@tangle-network/agent-eval 0.128.2 → 0.130.0

package/dist/chunk-IHQDPH7D.js

@@ -1,171 +0,0 @@
-import {
-  ROLLOUT_SCHEMA
-} from "./chunk-UWZZKKU7.js";
-import {
-  buildTrajectory
-} from "./chunk-RZTMDUO7.js";
-import {
-  runTaskScore
-} from "./chunk-2JX3CFMB.js";
-import {
-  ValidationError
-} from "./chunk-ONWEPEDO.js";
-
-// src/rollout/mint.ts
-var asText = (v, scrub) => {
-  const s = typeof v === "string" ? v : JSON.stringify(v);
-  return scrub(s ?? "");
-};
-function projectStep(span, scrub) {
-  const base = {
-    kind: span.kind,
-    name: scrub(span.name),
-    status: span.status,
-    durationMs: span.endedAt !== void 0 ? span.endedAt - span.startedAt : void 0
-  };
-  if (span.kind === "llm") {
-    const llm = span;
-    const last = llm.messages[llm.messages.length - 1];
-    if (last) base.input = scrub(last.content);
-    if (llm.output !== void 0) base.output = scrub(llm.output);
-  } else if (span.kind === "tool") {
-    const tool = span;
-    base.input = asText(tool.args, scrub);
-    if (tool.result !== void 0) base.output = asText(tool.result, scrub);
-  }
-  return base;
-}
-function finalConversation(spans, scrub) {
-  const llms = spans.filter((s) => s.kind === "llm");
-  const last = llms[llms.length - 1];
-  if (!last) return [];
-  const messages = last.messages.map((m) => ({
-    role: m.role,
-    content: scrub(m.content)
-  }));
-  if (last.output !== void 0 && last.output !== "") {
-    messages.push({ role: "assistant", content: scrub(last.output) });
-  }
-  return messages;
-}
-function scoredReward(record) {
-  const score = runTaskScore(record);
-  if (score === void 0) {
-    throw new ValidationError(`Cannot mint rollout for run ${record.runId}: task score is missing`);
-  }
-  const gated = record.outcome.realness?.gated === true;
-  return {
-    reward: gated ? 0 : score,
-    gated,
-    source: record.outcome.holdoutScore !== void 0 ? "run-record/holdout-score" : "run-record/search-score"
-  };
-}
-function rolloutReward(record) {
-  const { reward, gated } = scoredReward(record);
-  return { reward, gated };
-}
-var SPLIT_FROM_TAG = {
-  search: "search",
-  dev: "dev",
-  holdout: "holdout"
-};
-function mintLine(record, steps, messages, options, capturedAt, gap) {
-  const { reward, gated, source } = scoredReward(record);
-  const uncaptured = record.costProvenance.kind === "uncaptured";
-  const terminalOutcome = record.terminalOutcome;
-  const isCompleted = terminalOutcome === "succeeded" || terminalOutcome === "failed";
-  const isTruncated = terminalOutcome === "cancelled" || terminalOutcome === "incomplete";
-  const terminalError = terminalOutcome === "failed" || terminalOutcome === "cancelled" || terminalOutcome === "incomplete" ? record.terminalFailureReason ?? `run ended ${terminalOutcome}` : null;
-  return {
-    schema: ROLLOUT_SCHEMA,
-    rollout_id: record.runId,
-    parent_rollout_id: null,
-    run_id: record.runId,
-    experiment_id: record.experimentId,
-    candidate_id: record.candidateId,
-    generation: null,
-    candidate_index: null,
-    role: options.role ?? "agent",
-    task: {
-      suite: options.suite ?? record.experimentId,
-      instance_id: record.scenarioId,
-      split: SPLIT_FROM_TAG[record.splitTag],
-      seed: record.seed,
-      rep: 0
-    },
-    policy: {
-      harness: null,
-      harness_version: null,
-      model: record.model,
-      provider: null,
-      profile_commit: record.commitSha,
-      prompt_hash: record.promptHash,
-      config_hash: record.configHash,
-      agent_profile_cell_id: record.agentProfile?.cellId ?? null,
-      sampling: null
-    },
-    messages,
-    tool_defs: [],
-    ...steps.length > 0 ? { steps } : {},
-    outcome: {
-      reward,
-      reward_source: source,
-      verdict: null,
-      metrics: { ...record.outcome.raw },
-      is_completed: isCompleted,
-      is_truncated: isTruncated,
-      error: terminalError,
-      realness_gated: gated
-    },
-    cost: {
-      usd: uncaptured ? null : record.costUsd,
-      tokens_in: record.tokenUsage.input,
-      tokens_out: record.tokenUsage.output,
-      tokens_reasoning: record.tokenUsage.reasoning ?? null,
-      cache_read: record.tokenUsage.cached ?? null,
-      cache_write: record.tokenUsage.cacheWrite ?? null,
-      wall_s: Math.round(record.wallMs / 1e3)
-    },
-    artifacts: { patch_path: null, run_dir: null, transcript_ref: null },
-    provenance: {
-      captured_at: capturedAt,
-      capture: "mint",
-      ...gap !== void 0 ? { gap } : {}
-    }
-  };
-}
-async function mintRolloutRows(records, store, options = {}) {
-  const scrub = options.scrub ?? ((t) => t);
-  const capturedAt = (options.now?.() ?? /* @__PURE__ */ new Date()).toISOString();
-  const rows = [];
-  const missingTraces = [];
-  for (const record of records) {
-    const trajectory = await buildTrajectory(store, record.runId);
-    if (trajectory.steps.length === 0) {
-      missingTraces.push(record.runId);
-      rows.push(
-        mintLine(record, [], [], options, capturedAt, "no trace spans recorded for this runId")
-      );
-      continue;
-    }
-    let steps = trajectory.steps.map((s) => projectStep(s.span, scrub));
-    if (options.maxSteps !== void 0 && steps.length > options.maxSteps) {
-      const head = Math.ceil(options.maxSteps / 2);
-      const tail = options.maxSteps - head;
-      steps = [...steps.slice(0, head), ...steps.slice(steps.length - tail)];
-    }
-    const conversation = finalConversation(
-      trajectory.steps.map((s) => s.span),
-      scrub
-    );
-    const gap = conversation.length === 0 ? "trace has no llm spans \u2014 no conversation to inline" : void 0;
-    rows.push(mintLine(record, steps, conversation, options, capturedAt, gap));
-  }
-  return { rows, missingTraces };
-}
-
-export {
-  rolloutReward,
-  mintRolloutRows
-};
-//# sourceMappingURL=chunk-IHQDPH7D.js.map

package/dist/chunk-IHQDPH7D.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/rollout/mint.ts"],"sourcesContent":["/**\n * Rollout minting — `tangle.rollout.v1` lines joined from the records the\n * substrate ALREADY keeps. There is no separate rollout store: a rollout\n * is the JOIN of a RunRecord (identity, provenance, cost, outcome) with\n * its trace (spans share `runId`), projected into the canonical line.\n *\n * Composition, not duplication:\n *   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)\n *   - step structure          → `buildTrajectory` over the shared TraceStore\n *   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)\n *   - PRM / reward-model      → `reward-model-export.ts`\n *\n * Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true\n * is never exported with a positive reward — the gate travels into the\n * training data (`reward` forced to 0, `realness_gated: true`), so a\n * fine-tune cannot learn from gamed successes.\n *\n * Records without spans become labeled GAP LINES (messages: [],\n * provenance.gap) — present in the output AND surfaced in\n * `missingTraces`; a capture gap is a finding, never a silent omission.\n */\n\nimport { ValidationError } from '../errors'\nimport { type RunRecord, runTaskScore } from '../run-record'\nimport type { LlmSpan, Message, Span, ToolSpan } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport { buildTrajectory } from '../trajectory'\nimport {\n  type ChatMessage,\n  ROLLOUT_SCHEMA,\n  type RolloutLine,\n  type RolloutRole,\n  type RolloutSplit,\n  type RolloutStep,\n} from './schema'\n\n/** Redactor applied to every exported string (secrets, PII). Identity by default. */\nexport type RolloutScrubber = (text: string) => string\n\nexport interface MintRolloutOptions {\n  scrub?: RolloutScrubber\n  /** Cap steps per line (longest runs first drop middle steps). Default: no cap. */\n  maxSteps?: number\n  /** Role recorded on every minted line. Default 'agent' (a solo eval run). */\n  role?: RolloutRole\n  /** Task suite label. Default: the record's `experimentId`. */\n  suite?: string\n  /** Injected clock for deterministic output. */\n  now?: () => Date\n}\n\nexport interface MintRolloutResult {\n  rows: RolloutLine[]\n  /** runIds that had a RunRecord but no spans — emitted as gap lines AND listed here. */\n  missingTraces: string[]\n}\n\nconst asText = (v: unknown, scrub: RolloutScrubber): string => {\n  const s = typeof v === 'string' ? v : JSON.stringify(v)\n  return scrub(s ?? '')\n}\n\nfunction projectStep(span: Span, scrub: RolloutScrubber): RolloutStep {\n  const base: RolloutStep = {\n    kind: span.kind,\n    name: scrub(span.name),\n    status: span.status,\n    durationMs: span.endedAt !== undefined ? span.endedAt - span.startedAt : undefined,\n  }\n  if (span.kind === 'llm') {\n    const llm = span as LlmSpan\n    const last = llm.messages[llm.messages.length - 1]\n    if (last) base.input = scrub(last.content)\n    if (llm.output !== undefined) base.output = scrub(llm.output)\n  } else if (span.kind === 'tool') {\n    const tool = span as ToolSpan\n    base.input = asText(tool.args, scrub)\n    if (tool.result !== undefined) base.output = asText(tool.result, scrub)\n  }\n  return base\n}\n\n/** The final llm span's history + output is the completed conversation. */\nfunction finalConversation(spans: Span[], scrub: RolloutScrubber): ChatMessage[] {\n  const llms = spans.filter((s): s is LlmSpan => s.kind === 'llm')\n  const last = llms[llms.length - 1]\n  if (!last) return []\n  const messages: ChatMessage[] = last.messages.map((m: Message) => ({\n    role: m.role,\n    content: scrub(m.content),\n  }))\n  if (last.output !== undefined && last.output !== '') {\n    messages.push({ role: 'assistant', content: scrub(last.output) })\n  }\n  return messages\n}\n\nfunction scoredReward(record: RunRecord): {\n  reward: number\n  gated: boolean\n  source: 'run-record/holdout-score' | 'run-record/search-score'\n} {\n  const score = runTaskScore(record)\n  if (score === undefined) {\n    throw new ValidationError(`Cannot mint rollout for run ${record.runId}: task score is missing`)\n  }\n  const gated = record.outcome.realness?.gated === true\n  return {\n    reward: gated ? 0 : score,\n    gated,\n    source:\n      record.outcome.holdoutScore !== undefined\n        ? 'run-record/holdout-score'\n        : 'run-record/search-score',\n  }\n}\n\nexport function rolloutReward(record: RunRecord): { reward: number; gated: boolean } {\n  const { reward, gated } = scoredReward(record)\n  return { reward, gated }\n}\n\nconst SPLIT_FROM_TAG: Record<RunRecord['splitTag'], RolloutSplit> = {\n  search: 'search',\n  dev: 'dev',\n  holdout: 'holdout',\n}\n\nfunction mintLine(\n  record: RunRecord,\n  steps: RolloutStep[],\n  messages: ChatMessage[],\n  options: MintRolloutOptions,\n  capturedAt: string,\n  gap?: string,\n): RolloutLine {\n  const { reward, gated, source } = scoredReward(record)\n  const uncaptured = record.costProvenance.kind === 'uncaptured'\n  const terminalOutcome = record.terminalOutcome\n  const isCompleted = terminalOutcome === 'succeeded' || terminalOutcome === 'failed'\n  const isTruncated = terminalOutcome === 'cancelled' || terminalOutcome === 'incomplete'\n  const terminalError =\n    terminalOutcome === 'failed' ||\n    terminalOutcome === 'cancelled' ||\n    terminalOutcome === 'incomplete'\n      ? (record.terminalFailureReason ?? `run ended ${terminalOutcome}`)\n      : null\n  return {\n    schema: ROLLOUT_SCHEMA,\n    rollout_id: record.runId,\n    parent_rollout_id: null,\n    run_id: record.runId,\n    experiment_id: record.experimentId,\n    candidate_id: record.candidateId,\n    generation: null,\n    candidate_index: null,\n    role: options.role ?? 'agent',\n    task: {\n      suite: options.suite ?? record.experimentId,\n      instance_id: record.scenarioId,\n      split: SPLIT_FROM_TAG[record.splitTag],\n      seed: record.seed,\n      rep: 0,\n    },\n    policy: {\n      harness: null,\n      harness_version: null,\n      model: record.model,\n      provider: null,\n      profile_commit: record.commitSha,\n      prompt_hash: record.promptHash,\n      config_hash: record.configHash,\n      agent_profile_cell_id: record.agentProfile?.cellId ?? null,\n      sampling: null,\n    },\n    messages,\n    tool_defs: [],\n    ...(steps.length > 0 ? { steps } : {}),\n    outcome: {\n      reward,\n      reward_source: source,\n      verdict: null,\n      metrics: { ...record.outcome.raw },\n      is_completed: isCompleted,\n      is_truncated: isTruncated,\n      error: terminalError,\n      realness_gated: gated,\n    },\n    cost: {\n      usd: uncaptured ? null : record.costUsd,\n      tokens_in: record.tokenUsage.input,\n      tokens_out: record.tokenUsage.output,\n      tokens_reasoning: record.tokenUsage.reasoning ?? null,\n      cache_read: record.tokenUsage.cached ?? null,\n      cache_write: record.tokenUsage.cacheWrite ?? null,\n      wall_s: Math.round(record.wallMs / 1000),\n    },\n    artifacts: { patch_path: null, run_dir: null, transcript_ref: null },\n    provenance: {\n      captured_at: capturedAt,\n      capture: 'mint',\n      ...(gap !== undefined ? { gap } : {}),\n    },\n  }\n}\n\n/**\n * Join RunRecords with their traces into canonical rollout lines. Records\n * without spans are emitted as labeled gap lines and reported in\n * `missingTraces`. Execution-only records without a task score are rejected\n * because a missing training label is not a zero reward.\n */\nexport async function mintRolloutRows(\n  records: RunRecord[],\n  store: TraceStore,\n  options: MintRolloutOptions = {},\n): Promise<MintRolloutResult> {\n  const scrub = options.scrub ?? ((t) => t)\n  const capturedAt = (options.now?.() ?? new Date()).toISOString()\n  const rows: RolloutLine[] = []\n  const missingTraces: string[] = []\n  for (const record of records) {\n    const trajectory = await buildTrajectory(store, record.runId)\n    if (trajectory.steps.length === 0) {\n      missingTraces.push(record.runId)\n      rows.push(\n        mintLine(record, [], [], options, capturedAt, 'no trace spans recorded for this runId'),\n      )\n      continue\n    }\n    let steps = trajectory.steps.map((s) => projectStep(s.span, scrub))\n    if (options.maxSteps !== undefined && steps.length > options.maxSteps) {\n      // Keep the head and tail — the middle of a long run is the least\n      // informative for outcome attribution.\n      const head = Math.ceil(options.maxSteps / 2)\n      const tail = options.maxSteps - head\n      steps = [...steps.slice(0, head), ...steps.slice(steps.length - tail)]\n    }\n    const conversation = finalConversation(\n      trajectory.steps.map((s) => s.span),\n      scrub,\n    )\n    const gap =\n      conversation.length === 0 ? 'trace has no llm spans — no conversation to inline' : undefined\n    rows.push(mintLine(record, steps, conversation, options, capturedAt, gap))\n  }\n  return { rows, missingTraces }\n}\n"],"mappings":";;;;;;;;;;;;;;AAyDA,IAAM,SAAS,CAAC,GAAY,UAAmC;AAC7D,QAAM,IAAI,OAAO,MAAM,WAAW,IAAI,KAAK,UAAU,CAAC;AACtD,SAAO,MAAM,KAAK,EAAE;AACtB;AAEA,SAAS,YAAY,MAAY,OAAqC;AACpE,QAAM,OAAoB;AAAA,IACxB,MAAM,KAAK;AAAA,IACX,MAAM,MAAM,KAAK,IAAI;AAAA,IACrB,QAAQ,KAAK;AAAA,IACb,YAAY,KAAK,YAAY,SAAY,KAAK,UAAU,KAAK,YAAY;AAAA,EAC3E;AACA,MAAI,KAAK,SAAS,OAAO;AACvB,UAAM,MAAM;AACZ,UAAM,OAAO,IAAI,SAAS,IAAI,SAAS,SAAS,CAAC;AACjD,QAAI,KAAM,MAAK,QAAQ,MAAM,KAAK,OAAO;AACzC,QAAI,IAAI,WAAW,OAAW,MAAK,SAAS,MAAM,IAAI,MAAM;AAAA,EAC9D,WAAW,KAAK,SAAS,QAAQ;AAC/B,UAAM,OAAO;AACb,SAAK,QAAQ,OAAO,KAAK,MAAM,KAAK;AACpC,QAAI,KAAK,WAAW,OAAW,MAAK,SAAS,OAAO,KAAK,QAAQ,KAAK;AAAA,EACxE;AACA,SAAO;AACT;AAGA,SAAS,kBAAkB,OAAe,OAAuC;AAC/E,QAAM,OAAO,MAAM,OAAO,CAAC,MAAoB,EAAE,SAAS,KAAK;AAC/D,QAAM,OAAO,KAAK,KAAK,SAAS,CAAC;AACjC,MAAI,CAAC,KAAM,QAAO,CAAC;AACnB,QAAM,WAA0B,KAAK,SAAS,IAAI,CAAC,OAAgB;AAAA,IACjE,MAAM,EAAE;AAAA,IACR,SAAS,MAAM,EAAE,OAAO;AAAA,EAC1B,EAAE;AACF,MAAI,KAAK,WAAW,UAAa,KAAK,WAAW,IAAI;AACnD,aAAS,KAAK,EAAE,MAAM,aAAa,SAAS,MAAM,KAAK,MAAM,EAAE,CAAC;AAAA,EAClE;AACA,SAAO;AACT;AAEA,SAAS,aAAa,QAIpB;AACA,QAAM,QAAQ,aAAa,MAAM;AACjC,MAAI,UAAU,QAAW;AACvB,UAAM,IAAI,gBAAgB,+BAA+B,OAAO,KAAK,yBAAyB;AAAA,EAChG;AACA,QAAM,QAAQ,OAAO,QAAQ,UAAU,UAAU;AACjD,SAAO;AAAA,IACL,QAAQ,QAAQ,IAAI;AAAA,IACpB;AAAA,IACA,QACE,OAAO,QAAQ,iBAAiB,SAC5B,6BACA;AAAA,EACR;AACF;AAEO,SAAS,cAAc,QAAuD;AACnF,QAAM,EAAE,QAAQ,MAAM,IAAI,aAAa,MAAM;AAC7C,SAAO,EAAE,QAAQ,MAAM;AACzB;AAEA,IAAM,iBAA8D;AAAA,EAClE,QAAQ;AAAA,EACR,KAAK;AAAA,EACL,SAAS;AACX;AAEA,SAAS,SACP,QACA,OACA,UACA,SACA,YACA,KACa;AACb,QAAM,EAAE,QAAQ,OAAO,OAAO,IAAI,aAAa,MAAM;AACrD,QAAM,aAAa,OAAO,eAAe,SAAS;AAClD,QAAM,kBAAkB,OAAO;AAC/B,QAAM,cAAc,oBAAoB,eAAe,oBAAoB;AAC3E,QAAM,cAAc,oBAAoB,eAAe,oBAAoB;AAC3E,QAAM,gBACJ,oBAAoB,YACpB,oBAAoB,eACpB,oBAAoB,eACf,OAAO,yBAAyB,aAAa,eAAe,KAC7D;AACN,SAAO;AAAA,IACL,QAAQ;AAAA,IACR,YAAY,OAAO;AAAA,IACnB,mBAAmB;AAAA,IACnB,QAAQ,OAAO;AAAA,IACf,eAAe,OAAO;AAAA,IACtB,cAAc,OAAO;AAAA,IACrB,YAAY;AAAA,IACZ,iBAAiB;AAAA,IACjB,MAAM,QAAQ,QAAQ;AAAA,IACtB,MAAM;AAAA,MACJ,OAAO,QAAQ,SAAS,OAAO;AAAA,MAC/B,aAAa,OAAO;AAAA,MACpB,OAAO,eAAe,OAAO,QAAQ;AAAA,MACrC,MAAM,OAAO;AAAA,MACb,KAAK;AAAA,IACP;AAAA,IACA,QAAQ;AAAA,MACN,SAAS;AAAA,MACT,iBAAiB;AAAA,MACjB,OAAO,OAAO;AAAA,MACd,UAAU;AAAA,MACV,gBAAgB,OAAO;AAAA,MACvB,aAAa,OAAO;AAAA,MACpB,aAAa,OAAO;AAAA,MACpB,uBAAuB,OAAO,cAAc,UAAU;AAAA,MACtD,UAAU;AAAA,IACZ;AAAA,IACA;AAAA,IACA,WAAW,CAAC;AAAA,IACZ,GAAI,MAAM,SAAS,IAAI,EAAE,MAAM,IAAI,CAAC;AAAA,IACpC,SAAS;AAAA,MACP;AAAA,MACA,eAAe;AAAA,MACf,SAAS;AAAA,MACT,SAAS,EAAE,GAAG,OAAO,QAAQ,IAAI;AAAA,MACjC,cAAc;AAAA,MACd,cAAc;AAAA,MACd,OAAO;AAAA,MACP,gBAAgB;AAAA,IAClB;AAAA,IACA,MAAM;AAAA,MACJ,KAAK,aAAa,OAAO,OAAO;AAAA,MAChC,WAAW,OAAO,WAAW;AAAA,MAC7B,YAAY,OAAO,WAAW;AAAA,MAC9B,kBAAkB,OAAO,WAAW,aAAa;AAAA,MACjD,YAAY,OAAO,WAAW,UAAU;AAAA,MACxC,aAAa,OAAO,WAAW,cAAc;AAAA,MAC7C,QAAQ,KAAK,MAAM,OAAO,SAAS,GAAI;AAAA,IACzC;AAAA,IACA,WAAW,EAAE,YAAY,MAAM,SAAS,MAAM,gBAAgB,KAAK;AAAA,IACnE,YAAY;AAAA,MACV,aAAa;AAAA,MACb,SAAS;AAAA,MACT,GAAI,QAAQ,SAAY,EAAE,IAAI,IAAI,CAAC;AAAA,IACrC;AAAA,EACF;AACF;AAQA,eAAsB,gBACpB,SACA,OACA,UAA8B,CAAC,GACH;AAC5B,QAAM,QAAQ,QAAQ,UAAU,CAAC,MAAM;AACvC,QAAM,cAAc,QAAQ,MAAM,KAAK,oBAAI,KAAK,GAAG,YAAY;AAC/D,QAAM,OAAsB,CAAC;AAC7B,QAAM,gBAA0B,CAAC;AACjC,aAAW,UAAU,SAAS;AAC5B,UAAM,aAAa,MAAM,gBAAgB,OAAO,OAAO,KAAK;AAC5D,QAAI,WAAW,MAAM,WAAW,GAAG;AACjC,oBAAc,KAAK,OAAO,KAAK;AAC/B,WAAK;AAAA,QACH,SAAS,QAAQ,CAAC,GAAG,CAAC,GAAG,SAAS,YAAY,wCAAwC;AAAA,MACxF;AACA;AAAA,IACF;AACA,QAAI,QAAQ,WAAW,MAAM,IAAI,CAAC,MAAM,YAAY,EAAE,MAAM,KAAK,CAAC;AAClE,QAAI,QAAQ,aAAa,UAAa,MAAM,SAAS,QAAQ,UAAU;AAGrE,YAAM,OAAO,KAAK,KAAK,QAAQ,WAAW,CAAC;AAC3C,YAAM,OAAO,QAAQ,WAAW;AAChC,cAAQ,CAAC,GAAG,MAAM,MAAM,GAAG,IAAI,GAAG,GAAG,MAAM,MAAM,MAAM,SAAS,IAAI,CAAC;AAAA,IACvE;AACA,UAAM,eAAe;AAAA,MACnB,WAAW,MAAM,IAAI,CAAC,MAAM,EAAE,IAAI;AAAA,MAClC;AAAA,IACF;AACA,UAAM,MACJ,aAAa,WAAW,IAAI,4DAAuD;AACrF,SAAK,KAAK,SAAS,QAAQ,OAAO,cAAc,SAAS,YAAY,GAAG,CAAC;AAAA,EAC3E;AACA,SAAO,EAAE,MAAM,cAAc;AAC/B;","names":[]}

package/dist/chunk-JHCHEVET.js

@@ -1,274 +0,0 @@
-// src/matrix/aggregation.ts
-function flattenRuns(cells) {
-  const rows = [];
-  for (const { cell, runs } of cells) {
-    for (const result of runs) rows.push({ cell, result });
-  }
-  return rows;
-}
-function quantile(sorted, q) {
-  if (sorted.length === 0) return 0;
-  if (sorted.length === 1) return sorted[0];
-  const pos = (sorted.length - 1) * q;
-  const lo = Math.floor(pos);
-  const hi = Math.ceil(pos);
-  if (lo === hi) return sorted[lo];
-  const frac = pos - lo;
-  return sorted[lo] * (1 - frac) + sorted[hi] * frac;
-}
-function summariseRows(rows, axisName, axisValue) {
-  if (rows.length === 0) {
-    return {
-      axisName,
-      axisValue,
-      cells: 0,
-      passRate: 0,
-      meanScore: 0,
-      p50Score: 0,
-      p90Score: 0,
-      totalCostUsd: 0,
-      meanDurationMs: 0
-    };
-  }
-  let pass = 0;
-  let scoreSum = 0;
-  let costSum = 0;
-  let durSum = 0;
-  const scores = [];
-  for (const { result } of rows) {
-    const errored = result.error !== void 0;
-    const score = errored ? 0 : result.verdict.score;
-    const valid = !errored && result.verdict.valid;
-    if (valid) pass++;
-    scoreSum += score;
-    scores.push(score);
-    costSum += result.costUsd;
-    durSum += result.durationMs;
-  }
-  scores.sort((a, b) => a - b);
-  return {
-    axisName,
-    axisValue,
-    cells: rows.length,
-    passRate: pass / rows.length,
-    meanScore: scoreSum / rows.length,
-    p50Score: quantile(scores, 0.5),
-    p90Score: quantile(scores, 0.9),
-    totalCostUsd: costSum,
-    meanDurationMs: durSum / rows.length
-  };
-}
-function bucketBy(rows, axisName, labelFor) {
-  const buckets = /* @__PURE__ */ new Map();
-  for (const row of rows) {
-    const slot = row.cell.axes[axisName];
-    if (!slot) continue;
-    const id = slot.id;
-    let arr = buckets.get(id);
-    if (!arr) {
-      arr = [];
-      buckets.set(id, arr);
-    }
-    arr.push(row);
-  }
-  const out = {};
-  for (const id of [...buckets.keys()].sort()) {
-    out[id] = summariseRows(buckets.get(id), axisName, labelFor(id));
-  }
-  return out;
-}
-function buildByAxis(cells, axes, aggregateBy) {
-  const rows = flattenRuns(cells);
-  const byName = new Map(axes.map((a) => [a.name, a]));
-  const byAxis = {};
-  for (const name of aggregateBy) {
-    const axis = byName.get(name);
-    const labelFor = (id) => {
-      if (!axis?.label) return id;
-      const found = axis.values.find((v) => v.id === id);
-      if (!found) return id;
-      return axis.label(found.value, id);
-    };
-    byAxis[name] = bucketBy(rows, name, labelFor);
-  }
-  return byAxis;
-}
-
-// src/matrix/runner.ts
-function cartesian(axes) {
-  if (axes.length === 0) return [{ axes: {} }];
-  for (const a of axes) if (a.values.length === 0) return [];
-  const out = [];
-  const idx = new Array(axes.length).fill(0);
-  while (true) {
-    const slot = {};
-    for (let i2 = 0; i2 < axes.length; i2++) {
-      const axis = axes[i2];
-      const v = axis.values[idx[i2]];
-      slot[axis.name] = { id: v.id, value: v.value };
-    }
-    out.push({ axes: slot });
-    let i = 0;
-    while (i < axes.length) {
-      const next = idx[i] + 1;
-      const axis = axes[i];
-      if (next < axis.values.length) {
-        idx[i] = next;
-        break;
-      }
-      idx[i] = 0;
-      i++;
-    }
-    if (i === axes.length) break;
-  }
-  return out;
-}
-function makeMatrixId() {
-  const t = Date.now().toString(36);
-  let r = "";
-  for (let i = 0; i < 8; i++) r += Math.floor(Math.random() * 16).toString(16);
-  return `mtx_${t}_${r}`;
-}
-function makeErrorResult(err) {
-  const e = err;
-  return {
-    output: void 0,
-    verdict: { valid: false, score: 0 },
-    costUsd: 0,
-    durationMs: 0,
-    error: {
-      message: typeof e?.message === "string" ? e.message : String(err),
-      kind: typeof e?.name === "string" ? e.name : "Error"
-    }
-  };
-}
-async function runAgentMatrix(opts) {
-  const startedAt = Date.now();
-  const reps = Math.max(1, opts.reps ?? 1);
-  const maxConcurrency = Math.max(1, opts.maxConcurrency ?? 4);
-  const costCeiling = opts.costCeiling ?? Number.POSITIVE_INFINITY;
-  const aggregateBy = opts.aggregateBy ?? opts.axes.map((a) => a.name);
-  const base = cartesian(opts.axes);
-  const filtered = opts.filter ? base.filter((c) => opts.filter(c)) : base;
-  const filteredOut = base.length - filtered.length;
-  const planned = [];
-  for (let i = 0; i < filtered.length; i++) {
-    for (let r = 0; r < reps; r++) {
-      planned.push({
-        axes: filtered[i].axes,
-        rep: r,
-        ordinal: i * reps + r
-      });
-    }
-  }
-  const cellRecords = [];
-  let cumulativeCost = 0;
-  let costCeilingReached = false;
-  let runsExecuted = 0;
-  let cellsUnscheduled = 0;
-  const aborted = () => opts.signal?.aborted === true;
-  let inFlight = 0;
-  let cursor = 0;
-  let resolveAll;
-  const done = new Promise((res) => {
-    resolveAll = res;
-  });
-  const pump = () => {
-    while (inFlight < maxConcurrency && cursor < planned.length) {
-      if (aborted() || costCeilingReached) {
-        const left = planned.length - cursor;
-        cellsUnscheduled += left;
-        cursor = planned.length;
-        break;
-      }
-      const cell = planned[cursor++];
-      inFlight++;
-      const record = { cell, runs: [] };
-      cellRecords.push(record);
-      const promise = (async () => {
-        try {
-          return await opts.runCell(cell);
-        } catch (err) {
-          return makeErrorResult(err);
-        }
-      })();
-      promise.then((result) => {
-        record.runs.push(result);
-        runsExecuted++;
-        cumulativeCost += result.costUsd;
-        if (cumulativeCost >= costCeiling && !costCeilingReached) {
-          costCeilingReached = true;
-          console.warn("[matrix] cost ceiling reached");
-        }
-        try {
-          opts.onCellComplete?.(cell, result);
-        } catch {
-        }
-        inFlight--;
-        if (cursor < planned.length) {
-          pump();
-        } else if (inFlight === 0) {
-          resolveAll?.();
-        }
-      });
-    }
-    if (cursor >= planned.length && inFlight === 0) resolveAll?.();
-  };
-  const onAbort = () => {
-    if (cursor < planned.length) {
-      cellsUnscheduled += planned.length - cursor;
-      cursor = planned.length;
-    }
-    if (inFlight === 0) resolveAll?.();
-  };
-  if (opts.signal) {
-    if (opts.signal.aborted) {
-      cellsUnscheduled = planned.length;
-      cursor = planned.length;
-      resolveAll?.();
-    } else {
-      opts.signal.addEventListener("abort", onAbort, { once: true });
-    }
-  }
-  if (planned.length === 0) {
-    resolveAll?.();
-  } else {
-    pump();
-  }
-  await done;
-  if (opts.signal) opts.signal.removeEventListener("abort", onAbort);
-  cellRecords.sort((a, b) => a.cell.ordinal - b.cell.ordinal);
-  let pass = 0;
-  let scoreSum = 0;
-  let totalCost = 0;
-  let runCount = 0;
-  for (const { runs } of cellRecords) {
-    for (const r of runs) {
-      runCount++;
-      const errored = r.error !== void 0;
-      if (!errored && r.verdict.valid) pass++;
-      scoreSum += errored ? 0 : r.verdict.score;
-      totalCost += r.costUsd;
-    }
-  }
-  const byAxis = buildByAxis(cellRecords, opts.axes, aggregateBy);
-  return {
-    cells: cellRecords,
-    byAxis,
-    summary: {
-      totalCells: planned.length,
-      runsExecuted,
-      cellsSkipped: cellsUnscheduled + filteredOut * reps,
-      overallPassRate: runCount === 0 ? 0 : pass / runCount,
-      overallMeanScore: runCount === 0 ? 0 : scoreSum / runCount,
-      totalCostUsd: totalCost,
-      durationMs: Date.now() - startedAt
-    },
-    matrixId: makeMatrixId()
-  };
-}
-
-export {
-  runAgentMatrix
-};
-//# sourceMappingURL=chunk-JHCHEVET.js.map

package/dist/chunk-JHCHEVET.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/matrix/aggregation.ts","../src/matrix/runner.ts"],"sourcesContent":["/**\n * Per-axis aggregation of cell runs into `AxisSummary` rows.\n *\n * Pure: consumes the final `cells: [{cell, runs}]` array and returns the\n * `byAxis` table. Error runs contribute 0 to passRate and meanScore. Cost\n * and duration always count — the budget was spent regardless.\n */\n\nimport type { AxisSummary, CellResult, MatrixAxis, MatrixCell, MatrixResult } from './types'\n\ninterface Row<Output> {\n  cell: MatrixCell\n  result: CellResult<Output>\n}\n\nfunction flattenRuns<Output>(cells: MatrixResult<Output>['cells']): Row<Output>[] {\n  const rows: Row<Output>[] = []\n  for (const { cell, runs } of cells) {\n    for (const result of runs) rows.push({ cell, result })\n  }\n  return rows\n}\n\nfunction quantile(sorted: number[], q: number): number {\n  if (sorted.length === 0) return 0\n  if (sorted.length === 1) return sorted[0] as number\n  const pos = (sorted.length - 1) * q\n  const lo = Math.floor(pos)\n  const hi = Math.ceil(pos)\n  if (lo === hi) return sorted[lo] as number\n  const frac = pos - lo\n  return (sorted[lo] as number) * (1 - frac) + (sorted[hi] as number) * frac\n}\n\nexport function summariseRows<Output>(\n  rows: Row<Output>[],\n  axisName: string,\n  axisValue: string,\n): AxisSummary {\n  if (rows.length === 0) {\n    return {\n      axisName,\n      axisValue,\n      cells: 0,\n      passRate: 0,\n      meanScore: 0,\n      p50Score: 0,\n      p90Score: 0,\n      totalCostUsd: 0,\n      meanDurationMs: 0,\n    }\n  }\n  let pass = 0\n  let scoreSum = 0\n  let costSum = 0\n  let durSum = 0\n  const scores: number[] = []\n  for (const { result } of rows) {\n    const errored = result.error !== undefined\n    const score = errored ? 0 : result.verdict.score\n    const valid = !errored && result.verdict.valid\n    if (valid) pass++\n    scoreSum += score\n    scores.push(score)\n    costSum += result.costUsd\n    durSum += result.durationMs\n  }\n  scores.sort((a, b) => a - b)\n  return {\n    axisName,\n    axisValue,\n    cells: rows.length,\n    passRate: pass / rows.length,\n    meanScore: scoreSum / rows.length,\n    p50Score: quantile(scores, 0.5),\n    p90Score: quantile(scores, 0.9),\n    totalCostUsd: costSum,\n    meanDurationMs: durSum / rows.length,\n  }\n}\n\nfunction bucketBy<Output>(\n  rows: Row<Output>[],\n  axisName: string,\n  labelFor: (id: string) => string,\n): Record<string, AxisSummary> {\n  const buckets = new Map<string, Row<Output>[]>()\n  for (const row of rows) {\n    const slot = row.cell.axes[axisName]\n    if (!slot) continue\n    const id = slot.id\n    let arr = buckets.get(id)\n    if (!arr) {\n      arr = []\n      buckets.set(id, arr)\n    }\n    arr.push(row)\n  }\n  const out: Record<string, AxisSummary> = {}\n  // Sorted keys for deterministic JSON serialisation.\n  for (const id of [...buckets.keys()].sort()) {\n    out[id] = summariseRows(buckets.get(id) as Row<Output>[], axisName, labelFor(id))\n  }\n  return out\n}\n\nexport function buildByAxis<Output>(\n  cells: MatrixResult<Output>['cells'],\n  axes: MatrixAxis<unknown>[],\n  aggregateBy: string[],\n): Record<string, Record<string, AxisSummary>> {\n  const rows = flattenRuns(cells)\n  const byName = new Map(axes.map((a) => [a.name, a]))\n  const byAxis: Record<string, Record<string, AxisSummary>> = {}\n  for (const name of aggregateBy) {\n    const axis = byName.get(name)\n    const labelFor = (id: string): string => {\n      if (!axis?.label) return id\n      const found = axis.values.find((v) => v.id === id)\n      if (!found) return id\n      return axis.label(found.value, id)\n    }\n    byAxis[name] = bucketBy(rows, name, labelFor)\n  }\n  return byAxis\n}\n","/**\n * N-axis cartesian runner.\n *\n * Expansion order: cartesian over `axes` in declared order, then `reps` as the\n * inner-most dim → `ordinal = (cartIdx * reps) + rep`. The returned\n * `cells[]` is sorted by `ordinal` so concurrent execution does not reorder\n * the output.\n *\n * Scheduling is a sliding window of in-flight promises capped at\n * `maxConcurrency`. The window stops admitting new cells when the cost\n * ceiling trips or the abort signal fires; in-flight cells finish.\n */\n\nimport { buildByAxis } from './aggregation'\nimport type {\n  CellResult,\n  MatrixAxis,\n  MatrixCell,\n  MatrixResult,\n  RunAgentMatrixOptions,\n} from './types'\n\ninterface BaseCell {\n  axes: Record<string, { id: string; value: unknown }>\n}\n\nfunction cartesian(axes: MatrixAxis<unknown>[]): BaseCell[] {\n  // Empty axes (`values=[]`) collapse the whole product to zero cells. An\n  // empty `axes` array yields a single empty-axes cell — degenerate but\n  // valid (caller is iterating only reps).\n  if (axes.length === 0) return [{ axes: {} }]\n  for (const a of axes) if (a.values.length === 0) return []\n  const out: BaseCell[] = []\n  const idx = new Array(axes.length).fill(0)\n  while (true) {\n    const slot: Record<string, { id: string; value: unknown }> = {}\n    for (let i = 0; i < axes.length; i++) {\n      const axis = axes[i] as MatrixAxis<unknown>\n      const v = axis.values[idx[i] as number] as { id: string; value: unknown }\n      slot[axis.name] = { id: v.id, value: v.value }\n    }\n    out.push({ axes: slot })\n    // Increment like an odometer, left-most axis is fastest.\n    let i = 0\n    while (i < axes.length) {\n      const next = (idx[i] as number) + 1\n      const axis = axes[i] as MatrixAxis<unknown>\n      if (next < axis.values.length) {\n        idx[i] = next\n        break\n      }\n      idx[i] = 0\n      i++\n    }\n    if (i === axes.length) break\n  }\n  return out\n}\n\nfunction makeMatrixId(): string {\n  // Stable id-like string: time + 8 random hex chars. Avoids node:crypto\n  // import to keep the matrix dep-free.\n  const t = Date.now().toString(36)\n  let r = ''\n  for (let i = 0; i < 8; i++) r += Math.floor(Math.random() * 16).toString(16)\n  return `mtx_${t}_${r}`\n}\n\nfunction makeErrorResult<Output>(err: unknown): CellResult<Output> {\n  const e = err as { message?: string; name?: string }\n  return {\n    output: undefined as unknown as Output,\n    verdict: { valid: false, score: 0 },\n    costUsd: 0,\n    durationMs: 0,\n    error: {\n      message: typeof e?.message === 'string' ? e.message : String(err),\n      kind: typeof e?.name === 'string' ? e.name : 'Error',\n    },\n  }\n}\n\nexport async function runAgentMatrix<Output>(\n  opts: RunAgentMatrixOptions<Output>,\n): Promise<MatrixResult<Output>> {\n  const startedAt = Date.now()\n  const reps = Math.max(1, opts.reps ?? 1)\n  const maxConcurrency = Math.max(1, opts.maxConcurrency ?? 4)\n  const costCeiling = opts.costCeiling ?? Number.POSITIVE_INFINITY\n  const aggregateBy = opts.aggregateBy ?? opts.axes.map((a) => a.name)\n\n  const base = cartesian(opts.axes)\n  const filtered = opts.filter\n    ? base.filter((c) => (opts.filter as (b: BaseCell) => boolean)(c))\n    : base\n  const filteredOut = base.length - filtered.length\n\n  const planned: MatrixCell[] = []\n  for (let i = 0; i < filtered.length; i++) {\n    for (let r = 0; r < reps; r++) {\n      planned.push({\n        axes: (filtered[i] as BaseCell).axes,\n        rep: r,\n        ordinal: i * reps + r,\n      })\n    }\n  }\n\n  const cellRecords: Array<{ cell: MatrixCell; runs: CellResult<Output>[] }> = []\n  let cumulativeCost = 0\n  let costCeilingReached = false\n  let runsExecuted = 0\n  let cellsUnscheduled = 0\n\n  const aborted = (): boolean => opts.signal?.aborted === true\n\n  // Per-run abort controller forwards the external signal so cell executors\n  // see cancellation. We don't expose it on `MatrixCell` — the signature on\n  // `runCell` per the public API is `(cell) => Promise<...>`. Executors that\n  // need cancellation use the external signal directly via closure.\n\n  let inFlight = 0\n  let cursor = 0\n  let resolveAll: (() => void) | undefined\n  const done = new Promise<void>((res) => {\n    resolveAll = res\n  })\n\n  const pump = (): void => {\n    while (inFlight < maxConcurrency && cursor < planned.length) {\n      if (aborted() || costCeilingReached) {\n        // Drain remaining as unscheduled.\n        const left = planned.length - cursor\n        cellsUnscheduled += left\n        cursor = planned.length\n        break\n      }\n      const cell = planned[cursor++] as MatrixCell\n      inFlight++\n      // Lazily allocate the record so cells appear in `cells[]` in any\n      // arrival order; we sort by ordinal at the end.\n      const record = { cell, runs: [] as CellResult<Output>[] }\n      cellRecords.push(record)\n      const promise: Promise<CellResult<Output>> = (async () => {\n        try {\n          return await opts.runCell(cell)\n        } catch (err) {\n          return makeErrorResult<Output>(err)\n        }\n      })()\n      promise.then((result) => {\n        record.runs.push(result)\n        runsExecuted++\n        cumulativeCost += result.costUsd\n        if (cumulativeCost >= costCeiling && !costCeilingReached) {\n          costCeilingReached = true\n          // eslint-disable-next-line no-console\n          console.warn('[matrix] cost ceiling reached')\n        }\n        try {\n          opts.onCellComplete?.(cell, result)\n        } catch {\n          // onCellComplete is observational — swallow throws so a noisy\n          // callback can't tank the run.\n        }\n        inFlight--\n        if (cursor < planned.length) {\n          pump()\n        } else if (inFlight === 0) {\n          resolveAll?.()\n        }\n      })\n    }\n    if (cursor >= planned.length && inFlight === 0) resolveAll?.()\n  }\n\n  const onAbort = (): void => {\n    // External abort: stop scheduling. In-flight cells finish; their\n    // executors observe `opts.signal.aborted` directly via closure.\n    if (cursor < planned.length) {\n      cellsUnscheduled += planned.length - cursor\n      cursor = planned.length\n    }\n    if (inFlight === 0) resolveAll?.()\n  }\n  if (opts.signal) {\n    if (opts.signal.aborted) {\n      cellsUnscheduled = planned.length\n      cursor = planned.length\n      resolveAll?.()\n    } else {\n      opts.signal.addEventListener('abort', onAbort, { once: true })\n    }\n  }\n\n  if (planned.length === 0) {\n    resolveAll?.()\n  } else {\n    pump()\n  }\n\n  await done\n  if (opts.signal) opts.signal.removeEventListener('abort', onAbort)\n\n  cellRecords.sort((a, b) => a.cell.ordinal - b.cell.ordinal)\n\n  let pass = 0\n  let scoreSum = 0\n  let totalCost = 0\n  let runCount = 0\n  for (const { runs } of cellRecords) {\n    for (const r of runs) {\n      runCount++\n      const errored = r.error !== undefined\n      if (!errored && r.verdict.valid) pass++\n      scoreSum += errored ? 0 : r.verdict.score\n      totalCost += r.costUsd\n    }\n  }\n\n  const byAxis = buildByAxis(cellRecords, opts.axes, aggregateBy)\n\n  return {\n    cells: cellRecords,\n    byAxis,\n    summary: {\n      totalCells: planned.length,\n      runsExecuted,\n      cellsSkipped: cellsUnscheduled + filteredOut * reps,\n      overallPassRate: runCount === 0 ? 0 : pass / runCount,\n      overallMeanScore: runCount === 0 ? 0 : scoreSum / runCount,\n      totalCostUsd: totalCost,\n      durationMs: Date.now() - startedAt,\n    },\n    matrixId: makeMatrixId(),\n  }\n}\n"],"mappings":";AAeA,SAAS,YAAoB,OAAqD;AAChF,QAAM,OAAsB,CAAC;AAC7B,aAAW,EAAE,MAAM,KAAK,KAAK,OAAO;AAClC,eAAW,UAAU,KAAM,MAAK,KAAK,EAAE,MAAM,OAAO,CAAC;AAAA,EACvD;AACA,SAAO;AACT;AAEA,SAAS,SAAS,QAAkB,GAAmB;AACrD,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,MAAI,OAAO,WAAW,EAAG,QAAO,OAAO,CAAC;AACxC,QAAM,OAAO,OAAO,SAAS,KAAK;AAClC,QAAM,KAAK,KAAK,MAAM,GAAG;AACzB,QAAM,KAAK,KAAK,KAAK,GAAG;AACxB,MAAI,OAAO,GAAI,QAAO,OAAO,EAAE;AAC/B,QAAM,OAAO,MAAM;AACnB,SAAQ,OAAO,EAAE,KAAgB,IAAI,QAAS,OAAO,EAAE,IAAe;AACxE;AAEO,SAAS,cACd,MACA,UACA,WACa;AACb,MAAI,KAAK,WAAW,GAAG;AACrB,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA,OAAO;AAAA,MACP,UAAU;AAAA,MACV,WAAW;AAAA,MACX,UAAU;AAAA,MACV,UAAU;AAAA,MACV,cAAc;AAAA,MACd,gBAAgB;AAAA,IAClB;AAAA,EACF;AACA,MAAI,OAAO;AACX,MAAI,WAAW;AACf,MAAI,UAAU;AACd,MAAI,SAAS;AACb,QAAM,SAAmB,CAAC;AAC1B,aAAW,EAAE,OAAO,KAAK,MAAM;AAC7B,UAAM,UAAU,OAAO,UAAU;AACjC,UAAM,QAAQ,UAAU,IAAI,OAAO,QAAQ;AAC3C,UAAM,QAAQ,CAAC,WAAW,OAAO,QAAQ;AACzC,QAAI,MAAO;AACX,gBAAY;AACZ,WAAO,KAAK,KAAK;AACjB,eAAW,OAAO;AAClB,cAAU,OAAO;AAAA,EACnB;AACA,SAAO,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC3B,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,OAAO,KAAK;AAAA,IACZ,UAAU,OAAO,KAAK;AAAA,IACtB,WAAW,WAAW,KAAK;AAAA,IAC3B,UAAU,SAAS,QAAQ,GAAG;AAAA,IAC9B,UAAU,SAAS,QAAQ,GAAG;AAAA,IAC9B,cAAc;AAAA,IACd,gBAAgB,SAAS,KAAK;AAAA,EAChC;AACF;AAEA,SAAS,SACP,MACA,UACA,UAC6B;AAC7B,QAAM,UAAU,oBAAI,IAA2B;AAC/C,aAAW,OAAO,MAAM;AACtB,UAAM,OAAO,IAAI,KAAK,KAAK,QAAQ;AACnC,QAAI,CAAC,KAAM;AACX,UAAM,KAAK,KAAK;AAChB,QAAI,MAAM,QAAQ,IAAI,EAAE;AACxB,QAAI,CAAC,KAAK;AACR,YAAM,CAAC;AACP,cAAQ,IAAI,IAAI,GAAG;AAAA,IACrB;AACA,QAAI,KAAK,GAAG;AAAA,EACd;AACA,QAAM,MAAmC,CAAC;AAE1C,aAAW,MAAM,CAAC,GAAG,QAAQ,KAAK,CAAC,EAAE,KAAK,GAAG;AAC3C,QAAI,EAAE,IAAI,cAAc,QAAQ,IAAI,EAAE,GAAoB,UAAU,SAAS,EAAE,CAAC;AAAA,EAClF;AACA,SAAO;AACT;AAEO,SAAS,YACd,OACA,MACA,aAC6C;AAC7C,QAAM,OAAO,YAAY,KAAK;AAC9B,QAAM,SAAS,IAAI,IAAI,KAAK,IAAI,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC;AACnD,QAAM,SAAsD,CAAC;AAC7D,aAAW,QAAQ,aAAa;AAC9B,UAAM,OAAO,OAAO,IAAI,IAAI;AAC5B,UAAM,WAAW,CAAC,OAAuB;AACvC,UAAI,CAAC,MAAM,MAAO,QAAO;AACzB,YAAM,QAAQ,KAAK,OAAO,KAAK,CAAC,MAAM,EAAE,OAAO,EAAE;AACjD,UAAI,CAAC,MAAO,QAAO;AACnB,aAAO,KAAK,MAAM,MAAM,OAAO,EAAE;AAAA,IACnC;AACA,WAAO,IAAI,IAAI,SAAS,MAAM,MAAM,QAAQ;AAAA,EAC9C;AACA,SAAO;AACT;;;ACnGA,SAAS,UAAU,MAAyC;AAI1D,MAAI,KAAK,WAAW,EAAG,QAAO,CAAC,EAAE,MAAM,CAAC,EAAE,CAAC;AAC3C,aAAW,KAAK,KAAM,KAAI,EAAE,OAAO,WAAW,EAAG,QAAO,CAAC;AACzD,QAAM,MAAkB,CAAC;AACzB,QAAM,MAAM,IAAI,MAAM,KAAK,MAAM,EAAE,KAAK,CAAC;AACzC,SAAO,MAAM;AACX,UAAM,OAAuD,CAAC;AAC9D,aAASA,KAAI,GAAGA,KAAI,KAAK,QAAQA,MAAK;AACpC,YAAM,OAAO,KAAKA,EAAC;AACnB,YAAM,IAAI,KAAK,OAAO,IAAIA,EAAC,CAAW;AACtC,WAAK,KAAK,IAAI,IAAI,EAAE,IAAI,EAAE,IAAI,OAAO,EAAE,MAAM;AAAA,IAC/C;AACA,QAAI,KAAK,EAAE,MAAM,KAAK,CAAC;AAEvB,QAAI,IAAI;AACR,WAAO,IAAI,KAAK,QAAQ;AACtB,YAAM,OAAQ,IAAI,CAAC,IAAe;AAClC,YAAM,OAAO,KAAK,CAAC;AACnB,UAAI,OAAO,KAAK,OAAO,QAAQ;AAC7B,YAAI,CAAC,IAAI;AACT;AAAA,MACF;AACA,UAAI,CAAC,IAAI;AACT;AAAA,IACF;AACA,QAAI,MAAM,KAAK,OAAQ;AAAA,EACzB;AACA,SAAO;AACT;AAEA,SAAS,eAAuB;AAG9B,QAAM,IAAI,KAAK,IAAI,EAAE,SAAS,EAAE;AAChC,MAAI,IAAI;AACR,WAAS,IAAI,GAAG,IAAI,GAAG,IAAK,MAAK,KAAK,MAAM,KAAK,OAAO,IAAI,EAAE,EAAE,SAAS,EAAE;AAC3E,SAAO,OAAO,CAAC,IAAI,CAAC;AACtB;AAEA,SAAS,gBAAwB,KAAkC;AACjE,QAAM,IAAI;AACV,SAAO;AAAA,IACL,QAAQ;AAAA,IACR,SAAS,EAAE,OAAO,OAAO,OAAO,EAAE;AAAA,IAClC,SAAS;AAAA,IACT,YAAY;AAAA,IACZ,OAAO;AAAA,MACL,SAAS,OAAO,GAAG,YAAY,WAAW,EAAE,UAAU,OAAO,GAAG;AAAA,MAChE,MAAM,OAAO,GAAG,SAAS,WAAW,EAAE,OAAO;AAAA,IAC/C;AAAA,EACF;AACF;AAEA,eAAsB,eACpB,MAC+B;AAC/B,QAAM,YAAY,KAAK,IAAI;AAC3B,QAAM,OAAO,KAAK,IAAI,GAAG,KAAK,QAAQ,CAAC;AACvC,QAAM,iBAAiB,KAAK,IAAI,GAAG,KAAK,kBAAkB,CAAC;AAC3D,QAAM,cAAc,KAAK,eAAe,OAAO;AAC/C,QAAM,cAAc,KAAK,eAAe,KAAK,KAAK,IAAI,CAAC,MAAM,EAAE,IAAI;AAEnE,QAAM,OAAO,UAAU,KAAK,IAAI;AAChC,QAAM,WAAW,KAAK,SAClB,KAAK,OAAO,CAAC,MAAO,KAAK,OAAoC,CAAC,CAAC,IAC/D;AACJ,QAAM,cAAc,KAAK,SAAS,SAAS;AAE3C,QAAM,UAAwB,CAAC;AAC/B,WAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,aAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,cAAQ,KAAK;AAAA,QACX,MAAO,SAAS,CAAC,EAAe;AAAA,QAChC,KAAK;AAAA,QACL,SAAS,IAAI,OAAO;AAAA,MACtB,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,cAAuE,CAAC;AAC9E,MAAI,iBAAiB;AACrB,MAAI,qBAAqB;AACzB,MAAI,eAAe;AACnB,MAAI,mBAAmB;AAEvB,QAAM,UAAU,MAAe,KAAK,QAAQ,YAAY;AAOxD,MAAI,WAAW;AACf,MAAI,SAAS;AACb,MAAI;AACJ,QAAM,OAAO,IAAI,QAAc,CAAC,QAAQ;AACtC,iBAAa;AAAA,EACf,CAAC;AAED,QAAM,OAAO,MAAY;AACvB,WAAO,WAAW,kBAAkB,SAAS,QAAQ,QAAQ;AAC3D,UAAI,QAAQ,KAAK,oBAAoB;AAEnC,cAAM,OAAO,QAAQ,SAAS;AAC9B,4BAAoB;AACpB,iBAAS,QAAQ;AACjB;AAAA,MACF;AACA,YAAM,OAAO,QAAQ,QAAQ;AAC7B;AAGA,YAAM,SAAS,EAAE,MAAM,MAAM,CAAC,EAA0B;AACxD,kBAAY,KAAK,MAAM;AACvB,YAAM,WAAwC,YAAY;AACxD,YAAI;AACF,iBAAO,MAAM,KAAK,QAAQ,IAAI;AAAA,QAChC,SAAS,KAAK;AACZ,iBAAO,gBAAwB,GAAG;AAAA,QACpC;AAAA,MACF,GAAG;AACH,cAAQ,KAAK,CAAC,WAAW;AACvB,eAAO,KAAK,KAAK,MAAM;AACvB;AACA,0BAAkB,OAAO;AACzB,YAAI,kBAAkB,eAAe,CAAC,oBAAoB;AACxD,+BAAqB;AAErB,kBAAQ,KAAK,+BAA+B;AAAA,QAC9C;AACA,YAAI;AACF,eAAK,iBAAiB,MAAM,MAAM;AAAA,QACpC,QAAQ;AAAA,QAGR;AACA;AACA,YAAI,SAAS,QAAQ,QAAQ;AAC3B,eAAK;AAAA,QACP,WAAW,aAAa,GAAG;AACzB,uBAAa;AAAA,QACf;AAAA,MACF,CAAC;AAAA,IACH;AACA,QAAI,UAAU,QAAQ,UAAU,aAAa,EAAG,cAAa;AAAA,EAC/D;AAEA,QAAM,UAAU,MAAY;AAG1B,QAAI,SAAS,QAAQ,QAAQ;AAC3B,0BAAoB,QAAQ,SAAS;AACrC,eAAS,QAAQ;AAAA,IACnB;AACA,QAAI,aAAa,EAAG,cAAa;AAAA,EACnC;AACA,MAAI,KAAK,QAAQ;AACf,QAAI,KAAK,OAAO,SAAS;AACvB,yBAAmB,QAAQ;AAC3B,eAAS,QAAQ;AACjB,mBAAa;AAAA,IACf,OAAO;AACL,WAAK,OAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AAAA,IAC/D;AAAA,EACF;AAEA,MAAI,QAAQ,WAAW,GAAG;AACxB,iBAAa;AAAA,EACf,OAAO;AACL,SAAK;AAAA,EACP;AAEA,QAAM;AACN,MAAI,KAAK,OAAQ,MAAK,OAAO,oBAAoB,SAAS,OAAO;AAEjE,cAAY,KAAK,CAAC,GAAG,MAAM,EAAE,KAAK,UAAU,EAAE,KAAK,OAAO;AAE1D,MAAI,OAAO;AACX,MAAI,WAAW;AACf,MAAI,YAAY;AAChB,MAAI,WAAW;AACf,aAAW,EAAE,KAAK,KAAK,aAAa;AAClC,eAAW,KAAK,MAAM;AACpB;AACA,YAAM,UAAU,EAAE,UAAU;AAC5B,UAAI,CAAC,WAAW,EAAE,QAAQ,MAAO;AACjC,kBAAY,UAAU,IAAI,EAAE,QAAQ;AACpC,mBAAa,EAAE;AAAA,IACjB;AAAA,EACF;AAEA,QAAM,SAAS,YAAY,aAAa,KAAK,MAAM,WAAW;AAE9D,SAAO;AAAA,IACL,OAAO;AAAA,IACP;AAAA,IACA,SAAS;AAAA,MACP,YAAY,QAAQ;AAAA,MACpB;AAAA,MACA,cAAc,mBAAmB,cAAc;AAAA,MAC/C,iBAAiB,aAAa,IAAI,IAAI,OAAO;AAAA,MAC7C,kBAAkB,aAAa,IAAI,IAAI,WAAW;AAAA,MAClD,cAAc;AAAA,MACd,YAAY,KAAK,IAAI,IAAI;AAAA,IAC3B;AAAA,IACA,UAAU,aAAa;AAAA,EACzB;AACF;","names":["i"]}

package/dist/chunk-K4DBDHLK.js

@@ -1,158 +0,0 @@
-// src/trace/attribute-vocabulary.ts
-import {
-  GEN_AI_INPUT_TOKEN_KEYS,
-  GEN_AI_MODEL_KEYS,
-  GEN_AI_OUTPUT_TOKEN_KEYS
-} from "@tangle-network/agent-core/telemetry";
-var OPENINFERENCE_SPAN_KIND = "openinference.span.kind";
-var LLM_MODEL_NAME = "llm.model_name";
-var LLM_INPUT_TOKENS = "llm.token_count.prompt";
-var LLM_OUTPUT_TOKENS = "llm.token_count.completion";
-var LLM_REASONING_TOKENS = "llm.token_count.reasoning";
-var LLM_CACHED_TOKENS = "llm.token_count.prompt_cache_hit";
-var LLM_CACHE_WRITE_TOKENS = "llm.token_count.prompt_cache_write";
-var LLM_CONTEXT_TOKENS = "tangle.llm.context_tokens";
-var LLM_COST_USD = "llm.cost_usd";
-var TOOL_NAME = "tool.name";
-var TOOL_ARGS_CAPTURED = "tool.args_captured";
-var TOOL_LATENCY_MS = "tool.latency_ms";
-var INPUT_VALUE = "input.value";
-var OUTPUT_VALUE = "output.value";
-var SPAN_KIND_ATTR_KEYS = [OPENINFERENCE_SPAN_KIND, "inference.observation_kind"];
-var LLM_MODEL_ATTR_KEYS = [
-  LLM_MODEL_NAME,
-  "inference.llm.model_name",
-  "llm.model",
-  ...GEN_AI_MODEL_KEYS,
-  "tangle.model"
-];
-var LLM_INPUT_TOKEN_ATTR_KEYS = [
-  LLM_INPUT_TOKENS,
-  "inference.llm.input_tokens",
-  "llm.input_tokens",
-  ...GEN_AI_INPUT_TOKEN_KEYS,
-  "tangle.tokens.in",
-  "tokens.in"
-];
-var LLM_OUTPUT_TOKEN_ATTR_KEYS = [
-  LLM_OUTPUT_TOKENS,
-  "inference.llm.output_tokens",
-  "llm.output_tokens",
-  ...GEN_AI_OUTPUT_TOKEN_KEYS,
-  "tangle.tokens.out",
-  "tokens.out"
-];
-var LLM_REASONING_TOKEN_ATTR_KEYS = [
-  LLM_REASONING_TOKENS,
-  "inference.llm.reasoning_tokens",
-  "llm.reasoning_tokens",
-  "gen_ai.usage.reasoning_tokens",
-  "gen_ai.usage.reasoning_output_tokens",
-  "reasoning_tokens",
-  "reasoning_output_tokens",
-  "tangle.tokens.reasoning",
-  "gen_ai.usage.output_tokens_details.reasoning_tokens",
-  "gen_ai.usage.completion_tokens_details.reasoning_tokens"
-];
-var LLM_CACHED_TOKEN_ATTR_KEYS = [
-  LLM_CACHED_TOKENS,
-  "inference.llm.cached_tokens",
-  "llm.cached_tokens",
-  "gen_ai.usage.cached_tokens",
-  "gen_ai.usage.prompt_tokens_details.cached_tokens",
-  "gen_ai.usage.input_tokens_details.cached_tokens",
-  "gen_ai.usage.cache_read_tokens",
-  "gen_ai.usage.cache_read_input_tokens",
-  "cache_read_tokens",
-  "cache_read_input_tokens",
-  "input_cache_read",
-  "tangle.tokens.cached"
-];
-var LLM_CACHE_WRITE_TOKEN_ATTR_KEYS = [
-  LLM_CACHE_WRITE_TOKENS,
-  "inference.llm.cache_write_tokens",
-  "llm.cache_write_tokens",
-  "gen_ai.usage.cache_creation_tokens",
-  "gen_ai.usage.cache_creation_input_tokens",
-  "cache_creation_tokens",
-  "cache_creation_input_tokens",
-  "input_cache_creation",
-  "tangle.tokens.cache_write"
-];
-var LLM_COST_ATTR_KEYS = [
-  LLM_COST_USD,
-  "inference.llm.cost.total",
-  "llm.cost.total",
-  "gen_ai.usage.cost",
-  "gen_ai.usage.cost_usd",
-  "tangle.cost.usd",
-  "cost.usd",
-  "cost"
-];
-var RUN_COST_ATTR_KEYS = ["tangle.cost.usd", "cost.usd"];
-var TOOL_NAME_ATTR_KEYS = [TOOL_NAME, "inference.tool.name"];
-function asNumber(value) {
-  if (typeof value === "number") return Number.isFinite(value) ? value : null;
-  if (typeof value === "string" && value.length > 0) {
-    const parsed = Number(value);
-    return Number.isFinite(parsed) ? parsed : null;
-  }
-  return null;
-}
-function firstNumberAttr(attributes, keys) {
-  for (const key of keys) {
-    const value = asNumber(attributes[key]);
-    if (value !== null) return value;
-  }
-  return null;
-}
-function contextInputTokens(usage) {
-  if (usage.inputTokens === void 0 || usage.inputTokens === null) return void 0;
-  return usage.inputTokens + (usage.cachedTokens ?? 0) + (usage.cacheWriteTokens ?? 0);
-}
-function applyLlmSpanOtlpAttributes(attributes, span) {
-  if (span.model !== void 0) attributes[LLM_MODEL_NAME] = span.model;
-  if (span.inputTokens !== void 0) attributes[LLM_INPUT_TOKENS] = span.inputTokens;
-  if (span.outputTokens !== void 0) attributes[LLM_OUTPUT_TOKENS] = span.outputTokens;
-  if (span.reasoningTokens !== void 0) attributes[LLM_REASONING_TOKENS] = span.reasoningTokens;
-  if (span.cachedTokens !== void 0) attributes[LLM_CACHED_TOKENS] = span.cachedTokens;
-  if (span.cacheWriteTokens !== void 0) {
-    attributes[LLM_CACHE_WRITE_TOKENS] = span.cacheWriteTokens;
-  }
-  const contextTokens = contextInputTokens(span);
-  if (contextTokens !== void 0) attributes[LLM_CONTEXT_TOKENS] = contextTokens;
-  if (span.costUsd !== void 0) attributes[LLM_COST_USD] = span.costUsd;
-  if (span.finishReason) attributes["llm.finish_reason"] = span.finishReason;
-}
-
-export {
-  OPENINFERENCE_SPAN_KIND,
-  LLM_MODEL_NAME,
-  LLM_INPUT_TOKENS,
-  LLM_OUTPUT_TOKENS,
-  LLM_REASONING_TOKENS,
-  LLM_CACHED_TOKENS,
-  LLM_CACHE_WRITE_TOKENS,
-  LLM_CONTEXT_TOKENS,
-  LLM_COST_USD,
-  TOOL_NAME,
-  TOOL_ARGS_CAPTURED,
-  TOOL_LATENCY_MS,
-  INPUT_VALUE,
-  OUTPUT_VALUE,
-  SPAN_KIND_ATTR_KEYS,
-  LLM_MODEL_ATTR_KEYS,
-  LLM_INPUT_TOKEN_ATTR_KEYS,
-  LLM_OUTPUT_TOKEN_ATTR_KEYS,
-  LLM_REASONING_TOKEN_ATTR_KEYS,
-  LLM_CACHED_TOKEN_ATTR_KEYS,
-  LLM_CACHE_WRITE_TOKEN_ATTR_KEYS,
-  LLM_COST_ATTR_KEYS,
-  RUN_COST_ATTR_KEYS,
-  TOOL_NAME_ATTR_KEYS,
-  asNumber,
-  firstNumberAttr,
-  contextInputTokens,
-  applyLlmSpanOtlpAttributes
-};
-//# sourceMappingURL=chunk-K4DBDHLK.js.map