@tangle-network/agent-eval 0.116.0 → 0.117.1

package/dist/chunk-NBSS5NDZ.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/analyst/ax-service.ts","../src/analyst/chat-client.ts","../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/skill-usage.ts","../src/run-critic.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["import type { AxAIService } from '@ax-llm/ax'\nimport { ai } from '@ax-llm/ax'\n\nexport interface CreateAnalystAiConfig {\n  /** OpenAI-compatible API key forwarded as `Authorization: Bearer`.\n   *  cli-bridge ignores the value on loopback but Ax requires a non-empty string. */\n  apiKey: string\n  /** OpenAI-compatible base URL — e.g. `https://router.tangle.tools/v1` or a\n   *  cli-bridge loopback. */\n  baseUrl: string\n  /** Model id forwarded to the analyst actor + responder. */\n  model: string\n  /** Ax provider name. Defaults to the OpenAI-compatible client. */\n  provider?: 'openai' | 'anthropic'\n}\n\n/**\n * Construct the `AxAIService` an analyst kind calls through\n * (`createTraceAnalystKind({ ai })`).\n *\n * Ax's `ai()` pins `config.model` to the OpenAI catalog enum, but every\n * OpenAI-compatible router an analyst points at (router.tangle.tools,\n * cli-bridge) accepts arbitrary model ids (claude-code/sonnet, openai/gpt-5.4,\n * …). Consumers were each re-rolling `ai({ name, apiKey, apiURL, config })`\n * behind an `as (a: any) => any` cast to dodge the enum; this is the one\n * canonical constructor so they don't have to — and don't take a direct\n * `@ax-llm/ax` dependency for it.\n */\nexport function createAnalystAi(config: CreateAnalystAiConfig): AxAIService {\n  return ai({\n    name: config.provider ?? 'openai',\n    apiKey: config.apiKey,\n    apiURL: config.baseUrl,\n    config: { model: config.model },\n  })\n}\n","/**\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\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}\n\n// ── Factory ─────────────────────────────────────────────────────────\n\nexport type CreateChatClientOpts =\n  | RouterTransportOpts\n  | CliBridgeTransportOpts\n  | DirectProviderTransportOpts\n  | SandboxSdkTransportOpts\n  | MockTransportOpts\n\ninterface BaseTransportOpts {\n  defaultModel?: string\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        } 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        } 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        } as LlmClientOptions),\n      )\n    case 'sandbox-sdk':\n      return {\n        transport: 'sandbox-sdk',\n        defaultModel: opts.defaultModel,\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        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    chat: async (req, callOpts) => {\n      const resolved = resolveModel(req, defaultModel)\n      // LlmClient.call doesn't accept an external AbortSignal today (it\n      // owns its own AbortController for the per-attempt timeout). We\n      // race the response against the caller's signal so awaiting code\n      // unblocks on abort. The in-flight HTTP request still runs to its\n      // own timeoutMs — when LlmClient grows a signal parameter, wire\n      // it directly here and drop the race.\n      const call = inner.call({\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      if (!callOpts?.signal) return await call\n      return await Promise.race([call, abortAsRejection(callOpts.signal)])\n    },\n  }\n}\n\nfunction abortAsRejection(signal: AbortSignal): Promise<never> {\n  if (signal.aborted) return Promise.reject(toAbortError(signal))\n  return new Promise<never>((_, reject) => {\n    signal.addEventListener('abort', () => reject(toAbortError(signal)), { once: true })\n  })\n}\n\nfunction toAbortError(signal: AbortSignal): Error {\n  const reason = (signal as { reason?: unknown }).reason\n  if (reason instanceof Error) return reason\n  const e = new Error('ChatClient.chat: aborted')\n  e.name = 'AbortError'\n  return e\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 * LockedJsonlAppender — mutex-serialized JSONL append helper for arbitrary\n * payloads. The reference-replay store does the same thing for typed\n * `ReferenceReplayRun` rows; this is the generic version used by\n * `MutationTelemetry`, `TrialTelemetry`, and any other consumer that wants\n * append-only durable telemetry without rolling its own lock.\n *\n * Locks are per absolute file path (process-local). Cross-process\n * concurrency is NOT addressed — that's an fcntl/flock problem.\n */\n\nimport { appendFileSync, existsSync, mkdirSync } from 'node:fs'\nimport { dirname } from 'node:path'\nimport { Mutex } from './concurrency'\n\nconst mutexes = new Map<string, Mutex>()\n\nfunction getMutex(path: string): Mutex {\n  let m = mutexes.get(path)\n  if (!m) {\n    m = new Mutex()\n    mutexes.set(path, m)\n  }\n  return m\n}\n\nexport class LockedJsonlAppender {\n  private readonly mutex: Mutex\n  constructor(public readonly path: string) {\n    this.mutex = getMutex(path)\n    if (!existsSync(dirname(path))) {\n      mkdirSync(dirname(path), { recursive: true })\n    }\n  }\n\n  async append(entry: unknown): Promise<void> {\n    const line = `${JSON.stringify(entry)}\\n`\n    await this.mutex.runExclusive(() => {\n      appendFileSync(this.path, line)\n    })\n  }\n}\n\n/** Reset all internal mutex state — tests only. */\nexport function resetLockedAppendersForTesting(): void {\n  mutexes.clear()\n}\n","/**\n * FindingsStore — durable persistence for AnalystFinding rows + a diff\n * helper so we can answer \"what changed since the last run?\" without\n * recomputing analysts.\n *\n * On-disk shape is JSONL: one finding per line, append-only, locked via\n * LockedJsonlAppender. Operators get crash-safety (no partial JSON),\n * cheap reads (sequential parse), and trivial backup (rsync the file).\n *\n * Reads are non-locking: a reader sees a consistent snapshot of all\n * fully-written lines and skips an incomplete trailing line if the\n * writer is mid-append. Cross-process locking is intentionally out of\n * scope (see locked-jsonl-appender.ts).\n *\n * The store is run-scoped: callers pass `runId` on append and on load,\n * which keeps multi-run files cleanly partitioned. The `diffFindings`\n * helper compares two run-id sets using stable `finding_id` semantics —\n * the diff is the cross-run signal the regression dashboard renders.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\n\nimport { LockedJsonlAppender } from '../locked-jsonl-appender'\nimport type { AnalystFinding } from './types'\n\n/**\n * One persisted row. We attach `run_id` on disk so a single file can\n * hold multiple runs and the diff helper can query without re-walking\n * separate files.\n */\nexport interface PersistedFinding extends AnalystFinding {\n  run_id: string\n}\n\nexport class FindingsStore {\n  private readonly appender: LockedJsonlAppender\n\n  constructor(public readonly path: string) {\n    this.appender = new LockedJsonlAppender(path)\n  }\n\n  async append(runId: string, findings: AnalystFinding[]): Promise<void> {\n    for (const f of findings) {\n      const row: PersistedFinding = { ...f, run_id: runId }\n      await this.appender.append(row)\n    }\n  }\n\n  /** Load every persisted finding. Discards malformed trailing lines silently. */\n  loadAll(): PersistedFinding[] {\n    if (!existsSync(this.path)) return []\n    const raw = readFileSync(this.path, 'utf8')\n    if (!raw) return []\n    const out: PersistedFinding[] = []\n    for (const line of raw.split('\\n')) {\n      if (!line) continue\n      try {\n        out.push(JSON.parse(line) as PersistedFinding)\n      } catch {\n        // Skip torn trailing line — the lock guarantees no torn lines\n        // mid-file, only at EOF when a writer is in-flight.\n      }\n    }\n    return out\n  }\n\n  /** Filter to a single run. */\n  loadRun(runId: string): PersistedFinding[] {\n    return this.loadAll().filter((r) => r.run_id === runId)\n  }\n}\n\n// ── Cross-run diff ──────────────────────────────────────────────────\n\nexport interface FindingsDiff {\n  /** New finding ids in `current` that weren't in `previous`. */\n  appeared: PersistedFinding[]\n  /** Finding ids in `previous` that aren't in `current`. */\n  disappeared: PersistedFinding[]\n  /** Same finding id present in both runs and unchanged per the materiality test. */\n  persisted: PersistedFinding[]\n  /**\n   * Same finding id in both runs but at least one non-identity field\n   * shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].\n   */\n  changed: Array<{ previous: PersistedFinding; current: PersistedFinding }>\n}\n\nexport interface DiffPolicy {\n  /**\n   * Predicate that decides whether two findings (same finding_id) count\n   * as a material change. Defaults to {@link defaultIsMaterial}: severity\n   * shift, confidence Δ > 0.05, or evidence count change. Compliance /\n   * perf consumers MAY supply a stricter predicate (e.g. rationale text\n   * diff, metric Δ thresholds).\n   */\n  isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean\n}\n\n/**\n * Default materiality test. Deliberately narrow so LLM-reword churn\n * doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.\n */\nexport function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean {\n  if (a.severity !== b.severity) return true\n  if (Math.abs((a.confidence ?? 0) - (b.confidence ?? 0)) > 0.05) return true\n  if (a.evidence_refs.length !== b.evidence_refs.length) return true\n  return false\n}\n\n/**\n * Diff two findings sets by stable finding_id. Callers typically load\n * the two run-id slices from the same store and pass them in.\n */\nexport function diffFindings(\n  previous: PersistedFinding[],\n  current: PersistedFinding[],\n  policy: DiffPolicy = {},\n): FindingsDiff {\n  const isMaterial = policy.isMaterial ?? defaultIsMaterial\n  const prevById = new Map(previous.map((f) => [f.finding_id, f]))\n  const curById = new Map(current.map((f) => [f.finding_id, f]))\n\n  const appeared: PersistedFinding[] = []\n  const disappeared: PersistedFinding[] = []\n  const persisted: PersistedFinding[] = []\n  const changed: FindingsDiff['changed'] = []\n\n  for (const [id, cur] of curById) {\n    const prev = prevById.get(id)\n    if (!prev) {\n      appeared.push(cur)\n      continue\n    }\n    if (isMaterial(prev, cur)) {\n      changed.push({ previous: prev, current: cur })\n    } else {\n      persisted.push(cur)\n    }\n  }\n  for (const [id, prev] of prevById) {\n    if (!curById.has(id)) disappeared.push(prev)\n  }\n  return { appeared, disappeared, persisted, changed }\n}\n","/**\n * Skill-usage analyst — a DETERMINISTIC `Analyst` over a Claude/Codex skill\n * library + its trace corpus. Unlike the trace-store kinds (failure-mode,\n * improvement, ...) this kind calls no LLM: it mines real usage and skill\n * structure and emits findings by rule.\n *\n * It exists because the naive \"Skill-tool invocation count\" lies low — it\n * misses orchestrated sub-dispatch (a leaf skill run BY /pursue or /governor\n * logs under the parent), slash-command entry, local-script bypass, and\n * on-disk artifacts. The 2026-05-30 skill audit found 39/53 skills at zero\n * direct invocations, yet only one was a genuine cut: the rest were\n * measurement-invisible or discovery-limited. This analyst encodes that\n * lesson as a multi-signal usage model so a cheap repeatable pass can keep\n * the library honest, and so the expensive audit workflow's verdicts can\n * GEPA-distill it toward agreement (see `gold/skill-verdicts.gold.jsonl`).\n *\n * Report-building (`buildSkillUsageReport`, an fs scan) is separated from\n * finding emission (`SkillUsageAnalyst.analyze`, pure) so the slow scan runs\n * once at the registry boundary and the rule logic stays unit-testable.\n */\n\nimport { type Dirent, existsSync, readdirSync, readFileSync, statSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type { Analyst, AnalystContext, AnalystFinding, AnalystSeverity } from '../types'\nimport { computeFindingId } from '../types'\n\n// ── Input model ──────────────────────────────────────────────────────\n\nexport type SkillKind = 'public' | 'private'\n\n/** One skill's multi-signal usage + structure. All counts are deterministic. */\nexport interface SkillUsageRecord {\n  name: string\n  kind: SkillKind\n  /** Absolute path to the skill's SKILL.md. */\n  path: string\n  lines: number\n  /** `\"skill\":\"<name>\"` Skill-tool invocations across the trace corpus. */\n  directInvocations: number\n  /** `<command-name>/<name>` slash invocations across the trace corpus. */\n  slashInvocations: number\n  /** Sibling skills whose SKILL.md dispatches to this one (`/<name>`). Proxy\n   *  for orchestrated sub-dispatch the per-skill counter cannot see. */\n  inboundRefs: number\n  /** On-disk artifacts attributable to the skill (e.g. `.evolve/<name>/**`). */\n  artifactCount: number\n  /** Tangle-private reference count in the body (leak signal for public skills). */\n  tanglePrivateRefs: number\n  hasReferencesDir: boolean\n  hasEvalsDir: boolean\n  /** Body mentions `skill-runs.jsonl` (visible to /reflect + /governor). */\n  logsRuns: boolean\n  /** Description carries an explicit `Triggers:` clause / trigger phrases. */\n  hasTriggerPhrases: boolean\n}\n\nexport interface SkillUsageReport {\n  generatedFromTraces: number\n  records: SkillUsageRecord[]\n}\n\nexport interface SkillUsageScanConfig {\n  /** Dirs holding `*.jsonl` transcripts (Claude `~/.claude/projects`, Codex sessions). */\n  transcriptDirs: string[]\n  /** Skill roots to scan; each dir directly under `root` with a `SKILL.md` is a skill. */\n  skillRoots: { root: string; kind: SkillKind }[]\n  /** Roots scanned for `<root>/.evolve/<skill>` artifact dirs. */\n  artifactRoots?: string[]\n  /** Token-prefixed mappings: skill name → extra artifact subpaths under an artifactRoot\n   *  (e.g. reflect → `.evolve/reflections`). Catches non-eponymous artifact dirs. */\n  artifactAliases?: Record<string, string[]>\n  /** Cap files read per transcript dir (bounds a huge corpus); 0 = unbounded. */\n  maxTranscriptsPerDir?: number\n}\n\n// ── Deterministic thresholds ─────────────────────────────────────────\n\n/** Anthropic's authoring guidance keeps SKILL.md short; past this with no\n *  `references/` split the body burns context budget every session. */\nconst BLOAT_LINE_THRESHOLD = 300\n\nconst TANGLE_PRIVATE_RE =\n  /\\b(cli-bridge|tangletools|ops-board|drew-gtr-pro|@tangle-network\\/|~\\/company|tangle\\.tools|gtm-agent)\\b|\\bkimi\\b|\\btcloud\\b/gi\nconst TRIGGER_RE = /triggers?\\s*[:-]/i\n\n// ── Report builder (fs scan — slow, runs once at the registry boundary) ──\n\nfunction listSkillDirs(root: string): { name: string; path: string }[] {\n  if (!existsSync(root)) return []\n  const out: { name: string; path: string }[] = []\n  for (const entry of readdirSync(root, { withFileTypes: true })) {\n    if (!entry.isDirectory() && !entry.isSymbolicLink()) continue\n    const skillMd = join(root, entry.name, 'SKILL.md')\n    if (existsSync(skillMd)) out.push({ name: entry.name, path: skillMd })\n  }\n  return out\n}\n\nfunction walkJsonl(dir: string, cap: number): string[] {\n  if (!existsSync(dir)) return []\n  const files: string[] = []\n  const stack = [dir]\n  while (stack.length) {\n    const cur = stack.pop()!\n    let entries: Dirent[]\n    try {\n      entries = readdirSync(cur, { withFileTypes: true })\n    } catch {\n      continue\n    }\n    for (const e of entries) {\n      const full = join(cur, e.name)\n      if (e.isDirectory()) stack.push(full)\n      else if (e.name.endsWith('.jsonl')) {\n        files.push(full)\n        if (cap > 0 && files.length >= cap) return files\n      }\n    }\n  }\n  return files\n}\n\nfunction frontmatterDescription(body: string): string {\n  const fm = /^---\\n([\\s\\S]*?)\\n---/.exec(body)\n  const block = fm?.[1] ?? ''\n  const m = /description:\\s*(.+)/i.exec(block)\n  return m?.[1] ?? ''\n}\n\nfunction countArtifacts(roots: string[], name: string, aliases: string[]): number {\n  let n = 0\n  for (const root of roots) {\n    const candidates = [join(root, '.evolve', name), ...aliases.map((a) => join(root, a))]\n    for (const dir of candidates) {\n      if (!existsSync(dir)) continue\n      try {\n        if (statSync(dir).isDirectory()) n += readdirSync(dir).length\n        else n += 1\n      } catch {\n        /* unreadable — skip */\n      }\n    }\n  }\n  return n\n}\n\n/** Scan the corpus + skill roots into a {@link SkillUsageReport}. Deterministic. */\nexport function buildSkillUsageReport(config: SkillUsageScanConfig): SkillUsageReport {\n  const skills = config.skillRoots.flatMap(({ root, kind }) =>\n    listSkillDirs(root).map((s) => ({ ...s, kind })),\n  )\n  const names = skills.map((s) => s.name)\n\n  // One pass over the corpus accumulating direct + slash counts per skill.\n  const direct = new Map<string, number>(names.map((n) => [n, 0]))\n  const slash = new Map<string, number>(names.map((n) => [n, 0]))\n  const skillRe = /\"skill\"\\s*:\\s*\"([a-z0-9_:-]+)\"/g\n  const cmdRe = /<command-name>\\/?([a-z0-9_:-]+)<\\/command-name>/g\n  let transcripts = 0\n  for (const dir of config.transcriptDirs) {\n    for (const file of walkJsonl(dir, config.maxTranscriptsPerDir ?? 0)) {\n      transcripts += 1\n      let data: string\n      try {\n        data = readFileSync(file, 'utf8')\n      } catch {\n        continue\n      }\n      for (const m of data.matchAll(skillRe)) {\n        const g = m[1]\n        if (!g) continue\n        const n = g.split(':').pop() ?? g\n        const prev = direct.get(n)\n        if (prev !== undefined) direct.set(n, prev + 1)\n      }\n      for (const m of data.matchAll(cmdRe)) {\n        const g = m[1]\n        if (g === undefined) continue\n        const prev = slash.get(g)\n        if (prev !== undefined) slash.set(g, prev + 1)\n      }\n    }\n  }\n\n  // Read each skill body once; compute structure + inbound refs across siblings.\n  const bodies = new Map<string, string>()\n  for (const s of skills) {\n    try {\n      bodies.set(s.name, readFileSync(s.path, 'utf8'))\n    } catch {\n      bodies.set(s.name, '')\n    }\n  }\n  const inbound = new Map<string, number>(names.map((n) => [n, 0]))\n  for (const target of names) {\n    const ref = new RegExp(`/${target}\\\\b|\\\\[\\\\[${target}\\\\]\\\\]`)\n    for (const s of skills) {\n      if (s.name === target) continue\n      if (ref.test(bodies.get(s.name) ?? '')) inbound.set(target, inbound.get(target)! + 1)\n    }\n  }\n\n  const records: SkillUsageRecord[] = skills.map((s) => {\n    const body = bodies.get(s.name) ?? ''\n    const dir = s.path.replace(/\\/SKILL\\.md$/, '')\n    return {\n      name: s.name,\n      kind: s.kind,\n      path: s.path,\n      lines: body ? body.split('\\n').length : 0,\n      directInvocations: direct.get(s.name) ?? 0,\n      slashInvocations: slash.get(s.name) ?? 0,\n      inboundRefs: inbound.get(s.name) ?? 0,\n      artifactCount: countArtifacts(\n        config.artifactRoots ?? [],\n        s.name,\n        config.artifactAliases?.[s.name] ?? [],\n      ),\n      tanglePrivateRefs: (body.match(TANGLE_PRIVATE_RE) ?? []).length,\n      hasReferencesDir: existsSync(join(dir, 'references')),\n      hasEvalsDir: existsSync(join(dir, 'evals')),\n      logsRuns: body.includes('skill-runs.jsonl'),\n      hasTriggerPhrases: TRIGGER_RE.test(frontmatterDescription(body) || body.slice(0, 600)),\n    }\n  })\n  return { generatedFromTraces: transcripts, records }\n}\n\n// ── Finding emission (pure — unit-testable, no LLM, no fs) ────────────\n\nconst ANALYST_ID = 'skill-usage'\n\nfunction finding(\n  area: string,\n  subject: string,\n  claim: string,\n  severity: AnalystSeverity,\n  confidence: number,\n  producedAt: string,\n  recommended: string,\n  evidenceUri: string,\n  rationale?: string,\n): AnalystFinding {\n  return {\n    schema_version: '1.0.0',\n    finding_id: computeFindingId({ analyst_id: ANALYST_ID, area, subject, claim }),\n    analyst_id: ANALYST_ID,\n    produced_at: producedAt,\n    severity,\n    area,\n    claim,\n    rationale,\n    evidence_refs: [{ kind: 'artifact', uri: evidenceUri }],\n    recommended_action: recommended,\n    confidence,\n    subject,\n  }\n}\n\n/** Pure rule pass over a report → findings. Exported for direct/unit use. */\nexport function emitSkillUsageFindings(\n  report: SkillUsageReport,\n  producedAt: string,\n): AnalystFinding[] {\n  const out: AnalystFinding[] = []\n  for (const r of report.records) {\n    const directTotal = r.directInvocations + r.slashInvocations\n    const trueUsage = directTotal + r.inboundRefs + r.artifactCount\n\n    // 1. Dead: no usage signal of ANY kind. The only real deprecation candidate.\n    if (trueUsage === 0) {\n      out.push(\n        finding(\n          'skill-usage',\n          r.name,\n          `Skill '${r.name}' has zero usage across all signals (direct, slash, inbound-refs, artifacts)`,\n          'high',\n          0.6,\n          producedAt,\n          'Confirm the skill covers a real recurring job; if not, deprecate. Zero true usage is the only deterministic deprecation candidate.',\n          r.path,\n          'No Skill-tool call, no slash invocation, no sibling dispatches to it, and no on-disk artifacts.',\n        ),\n      )\n    } else if (directTotal === 0 && r.inboundRefs + r.artifactCount > 0) {\n      // 2. Measurement-invisible: real use via orchestration/artifacts, never invoked directly.\n      out.push(\n        finding(\n          'skill-usage',\n          r.name,\n          `Skill '${r.name}' shows 0 direct invocations but is used via orchestration/artifacts (inbound=${r.inboundRefs}, artifacts=${r.artifactCount})`,\n          'info',\n          0.8,\n          producedAt,\n          'Do NOT treat as unused — usage is real but logged under parent skills or on disk. Strengthen direct-invocation discovery only if direct use is desired.',\n          r.path,\n          'The Skill-tool counter undercounts orchestrated/chained leaf skills.',\n        ),\n      )\n    }\n\n    // 3. Discovery gap: low direct use AND weak trigger surface.\n    if (directTotal <= 2 && !r.hasTriggerPhrases) {\n      out.push(\n        finding(\n          'discoverability',\n          r.name,\n          `Skill '${r.name}' is rarely invoked directly and its description has no explicit trigger phrases`,\n          'medium',\n          0.7,\n          producedAt,\n          'Add a `Triggers:` clause with verbatim user phrases to the frontmatter description so the model auto-invokes it.',\n          r.path,\n        ),\n      )\n    }\n\n    // 4. Public-repo leak.\n    if (r.kind === 'public' && r.tanglePrivateRefs > 0) {\n      out.push(\n        finding(\n          'safety',\n          r.name,\n          `Public skill '${r.name}' carries ${r.tanglePrivateRefs} Tangle-private reference(s)`,\n          'high',\n          0.75,\n          producedAt,\n          'Sanitize incidental internal refs (cli-bridge/kimi/tcloud/~company/private repos) or relocate to a private repo. Verify @tangle-network/* refs are to PUBLISHED packages before treating as a leak.',\n          r.path,\n        ),\n      )\n    }\n\n    // 5. Bloat / no progressive disclosure.\n    if (r.lines > BLOAT_LINE_THRESHOLD && !r.hasReferencesDir) {\n      out.push(\n        finding(\n          'maintainability',\n          r.name,\n          `Skill '${r.name}' is ${r.lines} lines with no references/ split (progressive disclosure)`,\n          'medium',\n          0.8,\n          producedAt,\n          `Split detail into references/ loaded on demand; keep SKILL.md a short overview. ${r.lines} lines load into every session's context budget.`,\n          r.path,\n        ),\n      )\n    }\n\n    // 6. No evals (Anthropic's \">=3 evals before docs\" rule).\n    if (!r.hasEvalsDir) {\n      out.push(\n        finding(\n          'data-quality',\n          r.name,\n          `Skill '${r.name}' ships no evals/`,\n          'low',\n          0.6,\n          producedAt,\n          'Add evals/evals.json with >=3 scenarios proving the skill beats baseline; gives regression coverage.',\n          r.path,\n        ),\n      )\n    }\n\n    // 7. No run logging → invisible to /reflect and /governor.\n    if (!r.logsRuns) {\n      out.push(\n        finding(\n          'observability',\n          r.name,\n          `Skill '${r.name}' never appends to .evolve/skill-runs.jsonl`,\n          'low',\n          0.55,\n          producedAt,\n          'Append one run line to .evolve/skill-runs.jsonl on completion, or declare it a non-logging leaf, so the self-improvement loop can see it ran.',\n          r.path,\n        ),\n      )\n    }\n  }\n  return out\n}\n\n// ── The Analyst ──────────────────────────────────────────────────────\n\nexport class SkillUsageAnalyst implements Analyst<SkillUsageReport> {\n  readonly id = ANALYST_ID\n  readonly description =\n    'Deterministic multi-signal skill-usage analysis: flags dead skills, measurement-invisible (orchestrated) usage, discovery gaps, public-repo leaks, bloat, missing evals, and missing run-logging.'\n  readonly inputKind = 'custom' as const\n  readonly cost = { kind: 'deterministic' as const, est_usd_per_run: 0 }\n  readonly version = '1.0.0'\n\n  async analyze(input: SkillUsageReport, ctx: AnalystContext): Promise<AnalystFinding[]> {\n    const producedAt = ctx.tags?.producedAt ?? new Date().toISOString()\n    ctx.log?.(\n      `skill-usage: ${input.records.length} skills over ${input.generatedFromTraces} transcripts`,\n    )\n    return emitSkillUsageFindings(input, producedAt)\n  }\n}\n\nexport const SKILL_USAGE_ANALYST = new SkillUsageAnalyst()\n","import { NotFoundError } from './errors'\nimport { aggregateRunScore, clamp01, type RunScore, type RunScoreWeights } from './run-score'\nimport type { Artifact, BudgetLedgerEntry, Run, Span, TraceEvent, TraceStore } from './trace'\n\nexport interface RunTrace {\n  run: Run\n  spans: Span[]\n  events: TraceEvent[]\n  artifacts: Artifact[]\n  budget: BudgetLedgerEntry[]\n}\n\nexport interface RunCriticOptions {\n  weights?: Partial<RunScoreWeights>\n  driftPatterns?: RegExp[]\n}\n\nconst DEFAULT_DRIFT_PATTERNS = [\n  /https?:\\/\\//i,\n  /\\btitle:\\s/i,\n  /\\bsummary:\\s/i,\n  /\\burl:\\s/i,\n  /\\bnpm package usage\\b/i,\n  /\\bnews\\b/i,\n]\n\nexport class RunCritic {\n  private readonly weights?: Partial<RunScoreWeights>\n  private readonly driftPatterns: RegExp[]\n\n  constructor(options: RunCriticOptions = {}) {\n    this.weights = options.weights\n    this.driftPatterns = options.driftPatterns ?? DEFAULT_DRIFT_PATTERNS\n  }\n\n  async score(store: TraceStore, runId: string): Promise<RunScore> {\n    const run = await store.getRun(runId)\n    if (!run) throw new NotFoundError(`run ${runId} not found`)\n    const [spans, events, artifacts, budget] = await Promise.all([\n      store.spans({ runId }),\n      store.events({ runId }),\n      store.artifacts(runId),\n      store.budget(runId),\n    ])\n    return this.scoreTrace({ run, spans, events, artifacts, budget })\n  }\n\n  scoreTrace(trace: RunTrace): RunScore {\n    const notes: string[] = []\n    const llmSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'llm' }> => s.kind === 'llm',\n    )\n    const toolSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'tool' }> => s.kind === 'tool',\n    )\n    const judgeSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'judge' }> => s.kind === 'judge',\n    )\n    const sandboxSpans = trace.spans.filter(\n      (s): s is Extract<Span, { kind: 'sandbox' }> => s.kind === 'sandbox',\n    )\n    const finalGateSpans = judgeSpans.filter(\n      (span) => span.dimension === 'final_gate' || span.attributes?.finalGate === true,\n    )\n\n    const success =\n      trace.run.outcome?.pass === true ? 1 : trace.run.status === 'completed' ? 0.5 : 0\n    if (!success) notes.push('run did not complete with pass=true')\n\n    const judgeAverage = judgeSpans.length\n      ? judgeSpans.reduce((sum, span) => sum + normalizeJudgeScore(span.score), 0) /\n        judgeSpans.length\n      : undefined\n    const outcomeScore =\n      typeof trace.run.outcome?.score === 'number'\n        ? clamp01(\n            trace.run.outcome.score > 1 ? trace.run.outcome.score / 100 : trace.run.outcome.score,\n          )\n        : undefined\n    const goalProgress = outcomeScore ?? judgeAverage ?? success\n\n    const successfulTools = toolSpans.filter((span) => span.status !== 'error').length\n    const toolUseQuality = toolSpans.length === 0 ? 0 : successfulTools / toolSpans.length\n    if (toolSpans.length === 0) notes.push('no tool spans recorded')\n\n    const patchEvidence =\n      trace.artifacts.length +\n      toolSpans.filter((span) => /write|edit|patch|apply/i.test(span.toolName)).length\n    const patchQuality = patchEvidence > 0 ? clamp01(patchEvidence / 4) : 0\n    if (!patchQuality) notes.push('no artifact or edit evidence recorded')\n\n    const sandboxTests = sandboxSpans.filter(\n      (span) => typeof span.testsTotal === 'number' && span.testsTotal > 0,\n    )\n    const testReality = sandboxTests.length\n      ? sandboxTests.reduce(\n          (sum, span) => sum + (span.testsPassed ?? 0) / Math.max(1, span.testsTotal ?? 1),\n          0,\n        ) / sandboxTests.length\n      : toolSpans.some((span) =>\n            /\\btest|vitest|pytest|jest|build|tsc\\b/i.test(JSON.stringify(span.args)),\n          )\n        ? 0.4\n        : 0\n    if (!testReality) notes.push('no real test/build evidence recorded')\n\n    const blockerSpans = judgeSpans.filter((span) => isBlockingJudge(span))\n    const finalGateBlockers = finalGateSpans.filter((span) => isBlockingJudge(span))\n    const finalGate = finalGateSpans.length ? (finalGateBlockers.length ? 0 : 1) : success\n    if (finalGateBlockers.length)\n      notes.push(`final gate blocked by ${finalGateBlockers.length} reviewer(s)`)\n    else if (!finalGateSpans.length) notes.push('no final gate judgment recorded')\n\n    const reviewerBlockers = judgeSpans.length ? blockerSpans.length / judgeSpans.length : 0\n    if (reviewerBlockers) notes.push(`detected ${blockerSpans.length} blocking reviewer signal(s)`)\n\n    const positiveGroundingSignals =\n      patchEvidence +\n      sandboxSpans.length +\n      llmSpans.filter((span) => looksRepoGrounded(span.output ?? '')).length\n    const driftSignals =\n      llmSpans.filter((span) => this.isDrift(span.output ?? '')).length +\n      trace.events.filter((event) => this.isDrift(JSON.stringify(event.payload))).length\n    const repoGroundedness =\n      positiveGroundingSignals + driftSignals === 0\n        ? 0\n        : positiveGroundingSignals / (positiveGroundingSignals + driftSignals)\n    const driftPenalty =\n      positiveGroundingSignals + driftSignals === 0\n        ? 0\n        : driftSignals / (positiveGroundingSignals + driftSignals)\n    if (driftSignals > 0) notes.push(`detected ${driftSignals} drift signal(s)`)\n\n    const costUsd = trace.budget.length\n      ? Math.max(\n          ...trace.budget\n            .filter((entry: BudgetLedgerEntry) => entry.dimension === 'usd')\n            .map((entry: BudgetLedgerEntry) => entry.consumed),\n          0,\n        )\n      : llmSpans.reduce((sum, span) => sum + (span.costUsd ?? 0), 0)\n    const wallSeconds =\n      trace.run.endedAt && trace.run.startedAt\n        ? Math.max(0, (trace.run.endedAt - trace.run.startedAt) / 1000)\n        : 0\n\n    return {\n      success,\n      goalProgress,\n      repoGroundedness,\n      driftPenalty,\n      toolUseQuality,\n      patchQuality,\n      testReality,\n      finalGate,\n      reviewerBlockers,\n      costUsd,\n      wallSeconds,\n      notes,\n    }\n  }\n\n  rank(score: RunScore): number {\n    return aggregateRunScore(score, this.weights)\n  }\n\n  private isDrift(text: string): boolean {\n    return this.driftPatterns.some((pattern) => pattern.test(text))\n  }\n}\n\nfunction normalizeJudgeScore(score: number): number {\n  return score > 1 ? clamp01(score / 10) : clamp01(score)\n}\n\nfunction looksRepoGrounded(text: string): boolean {\n  return /(?:src\\/|tests?\\/|package\\.json|tsconfig|\\.ts\\b|\\.tsx\\b|git status|pnpm |npm |vitest|pytest|jest)/i.test(\n    text,\n  )\n}\n\nfunction isBlockingJudge(span: Extract<Span, { kind: 'judge' }>): boolean {\n  return (\n    span.attributes?.blocking === true ||\n    span.attributes?.verdict === 'BLOCKING' ||\n    positiveNumber(span.attributes?.blockingFindings) ||\n    positiveNumber(span.attributes?.highFindings) ||\n    span.score <= 2\n  )\n}\n\nfunction positiveNumber(value: unknown): boolean {\n  return typeof value === 'number' && value > 0\n}\n","/**\n * Semantic concept judge — \"does the built artifact actually implement\n * the features the user asked for?\"\n *\n * Distinct from the domain/code/coherence judges in `judges.ts`:\n *   - those judges score free-form conversational agent outputs along\n *     quality dimensions (accuracy, depth, etc.)\n *   - this judge scores a *built artifact* (served HTML + source files)\n *     against an explicit list of expected concepts, returning per-concept\n *     {present, score 0-10, evidence, severity}.\n *\n * The judge is strict about distinguishing (a) a working implementation\n * from (b) a keyword-present stub. \"// TODO: mint button\" is NOT present.\n * Only real, functional, wired-up code counts.\n *\n * Use via {@link createSemanticConceptJudge} or directly via\n * {@link runSemanticConceptJudge}. Soft-fails (available=false) on LLM\n * or JSON-parse errors so the caller can treat that as \"layer skipped\"\n * rather than \"layer failed\" in a multi-layer pipeline.\n */\n\nimport { callLlmJson, type LlmClientOptions } from './llm-client'\nimport type { Severity } from './multi-layer-verifier'\n\n// ─── Types ──────────────────────────────────────────────────────────────\n\n/**\n * Implementation complexity class for weighted scoring.\n *\n * - `render` (default): the concept is a UI surface that displays static\n *   data — render a list, show a counter, lay out a button. Single-file\n *   work, no external integration.\n * - `integrate`: the concept requires wiring a real external system —\n *   wallet connect (wagmi + RainbowKit + chain config), payment provider\n *   (Stripe Elements + intent + webhook), an API client with auth.\n *   Multi-file, library-knowledge, runtime correctness matters.\n * - `compute`: the concept requires algorithmic work — solver, simulator,\n *   constraint propagation, ML inference. Correctness > UI polish.\n *\n * Default weights (when applied via `weightConcepts: 'complexity'`):\n *   render=1.0, integrate=2.0, compute=2.5\n *\n * Cross-vertical scoring without complexity weighting silently inflates\n * the rate of UI-heavy verticals (healthcare, fintech dashboards) vs\n * integration-heavy verticals (DeFi, wallets) — all concepts treated\n * equally even though the agent does 2-3x the work for `integrate`.\n */\nexport type ConceptComplexity = 'render' | 'integrate' | 'compute'\n\nexport interface ConceptSpec {\n  name: string\n  /** Short hints that help the judge; not used for matching. */\n  keywords?: string[]\n  /** Optional explicit weight; default 1.0. Overrides complexity-derived weight. */\n  weight?: number\n  /** Implementation complexity class. Default `render`. */\n  complexity?: ConceptComplexity\n}\n\nexport interface ConceptFinding {\n  concept: string\n  present: boolean\n  /** 0..10. 10 = production-ready; 7 = functional thin; 4 = partial; 0 = absent. */\n  score: number\n  evidence: string\n  severity: Severity\n}\n\nexport interface SemanticConceptJudgeInput {\n  /** Full natural-language prompt the agent was handed. */\n  userRequest: string\n  /** Rendered HTML the preview returns (UI artifacts). Optional. */\n  servedHtml?: string\n  /** Top-level source files from the agent's workdir. */\n  sourceFiles: Array<{ path: string; content: string }>\n  /** The expected concept list. */\n  expectedConcepts: ConceptSpec[]\n  /** Free-form metadata (id, difficulty) to inject into the prompt. */\n  artifactLabel?: string\n  artifactDescription?: string\n}\n\nexport interface SemanticConceptJudgeResult {\n  kind: 'semantic-concept'\n  version: string\n  /** Normalized 0..1 score — mean of per-concept scores / 10. */\n  score: number\n  presentCount: number\n  totalCount: number\n  findings: ConceptFinding[]\n  summary: string\n  durationMs: number\n  costUsd: number | null\n  /** False on LLM/JSON error — treat as \"skipped / unable to judge\" in pipelines. */\n  available: boolean\n  error?: string\n}\n\n/**\n * Score-aggregation strategy. `mean` averages 0-10 scores uniformly.\n * `complexity` applies the default weight table (render=1, integrate=2,\n * compute=2.5) unless a concept has an explicit `weight`. `explicit`\n * honors only `weight` (defaulting to 1 for unspecified).\n */\nexport type ConceptWeightStrategy = 'mean' | 'complexity' | 'explicit'\n\nexport const DEFAULT_COMPLEXITY_WEIGHTS: Record<ConceptComplexity, number> = {\n  render: 1.0,\n  integrate: 2.0,\n  compute: 2.5,\n}\n\nexport interface SemanticConceptJudgeOptions {\n  /** Model id to call. Default 'claude-sonnet-4-6' via agent-eval defaults. */\n  model?: string\n  /** Per-call timeout. Default 300s. */\n  timeoutMs?: number\n  /** Pipeline budget for the prompt (source blob truncation). Default 45000. */\n  maxSourceChars?: number\n  /** Per-file cap before inclusion. Default 20000. */\n  maxPerFileChars?: number\n  /** HTML cap. Default 30000. */\n  maxHtmlChars?: number\n  /** LlmClient config (baseUrl, apiKey, authHeader, …). */\n  llm?: LlmClientOptions\n  /**\n   * Score aggregation strategy. Default `mean` — uniform average across\n   * concepts. Cross-vertical comparisons should use `complexity` to\n   * neutralize the integrate-vs-render asymmetry.\n   */\n  weightConcepts?: ConceptWeightStrategy\n  /** Override the default complexity → weight table. */\n  complexityWeights?: Partial<Record<ConceptComplexity, number>>\n}\n\n// ─── Prompt assembly ────────────────────────────────────────────────────\n\nexport const SEMANTIC_CONCEPT_JUDGE_VERSION = 'semantic-concept-judge-v1-2026-04-24'\n\nconst DEFAULT_MAX_SOURCE = 45_000\nconst DEFAULT_MAX_HTML = 30_000\nconst DEFAULT_MAX_PER_FILE = 20_000\nconst DEFAULT_TIMEOUT = 300_000\nconst DEFAULT_MODEL = 'claude-sonnet-4-6'\n\nconst SEMANTIC_SCHEMA = {\n  type: 'object',\n  additionalProperties: false,\n  required: ['summary', 'concepts'],\n  properties: {\n    summary: { type: 'string', minLength: 20, maxLength: 600 },\n    concepts: {\n      type: 'array',\n      minItems: 1,\n      items: {\n        type: 'object',\n        additionalProperties: false,\n        required: ['concept', 'present', 'score', 'evidence', 'severity'],\n        properties: {\n          concept: { type: 'string', minLength: 1, maxLength: 120 },\n          present: { type: 'boolean' },\n          score: { type: 'number', minimum: 0, maximum: 10 },\n          evidence: { type: 'string', minLength: 5, maxLength: 400 },\n          severity: { type: 'string', enum: ['critical', 'major', 'minor', 'info'] },\n        },\n      },\n    },\n  },\n}\n\nfunction truncate(body: string, cap: number, label: string): string {\n  if (body.length <= cap) return body\n  return `${body.slice(0, cap)}\\n… [truncated ${body.length - cap} chars of ${label}]`\n}\n\nfunction buildPrompt(\n  input: SemanticConceptJudgeInput,\n  opts: Required<SemanticConceptJudgeOptions>,\n): string {\n  const sourceBlob = input.sourceFiles\n    .filter((f) => f.content.length <= opts.maxPerFileChars)\n    .map((f) => `--- FILE: ${f.path} ---\\n${f.content}`)\n    .join('\\n\\n')\n\n  const html = input.servedHtml ?? ''\n\n  return `You are a strict code-review judge evaluating whether an agent's 0-to-1 build actually implements the features the user asked for.\n\nYou MUST distinguish:\n  (a) WORKING code that implements the concept (rendered UI, wired handler, real API call),\n  (b) KEYWORD-PRESENT stub (comments mentioning the concept, variable names, TODOs),\n  (c) ABSENT (concept nowhere).\n\nA comment like \"// TODO: add mint button\" is NOT present — score 2-3. Only count a concept as present if there is real functional code: a rendered component, a call handler wired to state or a network call, a computed value actually used.\n\nUSER REQUEST (what the agent was asked to build):\n${input.userRequest}\n\n${input.artifactLabel ? `ARTIFACT METADATA:\\n  name: ${input.artifactLabel}\\n  description: ${input.artifactDescription ?? ''}\\n\\n` : ''}EXPECTED CONCEPTS (each must be graded independently):\n${input.expectedConcepts\n  .map(\n    (c, i) =>\n      `  ${i + 1}. \"${c.name}\"${c.keywords?.length ? ` — hints: [${c.keywords.slice(0, 6).join(' | ')}]` : ''}`,\n  )\n  .join('\\n')}\n\n${html ? `SERVED HTML (what the preview returns when hit):\\n${truncate(html, opts.maxHtmlChars, 'HTML')}\\n\\n` : ''}SOURCE FILES (the agent's workdir):\n${truncate(sourceBlob, opts.maxSourceChars, 'source')}\n\nFor EACH concept, return:\n  - concept: the concept name as given (match exactly)\n  - present: boolean — does a working implementation exist?\n  - score: 0-10 — 10 = production-ready; 7 = functional but thin; 4 = partial/stubbed; 2 = keyword-only comment; 0 = absent\n  - evidence: cite \"<file>:<line>\" or \"served-html:<selector>\" pointing at the strongest supporting code. If the concept is absent or stubbed, explain what's missing.\n  - severity:\n      \"info\" when present: true AND score >= 7\n      \"minor\" when present: true AND 4 <= score < 7\n      \"major\" when present: false OR score < 4\n      \"critical\" when the concept is not only absent but a core user flow depends on it\n\nAlso produce a \"summary\" (one sentence, 20-600 chars): overall verdict on whether this is a shippable implementation of the user request vs a keyword-dense placeholder.\n\nBE SKEPTICAL. Keyword matching already passed — your job is to catch what keyword matching misses. If the agent shipped a working build, say so. If it shipped a stub, say so. Don't grade on effort.\n\nReturn STRICT JSON. No prose outside the JSON.`\n}\n\n// ─── Runner ─────────────────────────────────────────────────────────────\n\n/**\n * Run the semantic concept judge. Soft-fails to available=false on\n * LLM/JSON errors — callers in a MultiLayerVerifier pipeline can treat\n * that as \"skip\" rather than \"fail.\"\n */\nexport async function runSemanticConceptJudge(\n  input: SemanticConceptJudgeInput,\n  options: SemanticConceptJudgeOptions = {},\n): Promise<SemanticConceptJudgeResult> {\n  const start = Date.now()\n  const totalCount = input.expectedConcepts.length\n\n  if (totalCount === 0) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount: 0,\n      findings: [],\n      summary: 'no expected concepts declared',\n      durationMs: 0,\n      costUsd: null,\n      available: false,\n      error: 'no expected concepts declared',\n    }\n  }\n\n  const opts: Required<SemanticConceptJudgeOptions> = {\n    model: options.model ?? DEFAULT_MODEL,\n    timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT,\n    maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE,\n    maxPerFileChars: options.maxPerFileChars ?? DEFAULT_MAX_PER_FILE,\n    maxHtmlChars: options.maxHtmlChars ?? DEFAULT_MAX_HTML,\n    llm: options.llm ?? {},\n    weightConcepts: options.weightConcepts ?? 'mean',\n    complexityWeights: { ...DEFAULT_COMPLEXITY_WEIGHTS, ...(options.complexityWeights ?? {}) },\n  }\n\n  // Build a name → weight map for aggregation. Mean strategy keeps every\n  // weight at 1 (uniform average). Complexity strategy reads the table\n  // and lets an explicit `weight` override. Explicit strategy uses ONLY\n  // the spec's `weight` (defaulting to 1).\n  const weightForConcept = (spec: ConceptSpec): number => {\n    if (opts.weightConcepts === 'mean') return 1\n    if (spec.weight != null) return spec.weight\n    if (opts.weightConcepts === 'complexity') {\n      return opts.complexityWeights[spec.complexity ?? 'render'] ?? 1\n    }\n    return 1\n  }\n  const weightByName = new Map<string, number>(\n    input.expectedConcepts.map((c) => [c.name, weightForConcept(c)]),\n  )\n\n  try {\n    const { value, result } = await callLlmJson<{\n      summary: string\n      concepts: ConceptFinding[]\n    }>(\n      {\n        model: opts.model,\n        messages: [\n          {\n            role: 'system',\n            content:\n              'You are a strict code-review judge. Return strict JSON only. No prose outside the JSON. A keyword in a comment is NOT a working implementation.',\n          },\n          { role: 'user', content: buildPrompt(input, opts) },\n        ],\n        jsonSchema: { name: 'semantic_concept_judge', schema: SEMANTIC_SCHEMA },\n        temperature: 0,\n        timeoutMs: opts.timeoutMs,\n      },\n      opts.llm,\n    )\n\n    if (!value?.concepts || !Array.isArray(value.concepts)) {\n      throw new Error('judge returned malformed response — expected array under \"concepts\"')\n    }\n\n    const findings: ConceptFinding[] = value.concepts.map((c) => ({\n      concept: String(c.concept),\n      present: Boolean(c.present),\n      score: Math.max(0, Math.min(10, Number(c.score ?? 0))),\n      evidence: String(c.evidence ?? ''),\n      severity: (['critical', 'major', 'minor', 'info'] as const).includes(c.severity)\n        ? c.severity\n        : 'info',\n    }))\n\n    const presentCount = findings.filter((f) => f.present && f.score >= 7).length\n    let weightSum = 0\n    let weightedScoreSum = 0\n    for (const f of findings) {\n      const w = weightByName.get(f.concept) ?? 1\n      weightSum += w\n      weightedScoreSum += w * f.score\n    }\n    const scoreAvg =\n      weightSum > 0\n        ? weightedScoreSum / weightSum\n        : findings.reduce((a, f) => a + f.score, 0) / Math.max(1, findings.length)\n\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: Number((scoreAvg / 10).toFixed(3)),\n      presentCount,\n      totalCount,\n      findings,\n      summary: String(value.summary ?? ''),\n      durationMs: Date.now() - start,\n      costUsd: result.costUsd ?? null,\n      available: true,\n    }\n  } catch (err) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount,\n      findings: [],\n      summary: '',\n      durationMs: Date.now() - start,\n      costUsd: null,\n      available: false,\n      error: err instanceof Error ? err.message : String(err),\n    }\n  }\n}\n\n/**\n * Factory: pin LLM options once, return a closure that accepts inputs.\n * Convenient for pipelines that want to share a single LlmClient config.\n */\nexport function createSemanticConceptJudge(\n  options: SemanticConceptJudgeOptions = {},\n): (input: SemanticConceptJudgeInput) => Promise<SemanticConceptJudgeResult> {\n  return (input) => runSemanticConceptJudge(input, options)\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AACA,SAAS,UAAU;AA2BZ,SAAS,gBAAgB,QAA4C;AAC1E,SAAO,GAAG;AAAA,IACR,MAAM,OAAO,YAAY;AAAA,IACzB,QAAQ,OAAO;AAAA,IACf,QAAQ,OAAO;AAAA,IACf,QAAQ,EAAE,OAAO,OAAO,MAAM;AAAA,EAChC,CAAC;AACH;;;ACqFO,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,QACf,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,QACzB,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,QACf,CAAqB;AAAA,MACvB;AAAA,IACF,KAAK;AACH,aAAO;AAAA,QACL,WAAW;AAAA,QACX,cAAc,KAAK;AAAA,QACnB,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,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,MAAM,OAAO,KAAK,aAAa;AAC7B,YAAM,WAAW,aAAa,KAAK,YAAY;AAO/C,YAAM,OAAO,MAAM,KAAK;AAAA,QACtB,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,CAAC;AACD,UAAI,CAAC,UAAU,OAAQ,QAAO,MAAM;AACpC,aAAO,MAAM,QAAQ,KAAK,CAAC,MAAM,iBAAiB,SAAS,MAAM,CAAC,CAAC;AAAA,IACrE;AAAA,EACF;AACF;AAEA,SAAS,iBAAiB,QAAqC;AAC7D,MAAI,OAAO,QAAS,QAAO,QAAQ,OAAO,aAAa,MAAM,CAAC;AAC9D,SAAO,IAAI,QAAe,CAAC,GAAG,WAAW;AACvC,WAAO,iBAAiB,SAAS,MAAM,OAAO,aAAa,MAAM,CAAC,GAAG,EAAE,MAAM,KAAK,CAAC;AAAA,EACrF,CAAC;AACH;AAEA,SAAS,aAAa,QAA4B;AAChD,QAAM,SAAU,OAAgC;AAChD,MAAI,kBAAkB,MAAO,QAAO;AACpC,QAAM,IAAI,IAAI,MAAM,0BAA0B;AAC9C,IAAE,OAAO;AACT,SAAO;AACT;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;;;AChNA,SAAS,gBAAgB,YAAY,iBAAiB;AACtD,SAAS,eAAe;AAGxB,IAAM,UAAU,oBAAI,IAAmB;AAEvC,SAAS,SAAS,MAAqB;AACrC,MAAI,IAAI,QAAQ,IAAI,IAAI;AACxB,MAAI,CAAC,GAAG;AACN,QAAI,IAAI,MAAM;AACd,YAAQ,IAAI,MAAM,CAAC;AAAA,EACrB;AACA,SAAO;AACT;AAEO,IAAM,sBAAN,MAA0B;AAAA,EAE/B,YAA4B,MAAc;AAAd;AAC1B,SAAK,QAAQ,SAAS,IAAI;AAC1B,QAAI,CAAC,WAAW,QAAQ,IAAI,CAAC,GAAG;AAC9B,gBAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AAAA,IAC9C;AAAA,EACF;AAAA,EAL4B;AAAA,EADX;AAAA,EAQjB,MAAM,OAAO,OAA+B;AAC1C,UAAM,OAAO,GAAG,KAAK,UAAU,KAAK,CAAC;AAAA;AACrC,UAAM,KAAK,MAAM,aAAa,MAAM;AAClC,qBAAe,KAAK,MAAM,IAAI;AAAA,IAChC,CAAC;AAAA,EACH;AACF;;;ACrBA,SAAS,cAAAA,aAAY,oBAAoB;AAclC,IAAM,gBAAN,MAAoB;AAAA,EAGzB,YAA4B,MAAc;AAAd;AAC1B,SAAK,WAAW,IAAI,oBAAoB,IAAI;AAAA,EAC9C;AAAA,EAF4B;AAAA,EAFX;AAAA,EAMjB,MAAM,OAAO,OAAe,UAA2C;AACrE,eAAW,KAAK,UAAU;AACxB,YAAM,MAAwB,EAAE,GAAG,GAAG,QAAQ,MAAM;AACpD,YAAM,KAAK,SAAS,OAAO,GAAG;AAAA,IAChC;AAAA,EACF;AAAA;AAAA,EAGA,UAA8B;AAC5B,QAAI,CAACC,YAAW,KAAK,IAAI,EAAG,QAAO,CAAC;AACpC,UAAM,MAAM,aAAa,KAAK,MAAM,MAAM;AAC1C,QAAI,CAAC,IAAK,QAAO,CAAC;AAClB,UAAM,MAA0B,CAAC;AACjC,eAAW,QAAQ,IAAI,MAAM,IAAI,GAAG;AAClC,UAAI,CAAC,KAAM;AACX,UAAI;AACF,YAAI,KAAK,KAAK,MAAM,IAAI,CAAqB;AAAA,MAC/C,QAAQ;AAAA,MAGR;AAAA,IACF;AACA,WAAO;AAAA,EACT;AAAA;AAAA,EAGA,QAAQ,OAAmC;AACzC,WAAO,KAAK,QAAQ,EAAE,OAAO,CAAC,MAAM,EAAE,WAAW,KAAK;AAAA,EACxD;AACF;AAiCO,SAAS,kBAAkB,GAAmB,GAA4B;AAC/E,MAAI,EAAE,aAAa,EAAE,SAAU,QAAO;AACtC,MAAI,KAAK,KAAK,EAAE,cAAc,MAAM,EAAE,cAAc,EAAE,IAAI,KAAM,QAAO;AACvE,MAAI,EAAE,cAAc,WAAW,EAAE,cAAc,OAAQ,QAAO;AAC9D,SAAO;AACT;AAMO,SAAS,aACd,UACA,SACA,SAAqB,CAAC,GACR;AACd,QAAM,aAAa,OAAO,cAAc;AACxC,QAAM,WAAW,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;AAC/D,QAAM,UAAU,IAAI,IAAI,QAAQ,IAAI,CAAC,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;AAE7D,QAAM,WAA+B,CAAC;AACtC,QAAM,cAAkC,CAAC;AACzC,QAAM,YAAgC,CAAC;AACvC,QAAM,UAAmC,CAAC;AAE1C,aAAW,CAAC,IAAI,GAAG,KAAK,SAAS;AAC/B,UAAM,OAAO,SAAS,IAAI,EAAE;AAC5B,QAAI,CAAC,MAAM;AACT,eAAS,KAAK,GAAG;AACjB;AAAA,IACF;AACA,QAAI,WAAW,MAAM,GAAG,GAAG;AACzB,cAAQ,KAAK,EAAE,UAAU,MAAM,SAAS,IAAI,CAAC;AAAA,IAC/C,OAAO;AACL,gBAAU,KAAK,GAAG;AAAA,IACpB;AAAA,EACF;AACA,aAAW,CAAC,IAAI,IAAI,KAAK,UAAU;AACjC,QAAI,CAAC,QAAQ,IAAI,EAAE,EAAG,aAAY,KAAK,IAAI;AAAA,EAC7C;AACA,SAAO,EAAE,UAAU,aAAa,WAAW,QAAQ;AACrD;;;AC3HA,SAAsB,cAAAC,aAAY,aAAa,gBAAAC,eAAc,gBAAgB;AAC7E,SAAS,YAAY;AAyDrB,IAAM,uBAAuB;AAE7B,IAAM,oBACJ;AACF,IAAM,aAAa;AAInB,SAAS,cAAc,MAAgD;AACrE,MAAI,CAACC,YAAW,IAAI,EAAG,QAAO,CAAC;AAC/B,QAAM,MAAwC,CAAC;AAC/C,aAAW,SAAS,YAAY,MAAM,EAAE,eAAe,KAAK,CAAC,GAAG;AAC9D,QAAI,CAAC,MAAM,YAAY,KAAK,CAAC,MAAM,eAAe,EAAG;AACrD,UAAM,UAAU,KAAK,MAAM,MAAM,MAAM,UAAU;AACjD,QAAIA,YAAW,OAAO,EAAG,KAAI,KAAK,EAAE,MAAM,MAAM,MAAM,MAAM,QAAQ,CAAC;AAAA,EACvE;AACA,SAAO;AACT;AAEA,SAAS,UAAU,KAAa,KAAuB;AACrD,MAAI,CAACA,YAAW,GAAG,EAAG,QAAO,CAAC;AAC9B,QAAM,QAAkB,CAAC;AACzB,QAAM,QAAQ,CAAC,GAAG;AAClB,SAAO,MAAM,QAAQ;AACnB,UAAM,MAAM,MAAM,IAAI;AACtB,QAAI;AACJ,QAAI;AACF,gBAAU,YAAY,KAAK,EAAE,eAAe,KAAK,CAAC;AAAA,IACpD,QAAQ;AACN;AAAA,IACF;AACA,eAAW,KAAK,SAAS;AACvB,YAAM,OAAO,KAAK,KAAK,EAAE,IAAI;AAC7B,UAAI,EAAE,YAAY,EAAG,OAAM,KAAK,IAAI;AAAA,eAC3B,EAAE,KAAK,SAAS,QAAQ,GAAG;AAClC,cAAM,KAAK,IAAI;AACf,YAAI,MAAM,KAAK,MAAM,UAAU,IAAK,QAAO;AAAA,MAC7C;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,uBAAuB,MAAsB;AACpD,QAAM,KAAK,wBAAwB,KAAK,IAAI;AAC5C,QAAM,QAAQ,KAAK,CAAC,KAAK;AACzB,QAAM,IAAI,uBAAuB,KAAK,KAAK;AAC3C,SAAO,IAAI,CAAC,KAAK;AACnB;AAEA,SAAS,eAAe,OAAiB,MAAc,SAA2B;AAChF,MAAI,IAAI;AACR,aAAW,QAAQ,OAAO;AACxB,UAAM,aAAa,CAAC,KAAK,MAAM,WAAW,IAAI,GAAG,GAAG,QAAQ,IAAI,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC;AACrF,eAAW,OAAO,YAAY;AAC5B,UAAI,CAACA,YAAW,GAAG,EAAG;AACtB,UAAI;AACF,YAAI,SAAS,GAAG,EAAE,YAAY,EAAG,MAAK,YAAY,GAAG,EAAE;AAAA,YAClD,MAAK;AAAA,MACZ,QAAQ;AAAA,MAER;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAGO,SAAS,sBAAsB,QAAgD;AACpF,QAAM,SAAS,OAAO,WAAW;AAAA,IAAQ,CAAC,EAAE,MAAM,KAAK,MACrD,cAAc,IAAI,EAAE,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,KAAK,EAAE;AAAA,EACjD;AACA,QAAM,QAAQ,OAAO,IAAI,CAAC,MAAM,EAAE,IAAI;AAGtC,QAAM,SAAS,IAAI,IAAoB,MAAM,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;AAC/D,QAAM,QAAQ,IAAI,IAAoB,MAAM,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;AAC9D,QAAM,UAAU;AAChB,QAAM,QAAQ;AACd,MAAI,cAAc;AAClB,aAAW,OAAO,OAAO,gBAAgB;AACvC,eAAW,QAAQ,UAAU,KAAK,OAAO,wBAAwB,CAAC,GAAG;AACnE,qBAAe;AACf,UAAI;AACJ,UAAI;AACF,eAAOC,cAAa,MAAM,MAAM;AAAA,MAClC,QAAQ;AACN;AAAA,MACF;AACA,iBAAW,KAAK,KAAK,SAAS,OAAO,GAAG;AACtC,cAAM,IAAI,EAAE,CAAC;AACb,YAAI,CAAC,EAAG;AACR,cAAM,IAAI,EAAE,MAAM,GAAG,EAAE,IAAI,KAAK;AAChC,cAAM,OAAO,OAAO,IAAI,CAAC;AACzB,YAAI,SAAS,OAAW,QAAO,IAAI,GAAG,OAAO,CAAC;AAAA,MAChD;AACA,iBAAW,KAAK,KAAK,SAAS,KAAK,GAAG;AACpC,cAAM,IAAI,EAAE,CAAC;AACb,YAAI,MAAM,OAAW;AACrB,cAAM,OAAO,MAAM,IAAI,CAAC;AACxB,YAAI,SAAS,OAAW,OAAM,IAAI,GAAG,OAAO,CAAC;AAAA,MAC/C;AAAA,IACF;AAAA,EACF;AAGA,QAAM,SAAS,oBAAI,IAAoB;AACvC,aAAW,KAAK,QAAQ;AACtB,QAAI;AACF,aAAO,IAAI,EAAE,MAAMA,cAAa,EAAE,MAAM,MAAM,CAAC;AAAA,IACjD,QAAQ;AACN,aAAO,IAAI,EAAE,MAAM,EAAE;AAAA,IACvB;AAAA,EACF;AACA,QAAM,UAAU,IAAI,IAAoB,MAAM,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC;AAChE,aAAW,UAAU,OAAO;AAC1B,UAAM,MAAM,IAAI,OAAO,IAAI,MAAM,aAAa,MAAM,QAAQ;AAC5D,eAAW,KAAK,QAAQ;AACtB,UAAI,EAAE,SAAS,OAAQ;AACvB,UAAI,IAAI,KAAK,OAAO,IAAI,EAAE,IAAI,KAAK,EAAE,EAAG,SAAQ,IAAI,QAAQ,QAAQ,IAAI,MAAM,IAAK,CAAC;AAAA,IACtF;AAAA,EACF;AAEA,QAAM,UAA8B,OAAO,IAAI,CAAC,MAAM;AACpD,UAAM,OAAO,OAAO,IAAI,EAAE,IAAI,KAAK;AACnC,UAAM,MAAM,EAAE,KAAK,QAAQ,gBAAgB,EAAE;AAC7C,WAAO;AAAA,MACL,MAAM,EAAE;AAAA,MACR,MAAM,EAAE;AAAA,MACR,MAAM,EAAE;AAAA,MACR,OAAO,OAAO,KAAK,MAAM,IAAI,EAAE,SAAS;AAAA,MACxC,mBAAmB,OAAO,IAAI,EAAE,IAAI,KAAK;AAAA,MACzC,kBAAkB,MAAM,IAAI,EAAE,IAAI,KAAK;AAAA,MACvC,aAAa,QAAQ,IAAI,EAAE,IAAI,KAAK;AAAA,MACpC,eAAe;AAAA,QACb,OAAO,iBAAiB,CAAC;AAAA,QACzB,EAAE;AAAA,QACF,OAAO,kBAAkB,EAAE,IAAI,KAAK,CAAC;AAAA,MACvC;AAAA,MACA,oBAAoB,KAAK,MAAM,iBAAiB,KAAK,CAAC,GAAG;AAAA,MACzD,kBAAkBD,YAAW,KAAK,KAAK,YAAY,CAAC;AAAA,MACpD,aAAaA,YAAW,KAAK,KAAK,OAAO,CAAC;AAAA,MAC1C,UAAU,KAAK,SAAS,kBAAkB;AAAA,MAC1C,mBAAmB,WAAW,KAAK,uBAAuB,IAAI,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;AAAA,IACvF;AAAA,EACF,CAAC;AACD,SAAO,EAAE,qBAAqB,aAAa,QAAQ;AACrD;AAIA,IAAM,aAAa;AAEnB,SAAS,QACP,MACA,SACA,OACA,UACA,YACA,YACA,aACA,aACA,WACgB;AAChB,SAAO;AAAA,IACL,gBAAgB;AAAA,IAChB,YAAY,iBAAiB,EAAE,YAAY,YAAY,MAAM,SAAS,MAAM,CAAC;AAAA,IAC7E,YAAY;AAAA,IACZ,aAAa;AAAA,IACb;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,eAAe,CAAC,EAAE,MAAM,YAAY,KAAK,YAAY,CAAC;AAAA,IACtD,oBAAoB;AAAA,IACpB;AAAA,IACA;AAAA,EACF;AACF;AAGO,SAAS,uBACd,QACA,YACkB;AAClB,QAAM,MAAwB,CAAC;AAC/B,aAAW,KAAK,OAAO,SAAS;AAC9B,UAAM,cAAc,EAAE,oBAAoB,EAAE;AAC5C,UAAM,YAAY,cAAc,EAAE,cAAc,EAAE;AAGlD,QAAI,cAAc,GAAG;AACnB,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF,WAAW,gBAAgB,KAAK,EAAE,cAAc,EAAE,gBAAgB,GAAG;AAEnE,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI,iFAAiF,EAAE,WAAW,eAAe,EAAE,aAAa;AAAA,UAC5I;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAGA,QAAI,eAAe,KAAK,CAAC,EAAE,mBAAmB;AAC5C,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAGA,QAAI,EAAE,SAAS,YAAY,EAAE,oBAAoB,GAAG;AAClD,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,iBAAiB,EAAE,IAAI,aAAa,EAAE,iBAAiB;AAAA,UACvD;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAGA,QAAI,EAAE,QAAQ,wBAAwB,CAAC,EAAE,kBAAkB;AACzD,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI,QAAQ,EAAE,KAAK;AAAA,UAC/B;AAAA,UACA;AAAA,UACA;AAAA,UACA,mFAAmF,EAAE,KAAK;AAAA,UAC1F,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAGA,QAAI,CAAC,EAAE,aAAa;AAClB,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAGA,QAAI,CAAC,EAAE,UAAU;AACf,UAAI;AAAA,QACF;AAAA,UACE;AAAA,UACA,EAAE;AAAA,UACF,UAAU,EAAE,IAAI;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,EAAE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAIO,IAAM,oBAAN,MAA6D;AAAA,EACzD,KAAK;AAAA,EACL,cACP;AAAA,EACO,YAAY;AAAA,EACZ,OAAO,EAAE,MAAM,iBAA0B,iBAAiB,EAAE;AAAA,EAC5D,UAAU;AAAA,EAEnB,MAAM,QAAQ,OAAyB,KAAgD;AACrF,UAAM,aAAa,IAAI,MAAM,eAAc,oBAAI,KAAK,GAAE,YAAY;AAClE,QAAI;AAAA,MACF,gBAAgB,MAAM,QAAQ,MAAM,gBAAgB,MAAM,mBAAmB;AAAA,IAC/E;AACA,WAAO,uBAAuB,OAAO,UAAU;AAAA,EACjD;AACF;AAEO,IAAM,sBAAsB,IAAI,kBAAkB;;;AClYzD,IAAM,yBAAyB;AAAA,EAC7B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEO,IAAM,YAAN,MAAgB;AAAA,EACJ;AAAA,EACA;AAAA,EAEjB,YAAY,UAA4B,CAAC,GAAG;AAC1C,SAAK,UAAU,QAAQ;AACvB,SAAK,gBAAgB,QAAQ,iBAAiB;AAAA,EAChD;AAAA,EAEA,MAAM,MAAM,OAAmB,OAAkC;AAC/D,UAAM,MAAM,MAAM,MAAM,OAAO,KAAK;AACpC,QAAI,CAAC,IAAK,OAAM,IAAI,cAAc,OAAO,KAAK,YAAY;AAC1D,UAAM,CAAC,OAAO,QAAQ,WAAW,MAAM,IAAI,MAAM,QAAQ,IAAI;AAAA,MAC3D,MAAM,MAAM,EAAE,MAAM,CAAC;AAAA,MACrB,MAAM,OAAO,EAAE,MAAM,CAAC;AAAA,MACtB,MAAM,UAAU,KAAK;AAAA,MACrB,MAAM,OAAO,KAAK;AAAA,IACpB,CAAC;AACD,WAAO,KAAK,WAAW,EAAE,KAAK,OAAO,QAAQ,WAAW,OAAO,CAAC;AAAA,EAClE;AAAA,EAEA,WAAW,OAA2B;AACpC,UAAM,QAAkB,CAAC;AACzB,UAAM,WAAW,MAAM,MAAM;AAAA,MAC3B,CAAC,MAA2C,EAAE,SAAS;AAAA,IACzD;AACA,UAAM,YAAY,MAAM,MAAM;AAAA,MAC5B,CAAC,MAA4C,EAAE,SAAS;AAAA,IAC1D;AACA,UAAM,aAAa,MAAM,MAAM;AAAA,MAC7B,CAAC,MAA6C,EAAE,SAAS;AAAA,IAC3D;AACA,UAAM,eAAe,MAAM,MAAM;AAAA,MAC/B,CAAC,MAA+C,EAAE,SAAS;AAAA,IAC7D;AACA,UAAM,iBAAiB,WAAW;AAAA,MAChC,CAAC,SAAS,KAAK,cAAc,gBAAgB,KAAK,YAAY,cAAc;AAAA,IAC9E;AAEA,UAAM,UACJ,MAAM,IAAI,SAAS,SAAS,OAAO,IAAI,MAAM,IAAI,WAAW,cAAc,MAAM;AAClF,QAAI,CAAC,QAAS,OAAM,KAAK,qCAAqC;AAE9D,UAAM,eAAe,WAAW,SAC5B,WAAW,OAAO,CAAC,KAAK,SAAS,MAAM,oBAAoB,KAAK,KAAK,GAAG,CAAC,IACzE,WAAW,SACX;AACJ,UAAM,eACJ,OAAO,MAAM,IAAI,SAAS,UAAU,WAChC;AAAA,MACE,MAAM,IAAI,QAAQ,QAAQ,IAAI,MAAM,IAAI,QAAQ,QAAQ,MAAM,MAAM,IAAI,QAAQ;AAAA,IAClF,IACA;AACN,UAAM,eAAe,gBAAgB,gBAAgB;AAErD,UAAM,kBAAkB,UAAU,OAAO,CAAC,SAAS,KAAK,WAAW,OAAO,EAAE;AAC5E,UAAM,iBAAiB,UAAU,WAAW,IAAI,IAAI,kBAAkB,UAAU;AAChF,QAAI,UAAU,WAAW,EAAG,OAAM,KAAK,wBAAwB;AAE/D,UAAM,gBACJ,MAAM,UAAU,SAChB,UAAU,OAAO,CAAC,SAAS,0BAA0B,KAAK,KAAK,QAAQ,CAAC,EAAE;AAC5E,UAAM,eAAe,gBAAgB,IAAI,QAAQ,gBAAgB,CAAC,IAAI;AACtE,QAAI,CAAC,aAAc,OAAM,KAAK,uCAAuC;AAErE,UAAM,eAAe,aAAa;AAAA,MAChC,CAAC,SAAS,OAAO,KAAK,eAAe,YAAY,KAAK,aAAa;AAAA,IACrE;AACA,UAAM,cAAc,aAAa,SAC7B,aAAa;AAAA,MACX,CAAC,KAAK,SAAS,OAAO,KAAK,eAAe,KAAK,KAAK,IAAI,GAAG,KAAK,cAAc,CAAC;AAAA,MAC/E;AAAA,IACF,IAAI,aAAa,SACjB,UAAU;AAAA,MAAK,CAAC,SACZ,yCAAyC,KAAK,KAAK,UAAU,KAAK,IAAI,CAAC;AAAA,IACzE,IACA,MACA;AACN,QAAI,CAAC,YAAa,OAAM,KAAK,sCAAsC;AAEnE,UAAM,eAAe,WAAW,OAAO,CAAC,SAAS,gBAAgB,IAAI,CAAC;AACtE,UAAM,oBAAoB,eAAe,OAAO,CAAC,SAAS,gBAAgB,IAAI,CAAC;AAC/E,UAAM,YAAY,eAAe,SAAU,kBAAkB,SAAS,IAAI,IAAK;AAC/E,QAAI,kBAAkB;AACpB,YAAM,KAAK,yBAAyB,kBAAkB,MAAM,cAAc;AAAA,aACnE,CAAC,eAAe,OAAQ,OAAM,KAAK,iCAAiC;AAE7E,UAAM,mBAAmB,WAAW,SAAS,aAAa,SAAS,WAAW,SAAS;AACvF,QAAI,iBAAkB,OAAM,KAAK,YAAY,aAAa,MAAM,8BAA8B;AAE9F,UAAM,2BACJ,gBACA,aAAa,SACb,SAAS,OAAO,CAAC,SAAS,kBAAkB,KAAK,UAAU,EAAE,CAAC,EAAE;AAClE,UAAM,eACJ,SAAS,OAAO,CAAC,SAAS,KAAK,QAAQ,KAAK,UAAU,EAAE,CAAC,EAAE,SAC3D,MAAM,OAAO,OAAO,CAAC,UAAU,KAAK,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,CAAC,EAAE;AAC9E,UAAM,mBACJ,2BAA2B,iBAAiB,IACxC,IACA,4BAA4B,2BAA2B;AAC7D,UAAM,eACJ,2BAA2B,iBAAiB,IACxC,IACA,gBAAgB,2BAA2B;AACjD,QAAI,eAAe,EAAG,OAAM,KAAK,YAAY,YAAY,kBAAkB;AAE3E,UAAM,UAAU,MAAM,OAAO,SACzB,KAAK;AAAA,MACH,GAAG,MAAM,OACN,OAAO,CAAC,UAA6B,MAAM,cAAc,KAAK,EAC9D,IAAI,CAAC,UAA6B,MAAM,QAAQ;AAAA,MACnD;AAAA,IACF,IACA,SAAS,OAAO,CAAC,KAAK,SAAS,OAAO,KAAK,WAAW,IAAI,CAAC;AAC/D,UAAM,cACJ,MAAM,IAAI,WAAW,MAAM,IAAI,YAC3B,KAAK,IAAI,IAAI,MAAM,IAAI,UAAU,MAAM,IAAI,aAAa,GAAI,IAC5D;AAEN,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA,EAEA,KAAK,OAAyB;AAC5B,WAAO,kBAAkB,OAAO,KAAK,OAAO;AAAA,EAC9C;AAAA,EAEQ,QAAQ,MAAuB;AACrC,WAAO,KAAK,cAAc,KAAK,CAAC,YAAY,QAAQ,KAAK,IAAI,CAAC;AAAA,EAChE;AACF;AAEA,SAAS,oBAAoB,OAAuB;AAClD,SAAO,QAAQ,IAAI,QAAQ,QAAQ,EAAE,IAAI,QAAQ,KAAK;AACxD;AAEA,SAAS,kBAAkB,MAAuB;AAChD,SAAO,qGAAqG;AAAA,IAC1G;AAAA,EACF;AACF;AAEA,SAAS,gBAAgB,MAAiD;AACxE,SACE,KAAK,YAAY,aAAa,QAC9B,KAAK,YAAY,YAAY,cAC7B,eAAe,KAAK,YAAY,gBAAgB,KAChD,eAAe,KAAK,YAAY,YAAY,KAC5C,KAAK,SAAS;AAElB;AAEA,SAAS,eAAe,OAAyB;AAC/C,SAAO,OAAO,UAAU,YAAY,QAAQ;AAC9C;;;ACvFO,IAAM,6BAAgE;AAAA,EAC3E,QAAQ;AAAA,EACR,WAAW;AAAA,EACX,SAAS;AACX;AA2BO,IAAM,iCAAiC;AAE9C,IAAM,qBAAqB;AAC3B,IAAM,mBAAmB;AACzB,IAAM,uBAAuB;AAC7B,IAAM,kBAAkB;AACxB,IAAM,gBAAgB;AAEtB,IAAM,kBAAkB;AAAA,EACtB,MAAM;AAAA,EACN,sBAAsB;AAAA,EACtB,UAAU,CAAC,WAAW,UAAU;AAAA,EAChC,YAAY;AAAA,IACV,SAAS,EAAE,MAAM,UAAU,WAAW,IAAI,WAAW,IAAI;AAAA,IACzD,UAAU;AAAA,MACR,MAAM;AAAA,MACN,UAAU;AAAA,MACV,OAAO;AAAA,QACL,MAAM;AAAA,QACN,sBAAsB;AAAA,QACtB,UAAU,CAAC,WAAW,WAAW,SAAS,YAAY,UAAU;AAAA,QAChE,YAAY;AAAA,UACV,SAAS,EAAE,MAAM,UAAU,WAAW,GAAG,WAAW,IAAI;AAAA,UACxD,SAAS,EAAE,MAAM,UAAU;AAAA,UAC3B,OAAO,EAAE,MAAM,UAAU,SAAS,GAAG,SAAS,GAAG;AAAA,UACjD,UAAU,EAAE,MAAM,UAAU,WAAW,GAAG,WAAW,IAAI;AAAA,UACzD,UAAU,EAAE,MAAM,UAAU,MAAM,CAAC,YAAY,SAAS,SAAS,MAAM,EAAE;AAAA,QAC3E;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,SAAS,MAAc,KAAa,OAAuB;AAClE,MAAI,KAAK,UAAU,IAAK,QAAO;AAC/B,SAAO,GAAG,KAAK,MAAM,GAAG,GAAG,CAAC;AAAA,oBAAkB,KAAK,SAAS,GAAG,aAAa,KAAK;AACnF;AAEA,SAAS,YACP,OACA,MACQ;AACR,QAAM,aAAa,MAAM,YACtB,OAAO,CAAC,MAAM,EAAE,QAAQ,UAAU,KAAK,eAAe,EACtD,IAAI,CAAC,MAAM,aAAa,EAAE,IAAI;AAAA,EAAS,EAAE,OAAO,EAAE,EAClD,KAAK,MAAM;AAEd,QAAM,OAAO,MAAM,cAAc;AAEjC,SAAO;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUP,MAAM,WAAW;AAAA;AAAA,EAEjB,MAAM,gBAAgB;AAAA,UAA+B,MAAM,aAAa;AAAA,iBAAoB,MAAM,uBAAuB,EAAE;AAAA;AAAA,IAAS,EAAE;AAAA,EACtI,MAAM,iBACL;AAAA,IACC,CAAC,GAAG,MACF,KAAK,IAAI,CAAC,MAAM,EAAE,IAAI,IAAI,EAAE,UAAU,SAAS,mBAAc,EAAE,SAAS,MAAM,GAAG,CAAC,EAAE,KAAK,KAAK,CAAC,MAAM,EAAE;AAAA,EAC3G,EACC,KAAK,IAAI,CAAC;AAAA;AAAA,EAEX,OAAO;AAAA,EAAqD,SAAS,MAAM,KAAK,cAAc,MAAM,CAAC;AAAA;AAAA,IAAS,EAAE;AAAA,EAChH,SAAS,YAAY,KAAK,gBAAgB,QAAQ,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAkBrD;AASA,eAAsB,wBACpB,OACA,UAAuC,CAAC,GACH;AACrC,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,aAAa,MAAM,iBAAiB;AAE1C,MAAI,eAAe,GAAG;AACpB,WAAO;AAAA,MACL,MAAM;AAAA,MACN,SAAS;AAAA,MACT,OAAO;AAAA,MACP,cAAc;AAAA,MACd,YAAY;AAAA,MACZ,UAAU,CAAC;AAAA,MACX,SAAS;AAAA,MACT,YAAY;AAAA,MACZ,SAAS;AAAA,MACT,WAAW;AAAA,MACX,OAAO;AAAA,IACT;AAAA,EACF;AAEA,QAAM,OAA8C;AAAA,IAClD,OAAO,QAAQ,SAAS;AAAA,IACxB,WAAW,QAAQ,aAAa;AAAA,IAChC,gBAAgB,QAAQ,kBAAkB;AAAA,IAC1C,iBAAiB,QAAQ,mBAAmB;AAAA,IAC5C,cAAc,QAAQ,gBAAgB;AAAA,IACtC,KAAK,QAAQ,OAAO,CAAC;AAAA,IACrB,gBAAgB,QAAQ,kBAAkB;AAAA,IAC1C,mBAAmB,EAAE,GAAG,4BAA4B,GAAI,QAAQ,qBAAqB,CAAC,EAAG;AAAA,EAC3F;AAMA,QAAM,mBAAmB,CAAC,SAA8B;AACtD,QAAI,KAAK,mBAAmB,OAAQ,QAAO;AAC3C,QAAI,KAAK,UAAU,KAAM,QAAO,KAAK;AACrC,QAAI,KAAK,mBAAmB,cAAc;AACxC,aAAO,KAAK,kBAAkB,KAAK,cAAc,QAAQ,KAAK;AAAA,IAChE;AACA,WAAO;AAAA,EACT;AACA,QAAM,eAAe,IAAI;AAAA,IACvB,MAAM,iBAAiB,IAAI,CAAC,MAAM,CAAC,EAAE,MAAM,iBAAiB,CAAC,CAAC,CAAC;AAAA,EACjE;AAEA,MAAI;AACF,UAAM,EAAE,OAAO,OAAO,IAAI,MAAM;AAAA,MAI9B;AAAA,QACE,OAAO,KAAK;AAAA,QACZ,UAAU;AAAA,UACR;AAAA,YACE,MAAM;AAAA,YACN,SACE;AAAA,UACJ;AAAA,UACA,EAAE,MAAM,QAAQ,SAAS,YAAY,OAAO,IAAI,EAAE;AAAA,QACpD;AAAA,QACA,YAAY,EAAE,MAAM,0BAA0B,QAAQ,gBAAgB;AAAA,QACtE,aAAa;AAAA,QACb,WAAW,KAAK;AAAA,MAClB;AAAA,MACA,KAAK;AAAA,IACP;AAEA,QAAI,CAAC,OAAO,YAAY,CAAC,MAAM,QAAQ,MAAM,QAAQ,GAAG;AACtD,YAAM,IAAI,MAAM,0EAAqE;AAAA,IACvF;AAEA,UAAM,WAA6B,MAAM,SAAS,IAAI,CAAC,OAAO;AAAA,MAC5D,SAAS,OAAO,EAAE,OAAO;AAAA,MACzB,SAAS,QAAQ,EAAE,OAAO;AAAA,MAC1B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,EAAE,SAAS,CAAC,CAAC,CAAC;AAAA,MACrD,UAAU,OAAO,EAAE,YAAY,EAAE;AAAA,MACjC,UAAW,CAAC,YAAY,SAAS,SAAS,MAAM,EAAY,SAAS,EAAE,QAAQ,IAC3E,EAAE,WACF;AAAA,IACN,EAAE;AAEF,UAAM,eAAe,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,EAAE,SAAS,CAAC,EAAE;AACvE,QAAI,YAAY;AAChB,QAAI,mBAAmB;AACvB,eAAW,KAAK,UAAU;AACxB,YAAM,IAAI,aAAa,IAAI,EAAE,OAAO,KAAK;AACzC,mBAAa;AACb,0BAAoB,IAAI,EAAE;AAAA,IAC5B;AACA,UAAM,WACJ,YAAY,IACR,mBAAmB,YACnB,SAAS,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,IAAI,GAAG,SAAS,MAAM;AAE7E,WAAO;AAAA,MACL,MAAM;AAAA,MACN,SAAS;AAAA,MACT,OAAO,QAAQ,WAAW,IAAI,QAAQ,CAAC,CAAC;AAAA,MACxC;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS,OAAO,MAAM,WAAW,EAAE;AAAA,MACnC,YAAY,KAAK,IAAI,IAAI;AAAA,MACzB,SAAS,OAAO,WAAW;AAAA,MAC3B,WAAW;AAAA,IACb;AAAA,EACF,SAAS,KAAK;AACZ,WAAO;AAAA,MACL,MAAM;AAAA,MACN,SAAS;AAAA,MACT,OAAO;AAAA,MACP,cAAc;AAAA,MACd;AAAA,MACA,UAAU,CAAC;AAAA,MACX,SAAS;AAAA,MACT,YAAY,KAAK,IAAI,IAAI;AAAA,MACzB,SAAS;AAAA,MACT,WAAW;AAAA,MACX,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,IACxD;AAAA,EACF;AACF;AAMO,SAAS,2BACd,UAAuC,CAAC,GACmC;AAC3E,SAAO,CAAC,UAAU,wBAAwB,OAAO,OAAO;AAC1D;","names":["existsSync","existsSync","existsSync","readFileSync","existsSync","readFileSync"]}

package/dist/chunk-NYFUT3B3.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/failure-taxonomy.ts","../src/tool-use-metrics.ts","../src/trajectory.ts","../src/baseline.ts"],"sourcesContent":["/**\n * Failure taxonomy — canonical classes + a default classifier.\n *\n * Every failed run should end up in a named class. The classifier here\n * is rule-based (fast, deterministic); an LLM fallback can be added by\n * the consumer for novel cases and trained into the rule base over time.\n *\n * Consumers call `classifyFailure(run, spans, events)` and persist the\n * returned class as `Run.outcome.failureClass`.\n */\n\nimport type { FailureClass, Run, Span, TraceEvent } from './trace/schema'\nimport { FAILURE_CLASSES } from './trace/schema'\n\nexport { FAILURE_CLASSES, type FailureClass }\n\nexport interface FailureContext {\n  run: Run\n  spans: Span[]\n  events: TraceEvent[]\n}\n\nexport interface FailureClassification {\n  failureClass: FailureClass\n  reason: string\n  triggerSpanId?: string\n  triggerEventId?: string\n}\n\n/** Ordered rules — first match wins. */\nexport interface FailureRule {\n  id: string\n  match: (ctx: FailureContext) => {\n    failureClass: FailureClass\n    reason: string\n    triggerSpanId?: string\n    triggerEventId?: string\n  } | null\n}\n\nexport const DEFAULT_RULES: FailureRule[] = [\n  // Outcome already named? Respect it.\n  {\n    id: 'explicit-outcome',\n    match: ({ run }) => {\n      const fc = run.outcome?.failureClass\n      if (fc && fc !== 'unknown')\n        return { failureClass: fc, reason: 'outcome.failureClass set explicitly' }\n      return null\n    },\n  },\n  {\n    id: 'knowledge-readiness-blocked',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'readiness_scored' &&\n          e.payload.passed === false,\n      )\n      return event\n        ? {\n            failureClass: 'knowledge_readiness_blocked',\n            reason: 'knowledge readiness report blocked execution',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'bad-integration-manifest',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          ((e.payload.kind === 'integration_manifest_validated' && e.payload.valid === false) ||\n            (e.payload.kind === 'integration_invoke_failed' &&\n              e.payload.code === 'manifest_invalid')),\n      )\n      return event\n        ? {\n            failureClass: 'bad_integration_manifest',\n            reason: 'integration manifest validation failed before launch',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'missing-integration-connection',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'integration_manifest_resolved' &&\n          hasResolutionStatus(e.payload, 'missing_connection'),\n      )\n      return event\n        ? {\n            failureClass: 'missing_integration_connection',\n            reason: 'required integration connection was missing',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'missing-integration-scope',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          ((e.payload.kind === 'integration_manifest_resolved' && hasMissingScopes(e.payload)) ||\n            (e.payload.kind === 'integration_invoke_failed' && e.payload.code === 'scope_denied')),\n      )\n      return event\n        ? {\n            failureClass: 'missing_integration_scope',\n            reason: 'integration grant or connection lacks required scopes',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'integration-approval-required',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          ((e.payload.kind === 'integration_invoke' && e.payload.status === 'approval_required') ||\n            (e.payload.kind === 'integration_invoke_failed' &&\n              e.payload.code === 'approval_required') ||\n            e.payload.kind === 'integration_approval_required'),\n      )\n      return event\n        ? {\n            failureClass: 'integration_approval_required',\n            reason: 'integration write paused for user approval',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'integration-auth-expired',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'integration_invoke_failed' &&\n          (e.payload.code === 'auth_expired' ||\n            e.payload.code === 'connection_not_active' ||\n            e.payload.code === 'capability_expired' ||\n            e.payload.status === 'expired'),\n      )\n      return event\n        ? {\n            failureClass: 'integration_auth_expired',\n            reason: 'integration connection or capability expired',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'unsafe-integration-write-denied',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'integration_invoke_failed' &&\n          (e.payload.code === 'unsafe_write_denied' ||\n            e.payload.code === 'policy_denied' ||\n            e.payload.code === 'action_denied'),\n      )\n      return event\n        ? {\n            failureClass: 'unsafe_integration_write_denied',\n            reason: 'integration write was denied by policy or capability scope',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'integration-provider-failure',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'integration_invoke_failed' &&\n          ![\n            'scope_denied',\n            'approval_required',\n            'auth_expired',\n            'connection_not_active',\n            'capability_expired',\n            'unsafe_write_denied',\n            'policy_denied',\n            'action_denied',\n            'manifest_invalid',\n          ].includes(String(e.payload.code)),\n      )\n      return event\n        ? {\n            failureClass: 'integration_provider_failure',\n            reason: 'integration provider invocation failed',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'missing-credentials',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'knowledge_gap' &&\n          e.payload.category === 'credential_or_secret',\n      )\n      return event\n        ? {\n            failureClass: 'missing_credentials',\n            reason: 'required credential or secret was missing',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'bad-retrieval',\n    match: ({ run, spans }) => {\n      if (run.outcome?.pass !== false) return null\n      const retrieval = spans.find(\n        (s) =>\n          s.kind === 'retrieval' && (s.hits.length === 0 || s.hits.every((hit) => hit.score <= 0)),\n      )\n      return retrieval\n        ? {\n            failureClass: 'bad_retrieval',\n            reason: 'retrieval returned no useful hits for a failed run',\n            triggerSpanId: retrieval.spanId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'insufficient-evidence',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'knowledge_gap' &&\n          e.payload.reason === 'insufficient_evidence',\n      )\n      return event\n        ? {\n            failureClass: 'insufficient_evidence',\n            reason: 'task proceeded with insufficient supporting evidence',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  {\n    id: 'contradictory-evidence',\n    match: ({ events }) => {\n      const event = events.find(\n        (e) =>\n          e.kind === 'custom' &&\n          e.payload.kind === 'knowledge_gap' &&\n          e.payload.reason === 'contradictory_evidence',\n      )\n      return event\n        ? {\n            failureClass: 'contradictory_evidence',\n            reason: 'supporting evidence contradicted itself',\n            triggerEventId: event.eventId,\n          }\n        : null\n    },\n  },\n  // Budget breach events\n  {\n    id: 'budget-breach',\n    match: ({ events }) => {\n      const breach = events.find((e) => e.kind === 'budget_breach')\n      return breach\n        ? {\n            failureClass: 'budget_exceeded',\n            reason: `budget breached on ${breach.payload.dimension ?? 'unknown dimension'}`,\n            triggerEventId: breach.eventId,\n          }\n        : null\n    },\n  },\n  // Policy violations\n  {\n    id: 'policy-violation',\n    match: ({ events }) => {\n      const e = events.find((x) => x.kind === 'policy_violation')\n      return e\n        ? {\n            failureClass: 'policy_violation',\n            reason: 'policy_violation event emitted',\n            triggerEventId: e.eventId,\n          }\n        : null\n    },\n  },\n  // Sandbox non-zero exit code\n  {\n    id: 'sandbox-failure',\n    match: ({ spans }) => {\n      const s = spans.find(\n        (x) => x.kind === 'sandbox' && typeof x.exitCode === 'number' && x.exitCode !== 0,\n      )\n      if (!s) return null\n      return {\n        failureClass: 'sandbox_failure',\n        reason: `sandbox exited ${(s as Extract<Span, { kind: 'sandbox' }>).exitCode}`,\n        triggerSpanId: s.spanId,\n      }\n    },\n  },\n  // Timeout: run aborted by external signal\n  {\n    id: 'timeout',\n    match: ({ run, events }) => {\n      if (run.status !== 'aborted') return null\n      const hasTimeout = events.some(\n        (e) =>\n          e.kind === 'error' &&\n          String(e.payload.reason ?? '')\n            .toLowerCase()\n            .includes('timeout'),\n      )\n      const note = (run.outcome?.notes ?? '').toLowerCase()\n      if (hasTimeout || note.includes('timeout') || note.includes('deadline')) {\n        return { failureClass: 'timeout', reason: 'timeout signal observed' }\n      }\n      return null\n    },\n  },\n  // Tool recovery failure: many consecutive tool errors on the same tool\n  {\n    id: 'tool-recovery-failure',\n    match: ({ spans }) => {\n      const tools = spans.filter((s) => s.kind === 'tool')\n      const byTool = new Map<string, Span[]>()\n      for (const t of tools) {\n        const name = (t as Extract<Span, { kind: 'tool' }>).toolName\n        const arr = byTool.get(name) ?? []\n        arr.push(t)\n        byTool.set(name, arr)\n      }\n      for (const [name, arr] of byTool) {\n        const errs = arr.filter((s) => s.status === 'error')\n        if (errs.length >= 3 && errs.length === arr.length) {\n          return {\n            failureClass: 'tool_recovery_failure',\n            reason: `${errs.length} consecutive errors on tool \"${name}\"`,\n            triggerSpanId: errs[errs.length - 1]!.spanId,\n          }\n        }\n      }\n      return null\n    },\n  },\n  // Tool selection error: the run failed and agent called zero tools despite having them\n  {\n    id: 'tool-selection-error',\n    match: ({ run, spans }) => {\n      if (run.outcome?.pass !== false) return null\n      const hasToolsAvailable = spans.some(\n        (s) =>\n          s.kind === 'agent' &&\n          (s.attributes?.toolsAvailable as number | undefined) !== undefined &&\n          (s.attributes?.toolsAvailable as number) > 0,\n      )\n      const tools = spans.filter((s) => s.kind === 'tool')\n      if (hasToolsAvailable && tools.length === 0) {\n        return {\n          failureClass: 'tool_selection_error',\n          reason: 'tools were available but none were called',\n        }\n      }\n      return null\n    },\n  },\n  // Format drift: scored by a judge with dimension='format' below threshold\n  {\n    id: 'format-drift',\n    match: ({ spans }) => {\n      const judge = spans.find(\n        (s) =>\n          s.kind === 'judge' &&\n          (s as Extract<Span, { kind: 'judge' }>).dimension === 'format' &&\n          (s as Extract<Span, { kind: 'judge' }>).score < 0.5,\n      )\n      return judge\n        ? {\n            failureClass: 'format_drift',\n            reason: 'format judge scored below 0.5',\n            triggerSpanId: judge.spanId,\n          }\n        : null\n    },\n  },\n]\n\nfunction hasResolutionStatus(payload: Record<string, unknown>, status: string): boolean {\n  if (status === 'missing_connection' && stringArray(payload.missingConnections).length > 0)\n    return true\n  return resolutionItems(payload).some((item) => item.status === status)\n}\n\nfunction hasMissingScopes(payload: Record<string, unknown>): boolean {\n  if (stringArray(payload.missingScopes).length > 0) return true\n  return resolutionItems(payload).some(\n    (item) => Array.isArray(item.missingScopes) && item.missingScopes.length > 0,\n  )\n}\n\nfunction resolutionItems(payload: Record<string, unknown>): Array<Record<string, unknown>> {\n  return [\n    ...records(payload.missing),\n    ...records(payload.optionalMissing),\n    ...records(payload.ready),\n  ]\n}\n\nfunction records(value: unknown): Array<Record<string, unknown>> {\n  if (!Array.isArray(value)) return []\n  return value.filter(\n    (item): item is Record<string, unknown> =>\n      Boolean(item) && typeof item === 'object' && !Array.isArray(item),\n  )\n}\n\nfunction stringArray(value: unknown): string[] {\n  return Array.isArray(value)\n    ? value.filter((item): item is string => typeof item === 'string')\n    : []\n}\n\n/** Classify the failure mode of a run using an ordered rule list. */\nexport function classifyFailure(\n  ctx: FailureContext,\n  rules: FailureRule[] = DEFAULT_RULES,\n): FailureClassification {\n  if (ctx.run.outcome?.pass !== false && ctx.run.status === 'completed') {\n    return { failureClass: 'success', reason: 'run completed with pass=true (or no explicit fail)' }\n  }\n  for (const rule of rules) {\n    const hit = rule.match(ctx)\n    if (hit) return hit\n  }\n  return { failureClass: 'unknown', reason: 'no rule matched; run failed for unclassified reason' }\n}\n","/**\n * Tool-use metrics — derived purely from trace data.\n *\n * No scoring assumptions: consumers supply optional ground-truth tool\n * selections per turn + optional \"information used downstream\" signals.\n * Without those, we still compute descriptive metrics (error rate,\n * retry rate, duplicate-call rate) that are useful on their own.\n */\n\nimport { argHash, groupBy, toolSpans } from './trace/query'\nimport type { Span } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface ToolUseMetrics {\n  runId: string\n  totalCalls: number\n  byTool: Record<string, ToolStats>\n  errorRate: number\n  /** Ratio of calls with identical (toolName, argHash) already seen earlier in the same run. */\n  duplicateRate: number\n  /** Ratio of error calls followed by ≥1 retry on same tool. */\n  retryRate: number\n  /** Optional: of the calls agent made, fraction the evaluator marked as \"correct selection\". */\n  selectionAccuracy?: number\n}\n\nexport interface ToolStats {\n  calls: number\n  errors: number\n  avgLatencyMs: number\n  duplicates: number\n}\n\nexport interface ToolUseOptions {\n  /** Map of spanId → whether the evaluator judged the tool selection correct. Optional. */\n  selectionLabels?: Record<string, boolean>\n}\n\nexport async function computeToolUseMetrics(\n  store: TraceStore,\n  runId: string,\n  options: ToolUseOptions = {},\n): Promise<ToolUseMetrics> {\n  const tools = await toolSpans(store, runId)\n  if (tools.length === 0) {\n    return { runId, totalCalls: 0, byTool: {}, errorRate: 0, duplicateRate: 0, retryRate: 0 }\n  }\n\n  const byTool: Record<string, ToolStats> = {}\n  let totalErrors = 0\n  let totalDuplicates = 0\n  const sortedTools = [...tools].sort((a, b) => a.startedAt - b.startedAt)\n  const seenSignatures = new Set<string>()\n\n  // duplicate detection + per-tool aggregation\n  for (const t of sortedTools) {\n    byTool[t.toolName] ??= { calls: 0, errors: 0, avgLatencyMs: 0, duplicates: 0 }\n    const stat = byTool[t.toolName]!\n    stat.calls += 1\n    if (t.status === 'error') {\n      stat.errors += 1\n      totalErrors += 1\n    }\n    if (typeof t.latencyMs === 'number') stat.avgLatencyMs += t.latencyMs\n    const sig = `${t.toolName}|${argHash(t.args)}`\n    if (seenSignatures.has(sig)) {\n      stat.duplicates += 1\n      totalDuplicates += 1\n    }\n    seenSignatures.add(sig)\n  }\n\n  for (const stat of Object.values(byTool)) {\n    stat.avgLatencyMs = stat.calls > 0 ? stat.avgLatencyMs / stat.calls : 0\n  }\n\n  // retry detection: per-tool chronological adjacency where error → next same-tool call\n  let retryOpportunities = 0\n  let retriesFollowed = 0\n  for (const [, arr] of groupBy(sortedTools, (t) => t.toolName)) {\n    for (let i = 0; i < arr.length; i++) {\n      if (arr[i]!.status !== 'error') continue\n      retryOpportunities += 1\n      if (arr[i + 1]) retriesFollowed += 1\n    }\n  }\n  const retryRate = retryOpportunities > 0 ? retriesFollowed / retryOpportunities : 0\n\n  let selectionAccuracy: number | undefined\n  if (options.selectionLabels) {\n    const labeled = sortedTools.filter((t) => t.spanId in options.selectionLabels!)\n    if (labeled.length > 0) {\n      selectionAccuracy =\n        labeled.filter((t) => options.selectionLabels![t.spanId]).length / labeled.length\n    }\n  }\n\n  return {\n    runId,\n    totalCalls: sortedTools.length,\n    byTool,\n    errorRate: totalErrors / sortedTools.length,\n    duplicateRate: totalDuplicates / sortedTools.length,\n    retryRate,\n    selectionAccuracy,\n  }\n}\n\nexport type { Span }\n","/**\n * Trajectory — ordered, structured view over a run's spans.\n *\n * A pure function `buildTrajectory(store, runId) → Trajectory` returns\n * a topologically ordered list of `TrajectoryStep` with parent-child\n * grouping collapsed into a single line-of-agent-work. Separate\n * analyzers (stuck-loop detection, waste ratio) live in\n * `pipelines/` and consume the trajectory.\n */\n\nimport type { Span, TraceEvent } from './trace/schema'\nimport type { TraceStore } from './trace/store'\n\nexport interface TrajectoryStep {\n  index: number\n  span: Span\n  /** Depth in the span tree from the root. 0 = top-level. */\n  depth: number\n  /** Events attached to this span. */\n  events: TraceEvent[]\n}\n\nexport interface Trajectory {\n  runId: string\n  steps: TrajectoryStep[]\n  llmTurns: number\n  toolCalls: number\n  judgeVerdicts: number\n  retrievals: number\n  totalDurationMs: number\n}\n\nexport async function buildTrajectory(store: TraceStore, runId: string): Promise<Trajectory> {\n  const spans = await store.spans({ runId })\n  const events = await store.events({ runId })\n  const childrenOf = new Map<string | undefined, Span[]>()\n  for (const s of spans) {\n    const arr = childrenOf.get(s.parentSpanId) ?? []\n    arr.push(s)\n    childrenOf.set(s.parentSpanId, arr)\n  }\n  // Sort children by startedAt so DFS yields chronological order within siblings.\n  for (const arr of childrenOf.values()) arr.sort((a, b) => a.startedAt - b.startedAt)\n\n  const eventsBySpan = new Map<string, TraceEvent[]>()\n  for (const e of events) {\n    if (!e.spanId) continue\n    const arr = eventsBySpan.get(e.spanId) ?? []\n    arr.push(e)\n    eventsBySpan.set(e.spanId, arr)\n  }\n\n  const steps: TrajectoryStep[] = []\n  const walk = (spanId: string | undefined, depth: number): void => {\n    const kids = childrenOf.get(spanId) ?? []\n    for (const child of kids) {\n      steps.push({\n        index: steps.length,\n        span: child,\n        depth,\n        events: eventsBySpan.get(child.spanId) ?? [],\n      })\n      walk(child.spanId, depth + 1)\n    }\n  }\n  walk(undefined, 0)\n\n  const llmTurns = steps.filter((s) => s.span.kind === 'llm').length\n  const toolCalls = steps.filter((s) => s.span.kind === 'tool').length\n  const judgeVerdicts = steps.filter((s) => s.span.kind === 'judge').length\n  const retrievals = steps.filter((s) => s.span.kind === 'retrieval').length\n\n  let totalDurationMs = 0\n  if (steps.length > 0) {\n    const starts = spans.map((s) => s.startedAt)\n    const ends = spans.map((s) => s.endedAt ?? s.startedAt)\n    totalDurationMs = Math.max(...ends) - Math.min(...starts)\n  }\n\n  return { runId, steps, llmTurns, toolCalls, judgeVerdicts, retrievals, totalDurationMs }\n}\n\n// Re-export core types for convenience so consumers don't import from two paths.\nexport type { Span, TraceEvent } from './trace/schema'\nexport type { TraceStore } from './trace/store'\n","/**\n * Baseline regression detection.\n *\n * Lifted from ADC baseline.ts. Every promotion-blocking signal boils down\n * to: \"is this run measurably worse than baseline?\" — with enough\n * statistical rigor to distinguish noise from drift.\n *\n * Uses:\n *   - Welch's t-test (unequal variance) for per-metric mean comparison\n *   - Cohen's d for effect size magnitude\n *   - IQR for stability flag (unstable samples can't be trusted for comparisons)\n *\n * Returns a structured verdict: improved | regressed | stable | unstable.\n */\n\nimport { cohensD } from './statistics'\n\nexport interface MetricSamples {\n  /** Stable metric key (e.g. \"overallScore\", \"firstTokenMs\"). */\n  metric: string\n  /** Whether higher values are better. */\n  higherIsBetter: boolean\n  baseline: number[]\n  candidate: number[]\n}\n\nexport interface MetricVerdict {\n  metric: string\n  baselineMean: number\n  candidateMean: number\n  delta: number\n  cohensD: number\n  welchT: number\n  welchDf: number\n  welchP: number\n  stable: boolean\n  /** IQR of the combined samples — used as a rough stability indicator. */\n  iqr: number\n  verdict: 'improved' | 'regressed' | 'stable' | 'unstable'\n}\n\nexport interface BaselineReport {\n  metrics: MetricVerdict[]\n  /** True if any critical metric regressed. */\n  hasRegression: boolean\n  /** True if any metric is unstable (too noisy to judge). */\n  hasUnstable: boolean\n}\n\nexport interface BaselineOptions {\n  /** Effect size threshold for meaningful delta (default 0.5 — medium effect). */\n  effectThreshold?: number\n  /** p-value threshold for statistical significance (default 0.05). */\n  alpha?: number\n  /** IQR/mean ratio above which samples are flagged unstable (default 0.30). */\n  unstableCvThreshold?: number\n}\n\n/**\n * Compare candidate samples against baseline per metric. Verdict logic:\n *   - unstable: IQR/|mean| > threshold on either set — not enough signal\n *   - improved: meaningful effect in the \"better\" direction AND p < alpha\n *   - regressed: meaningful effect in the \"worse\" direction AND p < alpha\n *   - stable: otherwise (no significant change)\n */\nexport function compareToBaseline(\n  samples: MetricSamples[],\n  options: BaselineOptions = {},\n): BaselineReport {\n  const effectThreshold = options.effectThreshold ?? 0.5\n  const alpha = options.alpha ?? 0.05\n  const cvThreshold = options.unstableCvThreshold ?? 0.3\n\n  const metrics: MetricVerdict[] = samples.map((s) => {\n    if (s.baseline.length < 2 || s.candidate.length < 2) {\n      throw new Error(`compareToBaseline: need ≥2 samples per side for \"${s.metric}\"`)\n    }\n    const bMean = mean(s.baseline)\n    const cMean = mean(s.candidate)\n    const delta = cMean - bMean\n    const d = cohensD(s.baseline, s.candidate) // positive = candidate higher\n    const { t, df, p } = welchsTTest(s.baseline, s.candidate)\n    // Stability is per-side: a comparison is trustworthy only when BOTH\n    // samples are internally consistent. Combining the sides would flag\n    // large-but-real deltas as \"unstable\" which is exactly what we want\n    // to detect.\n    const baselineIqr = iqr(s.baseline)\n    const candidateIqr = iqr(s.candidate)\n    const baselineStable = baselineIqr / Math.max(Math.abs(bMean), 1e-9) <= cvThreshold\n    const candidateStable = candidateIqr / Math.max(Math.abs(cMean), 1e-9) <= cvThreshold\n    const stable = baselineStable && candidateStable\n    const reportedIqr = Math.max(baselineIqr, candidateIqr)\n\n    let verdict: MetricVerdict['verdict']\n    if (!stable) {\n      verdict = 'unstable'\n    } else if (p < alpha && Math.abs(d) >= effectThreshold) {\n      const candidateIsBetter = s.higherIsBetter ? delta > 0 : delta < 0\n      verdict = candidateIsBetter ? 'improved' : 'regressed'\n    } else {\n      verdict = 'stable'\n    }\n\n    return {\n      metric: s.metric,\n      baselineMean: bMean,\n      candidateMean: cMean,\n      delta,\n      cohensD: d,\n      welchT: t,\n      welchDf: df,\n      welchP: p,\n      stable,\n      iqr: reportedIqr,\n      verdict,\n    }\n  })\n\n  return {\n    metrics,\n    hasRegression: metrics.some((m) => m.verdict === 'regressed'),\n    hasUnstable: metrics.some((m) => m.verdict === 'unstable'),\n  }\n}\n\nfunction mean(xs: number[]): number {\n  return xs.reduce((a, b) => a + b, 0) / xs.length\n}\n\n/** Inter-quartile range; 0 when the sample has no spread. */\nexport function iqr(xs: number[]): number {\n  if (xs.length === 0) return 0\n  const sorted = [...xs].sort((a, b) => a - b)\n  const q = (p: number) => {\n    const idx = p * (sorted.length - 1)\n    const lo = Math.floor(idx)\n    const hi = Math.ceil(idx)\n    return sorted[lo]! + (sorted[hi]! - sorted[lo]!) * (idx - lo)\n  }\n  return q(0.75) - q(0.25)\n}\n\n/**\n * Welch's t-test — unequal-variance two-sample t. Uses the same Student-t\n * CDF as `pairedTTest` (via incomplete beta); falls back to normal tail\n * when df is large.\n */\nexport function welchsTTest(a: number[], b: number[]): { t: number; df: number; p: number } {\n  if (a.length < 2 || b.length < 2) return { t: 0, df: 0, p: 1 }\n  const mA = mean(a)\n  const mB = mean(b)\n  const vA = variance(a, mA)\n  const vB = variance(b, mB)\n  const seSquared = vA / a.length + vB / b.length\n  if (seSquared === 0) return { t: mA === mB ? 0 : Infinity, df: 0, p: mA === mB ? 1 : 0 }\n  const t = (mB - mA) / Math.sqrt(seSquared)\n  const df =\n    (seSquared * seSquared) /\n    ((vA / a.length) ** 2 / (a.length - 1) + (vB / b.length) ** 2 / (b.length - 1))\n  const p = 2 * (1 - studentTCdf(Math.abs(t), df))\n  return { t, df, p }\n}\n\nfunction variance(xs: number[], m: number): number {\n  return xs.reduce((acc, x) => acc + (x - m) ** 2, 0) / (xs.length - 1)\n}\n\n// Re-used from statistics.ts via small local copy to avoid exporting internals.\nfunction studentTCdf(t: number, df: number): number {\n  if (df <= 0) return 0.5\n  if (df > 100) return normalCdf(t)\n  const x = df / (df + t * t)\n  const ib = incompleteBeta(x, df / 2, 0.5)\n  return t >= 0 ? 1 - 0.5 * ib : 0.5 * ib\n}\n\nfunction incompleteBeta(x: number, a: number, b: number): number {\n  if (x <= 0) return 0\n  if (x >= 1) return 1\n  const lnBeta = lnGamma(a) + lnGamma(b) - lnGamma(a + b)\n  const front = Math.exp(Math.log(x) * a + Math.log(1 - x) * b - lnBeta) / a\n  let c = 1\n  let d = 1 - ((a + b) * x) / (a + 1)\n  if (Math.abs(d) < 1e-30) d = 1e-30\n  d = 1 / d\n  let f = d\n  for (let m = 1; m <= 200; m++) {\n    const m2 = 2 * m\n    let num = (m * (b - m) * x) / ((a + m2 - 1) * (a + m2))\n    d = 1 + num * d\n    if (Math.abs(d) < 1e-30) d = 1e-30\n    c = 1 + num / c\n    if (Math.abs(c) < 1e-30) c = 1e-30\n    d = 1 / d\n    f *= d * c\n    num = -((a + m) * (a + b + m) * x) / ((a + m2) * (a + m2 + 1))\n    d = 1 + num * d\n    if (Math.abs(d) < 1e-30) d = 1e-30\n    c = 1 + num / c\n    if (Math.abs(c) < 1e-30) c = 1e-30\n    d = 1 / d\n    const delta = d * c\n    f *= delta\n    if (Math.abs(delta - 1) < 3e-7) break\n  }\n  return front * f\n}\n\nfunction lnGamma(z: number): number {\n  const coefs = [\n    0.99999999999980993, 676.5203681218851, -1259.1392167224028, 771.32342877765313,\n    -176.61502916214059, 12.507343278686905, -0.13857109526572012, 9.9843695780195716e-6,\n    1.5056327351493116e-7,\n  ]\n  if (z < 0.5) return Math.log(Math.PI / Math.sin(Math.PI * z)) - lnGamma(1 - z)\n  z -= 1\n  let x = coefs[0]!\n  for (let i = 1; i < 9; i++) x += coefs[i]! / (z + i)\n  const t = z + 7.5\n  return 0.5 * Math.log(2 * Math.PI) + (z + 0.5) * Math.log(t) - t + Math.log(x)\n}\n\nfunction normalCdf(x: number): number {\n  const a1 = 0.254829592\n  const a2 = -0.284496736\n  const a3 = 1.421413741\n  const a4 = -1.453152027\n  const a5 = 1.061405429\n  const p = 0.3275911\n  const sign = x < 0 ? -1 : 1\n  const absX = Math.abs(x)\n  const t = 1 / (1 + p * absX)\n  const y = 1 - ((((a5 * t + a4) * t + a3) * t + a2) * t + a1) * t * Math.exp((-absX * absX) / 2)\n  return 0.5 * (1 + sign * y)\n}\n"],"mappings":";;;;;;;;;;AAwCO,IAAM,gBAA+B;AAAA;AAAA,EAE1C;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,IAAI,MAAM;AAClB,YAAM,KAAK,IAAI,SAAS;AACxB,UAAI,MAAM,OAAO;AACf,eAAO,EAAE,cAAc,IAAI,QAAQ,sCAAsC;AAC3E,aAAO;AAAA,IACT;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,sBACnB,EAAE,QAAQ,WAAW;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,aACT,EAAE,QAAQ,SAAS,oCAAoC,EAAE,QAAQ,UAAU,SAC1E,EAAE,QAAQ,SAAS,+BAClB,EAAE,QAAQ,SAAS;AAAA,MAC3B;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,mCACnB,oBAAoB,EAAE,SAAS,oBAAoB;AAAA,MACvD;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,aACT,EAAE,QAAQ,SAAS,mCAAmC,iBAAiB,EAAE,OAAO,KAC/E,EAAE,QAAQ,SAAS,+BAA+B,EAAE,QAAQ,SAAS;AAAA,MAC5E;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,aACT,EAAE,QAAQ,SAAS,wBAAwB,EAAE,QAAQ,WAAW,uBAC/D,EAAE,QAAQ,SAAS,+BAClB,EAAE,QAAQ,SAAS,uBACrB,EAAE,QAAQ,SAAS;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,gCAClB,EAAE,QAAQ,SAAS,kBAClB,EAAE,QAAQ,SAAS,2BACnB,EAAE,QAAQ,SAAS,wBACnB,EAAE,QAAQ,WAAW;AAAA,MAC3B;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,gCAClB,EAAE,QAAQ,SAAS,yBAClB,EAAE,QAAQ,SAAS,mBACnB,EAAE,QAAQ,SAAS;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,+BACnB,CAAC;AAAA,UACC;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,QACF,EAAE,SAAS,OAAO,EAAE,QAAQ,IAAI,CAAC;AAAA,MACrC;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,mBACnB,EAAE,QAAQ,aAAa;AAAA,MAC3B;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,KAAK,MAAM,MAAM;AACzB,UAAI,IAAI,SAAS,SAAS,MAAO,QAAO;AACxC,YAAM,YAAY,MAAM;AAAA,QACtB,CAAC,MACC,EAAE,SAAS,gBAAgB,EAAE,KAAK,WAAW,KAAK,EAAE,KAAK,MAAM,CAAC,QAAQ,IAAI,SAAS,CAAC;AAAA,MAC1F;AACA,aAAO,YACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,eAAe,UAAU;AAAA,MAC3B,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,mBACnB,EAAE,QAAQ,WAAW;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,QAAQ,OAAO;AAAA,QACnB,CAAC,MACC,EAAE,SAAS,YACX,EAAE,QAAQ,SAAS,mBACnB,EAAE,QAAQ,WAAW;AAAA,MACzB;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,MAAM;AAAA,MACxB,IACA;AAAA,IACN;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,SAAS,OAAO,KAAK,CAAC,MAAM,EAAE,SAAS,eAAe;AAC5D,aAAO,SACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ,sBAAsB,OAAO,QAAQ,aAAa,mBAAmB;AAAA,QAC7E,gBAAgB,OAAO;AAAA,MACzB,IACA;AAAA,IACN;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,OAAO,MAAM;AACrB,YAAM,IAAI,OAAO,KAAK,CAAC,MAAM,EAAE,SAAS,kBAAkB;AAC1D,aAAO,IACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,gBAAgB,EAAE;AAAA,MACpB,IACA;AAAA,IACN;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,MAAM,MAAM;AACpB,YAAM,IAAI,MAAM;AAAA,QACd,CAAC,MAAM,EAAE,SAAS,aAAa,OAAO,EAAE,aAAa,YAAY,EAAE,aAAa;AAAA,MAClF;AACA,UAAI,CAAC,EAAG,QAAO;AACf,aAAO;AAAA,QACL,cAAc;AAAA,QACd,QAAQ,kBAAmB,EAAyC,QAAQ;AAAA,QAC5E,eAAe,EAAE;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,KAAK,OAAO,MAAM;AAC1B,UAAI,IAAI,WAAW,UAAW,QAAO;AACrC,YAAM,aAAa,OAAO;AAAA,QACxB,CAAC,MACC,EAAE,SAAS,WACX,OAAO,EAAE,QAAQ,UAAU,EAAE,EAC1B,YAAY,EACZ,SAAS,SAAS;AAAA,MACzB;AACA,YAAM,QAAQ,IAAI,SAAS,SAAS,IAAI,YAAY;AACpD,UAAI,cAAc,KAAK,SAAS,SAAS,KAAK,KAAK,SAAS,UAAU,GAAG;AACvE,eAAO,EAAE,cAAc,WAAW,QAAQ,0BAA0B;AAAA,MACtE;AACA,aAAO;AAAA,IACT;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,MAAM,MAAM;AACpB,YAAM,QAAQ,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM;AACnD,YAAM,SAAS,oBAAI,IAAoB;AACvC,iBAAW,KAAK,OAAO;AACrB,cAAM,OAAQ,EAAsC;AACpD,cAAM,MAAM,OAAO,IAAI,IAAI,KAAK,CAAC;AACjC,YAAI,KAAK,CAAC;AACV,eAAO,IAAI,MAAM,GAAG;AAAA,MACtB;AACA,iBAAW,CAAC,MAAM,GAAG,KAAK,QAAQ;AAChC,cAAM,OAAO,IAAI,OAAO,CAAC,MAAM,EAAE,WAAW,OAAO;AACnD,YAAI,KAAK,UAAU,KAAK,KAAK,WAAW,IAAI,QAAQ;AAClD,iBAAO;AAAA,YACL,cAAc;AAAA,YACd,QAAQ,GAAG,KAAK,MAAM,gCAAgC,IAAI;AAAA,YAC1D,eAAe,KAAK,KAAK,SAAS,CAAC,EAAG;AAAA,UACxC;AAAA,QACF;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,KAAK,MAAM,MAAM;AACzB,UAAI,IAAI,SAAS,SAAS,MAAO,QAAO;AACxC,YAAM,oBAAoB,MAAM;AAAA,QAC9B,CAAC,MACC,EAAE,SAAS,WACV,EAAE,YAAY,mBAA0C,UACxD,EAAE,YAAY,iBAA4B;AAAA,MAC/C;AACA,YAAM,QAAQ,MAAM,OAAO,CAAC,MAAM,EAAE,SAAS,MAAM;AACnD,UAAI,qBAAqB,MAAM,WAAW,GAAG;AAC3C,eAAO;AAAA,UACL,cAAc;AAAA,UACd,QAAQ;AAAA,QACV;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,OAAO,CAAC,EAAE,MAAM,MAAM;AACpB,YAAM,QAAQ,MAAM;AAAA,QAClB,CAAC,MACC,EAAE,SAAS,WACV,EAAuC,cAAc,YACrD,EAAuC,QAAQ;AAAA,MACpD;AACA,aAAO,QACH;AAAA,QACE,cAAc;AAAA,QACd,QAAQ;AAAA,QACR,eAAe,MAAM;AAAA,MACvB,IACA;AAAA,IACN;AAAA,EACF;AACF;AAEA,SAAS,oBAAoB,SAAkC,QAAyB;AACtF,MAAI,WAAW,wBAAwB,YAAY,QAAQ,kBAAkB,EAAE,SAAS;AACtF,WAAO;AACT,SAAO,gBAAgB,OAAO,EAAE,KAAK,CAAC,SAAS,KAAK,WAAW,MAAM;AACvE;AAEA,SAAS,iBAAiB,SAA2C;AACnE,MAAI,YAAY,QAAQ,aAAa,EAAE,SAAS,EAAG,QAAO;AAC1D,SAAO,gBAAgB,OAAO,EAAE;AAAA,IAC9B,CAAC,SAAS,MAAM,QAAQ,KAAK,aAAa,KAAK,KAAK,cAAc,SAAS;AAAA,EAC7E;AACF;AAEA,SAAS,gBAAgB,SAAkE;AACzF,SAAO;AAAA,IACL,GAAG,QAAQ,QAAQ,OAAO;AAAA,IAC1B,GAAG,QAAQ,QAAQ,eAAe;AAAA,IAClC,GAAG,QAAQ,QAAQ,KAAK;AAAA,EAC1B;AACF;AAEA,SAAS,QAAQ,OAAgD;AAC/D,MAAI,CAAC,MAAM,QAAQ,KAAK,EAAG,QAAO,CAAC;AACnC,SAAO,MAAM;AAAA,IACX,CAAC,SACC,QAAQ,IAAI,KAAK,OAAO,SAAS,YAAY,CAAC,MAAM,QAAQ,IAAI;AAAA,EACpE;AACF;AAEA,SAAS,YAAY,OAA0B;AAC7C,SAAO,MAAM,QAAQ,KAAK,IACtB,MAAM,OAAO,CAAC,SAAyB,OAAO,SAAS,QAAQ,IAC/D,CAAC;AACP;AAGO,SAAS,gBACd,KACA,QAAuB,eACA;AACvB,MAAI,IAAI,IAAI,SAAS,SAAS,SAAS,IAAI,IAAI,WAAW,aAAa;AACrE,WAAO,EAAE,cAAc,WAAW,QAAQ,qDAAqD;AAAA,EACjG;AACA,aAAW,QAAQ,OAAO;AACxB,UAAM,MAAM,KAAK,MAAM,GAAG;AAC1B,QAAI,IAAK,QAAO;AAAA,EAClB;AACA,SAAO,EAAE,cAAc,WAAW,QAAQ,sDAAsD;AAClG;;;ACvaA,eAAsB,sBACpB,OACA,OACA,UAA0B,CAAC,GACF;AACzB,QAAM,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC1C,MAAI,MAAM,WAAW,GAAG;AACtB,WAAO,EAAE,OAAO,YAAY,GAAG,QAAQ,CAAC,GAAG,WAAW,GAAG,eAAe,GAAG,WAAW,EAAE;AAAA,EAC1F;AAEA,QAAM,SAAoC,CAAC;AAC3C,MAAI,cAAc;AAClB,MAAI,kBAAkB;AACtB,QAAM,cAAc,CAAC,GAAG,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACvE,QAAM,iBAAiB,oBAAI,IAAY;AAGvC,aAAW,KAAK,aAAa;AAC3B,WAAO,EAAE,QAAQ,MAAM,EAAE,OAAO,GAAG,QAAQ,GAAG,cAAc,GAAG,YAAY,EAAE;AAC7E,UAAM,OAAO,OAAO,EAAE,QAAQ;AAC9B,SAAK,SAAS;AACd,QAAI,EAAE,WAAW,SAAS;AACxB,WAAK,UAAU;AACf,qBAAe;AAAA,IACjB;AACA,QAAI,OAAO,EAAE,cAAc,SAAU,MAAK,gBAAgB,EAAE;AAC5D,UAAM,MAAM,GAAG,EAAE,QAAQ,IAAI,QAAQ,EAAE,IAAI,CAAC;AAC5C,QAAI,eAAe,IAAI,GAAG,GAAG;AAC3B,WAAK,cAAc;AACnB,yBAAmB;AAAA,IACrB;AACA,mBAAe,IAAI,GAAG;AAAA,EACxB;AAEA,aAAW,QAAQ,OAAO,OAAO,MAAM,GAAG;AACxC,SAAK,eAAe,KAAK,QAAQ,IAAI,KAAK,eAAe,KAAK,QAAQ;AAAA,EACxE;AAGA,MAAI,qBAAqB;AACzB,MAAI,kBAAkB;AACtB,aAAW,CAAC,EAAE,GAAG,KAAK,QAAQ,aAAa,CAAC,MAAM,EAAE,QAAQ,GAAG;AAC7D,aAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK;AACnC,UAAI,IAAI,CAAC,EAAG,WAAW,QAAS;AAChC,4BAAsB;AACtB,UAAI,IAAI,IAAI,CAAC,EAAG,oBAAmB;AAAA,IACrC;AAAA,EACF;AACA,QAAM,YAAY,qBAAqB,IAAI,kBAAkB,qBAAqB;AAElF,MAAI;AACJ,MAAI,QAAQ,iBAAiB;AAC3B,UAAM,UAAU,YAAY,OAAO,CAAC,MAAM,EAAE,UAAU,QAAQ,eAAgB;AAC9E,QAAI,QAAQ,SAAS,GAAG;AACtB,0BACE,QAAQ,OAAO,CAAC,MAAM,QAAQ,gBAAiB,EAAE,MAAM,CAAC,EAAE,SAAS,QAAQ;AAAA,IAC/E;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA,YAAY,YAAY;AAAA,IACxB;AAAA,IACA,WAAW,cAAc,YAAY;AAAA,IACrC,eAAe,kBAAkB,YAAY;AAAA,IAC7C;AAAA,IACA;AAAA,EACF;AACF;;;AC1EA,eAAsB,gBAAgB,OAAmB,OAAoC;AAC3F,QAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;AACzC,QAAM,SAAS,MAAM,MAAM,OAAO,EAAE,MAAM,CAAC;AAC3C,QAAM,aAAa,oBAAI,IAAgC;AACvD,aAAW,KAAK,OAAO;AACrB,UAAM,MAAM,WAAW,IAAI,EAAE,YAAY,KAAK,CAAC;AAC/C,QAAI,KAAK,CAAC;AACV,eAAW,IAAI,EAAE,cAAc,GAAG;AAAA,EACpC;AAEA,aAAW,OAAO,WAAW,OAAO,EAAG,KAAI,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAEnF,QAAM,eAAe,oBAAI,IAA0B;AACnD,aAAW,KAAK,QAAQ;AACtB,QAAI,CAAC,EAAE,OAAQ;AACf,UAAM,MAAM,aAAa,IAAI,EAAE,MAAM,KAAK,CAAC;AAC3C,QAAI,KAAK,CAAC;AACV,iBAAa,IAAI,EAAE,QAAQ,GAAG;AAAA,EAChC;AAEA,QAAM,QAA0B,CAAC;AACjC,QAAM,OAAO,CAAC,QAA4B,UAAwB;AAChE,UAAM,OAAO,WAAW,IAAI,MAAM,KAAK,CAAC;AACxC,eAAW,SAAS,MAAM;AACxB,YAAM,KAAK;AAAA,QACT,OAAO,MAAM;AAAA,QACb,MAAM;AAAA,QACN;AAAA,QACA,QAAQ,aAAa,IAAI,MAAM,MAAM,KAAK,CAAC;AAAA,MAC7C,CAAC;AACD,WAAK,MAAM,QAAQ,QAAQ,CAAC;AAAA,IAC9B;AAAA,EACF;AACA,OAAK,QAAW,CAAC;AAEjB,QAAM,WAAW,MAAM,OAAO,CAAC,MAAM,EAAE,KAAK,SAAS,KAAK,EAAE;AAC5D,QAAM,YAAY,MAAM,OAAO,CAAC,MAAM,EAAE,KAAK,SAAS,MAAM,EAAE;AAC9D,QAAM,gBAAgB,MAAM,OAAO,CAAC,MAAM,EAAE,KAAK,SAAS,OAAO,EAAE;AACnE,QAAM,aAAa,MAAM,OAAO,CAAC,MAAM,EAAE,KAAK,SAAS,WAAW,EAAE;AAEpE,MAAI,kBAAkB;AACtB,MAAI,MAAM,SAAS,GAAG;AACpB,UAAM,SAAS,MAAM,IAAI,CAAC,MAAM,EAAE,SAAS;AAC3C,UAAM,OAAO,MAAM,IAAI,CAAC,MAAM,EAAE,WAAW,EAAE,SAAS;AACtD,sBAAkB,KAAK,IAAI,GAAG,IAAI,IAAI,KAAK,IAAI,GAAG,MAAM;AAAA,EAC1D;AAEA,SAAO,EAAE,OAAO,OAAO,UAAU,WAAW,eAAe,YAAY,gBAAgB;AACzF;;;ACfO,SAAS,kBACd,SACA,UAA2B,CAAC,GACZ;AAChB,QAAM,kBAAkB,QAAQ,mBAAmB;AACnD,QAAM,QAAQ,QAAQ,SAAS;AAC/B,QAAM,cAAc,QAAQ,uBAAuB;AAEnD,QAAM,UAA2B,QAAQ,IAAI,CAAC,MAAM;AAClD,QAAI,EAAE,SAAS,SAAS,KAAK,EAAE,UAAU,SAAS,GAAG;AACnD,YAAM,IAAI,MAAM,yDAAoD,EAAE,MAAM,GAAG;AAAA,IACjF;AACA,UAAM,QAAQ,KAAK,EAAE,QAAQ;AAC7B,UAAM,QAAQ,KAAK,EAAE,SAAS;AAC9B,UAAM,QAAQ,QAAQ;AACtB,UAAM,IAAI,QAAQ,EAAE,UAAU,EAAE,SAAS;AACzC,UAAM,EAAE,GAAG,IAAI,EAAE,IAAI,YAAY,EAAE,UAAU,EAAE,SAAS;AAKxD,UAAM,cAAc,IAAI,EAAE,QAAQ;AAClC,UAAM,eAAe,IAAI,EAAE,SAAS;AACpC,UAAM,iBAAiB,cAAc,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,IAAI,KAAK;AACxE,UAAM,kBAAkB,eAAe,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,IAAI,KAAK;AAC1E,UAAM,SAAS,kBAAkB;AACjC,UAAM,cAAc,KAAK,IAAI,aAAa,YAAY;AAEtD,QAAI;AACJ,QAAI,CAAC,QAAQ;AACX,gBAAU;AAAA,IACZ,WAAW,IAAI,SAAS,KAAK,IAAI,CAAC,KAAK,iBAAiB;AACtD,YAAM,oBAAoB,EAAE,iBAAiB,QAAQ,IAAI,QAAQ;AACjE,gBAAU,oBAAoB,aAAa;AAAA,IAC7C,OAAO;AACL,gBAAU;AAAA,IACZ;AAEA,WAAO;AAAA,MACL,QAAQ,EAAE;AAAA,MACV,cAAc;AAAA,MACd,eAAe;AAAA,MACf;AAAA,MACA,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,SAAS;AAAA,MACT,QAAQ;AAAA,MACR;AAAA,MACA,KAAK;AAAA,MACL;AAAA,IACF;AAAA,EACF,CAAC;AAED,SAAO;AAAA,IACL;AAAA,IACA,eAAe,QAAQ,KAAK,CAAC,MAAM,EAAE,YAAY,WAAW;AAAA,IAC5D,aAAa,QAAQ,KAAK,CAAC,MAAM,EAAE,YAAY,UAAU;AAAA,EAC3D;AACF;AAEA,SAAS,KAAK,IAAsB;AAClC,SAAO,GAAG,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAGO,SAAS,IAAI,IAAsB;AACxC,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,QAAM,SAAS,CAAC,GAAG,EAAE,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3C,QAAM,IAAI,CAAC,MAAc;AACvB,UAAM,MAAM,KAAK,OAAO,SAAS;AACjC,UAAM,KAAK,KAAK,MAAM,GAAG;AACzB,UAAM,KAAK,KAAK,KAAK,GAAG;AACxB,WAAO,OAAO,EAAE,KAAM,OAAO,EAAE,IAAK,OAAO,EAAE,MAAO,MAAM;AAAA,EAC5D;AACA,SAAO,EAAE,IAAI,IAAI,EAAE,IAAI;AACzB;AAOO,SAAS,YAAY,GAAa,GAAmD;AAC1F,MAAI,EAAE,SAAS,KAAK,EAAE,SAAS,EAAG,QAAO,EAAE,GAAG,GAAG,IAAI,GAAG,GAAG,EAAE;AAC7D,QAAM,KAAK,KAAK,CAAC;AACjB,QAAM,KAAK,KAAK,CAAC;AACjB,QAAM,KAAK,SAAS,GAAG,EAAE;AACzB,QAAM,KAAK,SAAS,GAAG,EAAE;AACzB,QAAM,YAAY,KAAK,EAAE,SAAS,KAAK,EAAE;AACzC,MAAI,cAAc,EAAG,QAAO,EAAE,GAAG,OAAO,KAAK,IAAI,UAAU,IAAI,GAAG,GAAG,OAAO,KAAK,IAAI,EAAE;AACvF,QAAM,KAAK,KAAK,MAAM,KAAK,KAAK,SAAS;AACzC,QAAM,KACH,YAAY,cACX,KAAK,EAAE,WAAW,KAAK,EAAE,SAAS,MAAM,KAAK,EAAE,WAAW,KAAK,EAAE,SAAS;AAC9E,QAAM,IAAI,KAAK,IAAI,YAAY,KAAK,IAAI,CAAC,GAAG,EAAE;AAC9C,SAAO,EAAE,GAAG,IAAI,EAAE;AACpB;AAEA,SAAS,SAAS,IAAc,GAAmB;AACjD,SAAO,GAAG,OAAO,CAAC,KAAK,MAAM,OAAO,IAAI,MAAM,GAAG,CAAC,KAAK,GAAG,SAAS;AACrE;AAGA,SAAS,YAAY,GAAW,IAAoB;AAClD,MAAI,MAAM,EAAG,QAAO;AACpB,MAAI,KAAK,IAAK,QAAO,UAAU,CAAC;AAChC,QAAM,IAAI,MAAM,KAAK,IAAI;AACzB,QAAM,KAAK,eAAe,GAAG,KAAK,GAAG,GAAG;AACxC,SAAO,KAAK,IAAI,IAAI,MAAM,KAAK,MAAM;AACvC;AAEA,SAAS,eAAe,GAAW,GAAW,GAAmB;AAC/D,MAAI,KAAK,EAAG,QAAO;AACnB,MAAI,KAAK,EAAG,QAAO;AACnB,QAAM,SAAS,QAAQ,CAAC,IAAI,QAAQ,CAAC,IAAI,QAAQ,IAAI,CAAC;AACtD,QAAM,QAAQ,KAAK,IAAI,KAAK,IAAI,CAAC,IAAI,IAAI,KAAK,IAAI,IAAI,CAAC,IAAI,IAAI,MAAM,IAAI;AACzE,MAAI,IAAI;AACR,MAAI,IAAI,KAAM,IAAI,KAAK,KAAM,IAAI;AACjC,MAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,MAAI,IAAI;AACR,MAAI,IAAI;AACR,WAAS,IAAI,GAAG,KAAK,KAAK,KAAK;AAC7B,UAAM,KAAK,IAAI;AACf,QAAI,MAAO,KAAK,IAAI,KAAK,MAAO,IAAI,KAAK,MAAM,IAAI;AACnD,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI;AACR,SAAK,IAAI;AACT,UAAM,GAAG,IAAI,MAAM,IAAI,IAAI,KAAK,OAAO,IAAI,OAAO,IAAI,KAAK;AAC3D,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI,MAAM;AACd,QAAI,KAAK,IAAI,CAAC,IAAI,MAAO,KAAI;AAC7B,QAAI,IAAI;AACR,UAAM,QAAQ,IAAI;AAClB,SAAK;AACL,QAAI,KAAK,IAAI,QAAQ,CAAC,IAAI,KAAM;AAAA,EAClC;AACA,SAAO,QAAQ;AACjB;AAEA,SAAS,QAAQ,GAAmB;AAClC,QAAM,QAAQ;AAAA,IACZ;AAAA,IAAqB;AAAA,IAAmB;AAAA,IAAqB;AAAA,IAC7D;AAAA,IAAqB;AAAA,IAAoB;AAAA,IAAsB;AAAA,IAC/D;AAAA,EACF;AACA,MAAI,IAAI,IAAK,QAAO,KAAK,IAAI,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,CAAC,CAAC,IAAI,QAAQ,IAAI,CAAC;AAC7E,OAAK;AACL,MAAI,IAAI,MAAM,CAAC;AACf,WAAS,IAAI,GAAG,IAAI,GAAG,IAAK,MAAK,MAAM,CAAC,KAAM,IAAI;AAClD,QAAM,IAAI,IAAI;AACd,SAAO,MAAM,KAAK,IAAI,IAAI,KAAK,EAAE,KAAK,IAAI,OAAO,KAAK,IAAI,CAAC,IAAI,IAAI,KAAK,IAAI,CAAC;AAC/E;AAEA,SAAS,UAAU,GAAmB;AACpC,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,KAAK;AACX,QAAM,IAAI;AACV,QAAM,OAAO,IAAI,IAAI,KAAK;AAC1B,QAAM,OAAO,KAAK,IAAI,CAAC;AACvB,QAAM,IAAI,KAAK,IAAI,IAAI;AACvB,QAAM,IAAI,QAAQ,KAAK,IAAI,MAAM,IAAI,MAAM,IAAI,MAAM,IAAI,MAAM,IAAI,KAAK,IAAK,CAAC,OAAO,OAAQ,CAAC;AAC9F,SAAO,OAAO,IAAI,OAAO;AAC3B;","names":[]}