // Deterministic arithmetic over spans, not a model judgment → certain.\n      confidence: 1,\n      evidence_refs: [\n        {\n          kind: 'metric',\n          uri: traceId\n            ? `metric://trace/${encodeURIComponent(traceId)}/efficiency/${sig.code}`\n            : `metric://efficiency/${sig.code}`,\n          excerpt: JSON.stringify(sig.evidence),\n        },\n      ],\n      recommended_action: RECOMMENDED_ACTION[sig.code],\n      metadata: {\n        deterministic: true,\n        evidence: sig.evidence,\n        ...(traceId ? { trace_id: traceId } : {}),\n      },\n      id_basis: sig.code,\n      ...(opts.producedAt ? { produced_at: opts.producedAt } : {}),\n    }),\n  )\n}\n\n/** The deterministic behavioral/efficiency analyst (no LLM, any-model). */\nexport function behavioralAnalyst(): Analyst<TraceAnalysisStore> {\n  return {\n    id: ANALYST_ID,\n    description:\n      'Deterministic behavioral/efficiency findings over OTLP spans — token-growth, output-decay, tool-monoculture, missing self-verification. Zero LLM; model-agnostic by construction.',\n    inputKind: 'trace-store',\n    cost: { kind: 'deterministic' },\n    version: '2.0.0',\n    async analyze(store) {\n      const overview = await store.getOverview()\n      const analyzedTraceIds = [...new Set(overview.sample_trace_ids)].sort()\n      const findingsById = new Map<\n        string,\n        { finding: AnalystFinding; traceIds: string[]; evidence: AnalystFinding['evidence_refs'] }\n      >()\n      for (const traceId of analyzedTraceIds) {\n        const viewed = await store.viewTrace({ trace_id: traceId })\n        if (viewed.trace_id !== traceId) {\n          throw new Error(\n            `behavioralAnalyst: requested trace '${traceId}', received '${viewed.trace_id}'`,\n          )\n        }\n        if (!viewed.spans) {\n          throw new Error(\n            `behavioralAnalyst: trace '${traceId}' is oversized; complete spans are required`,\n          )\n        }\n        const metrics = computeTraceMetrics(viewed.spans)\n        if (metrics.traceId !== null && metrics.traceId !== traceId) {\n          throw new Error(\n            `behavioralAnalyst: requested trace '${traceId}', received '${metrics.traceId}'`,\n          )\n        }\n        for (const finding of deriveEfficiencyFindings(metrics)) {\n          const current = findingsById.get(finding.finding_id)\n          if (!current) {\n            findingsById.set(finding.finding_id, {\n              finding,\n              traceIds: [traceId],\n              evidence: [...finding.evidence_refs],\n            })\n            continue\n          }\n          current.traceIds.push(traceId)\n          current.evidence.push(...finding.evidence_refs)\n        }\n      }\n      return [...findingsById.values()].map(({ finding, traceIds, evidence }) => ({\n        ...finding,\n        claim: AGGREGATE_CLAIM[finding.subject as SuboptimalCode](\n          traceIds.length,\n          analyzedTraceIds.length,\n        ),\n        rationale: `${traceIds.length}/${analyzedTraceIds.length} analyzed traces exhibited this pattern.`,\n        evidence_refs: evidence,\n        metadata: {\n          deterministic: true,\n          trace_ids: traceIds,\n          observed_trace_count: traceIds.length,\n          analyzed_trace_count: analyzedTraceIds.length,\n        },\n      }))\n    },\n  }\n}\n","/**\n * `buildDefaultAnalystRegistry` — the canonical analyst suite, so consumers\n * stop hand-wiring `new AnalystRegistry()` + per-kind `createTraceAnalystKind`.\n *\n * The deterministic `behavioralAnalyst` is ALWAYS registered (it needs no\n * model and is model-agnostic by construction). The agentic RLM kinds are\n * registered only when an `ai` service is supplied — so a caller with no LLM\n * still gets the full behavioral/efficiency diagnosis, and the substrate's\n * \"any model (including no model)\" guarantee holds at the suite level.\n */\n\nimport type { AxAIService } from '@ax-llm/ax'\nimport { behavioralAnalyst } from './behavioral-analyst'\nimport { createTraceAnalystKind, type TraceAnalystKindSpec } from './kind-factory'\nimport { DEFAULT_TRACE_ANALYST_KINDS } from './kinds'\nimport { AnalystRegistry, type AnalystRegistryOptions } from './registry'\n\nexport interface DefaultAnalystRegistryOptions {\n  /** Ax service for the agentic RLM kinds. Omit → only the deterministic analyst. */\n  ai?: AxAIService\n  /** Required unless `ai` was created by `createAnalystAi`. */\n  model?: string\n  /** Which agentic kinds to register when `ai` is present. Default = the shipped suite. */\n  kinds?: readonly TraceAnalystKindSpec[]\n  /** Set false to omit the deterministic behavioral analyst (default: include). */\n  includeBehavioral?: boolean\n  /** Forwarded to the AnalystRegistry constructor (signal, tags, priorFindings). */\n  registry?: AnalystRegistryOptions\n}\n\nexport function buildDefaultAnalystRegistry(\n  opts: DefaultAnalystRegistryOptions = {},\n): AnalystRegistry {\n  const registry = new AnalystRegistry(opts.registry)\n  if (opts.includeBehavioral !== false) {\n    registry.register(behavioralAnalyst())\n  }\n  if (opts.ai) {\n    const kinds = opts.kinds ?? DEFAULT_TRACE_ANALYST_KINDS\n    for (const spec of kinds) {\n      registry.register(createTraceAnalystKind(spec, { ai: opts.ai, model: opts.model }))\n    }\n  }\n  return registry\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;AA+HO,SAAS,iBAAiB,MAAwC;AACvE,UAAQ,KAAK,WAAW;AAAA,IACtB,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK,WAAW;AAAA,UACzB,QAAQ,KAAK;AAAA,UACb,YAAY,KAAK;AAAA,QACnB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK,WAAW;AAAA,UACzB,QAAQ,KAAK,UAAU;AAAA,UACvB,YAAY,KAAK;AAAA,QACnB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,KAAK;AAAA,QACL,KAAK;AAAA,QACL,IAAI,UAAU;AAAA,UACZ,SAAS,KAAK;AAAA,UACd,QAAQ,KAAK;AAAA,UACb,YAAY,KAAK;AAAA,QACnB,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,WAAW;AAAA,QACX,cAAc,KAAK;AAAA,QACnB,iBAAiB,KAAK;AAAA,QACtB,MAAM,OAAO,KAAK,aAAa,KAAK,KAAK,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;AAAA,MACzF;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,WAAW;AAAA,QACX,cAAc,KAAK;AAAA,QACnB,iBAAiB;AAAA,QACjB,MAAM,OAAO,KAAK,aAAa,KAAK,QAAQ,aAAa,KAAK,KAAK,YAAY,GAAG,QAAQ;AAAA,MAC5F;AAAA,EACJ;AACF;AAEA,SAAS,cACP,WACA,cACA,OACY;AACZ,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,iBAAiB,MAAM;AAAA,IACvB,MAAM,CAAC,KAAK,aAAa;AACvB,YAAM,WAAW,aAAa,KAAK,YAAY;AAC/C,YAAM,UAA0B;AAAA,QAC9B,OAAO,SAAS;AAAA,QAChB,UAAU,IAAI;AAAA,QACd,UAAU,IAAI;AAAA,QACd,YAAY,IAAI;AAAA,QAChB,aAAa,IAAI;AAAA,QACjB,WAAW,IAAI;AAAA,QACf,WAAW,IAAI;AAAA,MACjB;AACA,aAAO,MAAM,KAAK,SAAS;AAAA,QACzB,QAAQ,UAAU;AAAA,QAClB,gBAAgB,UAAU;AAAA,MAC5B,CAAC;AAAA,IACH;AAAA,EACF;AACF;AAEA,SAAS,aAAa,KAAkB,cAA+C;AACrF,MAAI,IAAI,MAAO,QAAO;AACtB,MAAI,CAAC,cAAc;AACjB,UAAM,IAAI;AAAA,MACR;AAAA,IAEF;AAAA,EACF;AACA,SAAO,EAAE,GAAG,KAAK,OAAO,aAAa;AACvC;;;AC9IA,IAAM,sBAAsB;AAE5B,IAAM,iBAAiB;AAMvB,IAAM,YACJ;AAEF,SAAS,IAAI,GAA2B;AACtC,SAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AACA,SAAS,QAAQ,OAAgC,MAAwC;AACvF,aAAW,OAAO,MAAM;AACtB,UAAM,QAAQ,IAAI,MAAM,GAAG,CAAC;AAC5B,QAAI,UAAU,KAAM,QAAO;AAAA,EAC7B;AACA,SAAO;AACT;AACA,SAAS,cAAc,GAAoC;AACzD,QAAM,eAAe,IAAI,EAAE,WAAW,kBAAkB,CAAC;AACzD,MAAI,iBAAiB,KAAM,QAAO;AAGlC,SACE,QAAQ,EAAE,YAAY,yBAAyB,KAAK,IAAI,EAAE,WAAW,wBAAwB,CAAC;AAElG;AACA,SAAS,eAAe,GAAoC;AAC1D,SACE,QAAQ,EAAE,YAAY,0BAA0B,KAChD,IAAI,EAAE,WAAW,yBAAyB,CAAC;AAE/C;AACA,SAAS,OAAO,GAAoC;AAClD,SAAO,IAAI,EAAE,WAAW,IAAI;AAC9B;AACA,SAAS,WAAW,GAAoC;AACtD,MAAI,EAAE,UAAW,QAAO,EAAE;AAC1B,aAAW,OAAO,qBAAqB;AACrC,UAAM,IAAI,EAAE,WAAW,GAAG;AAC1B,QAAI,OAAO,MAAM,YAAY,EAAE,SAAS,EAAG,QAAO;AAAA,EACpD;AACA,SAAO;AACT;AAMO,SAAS,oBAAoB,OAAuD;AACzF,QAAM,WAAW,IAAI,IAAI,MAAM,IAAI,CAAC,SAAS,KAAK,QAAQ,CAAC;AAC3D,MAAI,SAAS,OAAO,GAAG;AACrB,UAAM,IAAI;AAAA,MACR,gEAAgE,SAAS,IAAI;AAAA,IAC/E;AAAA,EACF;AACA,QAAM,UAAU,SAAS,OAAO,EAAE,KAAK,EAAE,SAAS;AAElD,QAAM,UAAyB,MAAM,IAAI,CAAC,UAAU;AAAA,IAClD;AAAA,IACA,OAAO,cAAc,IAAI;AAAA,IACzB,QAAQ,eAAe,IAAI;AAAA,IAC3B,MAAM,OAAO,IAAI;AAAA,EACnB,EAAE;AACF,QAAM,aAAa,QAAQ,OAAO,CAAC,WAAW,OAAO,KAAK,SAAS,KAAK;AACxE,QAAM,eACJ,WAAW,SAAS,IAChB,aACA,QAAQ,OAAO,CAAC,WAAW,OAAO,UAAU,QAAQ,OAAO,WAAW,IAAI;AAChF,QAAM,iBAAiB,oBAAoB,cAAc,KAAK;AAC9D,QAAM,kBAAkB,eAAe,CAAC;AACxC,QAAM,uBACJ,iBAAiB,qBAAqB,OAAO,CAAC,UAA2B,UAAU,IAAI,KAAK,CAAC;AAC/F,QAAM,wBACJ,iBAAiB,sBAAsB,OAAO,CAAC,UAA2B,UAAU,IAAI,KAAK,CAAC;AAChG,QAAM,gBAAwC,CAAC;AAC/C,MAAI,sBAAsB;AAE1B,aAAW,KAAK,OAAO;AACrB,UAAM,OAAO,WAAW,CAAC;AACzB,QAAI,MAAM;AACR,oBAAc,IAAI,KAAK,cAAc,IAAI,KAAK,KAAK;AACnD,UAAI,UAAU,KAAK,IAAI,EAAG,uBAAsB;AAAA,IAClD;AAAA,EACF;AAEA,QAAM,iBAAiB,OAAO,OAAO,aAAa,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC7E,QAAM,gBAAgB,OAAO,KAAK,aAAa,EAAE;AACjD,QAAM,qBAAqB,mBAAmB,IAAI,IAAI,gBAAgB;AAEtE,QAAM,UAA8B,CAAC;AACrC,QAAM,mBAAmB,oBAAI,IAAoB;AACjD,aAAW,YAAY,gBAAgB;AACrC,eAAW,UAAU,aAAa,QAAQ,GAAG;AAC3C,UAAI,iBAAiB,IAAI,OAAO,IAAI,EAAG;AACvC,uBAAiB,IAAI,OAAO,IAAI;AAChC,cAAQ,KAAK,MAAM;AAAA,IACrB;AAAA,EACF;AAEA,MAAI,kBAAkB,kBAAkB,kBAAkB,GAAG;AAC3D,UAAM,OAAO,OAAO,KAAK,aAAa,EAAE,CAAC;AACzC,YAAQ,KAAK;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,OAAO,cAAc,8BAA8B,IAAI;AAAA,MAC/D,UAAU,EAAE,MAAM,MAAM,OAAO,gBAAgB,gBAAgB,EAAE;AAAA,IACnE,CAAC;AAAA,EACH;AAEA,MAAI,kBAAkB,kBAAkB,CAAC,qBAAqB;AAC5D,YAAQ,KAAK;AAAA,MACX,MAAM;AAAA,MACN,UAAU;AAAA,MACV,QAAQ,GAAG,cAAc;AAAA,MACzB,UAAU,EAAE,YAAY,gBAAgB,oBAAoB,EAAE;AAAA,IAChE,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL;AAAA,IACA,cAAc,aAAa;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,oBACP,SACA,OAC2B;AAC3B,QAAM,YAAY,IAAI,IAAI,MAAM,IAAI,CAAC,SAAS,CAAC,KAAK,SAAS,IAAI,CAAC,CAAC;AACnE,QAAM,oBAAoB,kCAAkC,SAAS;AACrE,QAAM,gBAAgB,QAAQ,IAAI,CAAC,YAAY;AAAA,IAC7C;AAAA,IACA,GAAG,kBAAkB,OAAO,IAAI;AAAA,EAClC,EAAE;AACF,QAAM,cAAc,qBAAqB,aAAa;AACtD,QAAM,UAAU,oBAAI,IAAyD;AAC7E,aAAW,UAAU,eAAe;AAClC,UAAM,UAAU,YAAY,IAAI,OAAO,GAAG;AAC1C,UAAM,QAAQ,QAAQ,IAAI,OAAO,KAAK,EAAE,SAAS,OAAO,SAAS,SAAS,CAAC,EAAE;AAC7E,UAAM,QAAQ,KAAK,OAAO,MAAM;AAChC,YAAQ,IAAI,SAAS,KAAK;AAAA,EAC5B;AAEA,QAAM,YAAuC,CAAC;AAC9C,aAAW,EAAE,SAAS,SAAS,QAAQ,KAAK,QAAQ,OAAO,GAAG;AAC5D,UAAM,OAAO,gBAAgB,CAAC,GAAG,OAAO,EAAE,KAAK,mBAAmB,CAAC;AACnE,SAAK,QAAQ,CAAC,KAAK,UAAU;AAC3B,gBAAU,KAAK;AAAA,QACb,SAAS,KAAK,WAAW,IAAI,UAAU,GAAG,OAAO,IAAI,QAAQ,CAAC;AAAA,QAC9D,SAAS,IAAI,IAAI,CAAC,WAAW,OAAO,KAAK,OAAO;AAAA,QAChD,sBAAsB,IAAI,IAAI,CAAC,WAAW,OAAO,KAAK;AAAA,QACtD,uBAAuB,IAAI,IAAI,CAAC,WAAW,OAAO,MAAM;AAAA,MAC1D,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AAEA,SAAO,UAAU;AAAA,IACf,CAAC,GAAG,MACF,EAAE,QAAQ,SAAS,EAAE,QAAQ,UAC7B,EAAE,QAAQ,cAAc,EAAE,OAAO,KACjC,EAAE,QAAQ,CAAC,EAAG,cAAc,EAAE,QAAQ,CAAC,CAAE;AAAA,EAC7C;AACF;AASA,SAAS,kCAAkC,WAQzC;AACA,QAAM,QAAQ,oBAAI,IAA2B;AAC7C,SAAO,CAAC,SAAS;AACf,UAAM,WAAW,qBAAqB,KAAK,gBAAgB,WAAW,KAAK;AAC3E,UAAM,SAAS,SAAS;AACxB,UAAM,UAAU,SAAS,UACrB,QAAQ,SAAS,OAAO,KACxB,SAAS,kBACP,UAAU,SAAS,eAAe,KAClC,SACE,QAAQ,MAAM,KACd,KAAK,aACH,SAAS,KAAK,UAAU,KACxB,SAAS,KAAK,QAAQ;AAChC,UAAM,cAAc,SAAS,WAAW,SAAS,mBAAmB;AACpE,UAAM,WAAW,SAAS,aAAa,UAAU,IAAI,SAAS,UAAU,IAAI;AAC5E,UAAM,SAAS,gBAAgB,QAAQ,SAAS,eAAe;AAC/D,UAAM,YAAY,SAAS,OAAO;AAClC,WAAO;AAAA,MACL,KAAK,KAAK,UAAU,CAAC,SAAS,SAAS,KAAK,UAAU,SAAS,UAAU,CAAC;AAAA,MAC1E,UAAU;AAAA,MACV;AAAA,MACA,OAAO,YAAY,gBAAgB,UAAU,UAAU,IAAI;AAAA,MAC3D,KAAK,YAAY,gBAAgB,UAAU,QAAQ,IAAI;AAAA,IACzD;AAAA,EACF;AACF;AAEA,SAAS,qBACP,SACA,WACA,OACe;AACf,QAAM,QAAuB;AAAA,IAC3B,SAAS;AAAA,IACT,QAAQ;AAAA,IACR,iBAAiB;AAAA,IACjB,YAAY;AAAA,EACd;AACA,MAAI,CAAC,QAAS,QAAO;AAErB,QAAM,OAAiB,CAAC;AACxB,QAAM,YAAY,oBAAI,IAAoB;AAC1C,MAAI,YAA2B;AAC/B,MAAI,WAAW;AACf,SAAO,WAAW;AAChB,UAAM,SAAS,MAAM,IAAI,SAAS;AAClC,QAAI,QAAQ;AACV,iBAAW;AACX;AAAA,IACF;AACA,UAAM,aAAa,UAAU,IAAI,SAAS;AAC1C,QAAI,eAAe,QAAW;AAC5B,iBAAW;AAAA,QACT,SAAS;AAAA,QACT,QAAQ,CAAC,GAAG,KAAK,MAAM,UAAU,CAAC,EAAE,KAAK,EAAE,CAAC;AAAA,QAC5C,iBAAiB;AAAA,QACjB,YAAY,CAAC,GAAG,KAAK,MAAM,UAAU,CAAC,EAAE,KAAK,EAAE,CAAC;AAAA,MAClD;AACA;AAAA,IACF;AACA,UAAM,UAAU,UAAU,IAAI,SAAS;AACvC,QAAI,CAAC,SAAS;AACZ,iBAAW;AAAA,QACT,SAAS;AAAA,QACT,QAAQ;AAAA,QACR,iBAAiB;AAAA,QACjB,YAAY;AAAA,MACd;AACA;AAAA,IACF;AACA,QAAI,QAAQ,SAAS,SAAS;AAC5B,iBAAW;AAAA,QACT,SAAS,QAAQ;AAAA,QACjB,QAAQ;AAAA,QACR,iBAAiB;AAAA,QACjB,YAAY,QAAQ;AAAA,MACtB;AACA;AAAA,IACF;AACA,cAAU,IAAI,WAAW,KAAK,MAAM;AACpC,SAAK,KAAK,SAAS;AACnB,gBAAY,QAAQ;AAAA,EACtB;AAEA,WAAS,QAAQ,KAAK,SAAS,GAAG,SAAS,GAAG,SAAS,GAAG;AACxD,QAAI,SAAS,YAAY,QAAQ,SAAS,WAAW,MAAM;AACzD,iBAAW,EAAE,GAAG,UAAU,QAAQ,KAAK,KAAK,GAAI,YAAY,KAAK,KAAK,EAAG;AAAA,IAC3E,WACE,SAAS,gBAAgB,SAAS,WAAW,SAAS,mBAAmB,SAAS,SAClF;AACA,iBAAW,EAAE,GAAG,UAAU,YAAY,KAAK,KAAK,EAAG;AAAA,IACrD;AACA,UAAM,IAAI,KAAK,KAAK,GAAI,QAAQ;AAAA,EAClC;AACA,SAAO;AACT;AAEA,SAAS,oBAAoB,GAAgB,GAAwB;AACnE,QAAM,SAAS,gBAAgB,EAAE,KAAK,UAAU;AAChD,QAAM,SAAS,gBAAgB,EAAE,KAAK,UAAU;AAChD,MAAI,WAAW,QAAQ,WAAW,KAAM,QAAO;AAC/C,MAAI,WAAW,QAAQ,WAAW,KAAM,QAAO;AAC/C,MAAI,WAAW,QAAQ,WAAW,QAAQ,WAAW,OAAQ,QAAO,SAAS;AAC7E,MAAI,EAAE,SAAS,QAAQ,EAAE,SAAS,QAAQ,EAAE,SAAS,EAAE,KAAM,QAAO,EAAE,OAAO,EAAE;AAC/E,SAAO,EAAE,KAAK,QAAQ,cAAc,EAAE,KAAK,OAAO;AACpD;AAEA,SAAS,gBAAgB,SAAkD;AACzE,QAAM,OAAwB,CAAC;AAC/B,MAAI,SAAwB,CAAC;AAC7B,MAAI,UAAyB,CAAC;AAC9B,MAAI,aAAa,OAAO;AAExB,QAAM,cAAc,MAAM;AACxB,QAAI,OAAO,SAAS,EAAG,MAAK,KAAK,MAAM;AACvC,aAAS,CAAC;AAAA,EACZ;AACA,QAAM,eAAe,MAAM;AACzB,QAAI,QAAQ,WAAW,GAAG;AACxB,aAAO,KAAK,QAAQ,CAAC,CAAE;AAAA,IACzB,WAAW,QAAQ,SAAS,GAAG;AAC7B,kBAAY;AACZ,iBAAW,UAAU,QAAS,MAAK,KAAK,CAAC,MAAM,CAAC;AAAA,IAClD;AACA,cAAU,CAAC;AACX,iBAAa,OAAO;AAAA,EACtB;AAEA,aAAW,UAAU,SAAS;AAC5B,UAAM,QAAQ,gBAAgB,OAAO,KAAK,UAAU;AACpD,UAAM,MAAM,gBAAgB,OAAO,KAAK,QAAQ;AAChD,QAAI,UAAU,QAAQ,QAAQ,QAAQ,OAAO,KAAK,eAAe,KAAK,MAAM,OAAO;AACjF,mBAAa;AACb,kBAAY;AACZ,WAAK,KAAK,CAAC,MAAM,CAAC;AAClB;AAAA,IACF;AACA,QAAI,QAAQ,SAAS,KAAK,SAAS,WAAY,cAAa;AAC5D,YAAQ,KAAK,MAAM;AACnB,iBAAa,KAAK,IAAI,YAAY,GAAG;AAAA,EACvC;AACA,eAAa;AACb,cAAY;AACZ,SAAO;AACT;AAEA,SAAS,aAAa,UAAuD;AAC3E,QAAM,UAA8B,CAAC;AACrC,QAAM,SAAS,SAAS;AACxB,QAAM,UAAU,SAAS;AACzB,MAAI,OAAO,UAAU,KAAK,OAAO,MAAM,CAAC,UAA2B,UAAU,IAAI,GAAG;AAClF,UAAM,QAAQ,OAAO,CAAC;AACtB,UAAM,OAAO,OAAO,OAAO,SAAS,CAAC;AACrC,UAAM,cAAc,cAAc,QAAQ,CAAC,UAAU,YAAY,WAAW,QAAQ;AACpF,UAAM,iBAAiB,UAAU,KAAK,OAAO;AAC7C,UAAM,SAAS,iBAAiB,WAAW,QAAQ,IAAI,OAAO,QAAQ;AACtE,QAAI,eAAe,OAAO,SAAS,UAAU,qBAAqB;AAChE,YAAM,cAAc,iBAAiB,+BAA0B,GAAG,OAAO,QAAQ,CAAC,CAAC;AACnF,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ,yBAAyB,WAAW,KAAK,KAAK,SAAI,IAAI,YAAY,OAAO,MAAM;AAAA,QACvF,UAAU;AAAA,UACR;AAAA,UACA;AAAA,UACA,UAAU,iBAAiB,cAAc,OAAO,OAAO,QAAQ,CAAC,CAAC;AAAA,UACjE,OAAO,OAAO;AAAA,UACd,OAAO,SAAS;AAAA,UAChB,eAAe,SAAS,QAAQ,CAAC;AAAA,UACjC,cAAc,SAAS,QAAQ,SAAS,QAAQ,SAAS,CAAC;AAAA,QAC5D;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,MACE,OAAO,UAAU,KACjB,OAAO,WAAW,QAAQ,UAC1B,OAAO,MAAM,CAAC,UAA2B,UAAU,IAAI,KACvD,QAAQ,MAAM,CAAC,UAA2B,UAAU,IAAI,GACxD;AACA,UAAM,QAAQ,QAAQ,CAAC;AACvB,UAAM,OAAO,QAAQ,QAAQ,SAAS,CAAC;AACvC,UAAM,mBAAmB,cAAc,QAAQ,CAAC,UAAU,YAAY,WAAW,QAAQ;AACzF,UAAM,oBAAoB,cAAc,SAAS,CAAC,UAAU,YAAY,WAAW,QAAQ;AAC3F,UAAM,YAAY,OAAO,OAAO,SAAS,CAAC,IAAK,OAAO,CAAC;AACvD,QAAI,oBAAoB,aAAa,qBAAqB,OAAO,OAAO;AACtE,cAAQ,KAAK;AAAA,QACX,MAAM;AAAA,QACN,UAAU;AAAA,QACV,QAAQ,4BAA4B,KAAK,SAAI,IAAI,SAAS,QAAQ,MAAM;AAAA,QACxE,UAAU;AAAA,UACR;AAAA,UACA;AAAA,UACA,OAAO,QAAQ;AAAA,UACf,OAAO,SAAS;AAAA,UAChB,eAAe,SAAS,QAAQ,CAAC;AAAA,UACjC,cAAc,SAAS,QAAQ,SAAS,QAAQ,SAAS,CAAC;AAAA,QAC5D;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,cACP,QACA,WACS;AACT,SAAO,OAAO,MAAM,CAAC,EAAE,MAAM,CAAC,SAAS,UAAU,UAAU,OAAO,KAAK,GAAI,OAAO,CAAC;AACrF;;;ACncA,IAAM,qBAAqD;AAAA,EACzD,0BACE;AAAA,EACF,uBACE;AAAA,EACF,0BACE;AAAA,EACF,wBACE;AACJ;AAEA,IAAM,aAAa;AAEnB,IAAM,kBAA0F;AAAA,EAC9F,0BAA0B,CAAC,UAAU,aACnC,GAAG,QAAQ,IAAI,QAAQ;AAAA,EACzB,uBAAuB,CAAC,UAAU,aAChC,GAAG,QAAQ,IAAI,QAAQ;AAAA,EACzB,0BAA0B,CAAC,UAAU,aACnC,GAAG,QAAQ,IAAI,QAAQ;AAAA,EACzB,wBAAwB,CAAC,UAAU,aACjC,GAAG,QAAQ,IAAI,QAAQ;AAC3B;AAMO,SAAS,yBACd,SACA,OAAoD,CAAC,GACnC;AAClB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,UAAU,QAAQ;AACxB,SAAO,QAAQ,QAAQ;AAAA,IAAI,CAAC,QAC1B,YAAY;AAAA,MACV,YAAY;AAAA,MACZ,MAAM;AAAA,MACN,SAAS,IAAI;AAAA,MACb,OAAO,IAAI;AAAA,MACX,UAAU,IAAI;AAAA;AAAA,MAEd,YAAY;AAAA,MACZ,eAAe;AAAA,QACb;AAAA,UACE,MAAM;AAAA,UACN,KAAK,UACD,kBAAkB,mBAAmB,OAAO,CAAC,eAAe,IAAI,IAAI,KACpE,uBAAuB,IAAI,IAAI;AAAA,UACnC,SAAS,KAAK,UAAU,IAAI,QAAQ;AAAA,QACtC;AAAA,MACF;AAAA,MACA,oBAAoB,mBAAmB,IAAI,IAAI;AAAA,MAC/C,UAAU;AAAA,QACR,eAAe;AAAA,QACf,UAAU,IAAI;AAAA,QACd,GAAI,UAAU,EAAE,UAAU,QAAQ,IAAI,CAAC;AAAA,MACzC;AAAA,MACA,UAAU,IAAI;AAAA,MACd,GAAI,KAAK,aAAa,EAAE,aAAa,KAAK,WAAW,IAAI,CAAC;AAAA,IAC5D,CAAC;AAAA,EACH;AACF;AAGO,SAAS,oBAAiD;AAC/D,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS;AAAA,IACT,MAAM,QAAQ,OAAO;AACnB,YAAM,WAAW,MAAM,MAAM,YAAY;AACzC,YAAM,mBAAmB,CAAC,GAAG,IAAI,IAAI,SAAS,gBAAgB,CAAC,EAAE,KAAK;AACtE,YAAM,eAAe,oBAAI,IAGvB;AACF,iBAAW,WAAW,kBAAkB;AACtC,cAAM,SAAS,MAAM,MAAM,UAAU,EAAE,UAAU,QAAQ,CAAC;AAC1D,YAAI,OAAO,aAAa,SAAS;AAC/B,gBAAM,IAAI;AAAA,YACR,uCAAuC,OAAO,gBAAgB,OAAO,QAAQ;AAAA,UAC/E;AAAA,QACF;AACA,YAAI,CAAC,OAAO,OAAO;AACjB,gBAAM,IAAI;AAAA,YACR,6BAA6B,OAAO;AAAA,UACtC;AAAA,QACF;AACA,cAAM,UAAU,oBAAoB,OAAO,KAAK;AAChD,YAAI,QAAQ,YAAY,QAAQ,QAAQ,YAAY,SAAS;AAC3D,gBAAM,IAAI;AAAA,YACR,uCAAuC,OAAO,gBAAgB,QAAQ,OAAO;AAAA,UAC/E;AAAA,QACF;AACA,mBAAW,WAAW,yBAAyB,OAAO,GAAG;AACvD,gBAAM,UAAU,aAAa,IAAI,QAAQ,UAAU;AACnD,cAAI,CAAC,SAAS;AACZ,yBAAa,IAAI,QAAQ,YAAY;AAAA,cACnC;AAAA,cACA,UAAU,CAAC,OAAO;AAAA,cAClB,UAAU,CAAC,GAAG,QAAQ,aAAa;AAAA,YACrC,CAAC;AACD;AAAA,UACF;AACA,kBAAQ,SAAS,KAAK,OAAO;AAC7B,kBAAQ,SAAS,KAAK,GAAG,QAAQ,aAAa;AAAA,QAChD;AAAA,MACF;AACA,aAAO,CAAC,GAAG,aAAa,OAAO,CAAC,EAAE,IAAI,CAAC,EAAE,SAAS,UAAU,SAAS,OAAO;AAAA,QAC1E,GAAG;AAAA,QACH,OAAO,gBAAgB,QAAQ,OAAyB;AAAA,UACtD,SAAS;AAAA,UACT,iBAAiB;AAAA,QACnB;AAAA,QACA,WAAW,GAAG,SAAS,MAAM,IAAI,iBAAiB,MAAM;AAAA,QACxD,eAAe;AAAA,QACf,UAAU;AAAA,UACR,eAAe;AAAA,UACf,WAAW;AAAA,UACX,sBAAsB,SAAS;AAAA,UAC/B,sBAAsB,iBAAiB;AAAA,QACzC;AAAA,MACF,EAAE;AAAA,IACJ;AAAA,EACF;AACF;;;AC1HO,SAAS,4BACd,OAAsC,CAAC,GACtB;AACjB,QAAM,WAAW,IAAI,gBAAgB,KAAK,QAAQ;AAClD,MAAI,KAAK,sBAAsB,OAAO;AACpC,aAAS,SAAS,kBAAkB,CAAC;AAAA,EACvC;AACA,MAAI,KAAK,IAAI;AACX,UAAM,QAAQ,KAAK,SAAS;AAC5B,eAAW,QAAQ,OAAO;AACxB,eAAS,SAAS,uBAAuB,MAAM,EAAE,IAAI,KAAK,IAAI,OAAO,KAAK,MAAM,CAAC,CAAC;AAAA,IACpF;AAAA,EACF;AACA,SAAO;AACT;","names":[]}

package/dist/{chunk-TVXPJJ7H.js → chunk-H5UD2323.js}

@@ -6,7 +6,7 @@ import {
   LLM_OUTPUT_TOKEN_ATTR_KEYS,
   LLM_REASONING_TOKEN_ATTR_KEYS,
   RUN_COST_ATTR_KEYS
-} from "./chunk-BUBU3YHT.js";
+} from "./chunk-K4DBDHLK.js";
 
 // src/trace/execution-measurements.ts
 var TOKEN_MEASUREMENT_KEY_GROUPS = [
@@ -283,4 +283,4 @@ export {
   summarizeExecutionMeasurements,
   recordAggregateMeasurements
 };
-//# sourceMappingURL=chunk-TVXPJJ7H.js.map
+//# sourceMappingURL=chunk-H5UD2323.js.map

package/dist/{chunk-MLHRFWQK.js → chunk-JHOJHHU7.js}

@@ -61,6 +61,7 @@ var CostLedger = class {
   records = /* @__PURE__ */ new Map();
   activeCallIds = /* @__PURE__ */ new Set();
   lateCallIds = /* @__PURE__ */ new Set();
+  idleWaiters = /* @__PURE__ */ new Set();
   completedTasks = 0;
   revision = "memory";
   costLimitPersisted = false;
@@ -181,9 +182,27 @@ var CostLedger = class {
     } catch (error) {
       return { succeeded: false, callId: pending.callId, error: toError(error) };
     } finally {
-      if (!this.lateCallIds.has(pending.callId)) this.activeCallIds.delete(pending.callId);
+      if (!this.lateCallIds.has(pending.callId)) this.releaseActiveCall(pending.callId);
     }
   }
+  /** Wait until every call started by this ledger has produced a durable outcome. */
+  async waitForIdle(options = {}) {
+    if (this.activeCallIds.size === 0) return true;
+    const timeoutMs = options.timeoutMs ?? 5e3;
+    assertTimeout(timeoutMs, "waitForIdle.timeoutMs");
+    return new Promise((resolve) => {
+      let timer;
+      const finish = (settled) => {
+        this.idleWaiters.delete(onIdle);
+        if (timer !== void 0) clearTimeout(timer);
+        resolve(settled);
+      };
+      const onIdle = () => finish(true);
+      this.idleWaiters.add(onIdle);
+      timer = setTimeout(() => finish(false), timeoutMs);
+      if (this.activeCallIds.size === 0) onIdle();
+    });
+  }
   /** Settle a call left pending by a crashed process after reconciling with the provider. */
   reconcile(callId, observed, options = {}) {
     const pending = [...this.records.values()].find(
@@ -212,12 +231,16 @@ var CostLedger = class {
     );
     let inputTokens = 0;
     let outputTokens = 0;
+    let reasoningTokens = 0;
     let cachedTokens = 0;
+    let cacheWriteTokens = 0;
     let totalCostUsd = 0;
     for (const receipt of receipts) {
       inputTokens += receipt.inputTokens;
       outputTokens += receipt.outputTokens;
+      reasoningTokens += receipt.reasoningTokens ?? 0;
       cachedTokens += receipt.cachedTokens ?? 0;
+      cacheWriteTokens += receipt.cacheWriteTokens ?? 0;
       totalCostUsd += receipt.costUsd;
       if (receipt.costUnknown) {
         unpriced.add(receipt.model);
@@ -239,7 +262,9 @@ var CostLedger = class {
       rollup.calls += 1;
       rollup.inputTokens += receipt.inputTokens;
       rollup.outputTokens += receipt.outputTokens;
+      rollup.reasoningTokens = (rollup.reasoningTokens ?? 0) + (receipt.reasoningTokens ?? 0);
       rollup.cachedTokens += receipt.cachedTokens ?? 0;
+      rollup.cacheWriteTokens = (rollup.cacheWriteTokens ?? 0) + (receipt.cacheWriteTokens ?? 0);
       rollup.costUsd += receipt.costUsd;
       if (receipt.costUnknown) rollup.unpricedCalls += 1;
       if (receipt.usageUnknown) rollup.unknownUsageCalls += 1;
@@ -254,12 +279,14 @@ var CostLedger = class {
       reservedCostUsd: pending.reduce((sum, record) => sum + (record.maximumCostUsd ?? 0), 0),
       inputTokens,
       outputTokens,
+      reasoningTokens,
       cachedTokens,
+      cacheWriteTokens,
       totalCostUsd,
       byChannel: [...byChannel.values()].sort((a, b) => a.channel.localeCompare(b.channel)),
       unpricedModels: [...unpriced].sort(),
       fullyPriced: unpriced.size === 0,
-      usageComplete: receipts.every((receipt) => !receipt.usageUnknown),
+      usageComplete: pending.length === 0 && receipts.every((receipt) => !receipt.usageUnknown),
       accountingComplete: incompleteReasons.length === 0,
       incompleteReasons: [...new Set(incompleteReasons)]
     };
@@ -345,9 +372,15 @@ var CostLedger = class {
       }
     ).finally(() => {
       this.lateCallIds.delete(pending.callId);
-      this.activeCallIds.delete(pending.callId);
+      this.releaseActiveCall(pending.callId);
     });
   }
+  releaseActiveCall(callId) {
+    this.activeCallIds.delete(callId);
+    if (this.activeCallIds.size > 0) return;
+    for (const resolve of this.idleWaiters) resolve();
+    this.idleWaiters.clear();
+  }
   commitOutcome(pending, error, observed) {
     try {
       const receipt = this.commitReceipt(pending, observed, error.message);
@@ -420,10 +453,7 @@ var CostLedger = class {
     const callId = record.callId;
     const receipt = record.status === "settled" ? record : void 0;
     validateTransition(this.records, record);
-    const event = {
-      version: 1,
-      record: cloneRecord(record)
-    };
+    const event = record.status === "settled" && (record.reasoningTokens !== void 0 || record.cacheWriteTokens !== void 0) ? { version: 2, record: cloneReceipt(record) } : { version: 1, record: cloneRecord(record) };
     this.appendEvent(event, callId, receipt);
     this.records.set(callId, cloneRecord(record));
   }
@@ -459,7 +489,11 @@ function costForUsage(model, usage) {
   assertUsage(usage);
   if (!resolveModelPricing(model)) return { costUsd: 0, costUnknown: true };
   return {
-    costUsd: estimateCost(usage.inputTokens + (usage.cachedTokens ?? 0), usage.outputTokens, model),
+    costUsd: estimateCost(
+      usage.inputTokens + (usage.cachedTokens ?? 0) + (usage.cacheWriteTokens ?? 0),
+      usage.outputTokens,
+      model
+    ),
     costUnknown: false
   };
 }
@@ -505,7 +539,9 @@ function buildReceipt(pending, observed, error) {
       model: observed.model,
       inputTokens: observed.inputTokens,
       outputTokens: observed.outputTokens,
+      ...observed.reasoningTokens === void 0 ? {} : { reasoningTokens: observed.reasoningTokens },
       ...observed.cachedTokens === void 0 ? {} : { cachedTokens: observed.cachedTokens },
+      ...observed.cacheWriteTokens === void 0 ? {} : { cacheWriteTokens: observed.cacheWriteTokens },
       costUsd: costUnknown ? 0 : hasActual ? observed.actualCostUsd : estimated.costUsd,
       costUnknown,
       usageUnknown,
@@ -547,7 +583,7 @@ var PendingCostCallSchema = z.strictObject({
   status: z.literal("pending"),
   ...CostCallBaseShape
 });
-var CostReceiptSchema = z.strictObject({
+var CostReceiptBaseShape = {
   status: z.literal("settled"),
   ...CostCallBaseShape,
   inputTokens: TokenCount,
@@ -559,7 +595,17 @@ var CostReceiptSchema = z.strictObject({
   pricing: CostPricingSchema.optional(),
   actualCostUsd: NonNegative.optional(),
   error: z.string().optional()
-}).superRefine((receipt, ctx) => {
+};
+var LegacyCostReceiptSchema = z.strictObject(CostReceiptBaseShape).superRefine((receipt, ctx) => validateCostReceipt(receipt, ctx));
+var CostReceiptSchema = z.strictObject({
+  ...CostReceiptBaseShape,
+  reasoningTokens: TokenCount.optional(),
+  cacheWriteTokens: TokenCount.optional()
+}).superRefine((receipt, ctx) => validateCostReceipt(receipt, ctx));
+function validateCostReceipt(receipt, ctx) {
+  if (receipt.reasoningTokens !== void 0 && receipt.reasoningTokens > receipt.outputTokens) {
+    ctx.addIssue({ code: "custom", message: "reasoningTokens must not exceed outputTokens" });
+  }
   if (receipt.actualCostUsd !== void 0) {
     if (receipt.costUnknown || receipt.costUsd !== receipt.actualCostUsd) {
       ctx.addIssue({ code: "custom", message: "actual cost must be known and equal costUsd" });
@@ -592,11 +638,15 @@ var CostReceiptSchema = z.strictObject({
       message: `estimated cost ${receipt.costUsd} does not match pricing snapshot ${expected}`
     });
   }
-});
+}
 var CostLedgerEventSchema = z.union([
   z.strictObject({
     version: z.literal(1),
-    record: z.union([PendingCostCallSchema, CostReceiptSchema])
+    record: z.union([PendingCostCallSchema, LegacyCostReceiptSchema])
+  }),
+  z.strictObject({
+    version: z.literal(2),
+    record: CostReceiptSchema
   }),
   z.strictObject({
     version: z.literal(1),
@@ -618,8 +668,9 @@ function parseEvents(serialized) {
       lineNumber += 1;
       const event = CostLedgerEventSchema.parse(JSON.parse(line));
       if ("record" in event) {
-        validateTransition(records, event.record);
-        records.set(event.record.callId, cloneRecord(event.record));
+        const record = event.record;
+        validateTransition(records, record);
+        records.set(record.callId, cloneRecord(record));
       } else if ("completedTasks" in event) {
         completedTasks += event.completedTasks;
         if (!Number.isSafeInteger(completedTasks)) {
@@ -706,7 +757,7 @@ function cloneReceipt(receipt) {
   };
 }
 function costFromPricing(usage, pricing) {
-  return (usage.inputTokens + (usage.cachedTokens ?? 0)) / 1e3 * pricing.inputUsdPerThousand + usage.outputTokens / 1e3 * pricing.outputUsdPerThousand;
+  return (usage.inputTokens + (usage.cachedTokens ?? 0) + (usage.cacheWriteTokens ?? 0)) / 1e3 * pricing.inputUsdPerThousand + usage.outputTokens / 1e3 * pricing.outputUsdPerThousand;
 }
 function emptyRollup(channel) {
   return {
@@ -714,7 +765,9 @@ function emptyRollup(channel) {
     calls: 0,
     inputTokens: 0,
     outputTokens: 0,
+    reasoningTokens: 0,
     cachedTokens: 0,
+    cacheWriteTokens: 0,
     costUsd: 0,
     unpricedCalls: 0,
     unknownUsageCalls: 0
@@ -762,7 +815,16 @@ function validationMessage(cause) {
 function assertUsage(usage) {
   assertTokenCount(usage.inputTokens, "inputTokens");
   assertTokenCount(usage.outputTokens, "outputTokens");
+  if (usage.reasoningTokens !== void 0) {
+    assertTokenCount(usage.reasoningTokens, "reasoningTokens");
+    if (usage.reasoningTokens > usage.outputTokens) {
+      throw new ValidationError("CostLedger: reasoningTokens must not exceed outputTokens");
+    }
+  }
   if (usage.cachedTokens !== void 0) assertTokenCount(usage.cachedTokens, "cachedTokens");
+  if (usage.cacheWriteTokens !== void 0) {
+    assertTokenCount(usage.cacheWriteTokens, "cacheWriteTokens");
+  }
 }
 function assertTokenCount(value, name) {
   if (typeof value !== "number" || !Number.isInteger(value) || value < 0) {
@@ -778,6 +840,13 @@ function assertNonNegative(value, name) {
     );
   }
 }
+function assertTimeout(value, name) {
+  if (typeof value !== "number" || !Number.isSafeInteger(value) || value < 0 || value > 2147483647) {
+    throw new ValidationError(
+      `CostLedger: ${name} must be a non-negative safe integer no greater than 2147483647, got ${String(value)}`
+    );
+  }
+}
 function assertString(value, name) {
   if (typeof value !== "string" || value.trim().length === 0) {
     throw new ValidationError(`CostLedger: ${name} must be a non-empty string`);
@@ -795,4 +864,4 @@ export {
   modelPriceKey,
   costForUsage
 };
-//# sourceMappingURL=chunk-MLHRFWQK.js.map
+//# sourceMappingURL=chunk-JHOJHHU7.js.map