@tangle-network/agent-eval 0.80.0 → 0.81.0

package/dist/chunk-S42AWHMP.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/contract/intake/code-agent-session.ts"],"sourcesContent":["import { createHash } from 'node:crypto'\nimport { estimateCost, isModelPriced } from '../../metrics'\nimport type { RunRecord, RunSplitTag, RunTokenUsage } from '../../run-record'\n\nexport type CodeAgentSessionSource = 'codex' | 'claude-code' | 'opencode' | 'kimi-code' | 'pi'\n\nexport interface ParsedCodeAgentJsonl {\n  entries: unknown[]\n  malformedLines: number\n}\n\nexport interface CodeAgentSessionMetrics {\n  entries: number\n  userMessages: number\n  assistantMessages: number\n  reasoningItems: number\n  toolCalls: number\n  toolOutputs: number\n  toolErrors: number\n  patchAttempts: number\n  patchSuccesses: number\n  patchFailures: number\n  turnsStarted: number\n  turnsCompleted: number\n  turnsAborted: number\n  contextCompactions: number\n  prLinks: number\n  fileSnapshots: number\n  graphNodes: number\n  graphEdges: number\n  actionCandidates: number\n  verificationReports: number\n  completionDecisions: number\n  reliabilityRows: number\n  reliabilityLift: number\n  inputTokens: number\n  outputTokens: number\n  cachedTokens: number\n  observedCostUsd: number\n  wallMs: number\n  processScore: number\n}\n\nexport interface CodeAgentSessionDiagnostic {\n  source: CodeAgentSessionSource\n  sessionId: string\n  sourcePath?: string\n  entries: number\n  malformedLines: number\n  inferredScore: boolean\n  hasExplicitTerminalSignal: boolean\n  hasQualityLabel: boolean\n  hasTokenUsage: boolean\n  hasCost: boolean\n  warnings: string[]\n}\n\nexport interface CodeAgentSessionIntakeResult {\n  runs: RunRecord[]\n  diagnostics: CodeAgentSessionDiagnostic[]\n  metrics: CodeAgentSessionMetrics[]\n}\n\nexport interface CodeAgentSessionIntakeOptions {\n  entries: unknown[]\n  malformedLines?: number\n  sourcePath?: string\n  experimentId?: string\n  candidateId?: string\n  seed?: number\n  splitTag?: RunSplitTag\n  scenarioId?: string\n  model?: string\n  promptHash?: string\n  configHash?: string\n  commitSha?: string\n  score?: number\n}\n\nexport function parseCodeAgentJsonl(jsonl: string): ParsedCodeAgentJsonl {\n  const entries: unknown[] = []\n  let malformedLines = 0\n  for (const line of jsonl.split('\\n')) {\n    const trimmed = line.trim()\n    if (!trimmed) continue\n    try {\n      entries.push(JSON.parse(trimmed))\n    } catch {\n      malformedLines += 1\n    }\n  }\n  return { entries, malformedLines }\n}\n\nexport function fromCodexSession(\n  options: CodeAgentSessionIntakeOptions,\n): CodeAgentSessionIntakeResult {\n  return fromCodeAgentSession('codex', options)\n}\n\nexport function fromClaudeCodeSession(\n  options: CodeAgentSessionIntakeOptions,\n): CodeAgentSessionIntakeResult {\n  return fromCodeAgentSession('claude-code', options)\n}\n\nexport function fromOpenCodeSession(\n  options: CodeAgentSessionIntakeOptions,\n): CodeAgentSessionIntakeResult {\n  return fromCodeAgentSession('opencode', options)\n}\n\nexport function fromKimiCodeSession(\n  options: CodeAgentSessionIntakeOptions,\n): CodeAgentSessionIntakeResult {\n  return fromCodeAgentSession('kimi-code', options)\n}\n\nexport function fromPiSession(\n  options: CodeAgentSessionIntakeOptions,\n): CodeAgentSessionIntakeResult {\n  return fromCodeAgentSession('pi', options)\n}\n\nexport const fromPigraphSession = fromPiSession\n\nfunction fromCodeAgentSession(\n  source: CodeAgentSessionSource,\n  options: CodeAgentSessionIntakeOptions,\n): CodeAgentSessionIntakeResult {\n  const entries = options.entries.filter(isRecord)\n  if (entries.length === 0) {\n    return {\n      runs: [],\n      diagnostics: [\n        {\n          source,\n          sessionId: fallbackSessionId(source, options.sourcePath),\n          sourcePath: options.sourcePath,\n          entries: 0,\n          malformedLines: options.malformedLines ?? 0,\n          inferredScore: true,\n          hasExplicitTerminalSignal: false,\n          hasQualityLabel: false,\n          hasTokenUsage: false,\n          hasCost: false,\n          warnings: ['no parseable session entries'],\n        },\n      ],\n      metrics: [],\n    }\n  }\n\n  const metrics = metricsFor(source, entries)\n  const sessionId =\n    sessionIdFromEntries(source, entries) ?? fallbackSessionId(source, options.sourcePath)\n  const model = withSnapshot(options.model ?? modelFromEntries(source, entries) ?? source)\n  const tokenUsage: RunTokenUsage = {\n    input: metrics.inputTokens,\n    output: metrics.outputTokens,\n    ...(metrics.cachedTokens > 0 ? { cached: metrics.cachedTokens } : {}),\n  }\n  const estimatedCostUsd =\n    metrics.inputTokens > 0 || metrics.outputTokens > 0\n      ? estimateCost(metrics.inputTokens, metrics.outputTokens, model)\n      : 0\n  const costUsd = metrics.observedCostUsd > 0 ? metrics.observedCostUsd : estimatedCostUsd\n  const score = clamp01(options.score ?? metrics.processScore)\n  const promptHash =\n    options.promptHash ?? hashString(`prompt:${source}:${firstUserText(entries) ?? ''}`)\n  const configHash =\n    options.configHash ??\n    hashJson({\n      source,\n      model,\n      sourcePath: options.sourcePath,\n      cwd: cwdFromEntries(entries),\n      entryCount: entries.length,\n    })\n  const explicitTerminal = hasExplicitTerminalSignal(source, metrics)\n  const warnings = diagnosticsFor(metrics, {\n    model,\n    explicitTerminal,\n    malformedLines: options.malformedLines ?? 0,\n    scoreOverridden: options.score !== undefined,\n    costUsd,\n  })\n\n  const run: RunRecord = {\n    runId: `${source}:${sessionId}`,\n    experimentId: options.experimentId ?? `${source}-local-sessions`,\n    candidateId: options.candidateId ?? model,\n    seed: options.seed ?? stableSeed(sessionId),\n    model,\n    promptHash,\n    configHash,\n    commitSha: options.commitSha ?? 'local-session',\n    wallMs: metrics.wallMs,\n    costUsd,\n    tokenUsage,\n    outcome: {\n      holdoutScore: score,\n      raw: {\n        entries: metrics.entries,\n        user_messages: metrics.userMessages,\n        assistant_messages: metrics.assistantMessages,\n        reasoning_items: metrics.reasoningItems,\n        tool_calls: metrics.toolCalls,\n        tool_outputs: metrics.toolOutputs,\n        tool_errors: metrics.toolErrors,\n        patch_attempts: metrics.patchAttempts,\n        patch_successes: metrics.patchSuccesses,\n        patch_failures: metrics.patchFailures,\n        turns_started: metrics.turnsStarted,\n        turns_completed: metrics.turnsCompleted,\n        turns_aborted: metrics.turnsAborted,\n        context_compactions: metrics.contextCompactions,\n        pr_links: metrics.prLinks,\n        file_snapshots: metrics.fileSnapshots,\n        graph_nodes: metrics.graphNodes,\n        graph_edges: metrics.graphEdges,\n        action_candidates: metrics.actionCandidates,\n        verification_reports: metrics.verificationReports,\n        completion_decisions: metrics.completionDecisions,\n        reliability_rows: metrics.reliabilityRows,\n        reliability_lift: metrics.reliabilityLift,\n        input_tokens: metrics.inputTokens,\n        output_tokens: metrics.outputTokens,\n        cached_tokens: metrics.cachedTokens,\n        observed_cost_usd: metrics.observedCostUsd,\n        process_score: score,\n        inferred_score: options.score === undefined ? 1 : 0,\n        explicit_terminal_signal: explicitTerminal ? 1 : 0,\n        quality_label_present: options.score !== undefined ? 1 : 0,\n        cost_unknown: costUsd === 0 && !isModelPriced(model) ? 1 : 0,\n      },\n    },\n    splitTag: options.splitTag ?? 'holdout',\n    ...(options.scenarioId ? { scenarioId: options.scenarioId } : {}),\n    ...(metrics.toolErrors > 0 || metrics.turnsAborted > 0\n      ? { failureMode: metrics.turnsAborted > 0 ? 'turn_aborted' : 'tool_error' }\n      : {}),\n  }\n\n  return {\n    runs: [run],\n    diagnostics: [\n      {\n        source,\n        sessionId,\n        sourcePath: options.sourcePath,\n        entries: entries.length,\n        malformedLines: options.malformedLines ?? 0,\n        inferredScore: options.score === undefined,\n        hasExplicitTerminalSignal: explicitTerminal,\n        hasQualityLabel: options.score !== undefined,\n        hasTokenUsage: metrics.inputTokens > 0 || metrics.outputTokens > 0,\n        hasCost: costUsd > 0,\n        warnings,\n      },\n    ],\n    metrics: [metrics],\n  }\n}\n\nfunction metricsFor(\n  source: CodeAgentSessionSource,\n  entries: Record<string, unknown>[],\n): CodeAgentSessionMetrics {\n  switch (source) {\n    case 'codex':\n      return codexMetrics(entries)\n    case 'claude-code':\n      return claudeCodeMetrics(entries)\n    case 'opencode':\n      return openCodeMetrics(entries)\n    case 'kimi-code':\n      return kimiCodeMetrics(entries)\n    case 'pi':\n      return piMetrics(entries)\n  }\n}\n\nfunction codexMetrics(entries: Record<string, unknown>[]): CodeAgentSessionMetrics {\n  const metrics = emptyMetrics(entries.length)\n  let startedAt: number | undefined\n  let completedAt: number | undefined\n\n  for (const entry of entries) {\n    const payload = record(entry.payload) ?? {}\n    const entryType = stringField(entry, 'type')\n    const payloadType = stringField(payload, 'type')\n    const timestamp = timestampMs(entry.timestamp)\n    if (timestamp !== undefined) {\n      startedAt = startedAt === undefined ? timestamp : Math.min(startedAt, timestamp)\n      completedAt = completedAt === undefined ? timestamp : Math.max(completedAt, timestamp)\n    }\n\n    if (entryType === 'response_item') {\n      if (payloadType === 'function_call' || payloadType === 'custom_tool_call')\n        metrics.toolCalls += 1\n      if (payloadType === 'function_call_output' || payloadType === 'custom_tool_call_output')\n        metrics.toolOutputs += 1\n      if (payloadType === 'reasoning') metrics.reasoningItems += 1\n      if (payloadType === 'message' && stringField(payload, 'role') === 'assistant')\n        metrics.assistantMessages += 1\n    }\n\n    if (entryType === 'event_msg') {\n      if (payloadType === 'user_message') metrics.userMessages += 1\n      if (payloadType === 'agent_message') metrics.assistantMessages += 1\n      if (payloadType === 'task_started') {\n        metrics.turnsStarted += 1\n        const started = timestampMs(payload.started_at)\n        if (started !== undefined) {\n          startedAt = startedAt === undefined ? started : Math.min(startedAt, started)\n        }\n      }\n      if (payloadType === 'task_complete') {\n        metrics.turnsCompleted += 1\n        const completed = timestampMs(payload.completed_at)\n        if (completed !== undefined) {\n          completedAt = completedAt === undefined ? completed : Math.max(completedAt, completed)\n        }\n        const duration = numberField(payload, 'duration_ms')\n        if (duration !== undefined) metrics.wallMs = Math.max(metrics.wallMs, duration)\n      }\n      if (payloadType === 'turn_aborted') {\n        metrics.turnsAborted += 1\n        const completed = timestampMs(payload.completed_at)\n        if (completed !== undefined) {\n          completedAt = completedAt === undefined ? completed : Math.max(completedAt, completed)\n        }\n      }\n      if (payloadType === 'context_compacted') metrics.contextCompactions += 1\n      if (payloadType === 'patch_apply_end') {\n        metrics.patchAttempts += 1\n        if (payload.success === true) metrics.patchSuccesses += 1\n        else {\n          metrics.patchFailures += 1\n          metrics.toolErrors += 1\n        }\n      }\n      if (payloadType === 'token_count')\n        setCumulativeUsage(metrics, record(record(payload.info)?.total_token_usage))\n      const result = record(payload.result)\n      if (result && 'Err' in result) metrics.toolErrors += 1\n    }\n  }\n\n  if (metrics.wallMs === 0 && startedAt !== undefined && completedAt !== undefined) {\n    metrics.wallMs = Math.max(0, completedAt - startedAt)\n  }\n  metrics.processScore = codexProcessScore(metrics)\n  return metrics\n}\n\nfunction claudeCodeMetrics(entries: Record<string, unknown>[]): CodeAgentSessionMetrics {\n  const metrics = emptyMetrics(entries.length)\n  let startedAt: number | undefined\n  let completedAt: number | undefined\n\n  for (const entry of entries) {\n    const type = stringField(entry, 'type')\n    const timestamp = timestampMs(entry.timestamp)\n    if (timestamp !== undefined) {\n      startedAt = startedAt === undefined ? timestamp : Math.min(startedAt, timestamp)\n      completedAt = completedAt === undefined ? timestamp : Math.max(completedAt, timestamp)\n    }\n\n    if (type === 'user') metrics.userMessages += 1\n    if (type === 'assistant') metrics.assistantMessages += 1\n    if (type === 'pr-link') metrics.prLinks += 1\n    if (type === 'file-history-snapshot') metrics.fileSnapshots += 1\n\n    const message = record(entry.message)\n    if (message) addUsage(metrics, record(message.usage))\n    const content = Array.isArray(message?.content) ? message.content : []\n    for (const item of content) {\n      const part = record(item)\n      if (!part) continue\n      const partType = stringField(part, 'type')\n      if (partType === 'thinking') metrics.reasoningItems += 1\n      if (partType === 'tool_use') metrics.toolCalls += 1\n      if (partType === 'tool_result') {\n        metrics.toolOutputs += 1\n        if (part.is_error === true) metrics.toolErrors += 1\n      }\n    }\n  }\n\n  if (startedAt !== undefined && completedAt !== undefined)\n    metrics.wallMs = Math.max(0, completedAt - startedAt)\n  metrics.processScore = claudeProcessScore(metrics)\n  return metrics\n}\n\nfunction openCodeMetrics(entries: Record<string, unknown>[]): CodeAgentSessionMetrics {\n  const metrics = emptyMetrics(entries.length)\n  const messageUsage = emptyTokenTotals()\n  const partUsage = emptyTokenTotals()\n  let messageCost = 0\n  let partCost = 0\n  let startedAt: number | undefined\n  let completedAt: number | undefined\n\n  for (const entry of entries) {\n    const time = record(entry.time)\n    const created = timestampMs(time?.created)\n    const completed = timestampMs(time?.completed)\n    if (created !== undefined)\n      startedAt = startedAt === undefined ? created : Math.min(startedAt, created)\n    if (completed !== undefined)\n      completedAt = completedAt === undefined ? completed : Math.max(completedAt, completed)\n\n    const role = stringField(entry, 'role')\n    if (role === 'user') metrics.userMessages += 1\n    if (role === 'assistant') {\n      metrics.assistantMessages += 1\n      const finish = stringField(entry, 'finish')\n      if (finish === 'stop') metrics.turnsCompleted += 1\n      if (finish === 'error') metrics.turnsAborted += 1\n    }\n\n    const type = stringField(entry, 'type')\n    if (type === 'reasoning') metrics.reasoningItems += 1\n    if (type === 'tool') {\n      metrics.toolCalls += 1\n      const status = stringField(record(entry.state) ?? {}, 'status')\n      if (status === 'completed') metrics.toolOutputs += 1\n      if (status === 'error') {\n        metrics.toolOutputs += 1\n        metrics.toolErrors += 1\n      }\n    }\n    if (type === 'patch') {\n      metrics.patchAttempts += 1\n      metrics.patchSuccesses += 1\n    }\n\n    const cost = numberField(entry, 'cost')\n    if (record(entry.tokens) && role) {\n      addUsageTo(messageUsage, record(entry.tokens))\n      if (cost !== undefined) messageCost += cost\n    } else if (record(entry.tokens)) {\n      addUsageTo(partUsage, record(entry.tokens))\n      if (cost !== undefined) partCost += cost\n    }\n  }\n\n  const usage = messageUsage.input + messageUsage.output > 0 ? messageUsage : partUsage\n  metrics.inputTokens = usage.input\n  metrics.outputTokens = usage.output\n  metrics.cachedTokens = usage.cached\n  metrics.observedCostUsd = messageCost > 0 ? messageCost : partCost\n  if (metrics.wallMs === 0 && startedAt !== undefined && completedAt !== undefined)\n    metrics.wallMs = Math.max(0, completedAt - startedAt)\n  metrics.processScore = terminalProcessScore(metrics)\n  return metrics\n}\n\nfunction kimiCodeMetrics(entries: Record<string, unknown>[]): CodeAgentSessionMetrics {\n  const metrics = emptyMetrics(entries.length)\n  let startedAt: number | undefined\n  let completedAt: number | undefined\n\n  for (const entry of entries) {\n    const timestamp = timestampMs(entry.timestamp)\n    if (timestamp !== undefined) {\n      startedAt = startedAt === undefined ? timestamp : Math.min(startedAt, timestamp)\n      completedAt = completedAt === undefined ? timestamp : Math.max(completedAt, timestamp)\n    }\n\n    const role = stringField(entry, 'role')\n    if (role === 'user') metrics.userMessages += 1\n    if (role === 'assistant') metrics.assistantMessages += 1\n    if (role === '_usage') addUsage(metrics, entry)\n\n    const message = record(entry.message)\n    const messageType = message ? stringField(message, 'type') : undefined\n    const payload = record(message?.payload) ?? {}\n    if (messageType === 'TurnBegin') {\n      metrics.userMessages += 1\n      metrics.turnsStarted += 1\n    }\n    if (messageType === 'TurnEnd') metrics.turnsCompleted += 1\n    if (messageType === 'StepInterrupted') metrics.turnsAborted += 1\n    if (messageType === 'StepBegin') metrics.reasoningItems += 1\n    if (messageType === 'ContentPart') {\n      const payloadType = stringField(payload, 'type')\n      if (payloadType === 'think') metrics.reasoningItems += 1\n      if (payloadType === 'text') metrics.assistantMessages += 1\n    }\n    if (messageType === 'ToolCall') metrics.toolCalls += 1\n    if (messageType === 'ToolResult') {\n      metrics.toolOutputs += 1\n      if (record(payload.return_value)?.is_error === true) metrics.toolErrors += 1\n    }\n    if (messageType === 'StatusUpdate') addUsage(metrics, record(payload.token_usage))\n    if (messageType === 'Notification') {\n      const notification = record(payload.payload)\n      const exitCode = numberField(notification ?? {}, 'exit_code')\n      if (notification?.timed_out === true || (exitCode !== undefined && exitCode !== 0)) {\n        metrics.toolErrors += 1\n      }\n    }\n  }\n\n  if (startedAt !== undefined && completedAt !== undefined)\n    metrics.wallMs = Math.max(0, completedAt - startedAt)\n  metrics.processScore = terminalProcessScore(metrics)\n  return metrics\n}\n\nfunction piMetrics(entries: Record<string, unknown>[]): CodeAgentSessionMetrics {\n  const metrics = emptyMetrics(entries.length)\n  let bestReliabilityScore: number | undefined\n\n  for (const entry of entries) {\n    const nodes = Array.isArray(entry.nodes) ? entry.nodes : []\n    const edges = Array.isArray(entry.edges) ? entry.edges : []\n    metrics.graphNodes += nodes.length\n    metrics.graphEdges += edges.length\n\n    for (const node of nodes) {\n      const obj = record(node)\n      const ir = record(obj?.ir) ?? obj\n      const kind = stringField(ir ?? {}, 'kind')\n      if (kind === 'ActionCandidate') metrics.actionCandidates += 1\n      if (kind === 'ToolInvocation') metrics.toolCalls += 1\n      if (kind === 'ToolResult') metrics.toolOutputs += 1\n      if (kind === 'VerificationReport') metrics.verificationReports += 1\n      if (kind === 'CompletionDecision') {\n        metrics.completionDecisions += 1\n        metrics.turnsCompleted += 1\n      }\n    }\n\n    const averageReliability = numberField(entry, 'averageNodeReliability')\n    if (averageReliability !== undefined)\n      bestReliabilityScore = maxOptional(bestReliabilityScore, averageReliability)\n    const pessimisticPath = numberField(entry, 'pessimisticPathEstimate')\n    if (pessimisticPath !== undefined)\n      bestReliabilityScore = maxOptional(bestReliabilityScore, pessimisticPath)\n\n    const rows = Array.isArray(entry.rows) ? entry.rows : []\n    metrics.reliabilityRows += rows.length\n    for (const row of rows) {\n      const obj = record(row)\n      if (!obj) continue\n      const lift = numberField(obj, 'lift')\n      if (lift !== undefined) metrics.reliabilityLift = Math.max(metrics.reliabilityLift, lift)\n      const validated = numberField(obj, 'validatedSuccessEstimate')\n      if (validated !== undefined)\n        bestReliabilityScore = maxOptional(bestReliabilityScore, validated)\n    }\n  }\n\n  metrics.processScore =\n    bestReliabilityScore !== undefined\n      ? clamp01(bestReliabilityScore)\n      : metrics.completionDecisions > 0\n        ? 1\n        : 0\n  return metrics\n}\n\nfunction emptyMetrics(entries: number): CodeAgentSessionMetrics {\n  return {\n    entries,\n    userMessages: 0,\n    assistantMessages: 0,\n    reasoningItems: 0,\n    toolCalls: 0,\n    toolOutputs: 0,\n    toolErrors: 0,\n    patchAttempts: 0,\n    patchSuccesses: 0,\n    patchFailures: 0,\n    turnsStarted: 0,\n    turnsCompleted: 0,\n    turnsAborted: 0,\n    contextCompactions: 0,\n    prLinks: 0,\n    fileSnapshots: 0,\n    graphNodes: 0,\n    graphEdges: 0,\n    actionCandidates: 0,\n    verificationReports: 0,\n    completionDecisions: 0,\n    reliabilityRows: 0,\n    reliabilityLift: 0,\n    inputTokens: 0,\n    outputTokens: 0,\n    cachedTokens: 0,\n    observedCostUsd: 0,\n    wallMs: 0,\n    processScore: 0,\n  }\n}\n\nfunction codexProcessScore(metrics: CodeAgentSessionMetrics): number {\n  const terminalTurns = metrics.turnsCompleted + metrics.turnsAborted\n  const base =\n    terminalTurns > 0\n      ? metrics.turnsCompleted / terminalTurns\n      : metrics.assistantMessages > 0\n        ? 0.5\n        : 0\n  return penalizeErrors(base, metrics)\n}\n\nfunction claudeProcessScore(metrics: CodeAgentSessionMetrics): number {\n  const base = metrics.prLinks > 0 ? 1 : metrics.assistantMessages > 0 ? 0.75 : 0\n  return penalizeErrors(base, metrics)\n}\n\nfunction terminalProcessScore(metrics: CodeAgentSessionMetrics): number {\n  const terminalTurns = metrics.turnsCompleted + metrics.turnsAborted\n  const base =\n    terminalTurns > 0\n      ? metrics.turnsCompleted / terminalTurns\n      : metrics.assistantMessages > 0\n        ? 0.75\n        : metrics.toolCalls > 0\n          ? 0.5\n          : 0\n  return penalizeErrors(base, metrics)\n}\n\nfunction penalizeErrors(base: number, metrics: CodeAgentSessionMetrics): number {\n  const operations = Math.max(1, metrics.toolCalls + metrics.patchAttempts)\n  const errorRate = Math.min(1, (metrics.toolErrors + metrics.patchFailures) / operations)\n  return clamp01(base * (1 - 0.5 * errorRate))\n}\n\nfunction emptyTokenTotals(): { input: number; output: number; cached: number } {\n  return { input: 0, output: 0, cached: 0 }\n}\n\nfunction addUsage(\n  metrics: CodeAgentSessionMetrics,\n  usage: Record<string, unknown> | null | undefined,\n): void {\n  const parsed = readUsage(usage)\n  metrics.inputTokens += parsed.input\n  metrics.outputTokens += parsed.output\n  metrics.cachedTokens += parsed.cached\n}\n\nfunction addUsageTo(\n  totals: { input: number; output: number; cached: number },\n  usage: Record<string, unknown> | null | undefined,\n): void {\n  const parsed = readUsage(usage)\n  totals.input += parsed.input\n  totals.output += parsed.output\n  totals.cached += parsed.cached\n}\n\nfunction setCumulativeUsage(\n  metrics: CodeAgentSessionMetrics,\n  usage: Record<string, unknown> | null | undefined,\n): void {\n  const parsed = readUsage(usage)\n  metrics.inputTokens = Math.max(metrics.inputTokens, parsed.input)\n  metrics.outputTokens = Math.max(metrics.outputTokens, parsed.output)\n  metrics.cachedTokens = Math.max(metrics.cachedTokens, parsed.cached)\n}\n\nfunction readUsage(usage: Record<string, unknown> | null | undefined): {\n  input: number\n  output: number\n  cached: number\n} {\n  if (!usage) return { input: 0, output: 0, cached: 0 }\n  return {\n    input: numericUsage(usage, [\n      'input',\n      'input_tokens',\n      'inputTokens',\n      'prompt_tokens',\n      'promptTokens',\n      'input_other',\n    ]),\n    output: numericUsage(usage, [\n      'output',\n      'output_tokens',\n      'outputTokens',\n      'completion_tokens',\n      'completionTokens',\n      'reasoning',\n      'reasoning_tokens',\n      'reasoningTokens',\n    ]),\n    cached: numericUsage(usage, [\n      'cache',\n      'cached_tokens',\n      'cachedTokens',\n      'cache_read_input_tokens',\n      'cacheReadInputTokens',\n      'cache_creation_input_tokens',\n      'cacheCreationInputTokens',\n      'input_cache_read',\n      'input_cache_creation',\n    ]),\n  }\n}\n\nfunction numericUsage(obj: Record<string, unknown>, names: string[]): number {\n  let total = 0\n  for (const name of names) {\n    const value = obj[name]\n    if (typeof value === 'number' && Number.isFinite(value)) total += value\n  }\n  return total\n}\n\nfunction diagnosticsFor(\n  metrics: CodeAgentSessionMetrics,\n  options: {\n    model: string\n    explicitTerminal: boolean\n    malformedLines: number\n    scoreOverridden: boolean\n    costUsd: number\n  },\n): string[] {\n  const warnings: string[] = []\n  if (!options.scoreOverridden) warnings.push('outcome score is inferred from process telemetry')\n  if (!options.explicitTerminal) warnings.push('no explicit terminal success/failure signal')\n  if (metrics.inputTokens === 0 && metrics.outputTokens === 0) warnings.push('missing token usage')\n  if (options.costUsd === 0 && !isModelPriced(options.model)) warnings.push('model pricing unknown')\n  if (options.malformedLines > 0)\n    warnings.push(`${options.malformedLines} malformed JSONL lines skipped`)\n  return warnings\n}\n\nfunction hasExplicitTerminalSignal(\n  source: CodeAgentSessionSource,\n  metrics: CodeAgentSessionMetrics,\n): boolean {\n  if (source === 'codex') return metrics.turnsCompleted + metrics.turnsAborted > 0\n  if (source === 'opencode' || source === 'kimi-code')\n    return metrics.turnsCompleted + metrics.turnsAborted > 0\n  if (source === 'pi') return metrics.completionDecisions > 0 || metrics.reliabilityRows > 0\n  return metrics.prLinks > 0 || metrics.toolErrors > 0\n}\n\nfunction sessionIdFromEntries(\n  source: CodeAgentSessionSource,\n  entries: Record<string, unknown>[],\n): string | undefined {\n  for (const entry of entries) {\n    if (source === 'codex') {\n      const payload = record(entry.payload)\n      const id = payload ? stringField(payload, 'id') : undefined\n      if (id) return id\n    }\n    const message = record(entry.message)\n    const payload = record(message?.payload)\n    const sessionId =\n      stringField(entry, 'sessionID') ??\n      stringField(entry, 'sessionId') ??\n      stringField(entry, 'session_id') ??\n      stringField(payload ?? {}, 'session_id')\n    if (sessionId) return sessionId\n  }\n  return undefined\n}\n\nfunction modelFromEntries(\n  source: CodeAgentSessionSource,\n  entries: Record<string, unknown>[],\n): string | undefined {\n  let providerModel: string | undefined\n  for (const entry of entries) {\n    if (source === 'codex') {\n      const payload = record(entry.payload)\n      const model = payload ? stringField(payload, 'model') : undefined\n      if (model) return model\n      providerModel =\n        providerModel ?? (payload ? stringField(payload, 'model_provider') : undefined)\n    }\n    const openCodeModel = stringField(entry, 'modelID')\n    if (openCodeModel) {\n      const provider = stringField(entry, 'providerID')\n      return provider ? `${provider}/${openCodeModel}` : openCodeModel\n    }\n    const modelObject = record(entry.model)\n    const nestedModel = stringField(modelObject ?? {}, 'modelID')\n    if (nestedModel) {\n      const provider = stringField(modelObject ?? {}, 'providerID')\n      return provider ? `${provider}/${nestedModel}` : nestedModel\n    }\n    const message = record(entry.message)\n    const model = message ? stringField(message, 'model') : undefined\n    if (model) return model\n  }\n  return providerModel\n}\n\nfunction cwdFromEntries(entries: Record<string, unknown>[]): string | undefined {\n  for (const entry of entries) {\n    const path = record(entry.path)\n    const cwd =\n      stringField(entry, 'cwd') ??\n      stringField(path ?? {}, 'cwd') ??\n      stringField(path ?? {}, 'root') ??\n      (record(entry.payload) ? stringField(record(entry.payload)!, 'cwd') : undefined)\n    if (cwd) return cwd\n  }\n  return undefined\n}\n\nfunction firstUserText(entries: Record<string, unknown>[]): string | undefined {\n  for (const entry of entries) {\n    const payload = record(entry.payload)\n    const payloadType = payload ? stringField(payload, 'type') : undefined\n    if (payloadType === 'user_message') return stringField(payload!, 'message')\n\n    if (stringField(entry, 'role') === 'user') {\n      const content = entry.content\n      if (typeof content === 'string') return content\n    }\n\n    const wirePayload = record(record(entry.message)?.payload)\n    const userInput = wirePayload ? stringField(wirePayload, 'user_input') : undefined\n    if (userInput) return userInput\n\n    const message = record(entry.message)\n    if (message && stringField(message, 'role') === 'user') {\n      const content = message.content\n      if (typeof content === 'string') return content\n      if (Array.isArray(content)) {\n        const text = content\n          .map((part) => {\n            const obj = record(part)\n            return obj ? stringField(obj, 'text') : undefined\n          })\n          .filter((part): part is string => part !== undefined)\n          .join('\\n')\n        if (text) return text\n      }\n    }\n  }\n  return undefined\n}\n\nfunction fallbackSessionId(source: CodeAgentSessionSource, sourcePath: string | undefined): string {\n  return hashString(`${source}:${sourcePath ?? 'unknown-session'}`).slice(\n    'sha256:'.length,\n    'sha256:'.length + 20,\n  )\n}\n\nfunction withSnapshot(model: string): string {\n  const trimmed = model.trim() || 'unknown'\n  if (trimmed.includes('@') || /-\\d{8}$/.test(trimmed) || /-\\d{4}-\\d{2}-\\d{2}$/.test(trimmed)) {\n    return trimmed\n  }\n  return `${trimmed}@observed-local`\n}\n\nfunction hashJson(value: unknown): string {\n  return hashString(JSON.stringify(value))\n}\n\nfunction hashString(value: string): string {\n  return `sha256:${createHash('sha256').update(value).digest('hex')}`\n}\n\nfunction stableSeed(value: string): number {\n  return createHash('sha256').update(value).digest().readUInt32BE(0)\n}\n\nfunction maxOptional(current: number | undefined, next: number): number {\n  return current === undefined ? next : Math.max(current, next)\n}\n\nfunction clamp01(value: number): number {\n  if (!Number.isFinite(value)) return 0\n  return Math.max(0, Math.min(1, value))\n}\n\nfunction isRecord(value: unknown): value is Record<string, unknown> {\n  return value !== null && typeof value === 'object' && !Array.isArray(value)\n}\n\nfunction record(value: unknown): Record<string, unknown> | null {\n  return isRecord(value) ? value : null\n}\n\nfunction stringField(obj: Record<string, unknown>, key: string): string | undefined {\n  const value = obj[key]\n  return typeof value === 'string' && value.length > 0 ? value : undefined\n}\n\nfunction numberField(obj: Record<string, unknown>, key: string): number | undefined {\n  const value = obj[key]\n  return typeof value === 'number' && Number.isFinite(value) ? value : undefined\n}\n\nfunction timestampMs(value: unknown): number | undefined {\n  if (typeof value === 'number' && Number.isFinite(value)) {\n    return value > 1_000_000_000_000 ? value : value * 1000\n  }\n  if (typeof value === 'string' && value.length > 0) {\n    const parsed = Date.parse(value)\n    return Number.isFinite(parsed) ? parsed : undefined\n  }\n  return undefined\n}\n"],"mappings":";;;;;;AAAA,SAAS,kBAAkB;AA+EpB,SAAS,oBAAoB,OAAqC;AACvE,QAAM,UAAqB,CAAC;AAC5B,MAAI,iBAAiB;AACrB,aAAW,QAAQ,MAAM,MAAM,IAAI,GAAG;AACpC,UAAM,UAAU,KAAK,KAAK;AAC1B,QAAI,CAAC,QAAS;AACd,QAAI;AACF,cAAQ,KAAK,KAAK,MAAM,OAAO,CAAC;AAAA,IAClC,QAAQ;AACN,wBAAkB;AAAA,IACpB;AAAA,EACF;AACA,SAAO,EAAE,SAAS,eAAe;AACnC;AAEO,SAAS,iBACd,SAC8B;AAC9B,SAAO,qBAAqB,SAAS,OAAO;AAC9C;AAEO,SAAS,sBACd,SAC8B;AAC9B,SAAO,qBAAqB,eAAe,OAAO;AACpD;AAEO,SAAS,oBACd,SAC8B;AAC9B,SAAO,qBAAqB,YAAY,OAAO;AACjD;AAEO,SAAS,oBACd,SAC8B;AAC9B,SAAO,qBAAqB,aAAa,OAAO;AAClD;AAEO,SAAS,cACd,SAC8B;AAC9B,SAAO,qBAAqB,MAAM,OAAO;AAC3C;AAEO,IAAM,qBAAqB;AAElC,SAAS,qBACP,QACA,SAC8B;AAC9B,QAAM,UAAU,QAAQ,QAAQ,OAAO,QAAQ;AAC/C,MAAI,QAAQ,WAAW,GAAG;AACxB,WAAO;AAAA,MACL,MAAM,CAAC;AAAA,MACP,aAAa;AAAA,QACX;AAAA,UACE;AAAA,UACA,WAAW,kBAAkB,QAAQ,QAAQ,UAAU;AAAA,UACvD,YAAY,QAAQ;AAAA,UACpB,SAAS;AAAA,UACT,gBAAgB,QAAQ,kBAAkB;AAAA,UAC1C,eAAe;AAAA,UACf,2BAA2B;AAAA,UAC3B,iBAAiB;AAAA,UACjB,eAAe;AAAA,UACf,SAAS;AAAA,UACT,UAAU,CAAC,8BAA8B;AAAA,QAC3C;AAAA,MACF;AAAA,MACA,SAAS,CAAC;AAAA,IACZ;AAAA,EACF;AAEA,QAAM,UAAU,WAAW,QAAQ,OAAO;AAC1C,QAAM,YACJ,qBAAqB,QAAQ,OAAO,KAAK,kBAAkB,QAAQ,QAAQ,UAAU;AACvF,QAAM,QAAQ,aAAa,QAAQ,SAAS,iBAAiB,QAAQ,OAAO,KAAK,MAAM;AACvF,QAAM,aAA4B;AAAA,IAChC,OAAO,QAAQ;AAAA,IACf,QAAQ,QAAQ;AAAA,IAChB,GAAI,QAAQ,eAAe,IAAI,EAAE,QAAQ,QAAQ,aAAa,IAAI,CAAC;AAAA,EACrE;AACA,QAAM,mBACJ,QAAQ,cAAc,KAAK,QAAQ,eAAe,IAC9C,aAAa,QAAQ,aAAa,QAAQ,cAAc,KAAK,IAC7D;AACN,QAAM,UAAU,QAAQ,kBAAkB,IAAI,QAAQ,kBAAkB;AACxE,QAAM,QAAQ,QAAQ,QAAQ,SAAS,QAAQ,YAAY;AAC3D,QAAM,aACJ,QAAQ,cAAc,WAAW,UAAU,MAAM,IAAI,cAAc,OAAO,KAAK,EAAE,EAAE;AACrF,QAAM,aACJ,QAAQ,cACR,SAAS;AAAA,IACP;AAAA,IACA;AAAA,IACA,YAAY,QAAQ;AAAA,IACpB,KAAK,eAAe,OAAO;AAAA,IAC3B,YAAY,QAAQ;AAAA,EACtB,CAAC;AACH,QAAM,mBAAmB,0BAA0B,QAAQ,OAAO;AAClE,QAAM,WAAW,eAAe,SAAS;AAAA,IACvC;AAAA,IACA;AAAA,IACA,gBAAgB,QAAQ,kBAAkB;AAAA,IAC1C,iBAAiB,QAAQ,UAAU;AAAA,IACnC;AAAA,EACF,CAAC;AAED,QAAM,MAAiB;AAAA,IACrB,OAAO,GAAG,MAAM,IAAI,SAAS;AAAA,IAC7B,cAAc,QAAQ,gBAAgB,GAAG,MAAM;AAAA,IAC/C,aAAa,QAAQ,eAAe;AAAA,IACpC,MAAM,QAAQ,QAAQ,WAAW,SAAS;AAAA,IAC1C;AAAA,IACA;AAAA,IACA;AAAA,IACA,WAAW,QAAQ,aAAa;AAAA,IAChC,QAAQ,QAAQ;AAAA,IAChB;AAAA,IACA;AAAA,IACA,SAAS;AAAA,MACP,cAAc;AAAA,MACd,KAAK;AAAA,QACH,SAAS,QAAQ;AAAA,QACjB,eAAe,QAAQ;AAAA,QACvB,oBAAoB,QAAQ;AAAA,QAC5B,iBAAiB,QAAQ;AAAA,QACzB,YAAY,QAAQ;AAAA,QACpB,cAAc,QAAQ;AAAA,QACtB,aAAa,QAAQ;AAAA,QACrB,gBAAgB,QAAQ;AAAA,QACxB,iBAAiB,QAAQ;AAAA,QACzB,gBAAgB,QAAQ;AAAA,QACxB,eAAe,QAAQ;AAAA,QACvB,iBAAiB,QAAQ;AAAA,QACzB,eAAe,QAAQ;AAAA,QACvB,qBAAqB,QAAQ;AAAA,QAC7B,UAAU,QAAQ;AAAA,QAClB,gBAAgB,QAAQ;AAAA,QACxB,aAAa,QAAQ;AAAA,QACrB,aAAa,QAAQ;AAAA,QACrB,mBAAmB,QAAQ;AAAA,QAC3B,sBAAsB,QAAQ;AAAA,QAC9B,sBAAsB,QAAQ;AAAA,QAC9B,kBAAkB,QAAQ;AAAA,QAC1B,kBAAkB,QAAQ;AAAA,QAC1B,cAAc,QAAQ;AAAA,QACtB,eAAe,QAAQ;AAAA,QACvB,eAAe,QAAQ;AAAA,QACvB,mBAAmB,QAAQ;AAAA,QAC3B,eAAe;AAAA,QACf,gBAAgB,QAAQ,UAAU,SAAY,IAAI;AAAA,QAClD,0BAA0B,mBAAmB,IAAI;AAAA,QACjD,uBAAuB,QAAQ,UAAU,SAAY,IAAI;AAAA,QACzD,cAAc,YAAY,KAAK,CAAC,cAAc,KAAK,IAAI,IAAI;AAAA,MAC7D;AAAA,IACF;AAAA,IACA,UAAU,QAAQ,YAAY;AAAA,IAC9B,GAAI,QAAQ,aAAa,EAAE,YAAY,QAAQ,WAAW,IAAI,CAAC;AAAA,IAC/D,GAAI,QAAQ,aAAa,KAAK,QAAQ,eAAe,IACjD,EAAE,aAAa,QAAQ,eAAe,IAAI,iBAAiB,aAAa,IACxE,CAAC;AAAA,EACP;AAEA,SAAO;AAAA,IACL,MAAM,CAAC,GAAG;AAAA,IACV,aAAa;AAAA,MACX;AAAA,QACE;AAAA,QACA;AAAA,QACA,YAAY,QAAQ;AAAA,QACpB,SAAS,QAAQ;AAAA,QACjB,gBAAgB,QAAQ,kBAAkB;AAAA,QAC1C,eAAe,QAAQ,UAAU;AAAA,QACjC,2BAA2B;AAAA,QAC3B,iBAAiB,QAAQ,UAAU;AAAA,QACnC,eAAe,QAAQ,cAAc,KAAK,QAAQ,eAAe;AAAA,QACjE,SAAS,UAAU;AAAA,QACnB;AAAA,MACF;AAAA,IACF;AAAA,IACA,SAAS,CAAC,OAAO;AAAA,EACnB;AACF;AAEA,SAAS,WACP,QACA,SACyB;AACzB,UAAQ,QAAQ;AAAA,IACd,KAAK;AACH,aAAO,aAAa,OAAO;AAAA,IAC7B,KAAK;AACH,aAAO,kBAAkB,OAAO;AAAA,IAClC,KAAK;AACH,aAAO,gBAAgB,OAAO;AAAA,IAChC,KAAK;AACH,aAAO,gBAAgB,OAAO;AAAA,IAChC,KAAK;AACH,aAAO,UAAU,OAAO;AAAA,EAC5B;AACF;AAEA,SAAS,aAAa,SAA6D;AACjF,QAAM,UAAU,aAAa,QAAQ,MAAM;AAC3C,MAAI;AACJ,MAAI;AAEJ,aAAW,SAAS,SAAS;AAC3B,UAAM,UAAU,OAAO,MAAM,OAAO,KAAK,CAAC;AAC1C,UAAM,YAAY,YAAY,OAAO,MAAM;AAC3C,UAAM,cAAc,YAAY,SAAS,MAAM;AAC/C,UAAM,YAAY,YAAY,MAAM,SAAS;AAC7C,QAAI,cAAc,QAAW;AAC3B,kBAAY,cAAc,SAAY,YAAY,KAAK,IAAI,WAAW,SAAS;AAC/E,oBAAc,gBAAgB,SAAY,YAAY,KAAK,IAAI,aAAa,SAAS;AAAA,IACvF;AAEA,QAAI,cAAc,iBAAiB;AACjC,UAAI,gBAAgB,mBAAmB,gBAAgB;AACrD,gBAAQ,aAAa;AACvB,UAAI,gBAAgB,0BAA0B,gBAAgB;AAC5D,gBAAQ,eAAe;AACzB,UAAI,gBAAgB,YAAa,SAAQ,kBAAkB;AAC3D,UAAI,gBAAgB,aAAa,YAAY,SAAS,MAAM,MAAM;AAChE,gBAAQ,qBAAqB;AAAA,IACjC;AAEA,QAAI,cAAc,aAAa;AAC7B,UAAI,gBAAgB,eAAgB,SAAQ,gBAAgB;AAC5D,UAAI,gBAAgB,gBAAiB,SAAQ,qBAAqB;AAClE,UAAI,gBAAgB,gBAAgB;AAClC,gBAAQ,gBAAgB;AACxB,cAAM,UAAU,YAAY,QAAQ,UAAU;AAC9C,YAAI,YAAY,QAAW;AACzB,sBAAY,cAAc,SAAY,UAAU,KAAK,IAAI,WAAW,OAAO;AAAA,QAC7E;AAAA,MACF;AACA,UAAI,gBAAgB,iBAAiB;AACnC,gBAAQ,kBAAkB;AAC1B,cAAM,YAAY,YAAY,QAAQ,YAAY;AAClD,YAAI,cAAc,QAAW;AAC3B,wBAAc,gBAAgB,SAAY,YAAY,KAAK,IAAI,aAAa,SAAS;AAAA,QACvF;AACA,cAAM,WAAW,YAAY,SAAS,aAAa;AACnD,YAAI,aAAa,OAAW,SAAQ,SAAS,KAAK,IAAI,QAAQ,QAAQ,QAAQ;AAAA,MAChF;AACA,UAAI,gBAAgB,gBAAgB;AAClC,gBAAQ,gBAAgB;AACxB,cAAM,YAAY,YAAY,QAAQ,YAAY;AAClD,YAAI,cAAc,QAAW;AAC3B,wBAAc,gBAAgB,SAAY,YAAY,KAAK,IAAI,aAAa,SAAS;AAAA,QACvF;AAAA,MACF;AACA,UAAI,gBAAgB,oBAAqB,SAAQ,sBAAsB;AACvE,UAAI,gBAAgB,mBAAmB;AACrC,gBAAQ,iBAAiB;AACzB,YAAI,QAAQ,YAAY,KAAM,SAAQ,kBAAkB;AAAA,aACnD;AACH,kBAAQ,iBAAiB;AACzB,kBAAQ,cAAc;AAAA,QACxB;AAAA,MACF;AACA,UAAI,gBAAgB;AAClB,2BAAmB,SAAS,OAAO,OAAO,QAAQ,IAAI,GAAG,iBAAiB,CAAC;AAC7E,YAAM,SAAS,OAAO,QAAQ,MAAM;AACpC,UAAI,UAAU,SAAS,OAAQ,SAAQ,cAAc;AAAA,IACvD;AAAA,EACF;AAEA,MAAI,QAAQ,WAAW,KAAK,cAAc,UAAa,gBAAgB,QAAW;AAChF,YAAQ,SAAS,KAAK,IAAI,GAAG,cAAc,SAAS;AAAA,EACtD;AACA,UAAQ,eAAe,kBAAkB,OAAO;AAChD,SAAO;AACT;AAEA,SAAS,kBAAkB,SAA6D;AACtF,QAAM,UAAU,aAAa,QAAQ,MAAM;AAC3C,MAAI;AACJ,MAAI;AAEJ,aAAW,SAAS,SAAS;AAC3B,UAAM,OAAO,YAAY,OAAO,MAAM;AACtC,UAAM,YAAY,YAAY,MAAM,SAAS;AAC7C,QAAI,cAAc,QAAW;AAC3B,kBAAY,cAAc,SAAY,YAAY,KAAK,IAAI,WAAW,SAAS;AAC/E,oBAAc,gBAAgB,SAAY,YAAY,KAAK,IAAI,aAAa,SAAS;AAAA,IACvF;AAEA,QAAI,SAAS,OAAQ,SAAQ,gBAAgB;AAC7C,QAAI,SAAS,YAAa,SAAQ,qBAAqB;AACvD,QAAI,SAAS,UAAW,SAAQ,WAAW;AAC3C,QAAI,SAAS,wBAAyB,SAAQ,iBAAiB;AAE/D,UAAM,UAAU,OAAO,MAAM,OAAO;AACpC,QAAI,QAAS,UAAS,SAAS,OAAO,QAAQ,KAAK,CAAC;AACpD,UAAM,UAAU,MAAM,QAAQ,SAAS,OAAO,IAAI,QAAQ,UAAU,CAAC;AACrE,eAAW,QAAQ,SAAS;AAC1B,YAAM,OAAO,OAAO,IAAI;AACxB,UAAI,CAAC,KAAM;AACX,YAAM,WAAW,YAAY,MAAM,MAAM;AACzC,UAAI,aAAa,WAAY,SAAQ,kBAAkB;AACvD,UAAI,aAAa,WAAY,SAAQ,aAAa;AAClD,UAAI,aAAa,eAAe;AAC9B,gBAAQ,eAAe;AACvB,YAAI,KAAK,aAAa,KAAM,SAAQ,cAAc;AAAA,MACpD;AAAA,IACF;AAAA,EACF;AAEA,MAAI,cAAc,UAAa,gBAAgB;AAC7C,YAAQ,SAAS,KAAK,IAAI,GAAG,cAAc,SAAS;AACtD,UAAQ,eAAe,mBAAmB,OAAO;AACjD,SAAO;AACT;AAEA,SAAS,gBAAgB,SAA6D;AACpF,QAAM,UAAU,aAAa,QAAQ,MAAM;AAC3C,QAAM,eAAe,iBAAiB;AACtC,QAAM,YAAY,iBAAiB;AACnC,MAAI,cAAc;AAClB,MAAI,WAAW;AACf,MAAI;AACJ,MAAI;AAEJ,aAAW,SAAS,SAAS;AAC3B,UAAM,OAAO,OAAO,MAAM,IAAI;AAC9B,UAAM,UAAU,YAAY,MAAM,OAAO;AACzC,UAAM,YAAY,YAAY,MAAM,SAAS;AAC7C,QAAI,YAAY;AACd,kBAAY,cAAc,SAAY,UAAU,KAAK,IAAI,WAAW,OAAO;AAC7E,QAAI,cAAc;AAChB,oBAAc,gBAAgB,SAAY,YAAY,KAAK,IAAI,aAAa,SAAS;AAEvF,UAAM,OAAO,YAAY,OAAO,MAAM;AACtC,QAAI,SAAS,OAAQ,SAAQ,gBAAgB;AAC7C,QAAI,SAAS,aAAa;AACxB,cAAQ,qBAAqB;AAC7B,YAAM,SAAS,YAAY,OAAO,QAAQ;AAC1C,UAAI,WAAW,OAAQ,SAAQ,kBAAkB;AACjD,UAAI,WAAW,QAAS,SAAQ,gBAAgB;AAAA,IAClD;AAEA,UAAM,OAAO,YAAY,OAAO,MAAM;AACtC,QAAI,SAAS,YAAa,SAAQ,kBAAkB;AACpD,QAAI,SAAS,QAAQ;AACnB,cAAQ,aAAa;AACrB,YAAM,SAAS,YAAY,OAAO,MAAM,KAAK,KAAK,CAAC,GAAG,QAAQ;AAC9D,UAAI,WAAW,YAAa,SAAQ,eAAe;AACnD,UAAI,WAAW,SAAS;AACtB,gBAAQ,eAAe;AACvB,gBAAQ,cAAc;AAAA,MACxB;AAAA,IACF;AACA,QAAI,SAAS,SAAS;AACpB,cAAQ,iBAAiB;AACzB,cAAQ,kBAAkB;AAAA,IAC5B;AAEA,UAAM,OAAO,YAAY,OAAO,MAAM;AACtC,QAAI,OAAO,MAAM,MAAM,KAAK,MAAM;AAChC,iBAAW,cAAc,OAAO,MAAM,MAAM,CAAC;AAC7C,UAAI,SAAS,OAAW,gBAAe;AAAA,IACzC,WAAW,OAAO,MAAM,MAAM,GAAG;AAC/B,iBAAW,WAAW,OAAO,MAAM,MAAM,CAAC;AAC1C,UAAI,SAAS,OAAW,aAAY;AAAA,IACtC;AAAA,EACF;AAEA,QAAM,QAAQ,aAAa,QAAQ,aAAa,SAAS,IAAI,eAAe;AAC5E,UAAQ,cAAc,MAAM;AAC5B,UAAQ,eAAe,MAAM;AAC7B,UAAQ,eAAe,MAAM;AAC7B,UAAQ,kBAAkB,cAAc,IAAI,cAAc;AAC1D,MAAI,QAAQ,WAAW,KAAK,cAAc,UAAa,gBAAgB;AACrE,YAAQ,SAAS,KAAK,IAAI,GAAG,cAAc,SAAS;AACtD,UAAQ,eAAe,qBAAqB,OAAO;AACnD,SAAO;AACT;AAEA,SAAS,gBAAgB,SAA6D;AACpF,QAAM,UAAU,aAAa,QAAQ,MAAM;AAC3C,MAAI;AACJ,MAAI;AAEJ,aAAW,SAAS,SAAS;AAC3B,UAAM,YAAY,YAAY,MAAM,SAAS;AAC7C,QAAI,cAAc,QAAW;AAC3B,kBAAY,cAAc,SAAY,YAAY,KAAK,IAAI,WAAW,SAAS;AAC/E,oBAAc,gBAAgB,SAAY,YAAY,KAAK,IAAI,aAAa,SAAS;AAAA,IACvF;AAEA,UAAM,OAAO,YAAY,OAAO,MAAM;AACtC,QAAI,SAAS,OAAQ,SAAQ,gBAAgB;AAC7C,QAAI,SAAS,YAAa,SAAQ,qBAAqB;AACvD,QAAI,SAAS,SAAU,UAAS,SAAS,KAAK;AAE9C,UAAM,UAAU,OAAO,MAAM,OAAO;AACpC,UAAM,cAAc,UAAU,YAAY,SAAS,MAAM,IAAI;AAC7D,UAAM,UAAU,OAAO,SAAS,OAAO,KAAK,CAAC;AAC7C,QAAI,gBAAgB,aAAa;AAC/B,cAAQ,gBAAgB;AACxB,cAAQ,gBAAgB;AAAA,IAC1B;AACA,QAAI,gBAAgB,UAAW,SAAQ,kBAAkB;AACzD,QAAI,gBAAgB,kBAAmB,SAAQ,gBAAgB;AAC/D,QAAI,gBAAgB,YAAa,SAAQ,kBAAkB;AAC3D,QAAI,gBAAgB,eAAe;AACjC,YAAM,cAAc,YAAY,SAAS,MAAM;AAC/C,UAAI,gBAAgB,QAAS,SAAQ,kBAAkB;AACvD,UAAI,gBAAgB,OAAQ,SAAQ,qBAAqB;AAAA,IAC3D;AACA,QAAI,gBAAgB,WAAY,SAAQ,aAAa;AACrD,QAAI,gBAAgB,cAAc;AAChC,cAAQ,eAAe;AACvB,UAAI,OAAO,QAAQ,YAAY,GAAG,aAAa,KAAM,SAAQ,cAAc;AAAA,IAC7E;AACA,QAAI,gBAAgB,eAAgB,UAAS,SAAS,OAAO,QAAQ,WAAW,CAAC;AACjF,QAAI,gBAAgB,gBAAgB;AAClC,YAAM,eAAe,OAAO,QAAQ,OAAO;AAC3C,YAAM,WAAW,YAAY,gBAAgB,CAAC,GAAG,WAAW;AAC5D,UAAI,cAAc,cAAc,QAAS,aAAa,UAAa,aAAa,GAAI;AAClF,gBAAQ,cAAc;AAAA,MACxB;AAAA,IACF;AAAA,EACF;AAEA,MAAI,cAAc,UAAa,gBAAgB;AAC7C,YAAQ,SAAS,KAAK,IAAI,GAAG,cAAc,SAAS;AACtD,UAAQ,eAAe,qBAAqB,OAAO;AACnD,SAAO;AACT;AAEA,SAAS,UAAU,SAA6D;AAC9E,QAAM,UAAU,aAAa,QAAQ,MAAM;AAC3C,MAAI;AAEJ,aAAW,SAAS,SAAS;AAC3B,UAAM,QAAQ,MAAM,QAAQ,MAAM,KAAK,IAAI,MAAM,QAAQ,CAAC;AAC1D,UAAM,QAAQ,MAAM,QAAQ,MAAM,KAAK,IAAI,MAAM,QAAQ,CAAC;AAC1D,YAAQ,cAAc,MAAM;AAC5B,YAAQ,cAAc,MAAM;AAE5B,eAAW,QAAQ,OAAO;AACxB,YAAM,MAAM,OAAO,IAAI;AACvB,YAAM,KAAK,OAAO,KAAK,EAAE,KAAK;AAC9B,YAAM,OAAO,YAAY,MAAM,CAAC,GAAG,MAAM;AACzC,UAAI,SAAS,kBAAmB,SAAQ,oBAAoB;AAC5D,UAAI,SAAS,iBAAkB,SAAQ,aAAa;AACpD,UAAI,SAAS,aAAc,SAAQ,eAAe;AAClD,UAAI,SAAS,qBAAsB,SAAQ,uBAAuB;AAClE,UAAI,SAAS,sBAAsB;AACjC,gBAAQ,uBAAuB;AAC/B,gBAAQ,kBAAkB;AAAA,MAC5B;AAAA,IACF;AAEA,UAAM,qBAAqB,YAAY,OAAO,wBAAwB;AACtE,QAAI,uBAAuB;AACzB,6BAAuB,YAAY,sBAAsB,kBAAkB;AAC7E,UAAM,kBAAkB,YAAY,OAAO,yBAAyB;AACpE,QAAI,oBAAoB;AACtB,6BAAuB,YAAY,sBAAsB,eAAe;AAE1E,UAAM,OAAO,MAAM,QAAQ,MAAM,IAAI,IAAI,MAAM,OAAO,CAAC;AACvD,YAAQ,mBAAmB,KAAK;AAChC,eAAW,OAAO,MAAM;AACtB,YAAM,MAAM,OAAO,GAAG;AACtB,UAAI,CAAC,IAAK;AACV,YAAM,OAAO,YAAY,KAAK,MAAM;AACpC,UAAI,SAAS,OAAW,SAAQ,kBAAkB,KAAK,IAAI,QAAQ,iBAAiB,IAAI;AACxF,YAAM,YAAY,YAAY,KAAK,0BAA0B;AAC7D,UAAI,cAAc;AAChB,+BAAuB,YAAY,sBAAsB,SAAS;AAAA,IACtE;AAAA,EACF;AAEA,UAAQ,eACN,yBAAyB,SACrB,QAAQ,oBAAoB,IAC5B,QAAQ,sBAAsB,IAC5B,IACA;AACR,SAAO;AACT;AAEA,SAAS,aAAa,SAA0C;AAC9D,SAAO;AAAA,IACL;AAAA,IACA,cAAc;AAAA,IACd,mBAAmB;AAAA,IACnB,gBAAgB;AAAA,IAChB,WAAW;AAAA,IACX,aAAa;AAAA,IACb,YAAY;AAAA,IACZ,eAAe;AAAA,IACf,gBAAgB;AAAA,IAChB,eAAe;AAAA,IACf,cAAc;AAAA,IACd,gBAAgB;AAAA,IAChB,cAAc;AAAA,IACd,oBAAoB;AAAA,IACpB,SAAS;AAAA,IACT,eAAe;AAAA,IACf,YAAY;AAAA,IACZ,YAAY;AAAA,IACZ,kBAAkB;AAAA,IAClB,qBAAqB;AAAA,IACrB,qBAAqB;AAAA,IACrB,iBAAiB;AAAA,IACjB,iBAAiB;AAAA,IACjB,aAAa;AAAA,IACb,cAAc;AAAA,IACd,cAAc;AAAA,IACd,iBAAiB;AAAA,IACjB,QAAQ;AAAA,IACR,cAAc;AAAA,EAChB;AACF;AAEA,SAAS,kBAAkB,SAA0C;AACnE,QAAM,gBAAgB,QAAQ,iBAAiB,QAAQ;AACvD,QAAM,OACJ,gBAAgB,IACZ,QAAQ,iBAAiB,gBACzB,QAAQ,oBAAoB,IAC1B,MACA;AACR,SAAO,eAAe,MAAM,OAAO;AACrC;AAEA,SAAS,mBAAmB,SAA0C;AACpE,QAAM,OAAO,QAAQ,UAAU,IAAI,IAAI,QAAQ,oBAAoB,IAAI,OAAO;AAC9E,SAAO,eAAe,MAAM,OAAO;AACrC;AAEA,SAAS,qBAAqB,SAA0C;AACtE,QAAM,gBAAgB,QAAQ,iBAAiB,QAAQ;AACvD,QAAM,OACJ,gBAAgB,IACZ,QAAQ,iBAAiB,gBACzB,QAAQ,oBAAoB,IAC1B,OACA,QAAQ,YAAY,IAClB,MACA;AACV,SAAO,eAAe,MAAM,OAAO;AACrC;AAEA,SAAS,eAAe,MAAc,SAA0C;AAC9E,QAAM,aAAa,KAAK,IAAI,GAAG,QAAQ,YAAY,QAAQ,aAAa;AACxE,QAAM,YAAY,KAAK,IAAI,IAAI,QAAQ,aAAa,QAAQ,iBAAiB,UAAU;AACvF,SAAO,QAAQ,QAAQ,IAAI,MAAM,UAAU;AAC7C;AAEA,SAAS,mBAAsE;AAC7E,SAAO,EAAE,OAAO,GAAG,QAAQ,GAAG,QAAQ,EAAE;AAC1C;AAEA,SAAS,SACP,SACA,OACM;AACN,QAAM,SAAS,UAAU,KAAK;AAC9B,UAAQ,eAAe,OAAO;AAC9B,UAAQ,gBAAgB,OAAO;AAC/B,UAAQ,gBAAgB,OAAO;AACjC;AAEA,SAAS,WACP,QACA,OACM;AACN,QAAM,SAAS,UAAU,KAAK;AAC9B,SAAO,SAAS,OAAO;AACvB,SAAO,UAAU,OAAO;AACxB,SAAO,UAAU,OAAO;AAC1B;AAEA,SAAS,mBACP,SACA,OACM;AACN,QAAM,SAAS,UAAU,KAAK;AAC9B,UAAQ,cAAc,KAAK,IAAI,QAAQ,aAAa,OAAO,KAAK;AAChE,UAAQ,eAAe,KAAK,IAAI,QAAQ,cAAc,OAAO,MAAM;AACnE,UAAQ,eAAe,KAAK,IAAI,QAAQ,cAAc,OAAO,MAAM;AACrE;AAEA,SAAS,UAAU,OAIjB;AACA,MAAI,CAAC,MAAO,QAAO,EAAE,OAAO,GAAG,QAAQ,GAAG,QAAQ,EAAE;AACpD,SAAO;AAAA,IACL,OAAO,aAAa,OAAO;AAAA,MACzB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,IACD,QAAQ,aAAa,OAAO;AAAA,MAC1B;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,IACD,QAAQ,aAAa,OAAO;AAAA,MAC1B;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AACF;AAEA,SAAS,aAAa,KAA8B,OAAyB;AAC3E,MAAI,QAAQ;AACZ,aAAW,QAAQ,OAAO;AACxB,UAAM,QAAQ,IAAI,IAAI;AACtB,QAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,EAAG,UAAS;AAAA,EACpE;AACA,SAAO;AACT;AAEA,SAAS,eACP,SACA,SAOU;AACV,QAAM,WAAqB,CAAC;AAC5B,MAAI,CAAC,QAAQ,gBAAiB,UAAS,KAAK,kDAAkD;AAC9F,MAAI,CAAC,QAAQ,iBAAkB,UAAS,KAAK,6CAA6C;AAC1F,MAAI,QAAQ,gBAAgB,KAAK,QAAQ,iBAAiB,EAAG,UAAS,KAAK,qBAAqB;AAChG,MAAI,QAAQ,YAAY,KAAK,CAAC,cAAc,QAAQ,KAAK,EAAG,UAAS,KAAK,uBAAuB;AACjG,MAAI,QAAQ,iBAAiB;AAC3B,aAAS,KAAK,GAAG,QAAQ,cAAc,gCAAgC;AACzE,SAAO;AACT;AAEA,SAAS,0BACP,QACA,SACS;AACT,MAAI,WAAW,QAAS,QAAO,QAAQ,iBAAiB,QAAQ,eAAe;AAC/E,MAAI,WAAW,cAAc,WAAW;AACtC,WAAO,QAAQ,iBAAiB,QAAQ,eAAe;AACzD,MAAI,WAAW,KAAM,QAAO,QAAQ,sBAAsB,KAAK,QAAQ,kBAAkB;AACzF,SAAO,QAAQ,UAAU,KAAK,QAAQ,aAAa;AACrD;AAEA,SAAS,qBACP,QACA,SACoB;AACpB,aAAW,SAAS,SAAS;AAC3B,QAAI,WAAW,SAAS;AACtB,YAAMA,WAAU,OAAO,MAAM,OAAO;AACpC,YAAM,KAAKA,WAAU,YAAYA,UAAS,IAAI,IAAI;AAClD,UAAI,GAAI,QAAO;AAAA,IACjB;AACA,UAAM,UAAU,OAAO,MAAM,OAAO;AACpC,UAAM,UAAU,OAAO,SAAS,OAAO;AACvC,UAAM,YACJ,YAAY,OAAO,WAAW,KAC9B,YAAY,OAAO,WAAW,KAC9B,YAAY,OAAO,YAAY,KAC/B,YAAY,WAAW,CAAC,GAAG,YAAY;AACzC,QAAI,UAAW,QAAO;AAAA,EACxB;AACA,SAAO;AACT;AAEA,SAAS,iBACP,QACA,SACoB;AACpB,MAAI;AACJ,aAAW,SAAS,SAAS;AAC3B,QAAI,WAAW,SAAS;AACtB,YAAM,UAAU,OAAO,MAAM,OAAO;AACpC,YAAMC,SAAQ,UAAU,YAAY,SAAS,OAAO,IAAI;AACxD,UAAIA,OAAO,QAAOA;AAClB,sBACE,kBAAkB,UAAU,YAAY,SAAS,gBAAgB,IAAI;AAAA,IACzE;AACA,UAAM,gBAAgB,YAAY,OAAO,SAAS;AAClD,QAAI,eAAe;AACjB,YAAM,WAAW,YAAY,OAAO,YAAY;AAChD,aAAO,WAAW,GAAG,QAAQ,IAAI,aAAa,KAAK;AAAA,IACrD;AACA,UAAM,cAAc,OAAO,MAAM,KAAK;AACtC,UAAM,cAAc,YAAY,eAAe,CAAC,GAAG,SAAS;AAC5D,QAAI,aAAa;AACf,YAAM,WAAW,YAAY,eAAe,CAAC,GAAG,YAAY;AAC5D,aAAO,WAAW,GAAG,QAAQ,IAAI,WAAW,KAAK;AAAA,IACnD;AACA,UAAM,UAAU,OAAO,MAAM,OAAO;AACpC,UAAM,QAAQ,UAAU,YAAY,SAAS,OAAO,IAAI;AACxD,QAAI,MAAO,QAAO;AAAA,EACpB;AACA,SAAO;AACT;AAEA,SAAS,eAAe,SAAwD;AAC9E,aAAW,SAAS,SAAS;AAC3B,UAAM,OAAO,OAAO,MAAM,IAAI;AAC9B,UAAM,MACJ,YAAY,OAAO,KAAK,KACxB,YAAY,QAAQ,CAAC,GAAG,KAAK,KAC7B,YAAY,QAAQ,CAAC,GAAG,MAAM,MAC7B,OAAO,MAAM,OAAO,IAAI,YAAY,OAAO,MAAM,OAAO,GAAI,KAAK,IAAI;AACxE,QAAI,IAAK,QAAO;AAAA,EAClB;AACA,SAAO;AACT;AAEA,SAAS,cAAc,SAAwD;AAC7E,aAAW,SAAS,SAAS;AAC3B,UAAM,UAAU,OAAO,MAAM,OAAO;AACpC,UAAM,cAAc,UAAU,YAAY,SAAS,MAAM,IAAI;AAC7D,QAAI,gBAAgB,eAAgB,QAAO,YAAY,SAAU,SAAS;AAE1E,QAAI,YAAY,OAAO,MAAM,MAAM,QAAQ;AACzC,YAAM,UAAU,MAAM;AACtB,UAAI,OAAO,YAAY,SAAU,QAAO;AAAA,IAC1C;AAEA,UAAM,cAAc,OAAO,OAAO,MAAM,OAAO,GAAG,OAAO;AACzD,UAAM,YAAY,cAAc,YAAY,aAAa,YAAY,IAAI;AACzE,QAAI,UAAW,QAAO;AAEtB,UAAM,UAAU,OAAO,MAAM,OAAO;AACpC,QAAI,WAAW,YAAY,SAAS,MAAM,MAAM,QAAQ;AACtD,YAAM,UAAU,QAAQ;AACxB,UAAI,OAAO,YAAY,SAAU,QAAO;AACxC,UAAI,MAAM,QAAQ,OAAO,GAAG;AAC1B,cAAM,OAAO,QACV,IAAI,CAAC,SAAS;AACb,gBAAM,MAAM,OAAO,IAAI;AACvB,iBAAO,MAAM,YAAY,KAAK,MAAM,IAAI;AAAA,QAC1C,CAAC,EACA,OAAO,CAAC,SAAyB,SAAS,MAAS,EACnD,KAAK,IAAI;AACZ,YAAI,KAAM,QAAO;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,kBAAkB,QAAgC,YAAwC;AACjG,SAAO,WAAW,GAAG,MAAM,IAAI,cAAc,iBAAiB,EAAE,EAAE;AAAA,IAChE,UAAU;AAAA,IACV,UAAU,SAAS;AAAA,EACrB;AACF;AAEA,SAAS,aAAa,OAAuB;AAC3C,QAAM,UAAU,MAAM,KAAK,KAAK;AAChC,MAAI,QAAQ,SAAS,GAAG,KAAK,UAAU,KAAK,OAAO,KAAK,sBAAsB,KAAK,OAAO,GAAG;AAC3F,WAAO;AAAA,EACT;AACA,SAAO,GAAG,OAAO;AACnB;AAEA,SAAS,SAAS,OAAwB;AACxC,SAAO,WAAW,KAAK,UAAU,KAAK,CAAC;AACzC;AAEA,SAAS,WAAW,OAAuB;AACzC,SAAO,UAAU,WAAW,QAAQ,EAAE,OAAO,KAAK,EAAE,OAAO,KAAK,CAAC;AACnE;AAEA,SAAS,WAAW,OAAuB;AACzC,SAAO,WAAW,QAAQ,EAAE,OAAO,KAAK,EAAE,OAAO,EAAE,aAAa,CAAC;AACnE;AAEA,SAAS,YAAY,SAA6B,MAAsB;AACtE,SAAO,YAAY,SAAY,OAAO,KAAK,IAAI,SAAS,IAAI;AAC9D;AAEA,SAAS,QAAQ,OAAuB;AACtC,MAAI,CAAC,OAAO,SAAS,KAAK,EAAG,QAAO;AACpC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,KAAK,CAAC;AACvC;AAEA,SAAS,SAAS,OAAkD;AAClE,SAAO,UAAU,QAAQ,OAAO,UAAU,YAAY,CAAC,MAAM,QAAQ,KAAK;AAC5E;AAEA,SAAS,OAAO,OAAgD;AAC9D,SAAO,SAAS,KAAK,IAAI,QAAQ;AACnC;AAEA,SAAS,YAAY,KAA8B,KAAiC;AAClF,QAAM,QAAQ,IAAI,GAAG;AACrB,SAAO,OAAO,UAAU,YAAY,MAAM,SAAS,IAAI,QAAQ;AACjE;AAEA,SAAS,YAAY,KAA8B,KAAiC;AAClF,QAAM,QAAQ,IAAI,GAAG;AACrB,SAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,IAAI,QAAQ;AACvE;AAEA,SAAS,YAAY,OAAoC;AACvD,MAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,GAAG;AACvD,WAAO,QAAQ,OAAoB,QAAQ,QAAQ;AAAA,EACrD;AACA,MAAI,OAAO,UAAU,YAAY,MAAM,SAAS,GAAG;AACjD,UAAM,SAAS,KAAK,MAAM,KAAK;AAC/B,WAAO,OAAO,SAAS,MAAM,IAAI,SAAS;AAAA,EAC5C;AACA,SAAO;AACT;","names":["payload","model"]}

package/dist/chunk-VI2UW6B6.js

@@ -0,0 +1,162 @@
+// src/metrics.ts
+var MODEL_PRICING = {
+  "gpt-4o": { input: 25e-4, output: 0.01 },
+  "gpt-4o-mini": { input: 15e-5, output: 6e-4 },
+  "gpt-4-turbo": { input: 0.01, output: 0.03 },
+  "claude-sonnet-4-20250514": { input: 3e-3, output: 0.015 },
+  "claude-opus-4-20250514": { input: 0.015, output: 0.075 },
+  "claude-3-haiku-20240307": { input: 25e-5, output: 125e-5 }
+};
+var FAMILY_PRICING = [
+  [/claude.*opus/, { input: 0.015, output: 0.075 }],
+  [/claude.*haiku/, { input: 8e-4, output: 4e-3 }],
+  [/claude.*sonnet|claude-code|claude-sonnet/, { input: 3e-3, output: 0.015 }],
+  [/gpt-4o-mini/, { input: 15e-5, output: 6e-4 }],
+  [/gpt-5|gpt-4\.1|o[134]\b/, { input: 125e-5, output: 0.01 }],
+  [/gpt-4o|gpt-4/, { input: 25e-4, output: 0.01 }],
+  [/deepseek/, { input: 3e-4, output: 11e-4 }],
+  [/glm|zhipu|zai/, { input: 6e-4, output: 22e-4 }],
+  [/kimi|moonshot/, { input: 6e-4, output: 25e-4 }],
+  [/qwen/, { input: 4e-4, output: 12e-4 }],
+  [/gemini.*flash/, { input: 1e-4, output: 4e-4 }],
+  [/gemini/, { input: 125e-5, output: 5e-3 }],
+  [/llama/, { input: 2e-4, output: 6e-4 }]
+];
+function normalizeModelId(model) {
+  return (model.split("@")[0] ?? model).trim().toLowerCase();
+}
+function resolveModelPricing(model) {
+  if (MODEL_PRICING[model]) return MODEL_PRICING[model];
+  const id = normalizeModelId(model);
+  if (MODEL_PRICING[id]) return MODEL_PRICING[id];
+  for (const [pattern, price] of FAMILY_PRICING) {
+    if (pattern.test(id)) return price;
+  }
+  return null;
+}
+function isModelPriced(model) {
+  return resolveModelPricing(model) !== null;
+}
+var warnedUnpricedModels = /* @__PURE__ */ new Set();
+function estimateTokens(text) {
+  return Math.ceil(text.length / 4);
+}
+function estimateCost(inputTokens, outputTokens, model) {
+  const pricing = resolveModelPricing(model);
+  if (!pricing) {
+    if (!warnedUnpricedModels.has(model)) {
+      warnedUnpricedModels.add(model);
+      console.warn(
+        `estimateCost: no pricing for model "${model}" \u2014 returning 0; add it to MODEL_PRICING/FAMILY_PRICING (cost/Pareto axes will be blank until then)`
+      );
+    }
+    return 0;
+  }
+  return inputTokens / 1e3 * pricing.input + outputTokens / 1e3 * pricing.output;
+}
+var TokenCounter = class {
+  totalInput = 0;
+  totalOutput = 0;
+  totalCost = 0;
+  model;
+  constructor(model = "gpt-4o") {
+    this.model = model;
+  }
+  /** Record tokens for a turn, returns per-turn cost */
+  record(inputTokens, outputTokens) {
+    this.totalInput += inputTokens;
+    this.totalOutput += outputTokens;
+    const cost = estimateCost(inputTokens, outputTokens, this.model);
+    this.totalCost += cost;
+    return cost;
+  }
+  /** Estimate and record from raw text */
+  recordFromText(inputText, outputText) {
+    const inputTokens = estimateTokens(inputText);
+    const outputTokens = estimateTokens(outputText);
+    const cost = this.record(inputTokens, outputTokens);
+    return { inputTokens, outputTokens, cost };
+  }
+  getTotalInput() {
+    return this.totalInput;
+  }
+  getTotalOutput() {
+    return this.totalOutput;
+  }
+  getTotalCost() {
+    return this.totalCost;
+  }
+};
+var MetricsCollector = class {
+  client;
+  workspaceId;
+  metrics = [];
+  constructor(client, workspaceId) {
+    this.client = client;
+    this.workspaceId = workspaceId;
+  }
+  /** Collect metrics after a turn completes */
+  async collect(turn, responseLatencyMs, responseChars, codeBlocksProduced, blocksExtracted, completionCriteriaMet, completionCriteriaTotal, qualityScore, inputTokens = 0, outputTokens = 0, estimatedCostUsd = 0) {
+    const state = await this.getState();
+    const m = {
+      turn,
+      timestamp: (/* @__PURE__ */ new Date()).toISOString(),
+      tasks: state.tasks,
+      events: state.events,
+      proposals: state.proposals,
+      vaultFiles: state.vaultFiles.length,
+      responseLatencyMs,
+      responseChars,
+      codeBlocksProduced,
+      blocksExtracted,
+      qualityScore,
+      inputTokens,
+      outputTokens,
+      estimatedCostUsd,
+      totalCostUsd: estimatedCostUsd,
+      completionPercent: completionCriteriaTotal > 0 ? completionCriteriaMet / completionCriteriaTotal * 100 : 0
+    };
+    this.metrics.push(m);
+    return m;
+  }
+  /** Get current product state */
+  async getState() {
+    const [tasks, events, approvals, vaultFiles] = await Promise.all([
+      this.client.getTasks(this.workspaceId),
+      this.client.getEvents(this.workspaceId),
+      this.client.getApprovals(this.workspaceId),
+      this.client.getVaultTree(this.workspaceId)
+    ]);
+    return {
+      tasks: tasks.length,
+      events: events.length,
+      proposals: {
+        pending: approvals.filter((a) => a.status === "pending").length,
+        approved: approvals.filter((a) => a.status === "approved").length,
+        rejected: approvals.filter((a) => a.status === "rejected").length
+      },
+      vaultFiles,
+      codeBlocks: 0,
+      generations: 0
+    };
+  }
+  /** Get all collected metrics */
+  getMetrics() {
+    return [...this.metrics];
+  }
+  /** Get convergence curve (completion% over turns) */
+  getConvergenceCurve() {
+    return this.metrics.map((m) => m.completionPercent);
+  }
+};
+
+export {
+  MODEL_PRICING,
+  resolveModelPricing,
+  isModelPriced,
+  estimateTokens,
+  estimateCost,
+  TokenCounter,
+  MetricsCollector
+};
+//# sourceMappingURL=chunk-VI2UW6B6.js.map

package/dist/chunk-VI2UW6B6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/metrics.ts"],"sourcesContent":["import type { ProductClient } from './client'\nimport type { DriverState, TurnMetrics } from './types'\n\ninterface TokenPrice {\n  input: number\n  output: number\n}\n\n/** Per-1K token pricing for exact model ids. */\nexport const MODEL_PRICING: Record<string, TokenPrice> = {\n  'gpt-4o': { input: 0.0025, output: 0.01 },\n  'gpt-4o-mini': { input: 0.00015, output: 0.0006 },\n  'gpt-4-turbo': { input: 0.01, output: 0.03 },\n  'claude-sonnet-4-20250514': { input: 0.003, output: 0.015 },\n  'claude-opus-4-20250514': { input: 0.015, output: 0.075 },\n  'claude-3-haiku-20240307': { input: 0.00025, output: 0.00125 },\n}\n\n/** Family-level pricing fallbacks (per-1K), matched against a normalized id\n *  after exact lookup misses. Ordered — first match wins. Covers the model\n *  ids actually used through the Tangle router + cli-bridge harnesses\n *  (`claude-code/sonnet`, `opencode/zai-coding-plan/glm-5.1`,\n *  `kimi-code/kimi-k2.6`, `deepseek-v4-pro`, `anthropic/claude-sonnet-4-6`, …),\n *  none of which appear in the exact table above — without this they priced\n *  to a silent $0, blanking every cost/Pareto axis downstream. */\nconst FAMILY_PRICING: Array<[RegExp, TokenPrice]> = [\n  [/claude.*opus/, { input: 0.015, output: 0.075 }],\n  [/claude.*haiku/, { input: 0.0008, output: 0.004 }],\n  [/claude.*sonnet|claude-code|claude-sonnet/, { input: 0.003, output: 0.015 }],\n  [/gpt-4o-mini/, { input: 0.00015, output: 0.0006 }],\n  [/gpt-5|gpt-4\\.1|o[134]\\b/, { input: 0.00125, output: 0.01 }],\n  [/gpt-4o|gpt-4/, { input: 0.0025, output: 0.01 }],\n  [/deepseek/, { input: 0.0003, output: 0.0011 }],\n  [/glm|zhipu|zai/, { input: 0.0006, output: 0.0022 }],\n  [/kimi|moonshot/, { input: 0.0006, output: 0.0025 }],\n  [/qwen/, { input: 0.0004, output: 0.0012 }],\n  [/gemini.*flash/, { input: 0.0001, output: 0.0004 }],\n  [/gemini/, { input: 0.00125, output: 0.005 }],\n  [/llama/, { input: 0.0002, output: 0.0006 }],\n]\n\n/** Normalize a model id for pricing: drop a `@snapshot` suffix, lowercase,\n *  and keep the final harness/provider-prefixed segment so family regexes\n *  match (`opencode/zai-coding-plan/glm-5.1` → `glm-5.1`). */\nfunction normalizeModelId(model: string): string {\n  return (model.split('@')[0] ?? model).trim().toLowerCase()\n}\n\n/** Resolve pricing for a model id: exact table, then family fallback.\n *  Returns null when the id matches nothing (caller decides — never a\n *  silent-zero masquerading as a real $0 cost). */\nexport function resolveModelPricing(model: string): TokenPrice | null {\n  if (MODEL_PRICING[model]) return MODEL_PRICING[model]\n  const id = normalizeModelId(model)\n  if (MODEL_PRICING[id]) return MODEL_PRICING[id]\n  for (const [pattern, price] of FAMILY_PRICING) {\n    if (pattern.test(id)) return price\n  }\n  return null\n}\n\n/** True when `model` has known pricing (exact or family). Lets cost-aware\n *  callers distinguish a real $0 from an unpriced model. */\nexport function isModelPriced(model: string): boolean {\n  return resolveModelPricing(model) !== null\n}\n\nconst warnedUnpricedModels = new Set<string>()\n\n/** Estimate token count from string length (chars / 4 approximation) */\nexport function estimateTokens(text: string): number {\n  return Math.ceil(text.length / 4)\n}\n\n/** Calculate cost in USD from token counts and model. Unknown models warn\n *  once (not a silent zero) and return 0 so callers that ignore pricing keep\n *  working; cost-sensitive callers should gate on {@link isModelPriced}. */\nexport function estimateCost(inputTokens: number, outputTokens: number, model: string): number {\n  const pricing = resolveModelPricing(model)\n  if (!pricing) {\n    if (!warnedUnpricedModels.has(model)) {\n      warnedUnpricedModels.add(model)\n      console.warn(\n        `estimateCost: no pricing for model \"${model}\" — returning 0; add it to ` +\n          'MODEL_PRICING/FAMILY_PRICING (cost/Pareto axes will be blank until then)',\n      )\n    }\n    return 0\n  }\n  return (inputTokens / 1000) * pricing.input + (outputTokens / 1000) * pricing.output\n}\n\n/**\n * TokenCounter — accumulates token usage and cost across turns.\n */\nexport class TokenCounter {\n  private totalInput = 0\n  private totalOutput = 0\n  private totalCost = 0\n  private model: string\n\n  constructor(model = 'gpt-4o') {\n    this.model = model\n  }\n\n  /** Record tokens for a turn, returns per-turn cost */\n  record(inputTokens: number, outputTokens: number): number {\n    this.totalInput += inputTokens\n    this.totalOutput += outputTokens\n    const cost = estimateCost(inputTokens, outputTokens, this.model)\n    this.totalCost += cost\n    return cost\n  }\n\n  /** Estimate and record from raw text */\n  recordFromText(\n    inputText: string,\n    outputText: string,\n  ): { inputTokens: number; outputTokens: number; cost: number } {\n    const inputTokens = estimateTokens(inputText)\n    const outputTokens = estimateTokens(outputText)\n    const cost = this.record(inputTokens, outputTokens)\n    return { inputTokens, outputTokens, cost }\n  }\n\n  getTotalInput(): number {\n    return this.totalInput\n  }\n  getTotalOutput(): number {\n    return this.totalOutput\n  }\n  getTotalCost(): number {\n    return this.totalCost\n  }\n}\n\n/**\n * MetricsCollector — collects per-turn metrics from the product.\n *\n * After each turn, queries the product's APIs to measure state changes.\n */\nexport class MetricsCollector {\n  private client: ProductClient\n  private workspaceId: string\n  private metrics: TurnMetrics[] = []\n  constructor(client: ProductClient, workspaceId: string) {\n    this.client = client\n    this.workspaceId = workspaceId\n  }\n\n  /** Collect metrics after a turn completes */\n  async collect(\n    turn: number,\n    responseLatencyMs: number,\n    responseChars: number,\n    codeBlocksProduced: number,\n    blocksExtracted: number,\n    completionCriteriaMet: number,\n    completionCriteriaTotal: number,\n    qualityScore?: number,\n    inputTokens = 0,\n    outputTokens = 0,\n    estimatedCostUsd = 0,\n  ): Promise<TurnMetrics> {\n    const state = await this.getState()\n\n    const m: TurnMetrics = {\n      turn,\n      timestamp: new Date().toISOString(),\n      tasks: state.tasks,\n      events: state.events,\n      proposals: state.proposals,\n      vaultFiles: state.vaultFiles.length,\n      responseLatencyMs,\n      responseChars,\n      codeBlocksProduced,\n      blocksExtracted,\n      qualityScore,\n      inputTokens,\n      outputTokens,\n      estimatedCostUsd,\n      totalCostUsd: estimatedCostUsd,\n      completionPercent:\n        completionCriteriaTotal > 0 ? (completionCriteriaMet / completionCriteriaTotal) * 100 : 0,\n    }\n\n    this.metrics.push(m)\n    return m\n  }\n\n  /** Get current product state */\n  async getState(): Promise<DriverState> {\n    const [tasks, events, approvals, vaultFiles] = await Promise.all([\n      this.client.getTasks(this.workspaceId),\n      this.client.getEvents(this.workspaceId),\n      this.client.getApprovals(this.workspaceId),\n      this.client.getVaultTree(this.workspaceId),\n    ])\n\n    return {\n      tasks: tasks.length,\n      events: events.length,\n      proposals: {\n        pending: approvals.filter((a) => a.status === 'pending').length,\n        approved: approvals.filter((a) => a.status === 'approved').length,\n        rejected: approvals.filter((a) => a.status === 'rejected').length,\n      },\n      vaultFiles,\n      codeBlocks: 0,\n      generations: 0,\n    }\n  }\n\n  /** Get all collected metrics */\n  getMetrics(): TurnMetrics[] {\n    return [...this.metrics]\n  }\n\n  /** Get convergence curve (completion% over turns) */\n  getConvergenceCurve(): number[] {\n    return this.metrics.map((m) => m.completionPercent)\n  }\n}\n"],"mappings":";AASO,IAAM,gBAA4C;AAAA,EACvD,UAAU,EAAE,OAAO,OAAQ,QAAQ,KAAK;AAAA,EACxC,eAAe,EAAE,OAAO,OAAS,QAAQ,KAAO;AAAA,EAChD,eAAe,EAAE,OAAO,MAAM,QAAQ,KAAK;AAAA,EAC3C,4BAA4B,EAAE,OAAO,MAAO,QAAQ,MAAM;AAAA,EAC1D,0BAA0B,EAAE,OAAO,OAAO,QAAQ,MAAM;AAAA,EACxD,2BAA2B,EAAE,OAAO,OAAS,QAAQ,OAAQ;AAC/D;AASA,IAAM,iBAA8C;AAAA,EAClD,CAAC,gBAAgB,EAAE,OAAO,OAAO,QAAQ,MAAM,CAAC;AAAA,EAChD,CAAC,iBAAiB,EAAE,OAAO,MAAQ,QAAQ,KAAM,CAAC;AAAA,EAClD,CAAC,4CAA4C,EAAE,OAAO,MAAO,QAAQ,MAAM,CAAC;AAAA,EAC5E,CAAC,eAAe,EAAE,OAAO,OAAS,QAAQ,KAAO,CAAC;AAAA,EAClD,CAAC,2BAA2B,EAAE,OAAO,QAAS,QAAQ,KAAK,CAAC;AAAA,EAC5D,CAAC,gBAAgB,EAAE,OAAO,OAAQ,QAAQ,KAAK,CAAC;AAAA,EAChD,CAAC,YAAY,EAAE,OAAO,MAAQ,QAAQ,MAAO,CAAC;AAAA,EAC9C,CAAC,iBAAiB,EAAE,OAAO,MAAQ,QAAQ,MAAO,CAAC;AAAA,EACnD,CAAC,iBAAiB,EAAE,OAAO,MAAQ,QAAQ,MAAO,CAAC;AAAA,EACnD,CAAC,QAAQ,EAAE,OAAO,MAAQ,QAAQ,MAAO,CAAC;AAAA,EAC1C,CAAC,iBAAiB,EAAE,OAAO,MAAQ,QAAQ,KAAO,CAAC;AAAA,EACnD,CAAC,UAAU,EAAE,OAAO,QAAS,QAAQ,KAAM,CAAC;AAAA,EAC5C,CAAC,SAAS,EAAE,OAAO,MAAQ,QAAQ,KAAO,CAAC;AAC7C;AAKA,SAAS,iBAAiB,OAAuB;AAC/C,UAAQ,MAAM,MAAM,GAAG,EAAE,CAAC,KAAK,OAAO,KAAK,EAAE,YAAY;AAC3D;AAKO,SAAS,oBAAoB,OAAkC;AACpE,MAAI,cAAc,KAAK,EAAG,QAAO,cAAc,KAAK;AACpD,QAAM,KAAK,iBAAiB,KAAK;AACjC,MAAI,cAAc,EAAE,EAAG,QAAO,cAAc,EAAE;AAC9C,aAAW,CAAC,SAAS,KAAK,KAAK,gBAAgB;AAC7C,QAAI,QAAQ,KAAK,EAAE,EAAG,QAAO;AAAA,EAC/B;AACA,SAAO;AACT;AAIO,SAAS,cAAc,OAAwB;AACpD,SAAO,oBAAoB,KAAK,MAAM;AACxC;AAEA,IAAM,uBAAuB,oBAAI,IAAY;AAGtC,SAAS,eAAe,MAAsB;AACnD,SAAO,KAAK,KAAK,KAAK,SAAS,CAAC;AAClC;AAKO,SAAS,aAAa,aAAqB,cAAsB,OAAuB;AAC7F,QAAM,UAAU,oBAAoB,KAAK;AACzC,MAAI,CAAC,SAAS;AACZ,QAAI,CAAC,qBAAqB,IAAI,KAAK,GAAG;AACpC,2BAAqB,IAAI,KAAK;AAC9B,cAAQ;AAAA,QACN,uCAAuC,KAAK;AAAA,MAE9C;AAAA,IACF;AACA,WAAO;AAAA,EACT;AACA,SAAQ,cAAc,MAAQ,QAAQ,QAAS,eAAe,MAAQ,QAAQ;AAChF;AAKO,IAAM,eAAN,MAAmB;AAAA,EAChB,aAAa;AAAA,EACb,cAAc;AAAA,EACd,YAAY;AAAA,EACZ;AAAA,EAER,YAAY,QAAQ,UAAU;AAC5B,SAAK,QAAQ;AAAA,EACf;AAAA;AAAA,EAGA,OAAO,aAAqB,cAA8B;AACxD,SAAK,cAAc;AACnB,SAAK,eAAe;AACpB,UAAM,OAAO,aAAa,aAAa,cAAc,KAAK,KAAK;AAC/D,SAAK,aAAa;AAClB,WAAO;AAAA,EACT;AAAA;AAAA,EAGA,eACE,WACA,YAC6D;AAC7D,UAAM,cAAc,eAAe,SAAS;AAC5C,UAAM,eAAe,eAAe,UAAU;AAC9C,UAAM,OAAO,KAAK,OAAO,aAAa,YAAY;AAClD,WAAO,EAAE,aAAa,cAAc,KAAK;AAAA,EAC3C;AAAA,EAEA,gBAAwB;AACtB,WAAO,KAAK;AAAA,EACd;AAAA,EACA,iBAAyB;AACvB,WAAO,KAAK;AAAA,EACd;AAAA,EACA,eAAuB;AACrB,WAAO,KAAK;AAAA,EACd;AACF;AAOO,IAAM,mBAAN,MAAuB;AAAA,EACpB;AAAA,EACA;AAAA,EACA,UAAyB,CAAC;AAAA,EAClC,YAAY,QAAuB,aAAqB;AACtD,SAAK,SAAS;AACd,SAAK,cAAc;AAAA,EACrB;AAAA;AAAA,EAGA,MAAM,QACJ,MACA,mBACA,eACA,oBACA,iBACA,uBACA,yBACA,cACA,cAAc,GACd,eAAe,GACf,mBAAmB,GACG;AACtB,UAAM,QAAQ,MAAM,KAAK,SAAS;AAElC,UAAM,IAAiB;AAAA,MACrB;AAAA,MACA,YAAW,oBAAI,KAAK,GAAE,YAAY;AAAA,MAClC,OAAO,MAAM;AAAA,MACb,QAAQ,MAAM;AAAA,MACd,WAAW,MAAM;AAAA,MACjB,YAAY,MAAM,WAAW;AAAA,MAC7B;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,cAAc;AAAA,MACd,mBACE,0BAA0B,IAAK,wBAAwB,0BAA2B,MAAM;AAAA,IAC5F;AAEA,SAAK,QAAQ,KAAK,CAAC;AACnB,WAAO;AAAA,EACT;AAAA;AAAA,EAGA,MAAM,WAAiC;AACrC,UAAM,CAAC,OAAO,QAAQ,WAAW,UAAU,IAAI,MAAM,QAAQ,IAAI;AAAA,MAC/D,KAAK,OAAO,SAAS,KAAK,WAAW;AAAA,MACrC,KAAK,OAAO,UAAU,KAAK,WAAW;AAAA,MACtC,KAAK,OAAO,aAAa,KAAK,WAAW;AAAA,MACzC,KAAK,OAAO,aAAa,KAAK,WAAW;AAAA,IAC3C,CAAC;AAED,WAAO;AAAA,MACL,OAAO,MAAM;AAAA,MACb,QAAQ,OAAO;AAAA,MACf,WAAW;AAAA,QACT,SAAS,UAAU,OAAO,CAAC,MAAM,EAAE,WAAW,SAAS,EAAE;AAAA,QACzD,UAAU,UAAU,OAAO,CAAC,MAAM,EAAE,WAAW,UAAU,EAAE;AAAA,QAC3D,UAAU,UAAU,OAAO,CAAC,MAAM,EAAE,WAAW,UAAU,EAAE;AAAA,MAC7D;AAAA,MACA;AAAA,MACA,YAAY;AAAA,MACZ,aAAa;AAAA,IACf;AAAA,EACF;AAAA;AAAA,EAGA,aAA4B;AAC1B,WAAO,CAAC,GAAG,KAAK,OAAO;AAAA,EACzB;AAAA;AAAA,EAGA,sBAAgC;AAC9B,WAAO,KAAK,QAAQ,IAAI,CAAC,MAAM,EAAE,iBAAiB;AAAA,EACpD;AACF;","names":[]}

package/dist/{chunk-CF67I6QY.js → chunk-VIDQF3F5.js}

@@ -1,6 +1,6 @@
 import {
   callLlm
-} from "./chunk-IHDHUN2X.js";
+} from "./chunk-CVVHBFGN.js";
 import {
   TraceFileMissingError,
   buildTraceAnalystTools
@@ -1041,4 +1041,4 @@ export {
   DEFAULT_TRACE_ANALYST_KINDS,
   AnalystRegistry
 };
-//# sourceMappingURL=chunk-CF67I6QY.js.map
+//# sourceMappingURL=chunk-VIDQF3F5.js.map

package/dist/{chunk-XXNIODOM.js → chunk-WJL2NJXN.js}

@@ -1,9 +1,9 @@
 import {
   assertLlmRoute
-} from "./chunk-IHDHUN2X.js";
+} from "./chunk-CVVHBFGN.js";
 import {
   researchReport
-} from "./chunk-KX6F6NCG.js";
+} from "./chunk-Z7VFTS2J.js";
 import {
   RunIntegrityError,
   assertRunCaptured
@@ -329,4 +329,4 @@ function defaultRunId(params) {
 export {
   runEvalCampaign
 };
-//# sourceMappingURL=chunk-XXNIODOM.js.map
+//# sourceMappingURL=chunk-WJL2NJXN.js.map

package/dist/{chunk-LB2UOI5F.js → chunk-YGYXHNAQ.js}

@@ -5,158 +5,6 @@ import {
   ValidationError
 } from "./chunk-3BFEG2F6.js";
 
-// src/metrics.ts
-var MODEL_PRICING = {
-  "gpt-4o": { input: 25e-4, output: 0.01 },
-  "gpt-4o-mini": { input: 15e-5, output: 6e-4 },
-  "gpt-4-turbo": { input: 0.01, output: 0.03 },
-  "claude-sonnet-4-20250514": { input: 3e-3, output: 0.015 },
-  "claude-opus-4-20250514": { input: 0.015, output: 0.075 },
-  "claude-3-haiku-20240307": { input: 25e-5, output: 125e-5 }
-};
-var FAMILY_PRICING = [
-  [/claude.*opus/, { input: 0.015, output: 0.075 }],
-  [/claude.*haiku/, { input: 8e-4, output: 4e-3 }],
-  [/claude.*sonnet|claude-code|claude-sonnet/, { input: 3e-3, output: 0.015 }],
-  [/gpt-4o-mini/, { input: 15e-5, output: 6e-4 }],
-  [/gpt-5|gpt-4\.1|o[134]\b/, { input: 125e-5, output: 0.01 }],
-  [/gpt-4o|gpt-4/, { input: 25e-4, output: 0.01 }],
-  [/deepseek/, { input: 3e-4, output: 11e-4 }],
-  [/glm|zhipu|zai/, { input: 6e-4, output: 22e-4 }],
-  [/kimi|moonshot/, { input: 6e-4, output: 25e-4 }],
-  [/qwen/, { input: 4e-4, output: 12e-4 }],
-  [/gemini.*flash/, { input: 1e-4, output: 4e-4 }],
-  [/gemini/, { input: 125e-5, output: 5e-3 }],
-  [/llama/, { input: 2e-4, output: 6e-4 }]
-];
-function normalizeModelId(model) {
-  return (model.split("@")[0] ?? model).trim().toLowerCase();
-}
-function resolveModelPricing(model) {
-  if (MODEL_PRICING[model]) return MODEL_PRICING[model];
-  const id = normalizeModelId(model);
-  if (MODEL_PRICING[id]) return MODEL_PRICING[id];
-  for (const [pattern, price] of FAMILY_PRICING) {
-    if (pattern.test(id)) return price;
-  }
-  return null;
-}
-function isModelPriced(model) {
-  return resolveModelPricing(model) !== null;
-}
-var warnedUnpricedModels = /* @__PURE__ */ new Set();
-function estimateTokens(text) {
-  return Math.ceil(text.length / 4);
-}
-function estimateCost(inputTokens, outputTokens, model) {
-  const pricing = resolveModelPricing(model);
-  if (!pricing) {
-    if (!warnedUnpricedModels.has(model)) {
-      warnedUnpricedModels.add(model);
-      console.warn(
-        `estimateCost: no pricing for model "${model}" \u2014 returning 0; add it to MODEL_PRICING/FAMILY_PRICING (cost/Pareto axes will be blank until then)`
-      );
-    }
-    return 0;
-  }
-  return inputTokens / 1e3 * pricing.input + outputTokens / 1e3 * pricing.output;
-}
-var TokenCounter = class {
-  totalInput = 0;
-  totalOutput = 0;
-  totalCost = 0;
-  model;
-  constructor(model = "gpt-4o") {
-    this.model = model;
-  }
-  /** Record tokens for a turn, returns per-turn cost */
-  record(inputTokens, outputTokens) {
-    this.totalInput += inputTokens;
-    this.totalOutput += outputTokens;
-    const cost = estimateCost(inputTokens, outputTokens, this.model);
-    this.totalCost += cost;
-    return cost;
-  }
-  /** Estimate and record from raw text */
-  recordFromText(inputText, outputText) {
-    const inputTokens = estimateTokens(inputText);
-    const outputTokens = estimateTokens(outputText);
-    const cost = this.record(inputTokens, outputTokens);
-    return { inputTokens, outputTokens, cost };
-  }
-  getTotalInput() {
-    return this.totalInput;
-  }
-  getTotalOutput() {
-    return this.totalOutput;
-  }
-  getTotalCost() {
-    return this.totalCost;
-  }
-};
-var MetricsCollector = class {
-  client;
-  workspaceId;
-  metrics = [];
-  constructor(client, workspaceId) {
-    this.client = client;
-    this.workspaceId = workspaceId;
-  }
-  /** Collect metrics after a turn completes */
-  async collect(turn, responseLatencyMs, responseChars, codeBlocksProduced, blocksExtracted, completionCriteriaMet, completionCriteriaTotal, qualityScore, inputTokens = 0, outputTokens = 0, estimatedCostUsd = 0) {
-    const state = await this.getState();
-    const m = {
-      turn,
-      timestamp: (/* @__PURE__ */ new Date()).toISOString(),
-      tasks: state.tasks,
-      events: state.events,
-      proposals: state.proposals,
-      vaultFiles: state.vaultFiles.length,
-      responseLatencyMs,
-      responseChars,
-      codeBlocksProduced,
-      blocksExtracted,
-      qualityScore,
-      inputTokens,
-      outputTokens,
-      estimatedCostUsd,
-      totalCostUsd: estimatedCostUsd,
-      completionPercent: completionCriteriaTotal > 0 ? completionCriteriaMet / completionCriteriaTotal * 100 : 0
-    };
-    this.metrics.push(m);
-    return m;
-  }
-  /** Get current product state */
-  async getState() {
-    const [tasks, events, approvals, vaultFiles] = await Promise.all([
-      this.client.getTasks(this.workspaceId),
-      this.client.getEvents(this.workspaceId),
-      this.client.getApprovals(this.workspaceId),
-      this.client.getVaultTree(this.workspaceId)
-    ]);
-    return {
-      tasks: tasks.length,
-      events: events.length,
-      proposals: {
-        pending: approvals.filter((a) => a.status === "pending").length,
-        approved: approvals.filter((a) => a.status === "approved").length,
-        rejected: approvals.filter((a) => a.status === "rejected").length
-      },
-      vaultFiles,
-      codeBlocks: 0,
-      generations: 0
-    };
-  }
-  /** Get all collected metrics */
-  getMetrics() {
-    return [...this.metrics];
-  }
-  /** Get convergence curve (completion% over turns) */
-  getConvergenceCurve() {
-    return this.metrics.map((m) => m.completionPercent);
-  }
-};
-
 // src/completion-verifier.ts
 var STOPWORDS = /* @__PURE__ */ new Set([
   "the",
@@ -344,6 +192,51 @@ ${content.slice(0, maxContentChars)}`
     return parseCorrectnessResponse(raw);
   };
 }
+var TITLE_STOPWORDS = /* @__PURE__ */ new Set([
+  "the",
+  "a",
+  "an",
+  "and",
+  "or",
+  "for",
+  "to",
+  "of",
+  "in",
+  "on",
+  "with",
+  "review",
+  "update",
+  "new",
+  "proposed"
+]);
+function createTokenRecallChecker(opts = {}) {
+  const minRecall = opts.minRecall ?? 0.5;
+  const minLen = opts.minContentLength ?? 120;
+  return async (requirement, content) => {
+    const body = content.trim();
+    if (body.length < minLen)
+      return {
+        correct: false,
+        reason: `content too thin (${body.length} chars) to be the deliverable`
+      };
+    const titleTokens = requirement.title.toLowerCase().split(/[^a-z0-9]+/).filter((t) => t.length > 2 && !TITLE_STOPWORDS.has(t));
+    if (titleTokens.length === 0)
+      return {
+        correct: true,
+        reason: "requirement title has no significant tokens \u2014 structural match accepted"
+      };
+    const lower = body.toLowerCase();
+    const hits = titleTokens.filter((t) => lower.includes(t)).length;
+    const recall = hits / titleTokens.length;
+    return recall >= minRecall ? {
+      correct: true,
+      reason: `content recalls ${hits}/${titleTokens.length} requirement tokens`
+    } : {
+      correct: false,
+      reason: `content recalls only ${hits}/${titleTokens.length} requirement tokens`
+    };
+  };
+}
 
 // src/produced-state.ts
 function artifactKind(mimeType) {
@@ -396,17 +289,11 @@ function agentProfileHash(profile) {
 }
 
 export {
-  MODEL_PRICING,
-  resolveModelPricing,
-  isModelPriced,
-  estimateTokens,
-  estimateCost,
-  TokenCounter,
-  MetricsCollector,
   verifyCompletion,
   parseCorrectnessResponse,
   createLlmCorrectnessChecker,
+  createTokenRecallChecker,
   extractProducedState,
   agentProfileHash
 };
-//# sourceMappingURL=chunk-LB2UOI5F.js.map
+//# sourceMappingURL=chunk-YGYXHNAQ.js.map

package/dist/chunk-YGYXHNAQ.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/completion-verifier.ts","../src/produced-state.ts","../src/agent-profile.ts"],"sourcesContent":["/**\n * Completion verifier — the task-completion oracle.\n *\n * Answers the only eval question that is not a proxy: did the agent actually\n * COMPLETE the task — produce every required deliverable, persisted and\n * correct — rather than describe what should be done. A fluent transcript\n * that never produces the artifact scores zero here.\n *\n * Per requirement, a two-stage check:\n *   1. Structural — a produced item (vault artifact / approved proposal /\n *      tool call) of the right kind is matched against the requirement and\n *      carries non-empty content. Deterministic; no LLM.\n *   2. Correctness — only if structurally present AND the matched item\n *      carries content, one targeted check decides whether that item\n *      actually fulfils the requirement. A hallucinated artifact fails here;\n *      an absent one already failed stage 1.\n *\n * `completionRate` is satisfied / total. Quality dimensions are meaningless\n * on an incomplete task — callers gate on `fullyComplete` / `completionRate`\n * before scoring quality.\n */\n\nimport type { TCloud } from '@tangle-network/tcloud'\nimport type { Artifact } from './artifact-validator'\n\n/** What kind of produced state can satisfy a requirement structurally. */\nexport type SatisfiedBy = 'artifact' | 'proposal' | 'tool-call' | 'any'\n\nexport interface CompletionRequirement {\n  /** Stable id from the task gold (e.g. a persona's `expected_requirements[].req_id`). */\n  reqId: string\n  /** Human-readable description of the required deliverable. */\n  title: string\n  /** Optional kind/category hint, matched against a produced item's kind. */\n  category?: string\n  /** What produced state satisfies this requirement. Defaults to 'any'. */\n  satisfiedBy?: SatisfiedBy\n}\n\nexport interface TaskGold {\n  taskId: string\n  requirements: CompletionRequirement[]\n}\n\nexport interface ProducedProposal {\n  id: string\n  title: string\n  status: 'pending' | 'approved' | 'rejected'\n  /** Optional persisted body — when present, enables a correctness check. */\n  content?: string\n}\n\n/** Everything observable about what a run actually produced. */\nexport interface ProducedState {\n  /** Persisted vault artifacts. Reuses the shared `Artifact` shape. */\n  artifacts: Artifact[]\n  /** Proposals / filings the agent created. */\n  proposals: ProducedProposal[]\n  /** Names of tools the agent invoked. */\n  toolCalls: string[]\n}\n\nexport interface RequirementCheck {\n  reqId: string\n  title: string\n  /** A produced item of the right kind matched the requirement, non-empty. */\n  structurallyPresent: boolean\n  /**\n   * Whether the matched item actually fulfils the requirement. `null` when\n   * not structurally present, or when the matched item carries no content\n   * to assess.\n   */\n  correct: boolean | null\n  /** structurallyPresent && correct !== false. */\n  satisfied: boolean\n  /** Human-readable evidence for the verdict. */\n  evidence: string[]\n}\n\nexport interface CompletionVerdict {\n  taskId: string\n  requirements: RequirementCheck[]\n  /** satisfied / total requirements. */\n  completionRate: number\n  /** Every requirement satisfied. */\n  fullyComplete: boolean\n}\n\n/**\n * Decides whether a produced item's content actually fulfils a requirement.\n * Injected so the structural verifier stays pure and unit-testable; the\n * production implementation is `createLlmCorrectnessChecker`.\n */\nexport type CorrectnessChecker = (\n  requirement: CompletionRequirement,\n  content: string,\n) => Promise<{ correct: boolean; reason: string }>\n\nconst STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'of',\n  'for',\n  'and',\n  'or',\n  'to',\n  'in',\n  'on',\n  'with',\n  'by',\n])\n\nconst MATCH_THRESHOLD = 0.5\nconst MIN_CONTENT_CHARS = 50\n\nfunction tokens(s: string): Set<string> {\n  return new Set(\n    s\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 1 && !STOPWORDS.has(t)),\n  )\n}\n\n/**\n * Recall of the requirement's tokens within a candidate's identifying text.\n * Recall, not Jaccard — a candidate's path/id legitimately carries extra\n * tokens the requirement does not name.\n */\nfunction tokenRecall(requirementText: string, candidateText: string): number {\n  const req = tokens(requirementText)\n  if (req.size === 0) return 0\n  const cand = tokens(candidateText)\n  let hit = 0\n  for (const t of req) if (cand.has(t)) hit++\n  return hit / req.size\n}\n\ninterface Candidate {\n  reqIndex: number\n  /** Unique key for a produced item — each item satisfies at most one requirement. */\n  itemKey: string\n  score: number\n  evidence: string\n  /** Content to correctness-check, or null when the matched item has none. */\n  content: string | null\n}\n\nfunction artifactCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  artifacts: Artifact[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  artifacts.forEach((a, i) => {\n    if ((a.content ?? '').trim().length < MIN_CONTENT_CHARS) return\n    let score = tokenRecall(reqText, `${a.path ?? ''} ${a.kind}`)\n    if (req.category && a.kind && req.category.toLowerCase() === a.kind.toLowerCase()) {\n      score = Math.max(score, 1)\n    }\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `artifact:${i}`,\n      score,\n      evidence: `artifact '${a.path ?? a.kind}' matched (token recall ${score.toFixed(2)})`,\n      content: a.content ?? null,\n    })\n  })\n  return out\n}\n\nfunction proposalCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  proposals: ProducedProposal[],\n): Candidate[] {\n  const reqText = `${req.title} ${req.category ?? ''}`\n  const out: Candidate[] = []\n  for (const p of proposals) {\n    // Pending or rejected work is not a completed deliverable.\n    if (p.status !== 'approved') continue\n    const score = tokenRecall(reqText, p.title)\n    if (score < MATCH_THRESHOLD) continue\n    const body = p.content ?? ''\n    out.push({\n      reqIndex,\n      itemKey: `proposal:${p.id}`,\n      score,\n      evidence: `approved proposal '${p.title}' matched (token recall ${score.toFixed(2)})`,\n      content: body.trim().length >= MIN_CONTENT_CHARS ? body : null,\n    })\n  }\n  return out\n}\n\nfunction toolCallCandidates(\n  req: CompletionRequirement,\n  reqIndex: number,\n  toolCalls: string[],\n): Candidate[] {\n  const out: Candidate[] = []\n  toolCalls.forEach((name, i) => {\n    const score = tokenRecall(req.title, name)\n    if (score < MATCH_THRESHOLD) return\n    out.push({\n      reqIndex,\n      itemKey: `tool:${i}`,\n      score,\n      evidence: `tool call '${name}' matched (token recall ${score.toFixed(2)})`,\n      content: null,\n    })\n  })\n  return out\n}\n\n/**\n * Verify whether a run completed the task. `checkCorrectness` is injected —\n * `createLlmCorrectnessChecker` for production, a deterministic stub in tests.\n *\n * Throws on a gold spec with no requirements: an eval task that requires\n * nothing is a misconfiguration, not a vacuously-complete task.\n */\nexport async function verifyCompletion(\n  gold: TaskGold,\n  state: ProducedState,\n  checkCorrectness: CorrectnessChecker,\n): Promise<CompletionVerdict> {\n  if (gold.requirements.length === 0) {\n    throw new Error(\n      `verifyCompletion: task '${gold.taskId}' has no requirements — malformed gold spec`,\n    )\n  }\n\n  // Collect every above-threshold (requirement, produced-item) candidate, then\n  // assign greedily by descending score: each requirement and each produced\n  // item is used at most once. One deliverable fulfils one requirement.\n  const candidates: Candidate[] = []\n  gold.requirements.forEach((req, i) => {\n    const by = req.satisfiedBy ?? 'any'\n    if (by === 'artifact' || by === 'any') {\n      candidates.push(...artifactCandidates(req, i, state.artifacts))\n    }\n    if (by === 'proposal' || by === 'any') {\n      candidates.push(...proposalCandidates(req, i, state.proposals))\n    }\n    if (by === 'tool-call' || by === 'any') {\n      candidates.push(...toolCallCandidates(req, i, state.toolCalls))\n    }\n  })\n  candidates.sort((a, b) => b.score - a.score)\n\n  const assigned = new Map<number, Candidate>()\n  const itemTaken = new Set<string>()\n  for (const c of candidates) {\n    if (assigned.has(c.reqIndex) || itemTaken.has(c.itemKey)) continue\n    assigned.set(c.reqIndex, c)\n    itemTaken.add(c.itemKey)\n  }\n\n  const requirements: RequirementCheck[] = []\n  for (let i = 0; i < gold.requirements.length; i++) {\n    const req = gold.requirements[i]!\n    const match = assigned.get(i)\n    const evidence: string[] = []\n    let correct: boolean | null = null\n\n    if (match) {\n      evidence.push(match.evidence)\n      if (match.content !== null) {\n        const r = await checkCorrectness(req, match.content)\n        correct = r.correct\n        evidence.push(`correctness: ${r.correct ? 'pass' : 'fail'} — ${r.reason}`)\n      } else {\n        evidence.push('correctness: not assessed — matched item carries no content')\n      }\n    } else {\n      const by = req.satisfiedBy ?? 'any'\n      const kind = by === 'any' ? 'artifact/proposal/tool-call' : by\n      evidence.push(`no produced ${kind} matched this requirement`)\n    }\n\n    const structurallyPresent = match !== undefined\n    const satisfied = structurallyPresent && correct !== false\n    requirements.push({\n      reqId: req.reqId,\n      title: req.title,\n      structurallyPresent,\n      correct,\n      satisfied,\n      evidence,\n    })\n  }\n\n  const satisfiedCount = requirements.filter((r) => r.satisfied).length\n  return {\n    taskId: gold.taskId,\n    requirements,\n    completionRate: satisfiedCount / requirements.length,\n    fullyComplete: satisfiedCount === requirements.length,\n  }\n}\n\nexport interface LlmCorrectnessCheckerOpts {\n  model?: string\n  /** Max chars of artifact content sent to the checker. */\n  maxContentChars?: number\n}\n\n/** Parse the correctness checker's model response. Fails loud on a bad shape. */\nexport function parseCorrectnessResponse(raw: string): { correct: boolean; reason: string } {\n  const match = raw.match(/\\{[\\s\\S]*\\}/)\n  if (!match) {\n    throw new Error(`correctness checker: no JSON object in model response: ${raw.slice(0, 200)}`)\n  }\n  const parsed = JSON.parse(match[0]) as { correct?: unknown; reason?: unknown }\n  if (typeof parsed.correct !== 'boolean') {\n    throw new Error(`correctness checker: 'correct' is not a boolean in: ${match[0].slice(0, 200)}`)\n  }\n  return { correct: parsed.correct, reason: typeof parsed.reason === 'string' ? parsed.reason : '' }\n}\n\n/**\n * Production `CorrectnessChecker` — one LLM call per matched artifact,\n * deterministic (temperature 0), structured JSON out. Judges fulfilment\n * only: a plan, a gesture, or a description of what should be done does not\n * fulfil a requirement — the artifact must BE the deliverable.\n */\nexport function createLlmCorrectnessChecker(\n  tc: TCloud,\n  opts: LlmCorrectnessCheckerOpts = {},\n): CorrectnessChecker {\n  const model = opts.model ?? 'claude-sonnet-4-6'\n  const maxContentChars = opts.maxContentChars ?? 8000\n  return async (requirement, content) => {\n    const resp = await tc.chat({\n      model,\n      messages: [\n        {\n          role: 'system',\n          content:\n            'You verify whether a produced work artifact actually fulfils a stated requirement. Judge fulfilment only — is the deliverable substantively present and on-point — not polish. A plan to do it later, a vague gesture, or a description of what should be done does NOT fulfil a requirement; the artifact must BE the deliverable. Respond with a single JSON object: {\"correct\": boolean, \"reason\": string (<= 30 words)}.',\n        },\n        {\n          role: 'user',\n          content: `Requirement: ${requirement.title}\\n${\n            requirement.category ? `Category: ${requirement.category}\\n` : ''\n          }\\nProduced artifact:\\n${content.slice(0, maxContentChars)}`,\n        },\n      ],\n      temperature: 0,\n      maxTokens: 200,\n    })\n    const raw =\n      (resp as { choices?: { message?: { content?: string } }[] }).choices?.[0]?.message?.content ??\n      ''\n    return parseCorrectnessResponse(raw)\n  }\n}\n\n/** Stopwords for requirement-title tokenization — drops the imperative verbs\n *  ('review', 'update', …) common to deliverable titles so recall keys on the\n *  substantive nouns, not the boilerplate ask. */\nconst TITLE_STOPWORDS = new Set([\n  'the',\n  'a',\n  'an',\n  'and',\n  'or',\n  'for',\n  'to',\n  'of',\n  'in',\n  'on',\n  'with',\n  'review',\n  'update',\n  'new',\n  'proposed',\n])\n\n/**\n * Deterministic `CorrectnessChecker` — the no-LLM counterpart to\n * `createLlmCorrectnessChecker`. A produced item fulfils a requirement when its\n * content is substantive (≥ `minContentLength` chars) AND recalls ≥ `minRecall`\n * of the requirement title's significant tokens. No network — the default gate\n * for apps and tests without an LLM judge. Pass to `verifyCompletion` as the\n * checker.\n */\nexport function createTokenRecallChecker(\n  opts: { minRecall?: number; minContentLength?: number } = {},\n): CorrectnessChecker {\n  const minRecall = opts.minRecall ?? 0.5\n  const minLen = opts.minContentLength ?? 120\n  return async (requirement, content) => {\n    const body = content.trim()\n    if (body.length < minLen)\n      return {\n        correct: false,\n        reason: `content too thin (${body.length} chars) to be the deliverable`,\n      }\n    const titleTokens = requirement.title\n      .toLowerCase()\n      .split(/[^a-z0-9]+/)\n      .filter((t) => t.length > 2 && !TITLE_STOPWORDS.has(t))\n    if (titleTokens.length === 0)\n      return {\n        correct: true,\n        reason: 'requirement title has no significant tokens — structural match accepted',\n      }\n    const lower = body.toLowerCase()\n    const hits = titleTokens.filter((t) => lower.includes(t)).length\n    const recall = hits / titleTokens.length\n    return recall >= minRecall\n      ? {\n          correct: true,\n          reason: `content recalls ${hits}/${titleTokens.length} requirement tokens`,\n        }\n      : {\n          correct: false,\n          reason: `content recalls only ${hits}/${titleTokens.length} requirement tokens`,\n        }\n  }\n}\n","/**\n * Produced-state extraction — normalize a run's runtime event stream into the\n * typed `ProducedState` the completion oracle consumes.\n *\n * `ProducedState` answers \"what did the agent actually produce\" — vault\n * artifacts, proposals, tool calls. The runtime emits these as a stream of\n * events; this module is the single normalization point from that stream to\n * the shape `verifyCompletion` expects.\n *\n * Input is structurally typed (`RuntimeEventLike`) so this module does not\n * depend on agent-runtime — agent-runtime's `RuntimeStreamEvent` satisfies it\n * structurally. The `content` on `ArtifactEventLike` and the whole\n * `proposal_created` variant are the runtime-side enrichments this contract\n * requires; the runtime emits them, this module consumes them.\n */\n\nimport type { Artifact } from './artifact-validator'\nimport type { ProducedProposal, ProducedState } from './completion-verifier'\n\n/** A tool the agent invoked. */\nexport interface ToolCallEventLike {\n  type: 'tool_call'\n  toolName: string\n}\n\n/**\n * An artifact the agent produced. `content` is the enriched field — the\n * runtime's base `artifact` event carries only metadata; the completion\n * oracle needs the body to verify the deliverable, so the runtime emits it.\n */\nexport interface ArtifactEventLike {\n  type: 'artifact'\n  artifactId: string\n  name?: string\n  mimeType?: string\n  uri?: string\n  content?: string\n}\n\n/** A proposal / filing the agent created. */\nexport interface ProposalEventLike {\n  type: 'proposal_created'\n  proposalId: string\n  title: string\n  status?: 'pending' | 'approved' | 'rejected'\n}\n\n/**\n * The subset of runtime stream events `extractProducedState` consumes.\n * agent-runtime's full `RuntimeStreamEvent` union satisfies this structurally;\n * the `{ type: string }` catch-all keeps the input permissive so callers can\n * pass the whole unfiltered telemetry stream — unrecognized events are skipped.\n */\nexport type RuntimeEventLike =\n  | ToolCallEventLike\n  | ArtifactEventLike\n  | ProposalEventLike\n  | { type: string }\n\nfunction artifactKind(mimeType: string | undefined): string {\n  if (!mimeType) return 'file'\n  if (mimeType.includes('json')) return 'json'\n  if (mimeType.startsWith('text/')) return 'text'\n  return 'file'\n}\n\n/**\n * Normalize a run's runtime event stream into `ProducedState`.\n *\n * Pure and total — unrecognized event types are skipped. `toolCalls` is\n * deduplicated by name in first-seen order (completion cares about a tool's\n * presence, not its call count). An artifact with neither a name nor a uri\n * still yields an entry keyed by its `artifactId` so it is never silently\n * dropped; an artifact with no `content` yields empty content, which the\n * completion oracle's structural check then rejects on its own.\n */\nexport function extractProducedState(events: readonly RuntimeEventLike[]): ProducedState {\n  const artifacts: Artifact[] = []\n  const proposals: ProducedProposal[] = []\n  const toolCalls: string[] = []\n  const seenTools = new Set<string>()\n\n  for (const ev of events) {\n    if (ev.type === 'tool_call') {\n      const name = (ev as ToolCallEventLike).toolName\n      if (name && !seenTools.has(name)) {\n        seenTools.add(name)\n        toolCalls.push(name)\n      }\n    } else if (ev.type === 'artifact') {\n      const a = ev as ArtifactEventLike\n      artifacts.push({\n        kind: artifactKind(a.mimeType),\n        path: a.name ?? a.uri ?? a.artifactId,\n        content: a.content ?? '',\n      })\n    } else if (ev.type === 'proposal_created') {\n      const p = ev as ProposalEventLike\n      proposals.push({ id: p.proposalId, title: p.title, status: p.status ?? 'pending' })\n    }\n  }\n\n  return { artifacts, proposals, toolCalls }\n}\n","/**\n * @stable\n *\n * AgentProfile — the eval harness's unit of variation.\n *\n * A profile pins everything that changes agent behaviour for a benchmark\n * cell: the model, the active skills, the prompt version, the available\n * tools. Vary the profile — swap a model, add a skill — and re-run the suite\n * to benchmark the change. The scorecard keys a cell on\n * `(scenarioId, profileHash)`, so the model is not a separate axis: it lives\n * inside the profile, and two profiles with the same model but different\n * skills are different cells.\n *\n * `agentProfileHash` is the profile's behaviour identity. Two profiles that\n * produce the same agent behaviour share a hash (and a scorecard cell);\n * reordering `skills` or `tools` does not change it; the human-facing `id`\n * label does not affect it.\n */\n\nimport { createHash } from 'node:crypto'\nimport { ValidationError } from './errors'\nimport { canonicalize } from './pre-registration'\n\nexport interface AgentProfile {\n  /** Human-facing label, e.g. `sonnet-legal-skills-v3`. Not part of the hash. */\n  id: string\n  /** Model snapshot id this profile pins, e.g. `claude-sonnet-4-6@2025-04-15`. */\n  model: string\n  /** Skill ids/versions active in this profile — the primary behaviour lever. */\n  skills?: string[]\n  /** Prompt version identifier. */\n  promptVersion?: string\n  /** Tool ids available to the agent. */\n  tools?: string[]\n  /** Any other behaviour-bearing knobs that should fingerprint into the hash. */\n  metadata?: Record<string, string | number | boolean>\n}\n\n/**\n * Deterministic behaviour identity of a profile — a sha256 over the\n * behaviour-bearing fields. `skills` and `tools` are order-insensitive; the\n * `id` label is excluded. Throws on a profile with no `model` — an unkeyable\n * profile must fail loud rather than collapse into a blank-model cell.\n */\nexport function agentProfileHash(profile: AgentProfile): string {\n  if (typeof profile.model !== 'string' || profile.model.trim().length === 0) {\n    throw new ValidationError(`AgentProfile \"${profile.id}\" has no model — cannot hash`)\n  }\n  const behaviour = {\n    model: profile.model.trim(),\n    skills: [...(profile.skills ?? [])].sort(),\n    promptVersion: profile.promptVersion ?? null,\n    tools: [...(profile.tools ?? [])].sort(),\n    metadata: profile.metadata ?? {},\n  }\n  return createHash('sha256')\n    .update(JSON.stringify(canonicalize(behaviour)))\n    .digest('hex')\n}\n"],"mappings":";;;;;;;;AAkGA,IAAM,YAAY,oBAAI,IAAI;AAAA,EACxB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAED,IAAM,kBAAkB;AACxB,IAAM,oBAAoB;AAE1B,SAAS,OAAO,GAAwB;AACtC,SAAO,IAAI;AAAA,IACT,EACG,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,UAAU,IAAI,CAAC,CAAC;AAAA,EACpD;AACF;AAOA,SAAS,YAAY,iBAAyB,eAA+B;AAC3E,QAAM,MAAM,OAAO,eAAe;AAClC,MAAI,IAAI,SAAS,EAAG,QAAO;AAC3B,QAAM,OAAO,OAAO,aAAa;AACjC,MAAI,MAAM;AACV,aAAW,KAAK,IAAK,KAAI,KAAK,IAAI,CAAC,EAAG;AACtC,SAAO,MAAM,IAAI;AACnB;AAYA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,GAAG,MAAM;AAC1B,SAAK,EAAE,WAAW,IAAI,KAAK,EAAE,SAAS,kBAAmB;AACzD,QAAI,QAAQ,YAAY,SAAS,GAAG,EAAE,QAAQ,EAAE,IAAI,EAAE,IAAI,EAAE;AAC5D,QAAI,IAAI,YAAY,EAAE,QAAQ,IAAI,SAAS,YAAY,MAAM,EAAE,KAAK,YAAY,GAAG;AACjF,cAAQ,KAAK,IAAI,OAAO,CAAC;AAAA,IAC3B;AACA,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,CAAC;AAAA,MACtB;AAAA,MACA,UAAU,aAAa,EAAE,QAAQ,EAAE,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS,EAAE,WAAW;AAAA,IACxB,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,UAAU,GAAG,IAAI,KAAK,IAAI,IAAI,YAAY,EAAE;AAClD,QAAM,MAAmB,CAAC;AAC1B,aAAW,KAAK,WAAW;AAEzB,QAAI,EAAE,WAAW,WAAY;AAC7B,UAAM,QAAQ,YAAY,SAAS,EAAE,KAAK;AAC1C,QAAI,QAAQ,gBAAiB;AAC7B,UAAM,OAAO,EAAE,WAAW;AAC1B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,YAAY,EAAE,EAAE;AAAA,MACzB;AAAA,MACA,UAAU,sBAAsB,EAAE,KAAK,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MAClF,SAAS,KAAK,KAAK,EAAE,UAAU,oBAAoB,OAAO;AAAA,IAC5D,CAAC;AAAA,EACH;AACA,SAAO;AACT;AAEA,SAAS,mBACP,KACA,UACA,WACa;AACb,QAAM,MAAmB,CAAC;AAC1B,YAAU,QAAQ,CAAC,MAAM,MAAM;AAC7B,UAAM,QAAQ,YAAY,IAAI,OAAO,IAAI;AACzC,QAAI,QAAQ,gBAAiB;AAC7B,QAAI,KAAK;AAAA,MACP;AAAA,MACA,SAAS,QAAQ,CAAC;AAAA,MAClB;AAAA,MACA,UAAU,cAAc,IAAI,2BAA2B,MAAM,QAAQ,CAAC,CAAC;AAAA,MACvE,SAAS;AAAA,IACX,CAAC;AAAA,EACH,CAAC;AACD,SAAO;AACT;AASA,eAAsB,iBACpB,MACA,OACA,kBAC4B;AAC5B,MAAI,KAAK,aAAa,WAAW,GAAG;AAClC,UAAM,IAAI;AAAA,MACR,2BAA2B,KAAK,MAAM;AAAA,IACxC;AAAA,EACF;AAKA,QAAM,aAA0B,CAAC;AACjC,OAAK,aAAa,QAAQ,CAAC,KAAK,MAAM;AACpC,UAAM,KAAK,IAAI,eAAe;AAC9B,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,cAAc,OAAO,OAAO;AACrC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AACA,QAAI,OAAO,eAAe,OAAO,OAAO;AACtC,iBAAW,KAAK,GAAG,mBAAmB,KAAK,GAAG,MAAM,SAAS,CAAC;AAAA,IAChE;AAAA,EACF,CAAC;AACD,aAAW,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAE3C,QAAM,WAAW,oBAAI,IAAuB;AAC5C,QAAM,YAAY,oBAAI,IAAY;AAClC,aAAW,KAAK,YAAY;AAC1B,QAAI,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,IAAI,EAAE,OAAO,EAAG;AAC1D,aAAS,IAAI,EAAE,UAAU,CAAC;AAC1B,cAAU,IAAI,EAAE,OAAO;AAAA,EACzB;AAEA,QAAM,eAAmC,CAAC;AAC1C,WAAS,IAAI,GAAG,IAAI,KAAK,aAAa,QAAQ,KAAK;AACjD,UAAM,MAAM,KAAK,aAAa,CAAC;AAC/B,UAAM,QAAQ,SAAS,IAAI,CAAC;AAC5B,UAAM,WAAqB,CAAC;AAC5B,QAAI,UAA0B;AAE9B,QAAI,OAAO;AACT,eAAS,KAAK,MAAM,QAAQ;AAC5B,UAAI,MAAM,YAAY,MAAM;AAC1B,cAAM,IAAI,MAAM,iBAAiB,KAAK,MAAM,OAAO;AACnD,kBAAU,EAAE;AACZ,iBAAS,KAAK,gBAAgB,EAAE,UAAU,SAAS,MAAM,WAAM,EAAE,MAAM,EAAE;AAAA,MAC3E,OAAO;AACL,iBAAS,KAAK,kEAA6D;AAAA,MAC7E;AAAA,IACF,OAAO;AACL,YAAM,KAAK,IAAI,eAAe;AAC9B,YAAM,OAAO,OAAO,QAAQ,gCAAgC;AAC5D,eAAS,KAAK,eAAe,IAAI,2BAA2B;AAAA,IAC9D;AAEA,UAAM,sBAAsB,UAAU;AACtC,UAAM,YAAY,uBAAuB,YAAY;AACrD,iBAAa,KAAK;AAAA,MAChB,OAAO,IAAI;AAAA,MACX,OAAO,IAAI;AAAA,MACX;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAAA,EACH;AAEA,QAAM,iBAAiB,aAAa,OAAO,CAAC,MAAM,EAAE,SAAS,EAAE;AAC/D,SAAO;AAAA,IACL,QAAQ,KAAK;AAAA,IACb;AAAA,IACA,gBAAgB,iBAAiB,aAAa;AAAA,IAC9C,eAAe,mBAAmB,aAAa;AAAA,EACjD;AACF;AASO,SAAS,yBAAyB,KAAmD;AAC1F,QAAM,QAAQ,IAAI,MAAM,aAAa;AACrC,MAAI,CAAC,OAAO;AACV,UAAM,IAAI,MAAM,0DAA0D,IAAI,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EAC/F;AACA,QAAM,SAAS,KAAK,MAAM,MAAM,CAAC,CAAC;AAClC,MAAI,OAAO,OAAO,YAAY,WAAW;AACvC,UAAM,IAAI,MAAM,uDAAuD,MAAM,CAAC,EAAE,MAAM,GAAG,GAAG,CAAC,EAAE;AAAA,EACjG;AACA,SAAO,EAAE,SAAS,OAAO,SAAS,QAAQ,OAAO,OAAO,WAAW,WAAW,OAAO,SAAS,GAAG;AACnG;AAQO,SAAS,4BACd,IACA,OAAkC,CAAC,GACf;AACpB,QAAM,QAAQ,KAAK,SAAS;AAC5B,QAAM,kBAAkB,KAAK,mBAAmB;AAChD,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,OAAO,MAAM,GAAG,KAAK;AAAA,MACzB;AAAA,MACA,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SACE;AAAA,QACJ;AAAA,QACA;AAAA,UACE,MAAM;AAAA,UACN,SAAS,gBAAgB,YAAY,KAAK;AAAA,EACxC,YAAY,WAAW,aAAa,YAAY,QAAQ;AAAA,IAAO,EACjE;AAAA;AAAA,EAAyB,QAAQ,MAAM,GAAG,eAAe,CAAC;AAAA,QAC5D;AAAA,MACF;AAAA,MACA,aAAa;AAAA,MACb,WAAW;AAAA,IACb,CAAC;AACD,UAAM,MACH,KAA4D,UAAU,CAAC,GAAG,SAAS,WACpF;AACF,WAAO,yBAAyB,GAAG;AAAA,EACrC;AACF;AAKA,IAAM,kBAAkB,oBAAI,IAAI;AAAA,EAC9B;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUM,SAAS,yBACd,OAA0D,CAAC,GACvC;AACpB,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,SAAS,KAAK,oBAAoB;AACxC,SAAO,OAAO,aAAa,YAAY;AACrC,UAAM,OAAO,QAAQ,KAAK;AAC1B,QAAI,KAAK,SAAS;AAChB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ,qBAAqB,KAAK,MAAM;AAAA,MAC1C;AACF,UAAM,cAAc,YAAY,MAC7B,YAAY,EACZ,MAAM,YAAY,EAClB,OAAO,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,gBAAgB,IAAI,CAAC,CAAC;AACxD,QAAI,YAAY,WAAW;AACzB,aAAO;AAAA,QACL,SAAS;AAAA,QACT,QAAQ;AAAA,MACV;AACF,UAAM,QAAQ,KAAK,YAAY;AAC/B,UAAM,OAAO,YAAY,OAAO,CAAC,MAAM,MAAM,SAAS,CAAC,CAAC,EAAE;AAC1D,UAAM,SAAS,OAAO,YAAY;AAClC,WAAO,UAAU,YACb;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,mBAAmB,IAAI,IAAI,YAAY,MAAM;AAAA,IACvD,IACA;AAAA,MACE,SAAS;AAAA,MACT,QAAQ,wBAAwB,IAAI,IAAI,YAAY,MAAM;AAAA,IAC5D;AAAA,EACN;AACF;;;AC9WA,SAAS,aAAa,UAAsC;AAC1D,MAAI,CAAC,SAAU,QAAO;AACtB,MAAI,SAAS,SAAS,MAAM,EAAG,QAAO;AACtC,MAAI,SAAS,WAAW,OAAO,EAAG,QAAO;AACzC,SAAO;AACT;AAYO,SAAS,qBAAqB,QAAoD;AACvF,QAAM,YAAwB,CAAC;AAC/B,QAAM,YAAgC,CAAC;AACvC,QAAM,YAAsB,CAAC;AAC7B,QAAM,YAAY,oBAAI,IAAY;AAElC,aAAW,MAAM,QAAQ;AACvB,QAAI,GAAG,SAAS,aAAa;AAC3B,YAAM,OAAQ,GAAyB;AACvC,UAAI,QAAQ,CAAC,UAAU,IAAI,IAAI,GAAG;AAChC,kBAAU,IAAI,IAAI;AAClB,kBAAU,KAAK,IAAI;AAAA,MACrB;AAAA,IACF,WAAW,GAAG,SAAS,YAAY;AACjC,YAAM,IAAI;AACV,gBAAU,KAAK;AAAA,QACb,MAAM,aAAa,EAAE,QAAQ;AAAA,QAC7B,MAAM,EAAE,QAAQ,EAAE,OAAO,EAAE;AAAA,QAC3B,SAAS,EAAE,WAAW;AAAA,MACxB,CAAC;AAAA,IACH,WAAW,GAAG,SAAS,oBAAoB;AACzC,YAAM,IAAI;AACV,gBAAU,KAAK,EAAE,IAAI,EAAE,YAAY,OAAO,EAAE,OAAO,QAAQ,EAAE,UAAU,UAAU,CAAC;AAAA,IACpF;AAAA,EACF;AAEA,SAAO,EAAE,WAAW,WAAW,UAAU;AAC3C;;;ACpFA,SAAS,kBAAkB;AAyBpB,SAAS,iBAAiB,SAA+B;AAC9D,MAAI,OAAO,QAAQ,UAAU,YAAY,QAAQ,MAAM,KAAK,EAAE,WAAW,GAAG;AAC1E,UAAM,IAAI,gBAAgB,iBAAiB,QAAQ,EAAE,mCAA8B;AAAA,EACrF;AACA,QAAM,YAAY;AAAA,IAChB,OAAO,QAAQ,MAAM,KAAK;AAAA,IAC1B,QAAQ,CAAC,GAAI,QAAQ,UAAU,CAAC,CAAE,EAAE,KAAK;AAAA,IACzC,eAAe,QAAQ,iBAAiB;AAAA,IACxC,OAAO,CAAC,GAAI,QAAQ,SAAS,CAAC,CAAE,EAAE,KAAK;AAAA,IACvC,UAAU,QAAQ,YAAY,CAAC;AAAA,EACjC;AACA,SAAO,WAAW,QAAQ,EACvB,OAAO,KAAK,UAAU,aAAa,SAAS,CAAC,CAAC,EAC9C,OAAO,KAAK;AACjB;","names":[]}

package/dist/{chunk-KX6F6NCG.js → chunk-Z7VFTS2J.js}

@@ -5,7 +5,7 @@ import {
   pairedBootstrap,
   pairedMde,
   wilcoxonSignedRank
-} from "./chunk-ITBRCT73.js";
+} from "./chunk-IDVBLYCY.js";
 import {
   canonicalize,
   hashJson
@@ -877,4 +877,4 @@ export {
   RESEARCH_REPORT_HARD_PAIR_FLOOR,
   researchReport
 };
-//# sourceMappingURL=chunk-KX6F6NCG.js.map
+//# sourceMappingURL=chunk-Z7VFTS2J.js.map