@tangle-network/agent-eval 0.184.0 → 0.185.0

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.184.0",
+    "version": "0.185.0",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.1.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/otel-spans-CQY1pNiN.js

@@ -0,0 +1,275 @@
+import { s as ValidationError } from "./errors-Dngq5h35.js";
+import { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from "./trace-attributes.js";
+import { n as classifyOtlpSpanRole, r as isOtlpModelCall } from "./otlp-attributes-BpC_blLF.js";
+import { i as summarizeTraceErrors, n as recordAggregateMeasurements, r as summarizeExecutionMeasurements, t as readTaskFailureLabels } from "./task-failure-attributes-B34TiMBe.js";
+//#region src/contract/intake/otel-spans.ts
+/**
+* # `intake/otel-spans` — OTel `TraceSpanEvent[]` → `RunRecord[]`.
+*
+* Turns an existing observability stream into the substrate-canonical
+* `RunRecord` shape so consumers with logs but no eval discipline can
+* call `analyzeRuns()` against their production traffic immediately.
+*
+* Pivot rule: spans are grouped by `tangle.runId` (the same attribute the
+* hosted-tier wire format uses) or, when absent, by `traceId`. One group
+* becomes one `RunRecord`. The root span (no `parentSpanId`) supplies:
+*
+*   - `runId` (the group key)
+*   - `wallMs` from `endTimeUnixNano - startTimeUnixNano`
+*   - `model` from `gen_ai.request.model` / `llm.model` / `tangle.model`
+*   - task failure class and detail from explicit `tangle.task.*` attributes
+*   - cost from `cost.usd` / `gen_ai.usage.cost_usd` / `tangle.cost.usd`
+*   - token usage from model-call input, output, cache-read, and cache-write
+*     attributes without double-counting aggregate parent spans
+*   - task quality from an explicit `scoreForRun` callback or a designated
+*     evaluation attribute on a root / `EVALUATOR` span; `outcome.raw`
+*     collects every numeric attribute without promoting it to task quality.
+*
+* Errored tool, model, and child-agent spans contribute to execution-error
+* counts. Root process, guardrail, evaluator, propagated parent, and unknown
+* errors retain separate counters. Only one failed root can set
+* `RunRecord.terminalOutcome` and `RunRecord.terminalFailureReason`; a child
+* error cannot become a task failure.
+*/
+const TASK_SCORE_ATTR_KEYS = [
+	"gen_ai.evaluation.score.value",
+	"tangle.task.score",
+	"eval.score",
+	"tangle.score"
+];
+const MODEL_KEYS = [
+	"tangle.model",
+	...LLM_MODEL_ATTR_KEYS,
+	"model"
+];
+const PROMPT_HASH_KEYS = ["tangle.prompt_hash", "prompt.hash"];
+const CONFIG_HASH_KEYS = ["tangle.config_hash", "config.hash"];
+function fromOtelSpans(opts) {
+	const { spans, defaultSplit = "holdout", experimentId = "otel-corpus" } = opts;
+	const grouped = groupSpans(spans);
+	const runs = [];
+	for (const [groupKey, groupSpans] of grouped) {
+		const root = findRoot(groupSpans);
+		if (!root) continue;
+		const measurements = summarizeExecutionMeasurements(groupSpans.map((span) => ({
+			id: span.spanId,
+			...span.parentSpanId ? { parentId: span.parentSpanId } : {},
+			attributes: span.attributes,
+			modelCall: isExplicitModelCall(span),
+			aggregate: isExplicitAggregate(span)
+		})));
+		const callSpanIds = new Set(measurements.callSpanIds);
+		const callSpans = groupSpans.filter((span) => callSpanIds.has(span.spanId));
+		const wallMs = unixNanoDurationMs(root.startTimeUnixNano, root.endTimeUnixNano);
+		const model = readAttrString(callSpans, MODEL_KEYS) ?? readAttrString(groupSpans, MODEL_KEYS) ?? "unknown@unknown";
+		const capturedCost = (measurements.cost.complete ? measurements.cost.value : void 0) ?? measurements.aggregate?.costUsd;
+		const costUsd = capturedCost ?? null;
+		const scenarioId = readConsistentScenarioId(groupKey, groupSpans) ?? groupKey;
+		const promptHash = readAttrString(groupSpans, PROMPT_HASH_KEYS) ?? "sha256:unknown";
+		const configHash = readAttrString(groupSpans, CONFIG_HASH_KEYS) ?? "sha256:unknown";
+		const score = resolveTaskScore(groupKey, groupSpans, opts.scoreForRun);
+		const taskFailure = readTaskFailureLabels(groupSpans.filter((span) => !span.parentSpanId && isTerminalRootCandidate(span)), `fromOtelSpans: run '${groupKey}'`);
+		const rawNumeric = collectNumericAttrs(groupSpans);
+		const errorSummary = summarizeTraceErrors(groupSpans.map((span) => ({
+			id: spanIdentity(span),
+			...span.parentSpanId ? { parentId: parentIdentity(span) } : {},
+			role: errorRoleForSpan(span),
+			error: span.status?.code === "ERROR",
+			processRoot: !span.parentSpanId && isTerminalRootCandidate(span)
+		})));
+		rawNumeric.error_span_count = errorSummary.total;
+		rawNumeric.execution_error_count = errorSummary.execution;
+		rawNumeric.process_error_count = errorSummary.process;
+		rawNumeric.guardrail_error_count = errorSummary.guardrail;
+		rawNumeric.judge_error_count = errorSummary.evaluation;
+		rawNumeric.propagated_error_count = errorSummary.propagated;
+		rawNumeric.unclassified_error_count = errorSummary.unclassified;
+		rawNumeric.llm_span_count = measurements.modelCallCount;
+		if (measurements.cost.value !== void 0 && !measurements.cost.complete) rawNumeric.partial_observed_cost_usd = measurements.cost.value;
+		recordAggregateMeasurements(rawNumeric, measurements.aggregate);
+		const judgeScores = score !== void 0 ? {
+			perJudge: { "otel-derived": { score } },
+			perDimMean: { score },
+			composite: score
+		} : void 0;
+		const terminalOutcome = terminalOutcomeFromRoots(groupSpans);
+		const failedRoot = terminalOutcome === "failed" ? groupSpans.find((span) => !span.parentSpanId && isTerminalRootCandidate(span) && span.status?.code === "ERROR") : void 0;
+		const outcome = {
+			raw: rawNumeric,
+			...judgeScores ? { judgeScores } : {}
+		};
+		if (score !== void 0) if (defaultSplit === "holdout") outcome.holdoutScore = score;
+		else outcome.searchScore = score;
+		runs.push({
+			runId: groupKey,
+			experimentId,
+			candidateId: root.attributes["tangle.candidateId"] ?? "otel-default",
+			seed: 0,
+			model,
+			promptHash,
+			configHash,
+			commitSha: root.attributes["tangle.commit_sha"] ?? "unknown",
+			wallMs,
+			costUsd,
+			costProvenance: capturedCost === void 0 ? {
+				kind: "uncaptured",
+				usd: null
+			} : {
+				kind: "observed",
+				usd: capturedCost
+			},
+			tokenUsage: measurements.tokenUsage,
+			terminalOutcome,
+			...failedRoot ? { terminalFailureReason: failedRoot.status?.message ?? failedRoot.name } : {},
+			outcome,
+			...taskFailure,
+			splitTag: defaultSplit,
+			scenarioId
+		});
+	}
+	return runs;
+}
+function terminalOutcomeFromRoots(spans) {
+	const roots = spans.filter((span) => !span.parentSpanId && isTerminalRootCandidate(span));
+	if (roots.length !== 1) return "unknown";
+	if (roots[0].status?.code === "OK") return "succeeded";
+	if (roots[0].status?.code === "ERROR") return "failed";
+	return "unknown";
+}
+function isTerminalRootCandidate(span) {
+	const role = errorRoleForSpan(span);
+	if (role === "LLM" || role === "TOOL" || role === "GUARDRAIL" || role === "EVALUATOR") return false;
+	return true;
+}
+function readSpanKind(span) {
+	return readAttrString([span], [...SPAN_KIND_ATTR_KEYS, "span.kind"])?.toUpperCase();
+}
+function errorRoleForSpan(span) {
+	return classifyOtlpSpanRole({
+		kind: readSpanKind(span),
+		name: span.name,
+		attributes: span.attributes
+	});
+}
+function spanIdentity(span) {
+	return `${span.traceId}:${span.spanId}`;
+}
+function parentIdentity(span) {
+	return `${span.traceId}:${span.parentSpanId}`;
+}
+function isExplicitModelCall(span) {
+	return isOtlpModelCall({
+		kind: readSpanKind(span),
+		name: span.name,
+		attributes: span.attributes
+	});
+}
+function isExplicitAggregate(span) {
+	const kind = readSpanKind(span);
+	return kind !== void 0 && kind !== "LLM";
+}
+function groupSpans(spans) {
+	const m = /* @__PURE__ */ new Map();
+	for (const span of spans) {
+		const key = span["tangle.runId"] ?? span.traceId;
+		const list = m.get(key) ?? [];
+		list.push(span);
+		m.set(key, list);
+	}
+	return m;
+}
+function findRoot(group) {
+	const structuralRoots = group.filter((span) => !span.parentSpanId);
+	const terminalRoots = structuralRoots.filter(isTerminalRootCandidate);
+	return orderSpans(terminalRoots.length > 0 ? terminalRoots : structuralRoots.length > 0 ? structuralRoots : group)[0];
+}
+function readAttrString(spans, keys) {
+	for (const span of spans) for (const key of keys) {
+		const v = span.attributes[key];
+		if (typeof v === "string" && v.length > 0) return v;
+	}
+}
+function readConsistentScenarioId(runId, spans) {
+	const values = /* @__PURE__ */ new Set();
+	for (const span of spans) {
+		const topLevel = span["tangle.scenarioId"];
+		if (typeof topLevel === "string" && topLevel.length > 0) values.add(topLevel);
+		const attribute = span.attributes["tangle.scenarioId"];
+		if (typeof attribute === "string" && attribute.length > 0) values.add(attribute);
+	}
+	if (values.size > 1) throw new ValidationError(`fromOtelSpans: conflicting scenario ids for run '${runId}': ${[...values].sort().join(", ")}`);
+	return values.values().next().value;
+}
+function resolveTaskScore(runId, spans, scoreForRun) {
+	const orderedSpans = orderSpans(spans);
+	const sources = [];
+	if (scoreForRun) {
+		const supplied = scoreForRun(runId, orderedSpans);
+		if (supplied !== void 0) {
+			if (typeof supplied !== "number" || !Number.isFinite(supplied)) throw new ValidationError(`fromOtelSpans: scoreForRun returned a non-finite number for run '${runId}'`);
+			sources.push({
+				label: "scoreForRun",
+				value: supplied
+			});
+		}
+	}
+	for (const span of orderedSpans) {
+		const role = errorRoleForSpan(span);
+		if (span.parentSpanId && role !== "EVALUATOR") continue;
+		if (role === "EVALUATOR" && span.status?.code === "ERROR") continue;
+		for (const key of TASK_SCORE_ATTR_KEYS) {
+			if (!Object.hasOwn(span.attributes, key)) continue;
+			sources.push({
+				label: `span '${span.spanId}' attribute '${key}'`,
+				value: parseTaskScoreAttribute(runId, span.spanId, key, span.attributes[key])
+			});
+		}
+	}
+	if (sources.length === 0) return void 0;
+	sources.sort((left, right) => left.label.localeCompare(right.label));
+	const score = sources[0].value;
+	if (sources.some((source) => source.value !== score)) throw new ValidationError(`fromOtelSpans: conflicting task-quality scores for run '${runId}': ${sources.map((source) => `${source.label}=${source.value}`).join(", ")}`);
+	return score;
+}
+function parseTaskScoreAttribute(runId, spanId, key, value) {
+	const source = `span '${spanId}' attribute '${key}'`;
+	if (typeof value === "string") {
+		if (value.trim().length === 0) throw new ValidationError(`fromOtelSpans: ${source} is blank for run '${runId}'; task quality must be finite`);
+		const parsed = Number(value);
+		if (Number.isFinite(parsed)) return parsed;
+	} else if (typeof value === "number" && Number.isFinite(value)) return value;
+	throw new ValidationError(`fromOtelSpans: ${source} is not a finite task-quality score for run '${runId}'`);
+}
+function orderSpans(spans) {
+	return [...spans].sort((left, right) => compareUnixNano(left.startTimeUnixNano, right.startTimeUnixNano) || left.spanId.localeCompare(right.spanId));
+}
+function parseUnixNano(value) {
+	try {
+		return BigInt(value);
+	} catch {
+		throw new ValidationError(`fromOtelSpans: invalid Unix nanosecond timestamp '${value}'`);
+	}
+}
+function compareUnixNano(left, right) {
+	const leftValue = parseUnixNano(left);
+	const rightValue = parseUnixNano(right);
+	return leftValue < rightValue ? -1 : leftValue > rightValue ? 1 : 0;
+}
+function unixNanoDurationMs(start, end) {
+	const delta = parseUnixNano(end) - parseUnixNano(start);
+	if (delta <= 0n) return 0;
+	const wholeMs = delta / 1000000n;
+	const fractionalMs = delta % 1000000n;
+	const value = Number(wholeMs) + Number(fractionalMs) / 1e6;
+	if (!Number.isSafeInteger(Number(wholeMs))) throw new ValidationError("fromOtelSpans: span duration exceeds the safe millisecond range");
+	return value;
+}
+function collectNumericAttrs(spans) {
+	const raw = {};
+	for (const span of spans) for (const [k, v] of Object.entries(span.attributes)) if (typeof v === "number" && Number.isFinite(v)) raw[k] = v;
+	return raw;
+}
+//#endregion
+export { fromOtelSpans as t };
+
+//# sourceMappingURL=otel-spans-CQY1pNiN.js.map

package/dist/otel-spans-CQY1pNiN.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"otel-spans-CQY1pNiN.js","names":[],"sources":["../src/contract/intake/otel-spans.ts"],"sourcesContent":["/**\n * # `intake/otel-spans` — OTel `TraceSpanEvent[]` → `RunRecord[]`.\n *\n * Turns an existing observability stream into the substrate-canonical\n * `RunRecord` shape so consumers with logs but no eval discipline can\n * call `analyzeRuns()` against their production traffic immediately.\n *\n * Pivot rule: spans are grouped by `tangle.runId` (the same attribute the\n * hosted-tier wire format uses) or, when absent, by `traceId`. One group\n * becomes one `RunRecord`. The root span (no `parentSpanId`) supplies:\n *\n *   - `runId` (the group key)\n *   - `wallMs` from `endTimeUnixNano - startTimeUnixNano`\n *   - `model` from `gen_ai.request.model` / `llm.model` / `tangle.model`\n *   - task failure class and detail from explicit `tangle.task.*` attributes\n *   - cost from `cost.usd` / `gen_ai.usage.cost_usd` / `tangle.cost.usd`\n *   - token usage from model-call input, output, cache-read, and cache-write\n *     attributes without double-counting aggregate parent spans\n *   - task quality from an explicit `scoreForRun` callback or a designated\n *     evaluation attribute on a root / `EVALUATOR` span; `outcome.raw`\n *     collects every numeric attribute without promoting it to task quality.\n *\n * Errored tool, model, and child-agent spans contribute to execution-error\n * counts. Root process, guardrail, evaluator, propagated parent, and unknown\n * errors retain separate counters. Only one failed root can set\n * `RunRecord.terminalOutcome` and `RunRecord.terminalFailureReason`; a child\n * error cannot become a task failure.\n */\n\nimport { ValidationError } from '../../errors'\nimport type { TraceSpanEvent } from '../../hosted/types'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../../run-record'\nimport { summarizeTraceErrors, type TraceErrorRole } from '../../trace/error-classification'\nimport {\n  recordAggregateMeasurements,\n  summarizeExecutionMeasurements,\n} from '../../trace/execution-measurements'\nimport {\n  classifyOtlpSpanRole,\n  isOtlpModelCall,\n  LLM_MODEL_ATTR_KEYS,\n  SPAN_KIND_ATTR_KEYS,\n} from '../../trace/otlp-attributes'\nimport { readTaskFailureLabels } from '../../trace/task-failure-attributes'\n\nconst TASK_SCORE_ATTR_KEYS = [\n  'gen_ai.evaluation.score.value',\n  'tangle.task.score',\n  'eval.score',\n  'tangle.score',\n] as const\nconst MODEL_KEYS = ['tangle.model', ...LLM_MODEL_ATTR_KEYS, 'model']\nconst PROMPT_HASH_KEYS = ['tangle.prompt_hash', 'prompt.hash']\nconst CONFIG_HASH_KEYS = ['tangle.config_hash', 'config.hash']\n\nexport interface FromOtelSpansOptions {\n  spans: TraceSpanEvent[]\n  /** Default split tag for synthesized records. Defaults to `'holdout'`. */\n  defaultSplit?: RunSplitTag\n  /** Default `experimentId` when not present on any span. */\n  experimentId?: string\n  /**\n   * Explicit task-quality score for a logical run. The callback receives\n   * spans in deterministic time/id order. Its value must agree with any\n   * designated score attributes present on root or `EVALUATOR` spans.\n   */\n  scoreForRun?: (runId: string, spans: readonly TraceSpanEvent[]) => number | undefined\n}\n\nexport function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[] {\n  const { spans, defaultSplit = 'holdout', experimentId = 'otel-corpus' } = opts\n  const grouped = groupSpans(spans)\n\n  const runs: RunRecord[] = []\n  for (const [groupKey, groupSpans] of grouped) {\n    const root = findRoot(groupSpans)\n    if (!root) continue\n    const measurements = summarizeExecutionMeasurements(\n      groupSpans.map((span) => ({\n        id: span.spanId,\n        ...(span.parentSpanId ? { parentId: span.parentSpanId } : {}),\n        attributes: span.attributes,\n        modelCall: isExplicitModelCall(span),\n        aggregate: isExplicitAggregate(span),\n      })),\n    )\n    const callSpanIds = new Set(measurements.callSpanIds)\n    const callSpans = groupSpans.filter((span) => callSpanIds.has(span.spanId))\n    const wallMs = unixNanoDurationMs(root.startTimeUnixNano, root.endTimeUnixNano)\n    const model =\n      readAttrString(callSpans, MODEL_KEYS) ??\n      readAttrString(groupSpans, MODEL_KEYS) ??\n      'unknown@unknown'\n    const capturedCost =\n      (measurements.cost.complete ? measurements.cost.value : undefined) ??\n      measurements.aggregate?.costUsd\n    const costUsd = capturedCost ?? null\n    const scenarioId = readConsistentScenarioId(groupKey, groupSpans) ?? groupKey\n    const promptHash = readAttrString(groupSpans, PROMPT_HASH_KEYS) ?? 'sha256:unknown'\n    const configHash = readAttrString(groupSpans, CONFIG_HASH_KEYS) ?? 'sha256:unknown'\n    const score = resolveTaskScore(groupKey, groupSpans, opts.scoreForRun)\n    const taskFailure = readTaskFailureLabels(\n      groupSpans.filter((span) => !span.parentSpanId && isTerminalRootCandidate(span)),\n      `fromOtelSpans: run '${groupKey}'`,\n    )\n\n    const rawNumeric = collectNumericAttrs(groupSpans)\n    const errorSummary = summarizeTraceErrors(\n      groupSpans.map((span) => ({\n        id: spanIdentity(span),\n        ...(span.parentSpanId ? { parentId: parentIdentity(span) } : {}),\n        role: errorRoleForSpan(span),\n        error: span.status?.code === 'ERROR',\n        processRoot: !span.parentSpanId && isTerminalRootCandidate(span),\n      })),\n    )\n    rawNumeric.error_span_count = errorSummary.total\n    rawNumeric.execution_error_count = errorSummary.execution\n    rawNumeric.process_error_count = errorSummary.process\n    rawNumeric.guardrail_error_count = errorSummary.guardrail\n    rawNumeric.judge_error_count = errorSummary.evaluation\n    rawNumeric.propagated_error_count = errorSummary.propagated\n    rawNumeric.unclassified_error_count = errorSummary.unclassified\n    rawNumeric.llm_span_count = measurements.modelCallCount\n    if (measurements.cost.value !== undefined && !measurements.cost.complete) {\n      rawNumeric.partial_observed_cost_usd = measurements.cost.value\n    }\n    recordAggregateMeasurements(rawNumeric, measurements.aggregate)\n\n    const judgeScores: JudgeScoresRecord | undefined =\n      score !== undefined\n        ? {\n            perJudge: { 'otel-derived': { score } },\n            perDimMean: { score },\n            composite: score,\n          }\n        : undefined\n\n    const terminalOutcome = terminalOutcomeFromRoots(groupSpans)\n    const failedRoot =\n      terminalOutcome === 'failed'\n        ? groupSpans.find(\n            (span) =>\n              !span.parentSpanId && isTerminalRootCandidate(span) && span.status?.code === 'ERROR',\n          )\n        : undefined\n    const outcome: RunOutcome = {\n      raw: rawNumeric,\n      ...(judgeScores ? { judgeScores } : {}),\n    }\n    if (score !== undefined) {\n      if (defaultSplit === 'holdout') outcome.holdoutScore = score\n      else outcome.searchScore = score\n    }\n\n    runs.push({\n      runId: groupKey,\n      experimentId,\n      candidateId: (root.attributes['tangle.candidateId'] as string | undefined) ?? 'otel-default',\n      seed: 0,\n      model,\n      promptHash,\n      configHash,\n      commitSha: (root.attributes['tangle.commit_sha'] as string | undefined) ?? 'unknown',\n      wallMs,\n      costUsd,\n      costProvenance:\n        capturedCost === undefined\n          ? { kind: 'uncaptured', usd: null }\n          : { kind: 'observed', usd: capturedCost },\n      tokenUsage: measurements.tokenUsage,\n      terminalOutcome,\n      ...(failedRoot\n        ? { terminalFailureReason: failedRoot.status?.message ?? failedRoot.name }\n        : {}),\n      outcome,\n      ...taskFailure,\n      splitTag: defaultSplit,\n      scenarioId,\n    })\n  }\n  return runs\n}\n\nfunction terminalOutcomeFromRoots(spans: TraceSpanEvent[]): RunTerminalOutcome {\n  const roots = spans.filter((span) => !span.parentSpanId && isTerminalRootCandidate(span))\n  if (roots.length !== 1) return 'unknown'\n  if (roots[0]!.status?.code === 'OK') return 'succeeded'\n  if (roots[0]!.status?.code === 'ERROR') return 'failed'\n  return 'unknown'\n}\n\nfunction isTerminalRootCandidate(span: TraceSpanEvent): boolean {\n  const role = errorRoleForSpan(span)\n  if (role === 'LLM' || role === 'TOOL' || role === 'GUARDRAIL' || role === 'EVALUATOR') {\n    return false\n  }\n  return true\n}\n\nfunction readSpanKind(span: TraceSpanEvent): string | undefined {\n  return readAttrString([span], [...SPAN_KIND_ATTR_KEYS, 'span.kind'])?.toUpperCase()\n}\n\nfunction errorRoleForSpan(span: TraceSpanEvent): TraceErrorRole {\n  return classifyOtlpSpanRole({\n    kind: readSpanKind(span),\n    name: span.name,\n    attributes: span.attributes,\n  })\n}\n\nfunction spanIdentity(span: TraceSpanEvent): string {\n  return `${span.traceId}:${span.spanId}`\n}\n\nfunction parentIdentity(span: TraceSpanEvent): string {\n  return `${span.traceId}:${span.parentSpanId}`\n}\n\nfunction isExplicitModelCall(span: TraceSpanEvent): boolean {\n  return isOtlpModelCall({\n    kind: readSpanKind(span),\n    name: span.name,\n    attributes: span.attributes,\n  })\n}\n\nfunction isExplicitAggregate(span: TraceSpanEvent): boolean {\n  const kind = readSpanKind(span)\n  return kind !== undefined && kind !== 'LLM'\n}\n\n// ── Internal helpers ────────────────────────────────────────────────\n\nfunction groupSpans(spans: TraceSpanEvent[]): Map<string, TraceSpanEvent[]> {\n  const m = new Map<string, TraceSpanEvent[]>()\n  for (const span of spans) {\n    const key = (span['tangle.runId'] as string | undefined) ?? span.traceId\n    const list = m.get(key) ?? []\n    list.push(span)\n    m.set(key, list)\n  }\n  return m\n}\n\nfunction findRoot(group: TraceSpanEvent[]): TraceSpanEvent | undefined {\n  const structuralRoots = group.filter((span) => !span.parentSpanId)\n  const terminalRoots = structuralRoots.filter(isTerminalRootCandidate)\n  const pool =\n    terminalRoots.length > 0 ? terminalRoots : structuralRoots.length > 0 ? structuralRoots : group\n  return orderSpans(pool)[0]\n}\n\nfunction readAttrString(spans: TraceSpanEvent[], keys: string[]): string | undefined {\n  for (const span of spans) {\n    for (const key of keys) {\n      const v = span.attributes[key]\n      if (typeof v === 'string' && v.length > 0) return v\n    }\n  }\n  return undefined\n}\n\nfunction readConsistentScenarioId(\n  runId: string,\n  spans: readonly TraceSpanEvent[],\n): string | undefined {\n  const values = new Set<string>()\n  for (const span of spans) {\n    const topLevel = span['tangle.scenarioId']\n    if (typeof topLevel === 'string' && topLevel.length > 0) values.add(topLevel)\n    const attribute = span.attributes['tangle.scenarioId']\n    if (typeof attribute === 'string' && attribute.length > 0) values.add(attribute)\n  }\n  if (values.size > 1) {\n    throw new ValidationError(\n      `fromOtelSpans: conflicting scenario ids for run '${runId}': ${[...values].sort().join(', ')}`,\n    )\n  }\n  return values.values().next().value\n}\n\ninterface TaskScoreSource {\n  label: string\n  value: number\n}\n\nfunction resolveTaskScore(\n  runId: string,\n  spans: TraceSpanEvent[],\n  scoreForRun: FromOtelSpansOptions['scoreForRun'],\n): number | undefined {\n  const orderedSpans = orderSpans(spans)\n  const sources: TaskScoreSource[] = []\n\n  if (scoreForRun) {\n    const supplied: unknown = scoreForRun(runId, orderedSpans)\n    if (supplied !== undefined) {\n      if (typeof supplied !== 'number' || !Number.isFinite(supplied)) {\n        throw new ValidationError(\n          `fromOtelSpans: scoreForRun returned a non-finite number for run '${runId}'`,\n        )\n      }\n      sources.push({ label: 'scoreForRun', value: supplied })\n    }\n  }\n\n  for (const span of orderedSpans) {\n    const role = errorRoleForSpan(span)\n    if (span.parentSpanId && role !== 'EVALUATOR') continue\n    if (role === 'EVALUATOR' && span.status?.code === 'ERROR') continue\n    for (const key of TASK_SCORE_ATTR_KEYS) {\n      if (!Object.hasOwn(span.attributes, key)) continue\n      sources.push({\n        label: `span '${span.spanId}' attribute '${key}'`,\n        value: parseTaskScoreAttribute(runId, span.spanId, key, span.attributes[key]),\n      })\n    }\n  }\n\n  if (sources.length === 0) return undefined\n  sources.sort((left, right) => left.label.localeCompare(right.label))\n  const score = sources[0]!.value\n  if (sources.some((source) => source.value !== score)) {\n    const details = sources.map((source) => `${source.label}=${source.value}`).join(', ')\n    throw new ValidationError(\n      `fromOtelSpans: conflicting task-quality scores for run '${runId}': ${details}`,\n    )\n  }\n  return score\n}\n\nfunction parseTaskScoreAttribute(\n  runId: string,\n  spanId: string,\n  key: string,\n  value: unknown,\n): number {\n  const source = `span '${spanId}' attribute '${key}'`\n  if (typeof value === 'string') {\n    if (value.trim().length === 0) {\n      throw new ValidationError(\n        `fromOtelSpans: ${source} is blank for run '${runId}'; task quality must be finite`,\n      )\n    }\n    const parsed = Number(value)\n    if (Number.isFinite(parsed)) return parsed\n  } else if (typeof value === 'number' && Number.isFinite(value)) {\n    return value\n  }\n  throw new ValidationError(\n    `fromOtelSpans: ${source} is not a finite task-quality score for run '${runId}'`,\n  )\n}\n\nfunction orderSpans(spans: readonly TraceSpanEvent[]): TraceSpanEvent[] {\n  return [...spans].sort(\n    (left, right) =>\n      compareUnixNano(left.startTimeUnixNano, right.startTimeUnixNano) ||\n      left.spanId.localeCompare(right.spanId),\n  )\n}\n\nfunction parseUnixNano(value: string): bigint {\n  try {\n    return BigInt(value)\n  } catch {\n    throw new ValidationError(`fromOtelSpans: invalid Unix nanosecond timestamp '${value}'`)\n  }\n}\n\nfunction compareUnixNano(left: string, right: string): number {\n  const leftValue = parseUnixNano(left)\n  const rightValue = parseUnixNano(right)\n  return leftValue < rightValue ? -1 : leftValue > rightValue ? 1 : 0\n}\n\nfunction unixNanoDurationMs(start: string, end: string): number {\n  const delta = parseUnixNano(end) - parseUnixNano(start)\n  if (delta <= 0n) return 0\n  const wholeMs = delta / 1_000_000n\n  const fractionalMs = delta % 1_000_000n\n  const value = Number(wholeMs) + Number(fractionalMs) / 1_000_000\n  if (!Number.isSafeInteger(Number(wholeMs))) {\n    throw new ValidationError('fromOtelSpans: span duration exceeds the safe millisecond range')\n  }\n  return value\n}\n\nfunction collectNumericAttrs(spans: TraceSpanEvent[]): Record<string, number> {\n  const raw: Record<string, number> = {}\n  for (const span of spans) {\n    for (const [k, v] of Object.entries(span.attributes)) {\n      if (typeof v === 'number' && Number.isFinite(v)) raw[k] = v\n    }\n  }\n  return raw\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmDA,MAAM,uBAAuB;CAC3B;CACA;CACA;CACA;AACF;AACA,MAAM,aAAa;CAAC;CAAgB,GAAG;CAAqB;AAAO;AACnE,MAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAC7D,MAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAgB7D,SAAgB,cAAc,MAAyC;CACrE,MAAM,EAAE,OAAO,eAAe,WAAW,eAAe,kBAAkB;CAC1E,MAAM,UAAU,WAAW,KAAK;CAEhC,MAAM,OAAoB,CAAC;CAC3B,KAAK,MAAM,CAAC,UAAU,eAAe,SAAS;EAC5C,MAAM,OAAO,SAAS,UAAU;EAChC,IAAI,CAAC,MAAM;EACX,MAAM,eAAe,+BACnB,WAAW,KAAK,UAAU;GACxB,IAAI,KAAK;GACT,GAAI,KAAK,eAAe,EAAE,UAAU,KAAK,aAAa,IAAI,CAAC;GAC3D,YAAY,KAAK;GACjB,WAAW,oBAAoB,IAAI;GACnC,WAAW,oBAAoB,IAAI;EACrC,EAAE,CACJ;EACA,MAAM,cAAc,IAAI,IAAI,aAAa,WAAW;EACpD,MAAM,YAAY,WAAW,QAAQ,SAAS,YAAY,IAAI,KAAK,MAAM,CAAC;EAC1E,MAAM,SAAS,mBAAmB,KAAK,mBAAmB,KAAK,eAAe;EAC9E,MAAM,QACJ,eAAe,WAAW,UAAU,KACpC,eAAe,YAAY,UAAU,KACrC;EACF,MAAM,gBACH,aAAa,KAAK,WAAW,aAAa,KAAK,QAAQ,KAAA,MACxD,aAAa,WAAW;EAC1B,MAAM,UAAU,gBAAgB;EAChC,MAAM,aAAa,yBAAyB,UAAU,UAAU,KAAK;EACrE,MAAM,aAAa,eAAe,YAAY,gBAAgB,KAAK;EACnE,MAAM,aAAa,eAAe,YAAY,gBAAgB,KAAK;EACnE,MAAM,QAAQ,iBAAiB,UAAU,YAAY,KAAK,WAAW;EACrE,MAAM,cAAc,sBAClB,WAAW,QAAQ,SAAS,CAAC,KAAK,gBAAgB,wBAAwB,IAAI,CAAC,GAC/E,uBAAuB,SAAS,EAClC;EAEA,MAAM,aAAa,oBAAoB,UAAU;EACjD,MAAM,eAAe,qBACnB,WAAW,KAAK,UAAU;GACxB,IAAI,aAAa,IAAI;GACrB,GAAI,KAAK,eAAe,EAAE,UAAU,eAAe,IAAI,EAAE,IAAI,CAAC;GAC9D,MAAM,iBAAiB,IAAI;GAC3B,OAAO,KAAK,QAAQ,SAAS;GAC7B,aAAa,CAAC,KAAK,gBAAgB,wBAAwB,IAAI;EACjE,EAAE,CACJ;EACA,WAAW,mBAAmB,aAAa;EAC3C,WAAW,wBAAwB,aAAa;EAChD,WAAW,sBAAsB,aAAa;EAC9C,WAAW,wBAAwB,aAAa;EAChD,WAAW,oBAAoB,aAAa;EAC5C,WAAW,yBAAyB,aAAa;EACjD,WAAW,2BAA2B,aAAa;EACnD,WAAW,iBAAiB,aAAa;EACzC,IAAI,aAAa,KAAK,UAAU,KAAA,KAAa,CAAC,aAAa,KAAK,UAC9D,WAAW,4BAA4B,aAAa,KAAK;EAE3D,4BAA4B,YAAY,aAAa,SAAS;EAE9D,MAAM,cACJ,UAAU,KAAA,IACN;GACE,UAAU,EAAE,gBAAgB,EAAE,MAAM,EAAE;GACtC,YAAY,EAAE,MAAM;GACpB,WAAW;EACb,IACA,KAAA;EAEN,MAAM,kBAAkB,yBAAyB,UAAU;EAC3D,MAAM,aACJ,oBAAoB,WAChB,WAAW,MACR,SACC,CAAC,KAAK,gBAAgB,wBAAwB,IAAI,KAAK,KAAK,QAAQ,SAAS,OACjF,IACA,KAAA;EACN,MAAM,UAAsB;GAC1B,KAAK;GACL,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;EACvC;EACA,IAAI,UAAU,KAAA,GACZ,IAAI,iBAAiB,WAAW,QAAQ,eAAe;OAClD,QAAQ,cAAc;EAG7B,KAAK,KAAK;GACR,OAAO;GACP;GACA,aAAc,KAAK,WAAW,yBAAgD;GAC9E,MAAM;GACN;GACA;GACA;GACA,WAAY,KAAK,WAAW,wBAA+C;GAC3E;GACA;GACA,gBACE,iBAAiB,KAAA,IACb;IAAE,MAAM;IAAc,KAAK;GAAK,IAChC;IAAE,MAAM;IAAY,KAAK;GAAa;GAC5C,YAAY,aAAa;GACzB;GACA,GAAI,aACA,EAAE,uBAAuB,WAAW,QAAQ,WAAW,WAAW,KAAK,IACvE,CAAC;GACL;GACA,GAAG;GACH,UAAU;GACV;EACF,CAAC;CACH;CACA,OAAO;AACT;AAEA,SAAS,yBAAyB,OAA6C;CAC7E,MAAM,QAAQ,MAAM,QAAQ,SAAS,CAAC,KAAK,gBAAgB,wBAAwB,IAAI,CAAC;CACxF,IAAI,MAAM,WAAW,GAAG,OAAO;CAC/B,IAAI,MAAM,EAAE,CAAE,QAAQ,SAAS,MAAM,OAAO;CAC5C,IAAI,MAAM,EAAE,CAAE,QAAQ,SAAS,SAAS,OAAO;CAC/C,OAAO;AACT;AAEA,SAAS,wBAAwB,MAA+B;CAC9D,MAAM,OAAO,iBAAiB,IAAI;CAClC,IAAI,SAAS,SAAS,SAAS,UAAU,SAAS,eAAe,SAAS,aACxE,OAAO;CAET,OAAO;AACT;AAEA,SAAS,aAAa,MAA0C;CAC9D,OAAO,eAAe,CAAC,IAAI,GAAG,CAAC,GAAG,qBAAqB,WAAW,CAAC,CAAC,EAAE,YAAY;AACpF;AAEA,SAAS,iBAAiB,MAAsC;CAC9D,OAAO,qBAAqB;EAC1B,MAAM,aAAa,IAAI;EACvB,MAAM,KAAK;EACX,YAAY,KAAK;CACnB,CAAC;AACH;AAEA,SAAS,aAAa,MAA8B;CAClD,OAAO,GAAG,KAAK,QAAQ,GAAG,KAAK;AACjC;AAEA,SAAS,eAAe,MAA8B;CACpD,OAAO,GAAG,KAAK,QAAQ,GAAG,KAAK;AACjC;AAEA,SAAS,oBAAoB,MAA+B;CAC1D,OAAO,gBAAgB;EACrB,MAAM,aAAa,IAAI;EACvB,MAAM,KAAK;EACX,YAAY,KAAK;CACnB,CAAC;AACH;AAEA,SAAS,oBAAoB,MAA+B;CAC1D,MAAM,OAAO,aAAa,IAAI;CAC9B,OAAO,SAAS,KAAA,KAAa,SAAS;AACxC;AAIA,SAAS,WAAW,OAAwD;CAC1E,MAAM,oBAAI,IAAI,IAA8B;CAC5C,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,MAAO,KAAK,mBAA0C,KAAK;EACjE,MAAM,OAAO,EAAE,IAAI,GAAG,KAAK,CAAC;EAC5B,KAAK,KAAK,IAAI;EACd,EAAE,IAAI,KAAK,IAAI;CACjB;CACA,OAAO;AACT;AAEA,SAAS,SAAS,OAAqD;CACrE,MAAM,kBAAkB,MAAM,QAAQ,SAAS,CAAC,KAAK,YAAY;CACjE,MAAM,gBAAgB,gBAAgB,OAAO,uBAAuB;CAGpE,OAAO,WADL,cAAc,SAAS,IAAI,gBAAgB,gBAAgB,SAAS,IAAI,kBAAkB,KACtE,CAAC,CAAC;AAC1B;AAEA,SAAS,eAAe,OAAyB,MAAoC;CACnF,KAAK,MAAM,QAAQ,OACjB,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,IAAI,KAAK,WAAW;EAC1B,IAAI,OAAO,MAAM,YAAY,EAAE,SAAS,GAAG,OAAO;CACpD;AAGJ;AAEA,SAAS,yBACP,OACA,OACoB;CACpB,MAAM,yBAAS,IAAI,IAAY;CAC/B,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,WAAW,KAAK;EACtB,IAAI,OAAO,aAAa,YAAY,SAAS,SAAS,GAAG,OAAO,IAAI,QAAQ;EAC5E,MAAM,YAAY,KAAK,WAAW;EAClC,IAAI,OAAO,cAAc,YAAY,UAAU,SAAS,GAAG,OAAO,IAAI,SAAS;CACjF;CACA,IAAI,OAAO,OAAO,GAChB,MAAM,IAAI,gBACR,oDAAoD,MAAM,KAAK,CAAC,GAAG,MAAM,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,GAC7F;CAEF,OAAO,OAAO,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC;AAChC;AAOA,SAAS,iBACP,OACA,OACA,aACoB;CACpB,MAAM,eAAe,WAAW,KAAK;CACrC,MAAM,UAA6B,CAAC;CAEpC,IAAI,aAAa;EACf,MAAM,WAAoB,YAAY,OAAO,YAAY;EACzD,IAAI,aAAa,KAAA,GAAW;GAC1B,IAAI,OAAO,aAAa,YAAY,CAAC,OAAO,SAAS,QAAQ,GAC3D,MAAM,IAAI,gBACR,oEAAoE,MAAM,EAC5E;GAEF,QAAQ,KAAK;IAAE,OAAO;IAAe,OAAO;GAAS,CAAC;EACxD;CACF;CAEA,KAAK,MAAM,QAAQ,cAAc;EAC/B,MAAM,OAAO,iBAAiB,IAAI;EAClC,IAAI,KAAK,gBAAgB,SAAS,aAAa;EAC/C,IAAI,SAAS,eAAe,KAAK,QAAQ,SAAS,SAAS;EAC3D,KAAK,MAAM,OAAO,sBAAsB;GACtC,IAAI,CAAC,OAAO,OAAO,KAAK,YAAY,GAAG,GAAG;GAC1C,QAAQ,KAAK;IACX,OAAO,SAAS,KAAK,OAAO,eAAe,IAAI;IAC/C,OAAO,wBAAwB,OAAO,KAAK,QAAQ,KAAK,KAAK,WAAW,IAAI;GAC9E,CAAC;EACH;CACF;CAEA,IAAI,QAAQ,WAAW,GAAG,OAAO,KAAA;CACjC,QAAQ,MAAM,MAAM,UAAU,KAAK,MAAM,cAAc,MAAM,KAAK,CAAC;CACnE,MAAM,QAAQ,QAAQ,EAAE,CAAE;CAC1B,IAAI,QAAQ,MAAM,WAAW,OAAO,UAAU,KAAK,GAEjD,MAAM,IAAI,gBACR,2DAA2D,MAAM,KAFnD,QAAQ,KAAK,WAAW,GAAG,OAAO,MAAM,GAAG,OAAO,OAAO,CAAC,CAAC,KAAK,IAEF,GAC9E;CAEF,OAAO;AACT;AAEA,SAAS,wBACP,OACA,QACA,KACA,OACQ;CACR,MAAM,SAAS,SAAS,OAAO,eAAe,IAAI;CAClD,IAAI,OAAO,UAAU,UAAU;EAC7B,IAAI,MAAM,KAAK,CAAC,CAAC,WAAW,GAC1B,MAAM,IAAI,gBACR,kBAAkB,OAAO,qBAAqB,MAAM,+BACtD;EAEF,MAAM,SAAS,OAAO,KAAK;EAC3B,IAAI,OAAO,SAAS,MAAM,GAAG,OAAO;CACtC,OAAO,IAAI,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK,GAC3D,OAAO;CAET,MAAM,IAAI,gBACR,kBAAkB,OAAO,+CAA+C,MAAM,EAChF;AACF;AAEA,SAAS,WAAW,OAAoD;CACtE,OAAO,CAAC,GAAG,KAAK,CAAC,CAAC,MACf,MAAM,UACL,gBAAgB,KAAK,mBAAmB,MAAM,iBAAiB,KAC/D,KAAK,OAAO,cAAc,MAAM,MAAM,CAC1C;AACF;AAEA,SAAS,cAAc,OAAuB;CAC5C,IAAI;EACF,OAAO,OAAO,KAAK;CACrB,QAAQ;EACN,MAAM,IAAI,gBAAgB,qDAAqD,MAAM,EAAE;CACzF;AACF;AAEA,SAAS,gBAAgB,MAAc,OAAuB;CAC5D,MAAM,YAAY,cAAc,IAAI;CACpC,MAAM,aAAa,cAAc,KAAK;CACtC,OAAO,YAAY,aAAa,KAAK,YAAY,aAAa,IAAI;AACpE;AAEA,SAAS,mBAAmB,OAAe,KAAqB;CAC9D,MAAM,QAAQ,cAAc,GAAG,IAAI,cAAc,KAAK;CACtD,IAAI,SAAS,IAAI,OAAO;CACxB,MAAM,UAAU,QAAQ;CACxB,MAAM,eAAe,QAAQ;CAC7B,MAAM,QAAQ,OAAO,OAAO,IAAI,OAAO,YAAY,IAAI;CACvD,IAAI,CAAC,OAAO,cAAc,OAAO,OAAO,CAAC,GACvC,MAAM,IAAI,gBAAgB,iEAAiE;CAE7F,OAAO;AACT;AAEA,SAAS,oBAAoB,OAAiD;CAC5E,MAAM,MAA8B,CAAC;CACrC,KAAK,MAAM,QAAQ,OACjB,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,KAAK,UAAU,GACjD,IAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,GAAG,IAAI,KAAK;CAG9D,OAAO;AACT"}

package/dist/otlp-attributes-BpC_blLF.js

@@ -0,0 +1,83 @@
+import { INPUT_VALUE, LLM_MODEL_ATTR_KEYS, OUTPUT_VALUE, SPAN_KIND_ATTR_KEYS, TOOL_ARGS_CAPTURED, TOOL_LATENCY_MS, TOOL_NAME, TOOL_NAME_ATTR_KEYS } from "./trace-attributes.js";
+//#region src/trace/otlp-attributes.ts
+/** Canonical OpenInference-over-OTLP attribute vocabulary used at the trace boundary. */
+const TOOL_SPAN_ATTRIBUTE_KEYS = [
+	TOOL_NAME,
+	TOOL_ARGS_CAPTURED,
+	TOOL_LATENCY_MS,
+	INPUT_VALUE,
+	OUTPUT_VALUE
+];
+const EXPLICIT_SPAN_ROLES = /* @__PURE__ */ new Set([
+	"AGENT",
+	"CHAIN",
+	"EVALUATOR",
+	"GUARDRAIL",
+	"LLM",
+	"SPAN",
+	"TOOL"
+]);
+/**
+* Classify a span once for both measurement and error accounting.
+* An explicit OpenInference kind wins; untyped spans use the same tool and
+* model signals in online and offline intake.
+*/
+function classifyOtlpSpanRole(input) {
+	const explicitKind = input.kind ?? firstStringAttribute(input.attributes, SPAN_KIND_ATTR_KEYS);
+	if (explicitKind) {
+		const normalized = explicitKind.toUpperCase();
+		if (EXPLICIT_SPAN_ROLES.has(normalized)) return normalized;
+	}
+	if (firstStringAttribute(input.attributes, TOOL_NAME_ATTR_KEYS) !== void 0 || /^(?:function|tool)[.:/]/i.test(input.name)) return "TOOL";
+	const spanType = input.attributes["span.type"];
+	if (typeof spanType === "string" && spanType.toLowerCase() === "llm_request" || /(?:^|[.:/_-])(?:chat[._-]?completions?|llm)(?:$|[.:/_-])/i.test(input.name) || firstStringAttribute(input.attributes, LLM_MODEL_ATTR_KEYS) !== void 0 || typeof input.attributes["gen_ai.operation.name"] === "string") return "LLM";
+	return "UNKNOWN";
+}
+function isOtlpModelCall(input) {
+	return classifyOtlpSpanRole(input) === "LLM";
+}
+function toolSpanOtlpAttributes(span) {
+	const argsCaptured = span.argsCaptured !== false;
+	const attributes = {
+		[TOOL_NAME]: span.toolName,
+		[TOOL_ARGS_CAPTURED]: argsCaptured
+	};
+	if (span.latencyMs !== void 0) attributes[TOOL_LATENCY_MS] = span.latencyMs;
+	if (argsCaptured) attributes[INPUT_VALUE] = stringifyTraceValue(span.args);
+	if (span.result !== void 0) attributes[OUTPUT_VALUE] = stringifyTraceValue(span.result);
+	return attributes;
+}
+function applyToolSpanOtlpAttributes(attributes, span) {
+	for (const key of TOOL_SPAN_ATTRIBUTE_KEYS) delete attributes[key];
+	Object.assign(attributes, toolSpanOtlpAttributes(span));
+}
+function traceSpanKindToOpenInferenceKind(kind) {
+	switch (kind) {
+		case "llm": return "LLM";
+		case "tool": return "TOOL";
+		case "retrieval": return "CHAIN";
+		case "judge": return "EVALUATOR";
+		case "sandbox": return "CHAIN";
+		case "agent": return "AGENT";
+		default: return "SPAN";
+	}
+}
+function firstStringAttribute(attributes, keys) {
+	for (const key of keys) {
+		const value = attributes[key];
+		if (typeof value === "string" && value.length > 0) return value;
+	}
+}
+function stringifyTraceValue(value) {
+	if (value === void 0) return "null";
+	if (typeof value === "string") return value;
+	try {
+		return JSON.stringify(value) ?? String(value);
+	} catch {
+		return String(value);
+	}
+}
+//#endregion
+export { traceSpanKindToOpenInferenceKind as i, classifyOtlpSpanRole as n, isOtlpModelCall as r, applyToolSpanOtlpAttributes as t };
+
+//# sourceMappingURL=otlp-attributes-BpC_blLF.js.map

package/dist/otlp-attributes-BpC_blLF.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"otlp-attributes-BpC_blLF.js","names":[],"sources":["../src/trace/otlp-attributes.ts"],"sourcesContent":["/** Canonical OpenInference-over-OTLP attribute vocabulary used at the trace boundary. */\n\nimport {\n  INPUT_VALUE,\n  LLM_MODEL_ATTR_KEYS,\n  OUTPUT_VALUE,\n  SPAN_KIND_ATTR_KEYS,\n  TOOL_ARGS_CAPTURED,\n  TOOL_LATENCY_MS,\n  TOOL_NAME,\n  TOOL_NAME_ATTR_KEYS,\n} from './attribute-vocabulary'\nimport type { ToolSpan } from './schema'\n\nexport * from './attribute-vocabulary'\n\nconst TOOL_SPAN_ATTRIBUTE_KEYS = [\n  TOOL_NAME,\n  TOOL_ARGS_CAPTURED,\n  TOOL_LATENCY_MS,\n  INPUT_VALUE,\n  OUTPUT_VALUE,\n] as const\n\nexport type ToolSpanOtlpInput = Pick<\n  ToolSpan,\n  'toolName' | 'args' | 'argsCaptured' | 'result' | 'latencyMs'\n>\n\nexport type OtlpSpanRole =\n  | 'AGENT'\n  | 'CHAIN'\n  | 'EVALUATOR'\n  | 'GUARDRAIL'\n  | 'LLM'\n  | 'SPAN'\n  | 'TOOL'\n  | 'UNKNOWN'\n\nexport interface OtlpSpanRoleInput {\n  name: string\n  attributes: Record<string, unknown>\n  kind?: string | null\n}\n\nconst EXPLICIT_SPAN_ROLES = new Set<OtlpSpanRole>([\n  'AGENT',\n  'CHAIN',\n  'EVALUATOR',\n  'GUARDRAIL',\n  'LLM',\n  'SPAN',\n  'TOOL',\n])\n\n/**\n * Classify a span once for both measurement and error accounting.\n * An explicit OpenInference kind wins; untyped spans use the same tool and\n * model signals in online and offline intake.\n */\nexport function classifyOtlpSpanRole(input: OtlpSpanRoleInput): OtlpSpanRole {\n  const explicitKind = input.kind ?? firstStringAttribute(input.attributes, SPAN_KIND_ATTR_KEYS)\n  if (explicitKind) {\n    const normalized = explicitKind.toUpperCase() as OtlpSpanRole\n    if (EXPLICIT_SPAN_ROLES.has(normalized)) return normalized\n  }\n\n  if (\n    firstStringAttribute(input.attributes, TOOL_NAME_ATTR_KEYS) !== undefined ||\n    /^(?:function|tool)[.:/]/i.test(input.name)\n  ) {\n    return 'TOOL'\n  }\n\n  const spanType = input.attributes['span.type']\n  if (\n    (typeof spanType === 'string' && spanType.toLowerCase() === 'llm_request') ||\n    /(?:^|[.:/_-])(?:chat[._-]?completions?|llm)(?:$|[.:/_-])/i.test(input.name) ||\n    firstStringAttribute(input.attributes, LLM_MODEL_ATTR_KEYS) !== undefined ||\n    typeof input.attributes['gen_ai.operation.name'] === 'string'\n  ) {\n    return 'LLM'\n  }\n\n  return 'UNKNOWN'\n}\n\nexport function isOtlpModelCall(input: OtlpSpanRoleInput): boolean {\n  return classifyOtlpSpanRole(input) === 'LLM'\n}\n\nfunction toolSpanOtlpAttributes(\n  span: ToolSpanOtlpInput,\n): Record<string, string | number | boolean> {\n  const argsCaptured = span.argsCaptured !== false\n  const attributes: Record<string, string | number | boolean> = {\n    [TOOL_NAME]: span.toolName,\n    [TOOL_ARGS_CAPTURED]: argsCaptured,\n  }\n  if (span.latencyMs !== undefined) attributes[TOOL_LATENCY_MS] = span.latencyMs\n  if (argsCaptured) attributes[INPUT_VALUE] = stringifyTraceValue(span.args)\n  if (span.result !== undefined) attributes[OUTPUT_VALUE] = stringifyTraceValue(span.result)\n  return attributes\n}\n\nexport function applyToolSpanOtlpAttributes(\n  attributes: Record<string, unknown>,\n  span: ToolSpanOtlpInput,\n): void {\n  for (const key of TOOL_SPAN_ATTRIBUTE_KEYS) delete attributes[key]\n  Object.assign(attributes, toolSpanOtlpAttributes(span))\n}\n\nexport function traceSpanKindToOpenInferenceKind(kind: string): string {\n  switch (kind) {\n    case 'llm':\n      return 'LLM'\n    case 'tool':\n      return 'TOOL'\n    case 'retrieval':\n      return 'CHAIN'\n    case 'judge':\n      return 'EVALUATOR'\n    case 'sandbox':\n      return 'CHAIN'\n    case 'agent':\n      return 'AGENT'\n    default:\n      return 'SPAN'\n  }\n}\n\nfunction firstStringAttribute(\n  attributes: Record<string, unknown>,\n  keys: readonly string[],\n): string | undefined {\n  for (const key of keys) {\n    const value = attributes[key]\n    if (typeof value === 'string' && value.length > 0) return value\n  }\n  return undefined\n}\n\nfunction stringifyTraceValue(value: unknown): string {\n  if (value === undefined) return 'null'\n  if (typeof value === 'string') return value\n  try {\n    return JSON.stringify(value) ?? String(value)\n  } catch {\n    return String(value)\n  }\n}\n"],"mappings":";;;AAgBA,MAAM,2BAA2B;CAC/B;CACA;CACA;CACA;CACA;AACF;AAuBA,MAAM,sCAAsB,IAAI,IAAkB;CAChD;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;;;;;;AAOD,SAAgB,qBAAqB,OAAwC;CAC3E,MAAM,eAAe,MAAM,QAAQ,qBAAqB,MAAM,YAAY,mBAAmB;CAC7F,IAAI,cAAc;EAChB,MAAM,aAAa,aAAa,YAAY;EAC5C,IAAI,oBAAoB,IAAI,UAAU,GAAG,OAAO;CAClD;CAEA,IACE,qBAAqB,MAAM,YAAY,mBAAmB,MAAM,KAAA,KAChE,2BAA2B,KAAK,MAAM,IAAI,GAE1C,OAAO;CAGT,MAAM,WAAW,MAAM,WAAW;CAClC,IACG,OAAO,aAAa,YAAY,SAAS,YAAY,MAAM,iBAC5D,4DAA4D,KAAK,MAAM,IAAI,KAC3E,qBAAqB,MAAM,YAAY,mBAAmB,MAAM,KAAA,KAChE,OAAO,MAAM,WAAW,6BAA6B,UAErD,OAAO;CAGT,OAAO;AACT;AAEA,SAAgB,gBAAgB,OAAmC;CACjE,OAAO,qBAAqB,KAAK,MAAM;AACzC;AAEA,SAAS,uBACP,MAC2C;CAC3C,MAAM,eAAe,KAAK,iBAAiB;CAC3C,MAAM,aAAwD;GAC3D,YAAY,KAAK;GACjB,qBAAqB;CACxB;CACA,IAAI,KAAK,cAAc,KAAA,GAAW,WAAW,mBAAmB,KAAK;CACrE,IAAI,cAAc,WAAW,eAAe,oBAAoB,KAAK,IAAI;CACzE,IAAI,KAAK,WAAW,KAAA,GAAW,WAAW,gBAAgB,oBAAoB,KAAK,MAAM;CACzF,OAAO;AACT;AAEA,SAAgB,4BACd,YACA,MACM;CACN,KAAK,MAAM,OAAO,0BAA0B,OAAO,WAAW;CAC9D,OAAO,OAAO,YAAY,uBAAuB,IAAI,CAAC;AACxD;AAEA,SAAgB,iCAAiC,MAAsB;CACrE,QAAQ,MAAR;EACE,KAAK,OACH,OAAO;EACT,KAAK,QACH,OAAO;EACT,KAAK,aACH,OAAO;EACT,KAAK,SACH,OAAO;EACT,KAAK,WACH,OAAO;EACT,KAAK,SACH,OAAO;EACT,SACE,OAAO;CACX;AACF;AAEA,SAAS,qBACP,YACA,MACoB;CACpB,KAAK,MAAM,OAAO,MAAM;EACtB,MAAM,QAAQ,WAAW;EACzB,IAAI,OAAO,UAAU,YAAY,MAAM,SAAS,GAAG,OAAO;CAC5D;AAEF;AAEA,SAAS,oBAAoB,OAAwB;CACnD,IAAI,UAAU,KAAA,GAAW,OAAO;CAChC,IAAI,OAAO,UAAU,UAAU,OAAO;CACtC,IAAI;EACF,OAAO,KAAK,UAAU,KAAK,KAAK,OAAO,KAAK;CAC9C,QAAQ;EACN,OAAO,OAAO,KAAK;CACrB;AACF"}

package/dist/{run-record-BhMVsHR1.js → paired-promotion-decision-ClJh_9nK.js}

@@ -1,6 +1,5 @@
 import { g as pairedRiskDifferenceScore, h as pairedRiskDifferenceExact, p as pairedBinaryScale } from "./paired-arms-JSDcs2zx.js";
 import { a as pairedSignTest, i as pairedDeltaTieFraction, r as pairedBootstrap } from "./paired-tests-C8iCsioC.js";
-import { c as validateRunRecord } from "./run-record-tg_5H2IK.js";
 //#region src/paired-delta-test.ts
 /** Smallest all-positive sample that can clear a one-sided exact sign test. */
 function minimumPairsForPairedDeltaTest(confidence = .95) {
@@ -247,218 +246,6 @@ function fmt(x) {
 	return x.toFixed(4);
 }
 //#endregion
-//#region src/campaign/run-record.ts
-/**
-* A campaign cell carried a judge score without a `dimensions` record.
-*
-* Two exported types share the name `JudgeScore`: the campaign verdict
-* (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)
-* and the root export's flat per-dimension row (`{ judgeName, dimension, score }`
-* from `@tangle-network/agent-eval`). Campaign aggregation accepts only the
-* campaign shape; the flat shape previously crashed here with an opaque
-* TypeError deep inside aggregation.
-*/
-var CampaignJudgeScoreShapeError = class extends TypeError {
-	constructor(judgeName) {
-		super(`campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. Use the campaign JudgeScore ({ dimensions, composite, notes }) from '@tangle-network/agent-eval/campaign'; the root export's JudgeScore ({ judgeName, dimension, score }) is a different type with the same name.`);
-		this.name = "CampaignJudgeScoreShapeError";
-	}
-};
-/**
-* Project one campaign cell into the canonical run format.
-*
-* A dispatch error establishes terminal execution failure. A judge error only
-* establishes that quality measurement failed after dispatch completed.
-* Failures without a stage remain unknown. No failure becomes a zero-quality
-* label.
-*/
-function campaignCellToRunRecord(cell, options) {
-	const quality = projectCampaignCellQuality(cell);
-	const execution = campaignCellExecutionEvidence(cell);
-	const judgeErrorCount = Math.max(quality.raw.judge_error_count ?? 0, execution.judgeErrorCount ?? 0);
-	const cellCostProvenance = campaignCellCostProvenance(cell);
-	const costProvenance = cellCostProvenance.kind === "uncaptured" && options.defaultCostUsd !== void 0 ? {
-		kind: "estimated",
-		usd: options.defaultCostUsd
-	} : cellCostProvenance;
-	const costUsd = costProvenance.kind === "uncaptured" ? null : costProvenance.usd;
-	const raw = {
-		...finiteMetrics(options.raw),
-		...quality.raw,
-		rep: cell.rep,
-		duration_ms: cell.durationMs,
-		...costUsd === null ? {} : { cost_usd: costUsd },
-		...cellCostProvenance.kind === "uncaptured" ? { cost_known_subtotal_usd: cell.costUsd } : {},
-		cost_observed: costProvenance.kind === "observed" ? 1 : 0,
-		cost_estimated: costProvenance.kind === "estimated" ? 1 : 0,
-		cost_uncaptured: costProvenance.kind === "uncaptured" ? 1 : 0,
-		tokens_input: cell.tokenUsage.input,
-		tokens_output: cell.tokenUsage.output,
-		tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,
-		latency_ms: cell.durationMs,
-		...execution.executionErrorCount === void 0 ? {} : { execution_error_count: execution.executionErrorCount },
-		...judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {},
-		...execution.unclassifiedErrorCount === void 0 ? {} : { unclassified_error_count: execution.unclassifiedErrorCount }
-	};
-	if (typeof cell.generation === "number") raw.generation = cell.generation;
-	if (cell.tokenUsage.reasoning !== void 0) raw.tokens_reasoning = cell.tokenUsage.reasoning;
-	if (cell.tokenUsage.cached !== void 0) raw.tokens_cached = cell.tokenUsage.cached;
-	if (cell.tokenUsage.cacheWrite !== void 0) raw.tokens_cache_write = cell.tokenUsage.cacheWrite;
-	if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd;
-	if (costUsd !== null && quality.score !== void 0 && quality.score > .01) raw.cost_per_quality = costUsd / quality.score;
-	const outcome = {
-		raw,
-		...quality.judgeScores ? { judgeScores: quality.judgeScores } : {}
-	};
-	if (quality.score !== void 0) if (options.splitTag === "holdout") outcome.holdoutScore = quality.score;
-	else outcome.searchScore = quality.score;
-	return validateRunRecord({
-		runId: options.runId,
-		experimentId: options.experimentId,
-		candidateId: options.candidateId,
-		seed: options.seed ?? cell.seed,
-		model: options.model,
-		promptHash: options.promptHash,
-		configHash: options.configHash,
-		commitSha: options.commitSha,
-		wallMs: cell.durationMs,
-		costUsd,
-		costProvenance,
-		tokenUsage: { ...cell.tokenUsage },
-		terminalOutcome: execution.terminalOutcome,
-		...execution.terminalFailureReason ? { terminalFailureReason: execution.terminalFailureReason } : {},
-		outcome,
-		splitTag: options.splitTag,
-		scenarioId: options.scenarioId ?? cell.scenarioId,
-		...options.agentProfile ? { agentProfile: options.agentProfile } : {}
-	});
-}
-/**
-* Validate the cost fields that cross campaign cache and RunRecord boundaries.
-* `costUsd` is a known subtotal for uncaptured cells, but it must equal the
-* authoritative total whenever that total is observed or estimated.
-*/
-function campaignCellCostProvenance(cell) {
-	if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`);
-	const provenance = cell.costProvenance;
-	if (!provenance || typeof provenance !== "object") throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`);
-	if (provenance.kind === "uncaptured") {
-		if (provenance.usd !== null) throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`);
-		return {
-			kind: "uncaptured",
-			usd: null
-		};
-	}
-	if (provenance.kind !== "observed" && provenance.kind !== "estimated" || !Number.isFinite(provenance.usd) || provenance.usd < 0) throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`);
-	if (provenance.usd !== cell.costUsd) throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`);
-	return {
-		kind: provenance.kind,
-		usd: provenance.usd
-	};
-}
-function campaignCellExecutionEvidence(cell) {
-	if (cell.errorStage === "dispatch") return {
-		terminalOutcome: "failed",
-		executionErrorCount: 1,
-		...cell.error ? { terminalFailureReason: cell.error } : {}
-	};
-	if (cell.errorStage === "judge") return {
-		terminalOutcome: "succeeded",
-		executionErrorCount: 0,
-		judgeErrorCount: 1
-	};
-	if (!cell.error) return {
-		terminalOutcome: "succeeded",
-		executionErrorCount: 0
-	};
-	return {
-		terminalOutcome: "unknown",
-		unclassifiedErrorCount: 1
-	};
-}
-/**
-* Produce the only task-quality view used by campaign aggregates and exports.
-*
-* Successful judge results remain available for diagnosis after another judge
-* fails, but a task score exists only for an error-free cell whose reported
-* judge values are all finite.
-*/
-function projectCampaignCellQuality(cell) {
-	if (cell.errorStage === "dispatch") return {
-		successfulJudgeScores: {},
-		failedJudges: [],
-		raw: {}
-	};
-	const perJudge = {};
-	const successfulJudgeScores = {};
-	const dimensionValues = /* @__PURE__ */ new Map();
-	const composites = [];
-	const notes = [];
-	const failedJudges = new Set(cell.errorStage === "judge" ? [cell.errorJudge ?? "unknown-judge"] : []);
-	const raw = {};
-	for (const [judgeName, score] of Object.entries(cell.judgeScores)) {
-		const dimensionsShape = score.dimensions;
-		if (typeof dimensionsShape !== "object" || dimensionsShape === null || Array.isArray(dimensionsShape)) throw new CampaignJudgeScoreShapeError(judgeName);
-		const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite);
-		if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {
-			failedJudges.add(judgeName);
-			continue;
-		}
-		composites.push(score.composite);
-		successfulJudgeScores[judgeName] = score;
-		const dimensions = { ...score.dimensions };
-		perJudge[judgeName] = dimensions;
-		for (const [dimension, value] of Object.entries(dimensions)) {
-			raw[`${judgeName}.${dimension}`] = value;
-			const values = dimensionValues.get(dimension) ?? [];
-			values.push(value);
-			dimensionValues.set(dimension, values);
-		}
-		if (score.notes) notes.push(`${judgeName}: ${score.notes}`);
-		for (const failedJudge of score.failedJudges ?? []) failedJudges.add(`${judgeName}/${failedJudge}`);
-	}
-	if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size;
-	const sortedFailedJudges = [...failedJudges].sort();
-	if (composites.length === 0) return {
-		successfulJudgeScores,
-		failedJudges: sortedFailedJudges,
-		raw
-	};
-	const composite = mean(composites);
-	const perDimMean = Object.fromEntries([...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]));
-	const complete = cell.error === void 0 && cell.errorStage === void 0 && failedJudges.size === 0;
-	if (complete) raw.composite = composite;
-	return {
-		...complete ? { score: composite } : {},
-		raw,
-		successfulJudgeScores,
-		failedJudges: sortedFailedJudges,
-		judgeScores: {
-			perJudge,
-			perDimMean,
-			composite,
-			...sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {},
-			...notes.length > 0 ? { notes: notes.join(" | ") } : {}
-		}
-	};
-}
-/** Read the canonical task score without recomputing cell quality. */
-function campaignCellTaskScore(cell) {
-	return projectCampaignCellQuality(cell).score;
-}
-/** Read canonical successful judge dimensions without recomputing cell quality. */
-function campaignCellJudgeDimensions(cell) {
-	return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {};
-}
-function finiteMetrics(metrics) {
-	const finite = {};
-	for (const [key, value] of Object.entries(metrics ?? {})) if (Number.isFinite(value)) finite[key] = value;
-	return finite;
-}
-function mean(values) {
-	return values.reduce((sum, value) => sum + value, 0) / values.length;
-}
-//#endregion
-export { campaignCellToRunRecord as a, pairedDecisionShape as c, campaignCellTaskScore as i, minimumPairsForPairedDeltaTest as l, campaignCellExecutionEvidence as n, projectCampaignCellQuality as o, campaignCellJudgeDimensions as r, decidePairedPromotion as s, campaignCellCostProvenance as t, pairedDeltaTest as u };
+export { pairedDeltaTest as i, pairedDecisionShape as n, minimumPairsForPairedDeltaTest as r, decidePairedPromotion as t };
 
-//# sourceMappingURL=run-record-BhMVsHR1.js.map
+//# sourceMappingURL=paired-promotion-decision-ClJh_9nK.js.map