@tangle-network/agent-eval 0.129.0 → 0.130.1

package/dist/integrity-BzRbCHzi.js

@@ -0,0 +1,424 @@
+import { n as CaptureIntegrityError } from "./errors-8YnH8WlF.js";
+//#region src/trace/store.ts
+var InMemoryTraceStore = class {
+	runs = /* @__PURE__ */ new Map();
+	allSpans = [];
+	allEvents = [];
+	allArtifacts = [];
+	allBudget = [];
+	async appendRun(run) {
+		if (this.runs.has(run.runId)) throw new Error(`run ${run.runId} already exists`);
+		this.runs.set(run.runId, { ...run });
+	}
+	async updateRun(runId, patch) {
+		const existing = this.runs.get(runId);
+		if (!existing) throw new Error(`run ${runId} not found`);
+		this.runs.set(runId, {
+			...existing,
+			...patch
+		});
+	}
+	async appendSpan(span) {
+		this.allSpans.push({ ...span });
+	}
+	async updateSpan(spanId, patch) {
+		const idx = this.allSpans.findIndex((s) => s.spanId === spanId);
+		if (idx < 0) throw new Error(`span ${spanId} not found`);
+		this.allSpans[idx] = {
+			...this.allSpans[idx],
+			...patch
+		};
+	}
+	async appendEvent(event) {
+		this.allEvents.push({ ...event });
+	}
+	async appendArtifact(artifact) {
+		this.allArtifacts.push({ ...artifact });
+	}
+	async appendBudgetEntry(entry) {
+		this.allBudget.push({ ...entry });
+	}
+	async getRun(runId) {
+		const r = this.runs.get(runId);
+		return r ? { ...r } : void 0;
+	}
+	async listRuns(filter = {}) {
+		return [...this.runs.values()].filter((r) => matchesRun(r, filter));
+	}
+	async spans(filter = {}) {
+		return this.allSpans.filter((s) => matchesSpan(s, filter)).map((s) => ({ ...s }));
+	}
+	async events(filter = {}) {
+		return this.allEvents.filter((e) => matchesEvent(e, filter)).map((e) => ({ ...e }));
+	}
+	async budget(runId) {
+		return this.allBudget.filter((b) => b.runId === runId).map((b) => ({ ...b }));
+	}
+	async artifacts(runId) {
+		return this.allArtifacts.filter((a) => a.runId === runId).map((a) => ({ ...a }));
+	}
+};
+function matchesRun(r, f) {
+	if (f.scenarioId && r.scenarioId !== f.scenarioId) return false;
+	if (f.variantId && r.variantId !== f.variantId) return false;
+	if (f.status && r.status !== f.status) return false;
+	if (f.since !== void 0 && r.startedAt < f.since) return false;
+	if (f.until !== void 0 && r.startedAt > f.until) return false;
+	if (f.tag && r.tags?.[f.tag.key] !== f.tag.value) return false;
+	if (f.parentRunId && r.parentRunId !== f.parentRunId) return false;
+	if (f.projectId && r.projectId !== f.projectId) return false;
+	if (f.chatId && r.chatId !== f.chatId) return false;
+	if (f.layer && r.layer !== f.layer) return false;
+	return true;
+}
+function matchesSpan(s, f) {
+	if (f.runId && s.runId !== f.runId) return false;
+	if (f.parentSpanId && s.parentSpanId !== f.parentSpanId) return false;
+	if (f.kind && s.kind !== f.kind) return false;
+	if (f.name && s.name !== f.name) return false;
+	if (f.toolName && (s.kind !== "tool" || s.toolName !== f.toolName)) return false;
+	if (f.judgeId && (s.kind !== "judge" || s.judgeId !== f.judgeId)) return false;
+	if (f.since !== void 0 && s.startedAt < f.since) return false;
+	if (f.until !== void 0 && s.startedAt > f.until) return false;
+	return true;
+}
+function matchesEvent(e, f) {
+	if (f.runId && e.runId !== f.runId) return false;
+	if (f.spanId && e.spanId !== f.spanId) return false;
+	if (f.kind && e.kind !== f.kind) return false;
+	if (f.since !== void 0 && e.timestamp < f.since) return false;
+	if (f.until !== void 0 && e.timestamp > f.until) return false;
+	return true;
+}
+var FileSystemTraceStore = class {
+	dir;
+	maxBytes;
+	/** Lazy in-memory index for queries — populated on first read. */
+	index;
+	/** Memoized index build — concurrent first reads share one build, and an
+	*  append racing an in-flight load awaits this so its row isn't lost. */
+	indexPromise;
+	/** Strictly-increasing rollover stamp. Date.now() alone collides when two
+	*  rollovers land in the same millisecond, overwriting a rolled file. */
+	lastRolloverStamp = 0;
+	/**
+	* Per-file append serialization. stat → conditional-rename → appendFile is a
+	* read-modify-write on the active file; without a lock two concurrent appends
+	* to the same `name` can both pass the size check (exceeding maxBytes) or one
+	* can append to a file the other just renamed away. Each entry chains the
+	* next append behind the prior one for that file.
+	*/
+	appendLocks = /* @__PURE__ */ new Map();
+	constructor(options) {
+		this.dir = options.dir;
+		this.maxBytes = options.maxBytes ?? 32 * 1024 * 1024;
+	}
+	async ensureDir() {
+		await (await import("node:fs/promises")).mkdir(this.dir, { recursive: true });
+	}
+	async append(name, record) {
+		if (this.indexPromise) await this.indexPromise;
+		const result = (this.appendLocks.get(name) ?? Promise.resolve()).then(() => this.appendLocked(name, record));
+		const tail = result.then(() => {}, () => {});
+		this.appendLocks.set(name, tail);
+		try {
+			await result;
+		} finally {
+			if (this.appendLocks.get(name) === tail) this.appendLocks.delete(name);
+		}
+	}
+	async appendLocked(name, record) {
+		await this.ensureDir();
+		const fs = await import("node:fs/promises");
+		const path = await import("node:path");
+		const active = path.join(this.dir, `${name}.ndjson`);
+		try {
+			if ((await fs.stat(active)).size >= this.maxBytes) {
+				const stamp = Math.max(Date.now(), this.lastRolloverStamp + 1);
+				this.lastRolloverStamp = stamp;
+				const rolled = path.join(this.dir, `${name}.${stamp}.ndjson`);
+				await fs.rename(active, rolled);
+			}
+		} catch {}
+		await fs.appendFile(active, `${JSON.stringify(record)}\n`, "utf8");
+		if (this.index && !record?._update) await this.insertInto(name, record);
+	}
+	async insertInto(name, record) {
+		if (!this.index) return;
+		switch (name) {
+			case "runs":
+				await this.index.appendRun(record);
+				break;
+			case "spans":
+				await this.index.appendSpan(record);
+				break;
+			case "events":
+				await this.index.appendEvent(record);
+				break;
+			case "artifacts":
+				await this.index.appendArtifact(record);
+				break;
+			case "budget":
+				await this.index.appendBudgetEntry(record);
+				break;
+		}
+	}
+	load() {
+		this.indexPromise ??= this.buildIndex();
+		return this.indexPromise;
+	}
+	async buildIndex() {
+		const fs = await import("node:fs/promises");
+		const path = await import("node:path");
+		const store = new InMemoryTraceStore();
+		let entries;
+		try {
+			entries = await fs.readdir(this.dir);
+		} catch {
+			this.index = store;
+			return store;
+		}
+		entries.sort((a, b) => {
+			const baseA = a.split(".")[0];
+			const baseB = b.split(".")[0];
+			if (baseA !== baseB) return baseA < baseB ? -1 : 1;
+			const ta = a.match(/^[^.]+\.(\d+)\.ndjson$/);
+			const tb = b.match(/^[^.]+\.(\d+)\.ndjson$/);
+			return (ta ? Number(ta[1]) : Number.POSITIVE_INFINITY) - (tb ? Number(tb[1]) : Number.POSITIVE_INFINITY);
+		});
+		{
+			const runUpdates = [];
+			const spanUpdates = [];
+			for (const file of entries) {
+				if (!file.endsWith(".ndjson")) continue;
+				const full = path.join(this.dir, file);
+				const content = await fs.readFile(full, "utf8");
+				const base = file.split(".")[0];
+				const lines = content.split("\n");
+				for (let ln = 0; ln < lines.length; ln++) {
+					const line = lines[ln];
+					if (!line.trim()) continue;
+					let record;
+					try {
+						record = JSON.parse(line);
+					} catch (err) {
+						throw new Error(`FileSystemTraceStore: corrupt NDJSON in ${file} line ${ln + 1}: ${err instanceof Error ? err.message : String(err)} — ${line.slice(0, 120)}`);
+					}
+					if (base === "runs") {
+						if (record?._update) {
+							runUpdates.push(record);
+							continue;
+						}
+						try {
+							await store.appendRun(record);
+						} catch {
+							await store.updateRun(record.runId, record);
+						}
+					} else if (base === "spans") {
+						if (record?._update) {
+							spanUpdates.push(record);
+							continue;
+						}
+						await store.appendSpan(record);
+					} else if (base === "events") await store.appendEvent(record);
+					else if (base === "artifacts") await store.appendArtifact(record);
+					else if (base === "budget") await store.appendBudgetEntry(record);
+				}
+			}
+			for (const record of runUpdates) await store.updateRun(record.runId, record);
+			for (const record of spanUpdates) await store.updateSpan(record.spanId, record);
+		}
+		this.index = store;
+		return store;
+	}
+	async appendRun(run) {
+		await this.append("runs", run);
+	}
+	async updateRun(runId, patch) {
+		await this.append("runs", {
+			runId,
+			...patch,
+			_update: true
+		});
+		if (this.index) await this.index.updateRun(runId, patch);
+	}
+	async appendSpan(span) {
+		await this.append("spans", span);
+	}
+	async updateSpan(spanId, patch) {
+		await this.append("spans", {
+			spanId,
+			...patch,
+			_update: true
+		});
+		if (this.index) await this.index.updateSpan(spanId, patch);
+	}
+	async appendEvent(event) {
+		await this.append("events", event);
+	}
+	async appendArtifact(artifact) {
+		await this.append("artifacts", artifact);
+	}
+	async appendBudgetEntry(entry) {
+		await this.append("budget", entry);
+	}
+	async getRun(runId) {
+		return (await this.load()).getRun(runId);
+	}
+	async listRuns(filter) {
+		return (await this.load()).listRuns(filter);
+	}
+	async spans(filter) {
+		return (await this.load()).spans(filter);
+	}
+	async events(filter) {
+		return (await this.load()).events(filter);
+	}
+	async budget(runId) {
+		return (await this.load()).budget(runId);
+	}
+	async artifacts(runId) {
+		return (await this.load()).artifacts(runId);
+	}
+};
+//#endregion
+//#region src/trace/integrity.ts
+/**
+* Run-completion integrity check — at end of run, verify the expected event
+* types were actually captured. The point is the launch-review failure mode:
+* a run *appears* successful but the raw provider events were never written,
+* so a downstream reviewer can't reconstruct what happened.
+*
+* Pattern:
+*
+*   const report = await assertRunCaptured(store, runId, {
+*     llmSpansMin: 1,
+*     judgeSpansMin: 1,
+*     rawSink: providerSink,                  // must have ≥ 1 event for this run
+*     requireRawCoverageOfLlmSpans: true,     // every llm span has matching raw events
+*   })
+*   if (!report.ok) throwIfRunIncomplete(report)  // or mark run failed and continue
+*
+* The function is read-only on the store and returns a structured report;
+* the caller chooses the failure mode (throw, mark run failed, log warning).
+* `throwIfRunIncomplete` is the convenient strict mode.
+*/
+var RunIntegrityError = class extends CaptureIntegrityError {
+	report;
+	constructor(report) {
+		super(`Run ${report.runId} failed integrity check: ${report.issues.map((i) => i.code).join(", ")}`);
+		this.report = report;
+	}
+};
+async function assertRunCaptured(store, runId, expectations = {}) {
+	const issues = [];
+	const run = await store.getRun(runId);
+	if (!run) return {
+		ok: false,
+		runId,
+		llmSpanCount: 0,
+		judgeSpanCount: 0,
+		toolSpanCount: 0,
+		rawProviderEventCount: 0,
+		rawSpanCoverage: {
+			covered: 0,
+			total: 0
+		},
+		issues: [{
+			code: "no_run",
+			message: `Run ${runId} not found in store.`
+		}]
+	};
+	const spans = await store.spans({ runId });
+	const llmSpans = spans.filter((s) => s.kind === "llm");
+	const judgeSpans = spans.filter((s) => s.kind === "judge");
+	const toolSpans = spans.filter((s) => s.kind === "tool");
+	const llmMin = expectations.llmSpansMin ?? 0;
+	const judgeMin = expectations.judgeSpansMin ?? 0;
+	const toolMin = expectations.toolSpansMin ?? 0;
+	if (llmSpans.length < llmMin) issues.push({
+		code: "missing_llm_spans",
+		message: `Expected ≥ ${llmMin} LLM spans, found ${llmSpans.length}.`,
+		detail: {
+			expected: llmMin,
+			found: llmSpans.length
+		}
+	});
+	if (judgeSpans.length < judgeMin) issues.push({
+		code: "missing_judge_spans",
+		message: `Expected ≥ ${judgeMin} judge spans, found ${judgeSpans.length}.`,
+		detail: {
+			expected: judgeMin,
+			found: judgeSpans.length
+		}
+	});
+	if (toolSpans.length < toolMin) issues.push({
+		code: "missing_tool_spans",
+		message: `Expected ≥ ${toolMin} tool spans, found ${toolSpans.length}.`,
+		detail: {
+			expected: toolMin,
+			found: toolSpans.length
+		}
+	});
+	let rawEventCount = 0;
+	let coverage = {
+		covered: 0,
+		total: llmSpans.length
+	};
+	if (expectations.rawSink) if (!expectations.rawSink.list) issues.push({
+		code: "no_raw_sink",
+		message: "Provided rawSink does not implement list(); cannot verify capture."
+	});
+	else {
+		const events = await expectations.rawSink.list({ runId });
+		rawEventCount = events.length;
+		const rawMin = expectations.rawProviderEventsMin ?? 1;
+		if (rawEventCount < rawMin) issues.push({
+			code: "missing_raw_events",
+			message: `Expected ≥ ${rawMin} raw provider events, found ${rawEventCount}.`,
+			detail: {
+				expected: rawMin,
+				found: rawEventCount
+			}
+		});
+		if (expectations.requireRawCoverageOfLlmSpans) {
+			const requestEventsBySpan = new Set(events.filter((e) => e.direction === "request" && e.spanId).map((e) => e.spanId));
+			const orphaned = llmSpans.filter((s) => !requestEventsBySpan.has(s.spanId));
+			coverage = {
+				covered: llmSpans.length - orphaned.length,
+				total: llmSpans.length
+			};
+			if (orphaned.length > 0) issues.push({
+				code: "orphan_llm_span",
+				message: `${orphaned.length} LLM span(s) have no matching raw provider request event.`,
+				detail: { orphanedSpanIds: orphaned.map((s) => s.spanId) }
+			});
+		}
+	}
+	else if (expectations.requireRawCoverageOfLlmSpans || expectations.rawProviderEventsMin) issues.push({
+		code: "no_raw_sink",
+		message: "Raw coverage required but no rawSink supplied to the integrity check."
+	});
+	if (expectations.requireOutcome && (run.outcome === void 0 || run.outcome === null)) issues.push({
+		code: "missing_outcome",
+		message: `Run ${runId} has no outcome recorded.`
+	});
+	return {
+		ok: issues.length === 0,
+		runId,
+		llmSpanCount: llmSpans.length,
+		judgeSpanCount: judgeSpans.length,
+		toolSpanCount: toolSpans.length,
+		rawProviderEventCount: rawEventCount,
+		rawSpanCoverage: coverage,
+		issues
+	};
+}
+/** Strict mode: throws `RunIntegrityError` when the report isn't ok. */
+function throwIfRunIncomplete(report) {
+	if (!report.ok) throw new RunIntegrityError(report);
+}
+//#endregion
+export { InMemoryTraceStore as a, FileSystemTraceStore as i, assertRunCaptured as n, throwIfRunIncomplete as r, RunIntegrityError as t };
+
+//# sourceMappingURL=integrity-BzRbCHzi.js.map

package/dist/integrity-BzRbCHzi.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"integrity-BzRbCHzi.js","names":[],"sources":["../src/trace/store.ts","../src/trace/integrity.ts"],"sourcesContent":["/**\n * TraceStore — persistence + query over the TraceSchema v1 corpus.\n *\n * Two implementations ship in the core:\n *   - InMemoryTraceStore: dev + tests, fully in-process\n *   - FileSystemTraceStore: NDJSON append-only files per entity, suitable\n *     for long-running CI jobs; rolled over at 32MB\n *\n * Downstream adapters (DuckDB, Langfuse, R2 parquet) implement this same\n * interface — the rest of the framework is storage-agnostic.\n */\n\nimport type {\n  Artifact,\n  BudgetLedgerEntry,\n  EventKind,\n  Run,\n  RunStatus,\n  Span,\n  SpanKind,\n  TraceEvent,\n} from './schema'\n\nexport interface RunFilter {\n  scenarioId?: string\n  variantId?: string\n  status?: RunStatus\n  since?: number\n  until?: number\n  tag?: { key: string; value: string }\n  parentRunId?: string\n  projectId?: string\n  chatId?: string\n  layer?: import('./schema').RunLayer\n}\n\nexport interface SpanFilter {\n  runId?: string\n  parentSpanId?: string\n  kind?: SpanKind\n  name?: string\n  toolName?: string\n  judgeId?: string\n  since?: number\n  until?: number\n}\n\nexport interface EventFilter {\n  runId?: string\n  spanId?: string\n  kind?: EventKind\n  since?: number\n  until?: number\n}\n\nexport interface TraceStore {\n  appendRun(run: Run): Promise<void>\n  updateRun(runId: string, patch: Partial<Run>): Promise<void>\n  appendSpan(span: Span): Promise<void>\n  updateSpan(spanId: string, patch: Partial<Span>): Promise<void>\n  appendEvent(event: TraceEvent): Promise<void>\n  appendArtifact(artifact: Artifact): Promise<void>\n  appendBudgetEntry(entry: BudgetLedgerEntry): Promise<void>\n\n  getRun(runId: string): Promise<Run | undefined>\n  listRuns(filter?: RunFilter): Promise<Run[]>\n  spans(filter?: SpanFilter): Promise<Span[]>\n  events(filter?: EventFilter): Promise<TraceEvent[]>\n  budget(runId: string): Promise<BudgetLedgerEntry[]>\n  artifacts(runId: string): Promise<Artifact[]>\n}\n\n// ── In-memory ────────────────────────────────────────────────────────\n\nexport class InMemoryTraceStore implements TraceStore {\n  private runs = new Map<string, Run>()\n  private allSpans: Span[] = []\n  private allEvents: TraceEvent[] = []\n  private allArtifacts: Artifact[] = []\n  private allBudget: BudgetLedgerEntry[] = []\n\n  async appendRun(run: Run): Promise<void> {\n    if (this.runs.has(run.runId)) throw new Error(`run ${run.runId} already exists`)\n    this.runs.set(run.runId, { ...run })\n  }\n\n  async updateRun(runId: string, patch: Partial<Run>): Promise<void> {\n    const existing = this.runs.get(runId)\n    if (!existing) throw new Error(`run ${runId} not found`)\n    this.runs.set(runId, { ...existing, ...patch })\n  }\n\n  async appendSpan(span: Span): Promise<void> {\n    this.allSpans.push({ ...span })\n  }\n\n  async updateSpan(spanId: string, patch: Partial<Span>): Promise<void> {\n    const idx = this.allSpans.findIndex((s) => s.spanId === spanId)\n    if (idx < 0) throw new Error(`span ${spanId} not found`)\n    this.allSpans[idx] = { ...this.allSpans[idx], ...patch } as Span\n  }\n\n  async appendEvent(event: TraceEvent): Promise<void> {\n    this.allEvents.push({ ...event })\n  }\n\n  async appendArtifact(artifact: Artifact): Promise<void> {\n    this.allArtifacts.push({ ...artifact })\n  }\n\n  async appendBudgetEntry(entry: BudgetLedgerEntry): Promise<void> {\n    this.allBudget.push({ ...entry })\n  }\n\n  async getRun(runId: string): Promise<Run | undefined> {\n    const r = this.runs.get(runId)\n    return r ? { ...r } : undefined\n  }\n\n  async listRuns(filter: RunFilter = {}): Promise<Run[]> {\n    return [...this.runs.values()].filter((r) => matchesRun(r, filter))\n  }\n\n  async spans(filter: SpanFilter = {}): Promise<Span[]> {\n    return this.allSpans.filter((s) => matchesSpan(s, filter)).map((s) => ({ ...s }))\n  }\n\n  async events(filter: EventFilter = {}): Promise<TraceEvent[]> {\n    return this.allEvents.filter((e) => matchesEvent(e, filter)).map((e) => ({ ...e }))\n  }\n\n  async budget(runId: string): Promise<BudgetLedgerEntry[]> {\n    return this.allBudget.filter((b) => b.runId === runId).map((b) => ({ ...b }))\n  }\n\n  async artifacts(runId: string): Promise<Artifact[]> {\n    return this.allArtifacts.filter((a) => a.runId === runId).map((a) => ({ ...a }))\n  }\n}\n\nfunction matchesRun(r: Run, f: RunFilter): boolean {\n  if (f.scenarioId && r.scenarioId !== f.scenarioId) return false\n  if (f.variantId && r.variantId !== f.variantId) return false\n  if (f.status && r.status !== f.status) return false\n  if (f.since !== undefined && r.startedAt < f.since) return false\n  if (f.until !== undefined && r.startedAt > f.until) return false\n  if (f.tag && r.tags?.[f.tag.key] !== f.tag.value) return false\n  if (f.parentRunId && r.parentRunId !== f.parentRunId) return false\n  if (f.projectId && r.projectId !== f.projectId) return false\n  if (f.chatId && r.chatId !== f.chatId) return false\n  if (f.layer && r.layer !== f.layer) return false\n  return true\n}\n\nfunction matchesSpan(s: Span, f: SpanFilter): boolean {\n  if (f.runId && s.runId !== f.runId) return false\n  if (f.parentSpanId && s.parentSpanId !== f.parentSpanId) return false\n  if (f.kind && s.kind !== f.kind) return false\n  if (f.name && s.name !== f.name) return false\n  if (f.toolName && (s.kind !== 'tool' || s.toolName !== f.toolName)) return false\n  if (f.judgeId && (s.kind !== 'judge' || s.judgeId !== f.judgeId)) return false\n  if (f.since !== undefined && s.startedAt < f.since) return false\n  if (f.until !== undefined && s.startedAt > f.until) return false\n  return true\n}\n\nfunction matchesEvent(e: TraceEvent, f: EventFilter): boolean {\n  if (f.runId && e.runId !== f.runId) return false\n  if (f.spanId && e.spanId !== f.spanId) return false\n  if (f.kind && e.kind !== f.kind) return false\n  if (f.since !== undefined && e.timestamp < f.since) return false\n  if (f.until !== undefined && e.timestamp > f.until) return false\n  return true\n}\n\n// ── Filesystem (NDJSON append-only, one file per entity) ─────────────\n\nexport interface FileSystemTraceStoreOptions {\n  dir: string\n  /** Roll over NDJSON files when they exceed this size in bytes. Default 32 MB. */\n  maxBytes?: number\n}\n\nexport class FileSystemTraceStore implements TraceStore {\n  private dir: string\n  private maxBytes: number\n  /** Lazy in-memory index for queries — populated on first read. */\n  private index?: InMemoryTraceStore\n  /** Memoized index build — concurrent first reads share one build, and an\n   *  append racing an in-flight load awaits this so its row isn't lost. */\n  private indexPromise?: Promise<InMemoryTraceStore>\n  /** Strictly-increasing rollover stamp. Date.now() alone collides when two\n   *  rollovers land in the same millisecond, overwriting a rolled file. */\n  private lastRolloverStamp = 0\n  /**\n   * Per-file append serialization. stat → conditional-rename → appendFile is a\n   * read-modify-write on the active file; without a lock two concurrent appends\n   * to the same `name` can both pass the size check (exceeding maxBytes) or one\n   * can append to a file the other just renamed away. Each entry chains the\n   * next append behind the prior one for that file.\n   */\n  private appendLocks = new Map<string, Promise<void>>()\n\n  constructor(options: FileSystemTraceStoreOptions) {\n    this.dir = options.dir\n    this.maxBytes = options.maxBytes ?? 32 * 1024 * 1024\n  }\n\n  private async ensureDir(): Promise<void> {\n    const fs = await import('node:fs/promises')\n    await fs.mkdir(this.dir, { recursive: true })\n  }\n\n  private async append(name: string, record: unknown): Promise<void> {\n    // If an index load is in flight, wait for it. The row must hit disk AFTER\n    // the load finished reading, so it lands in the completed index via the\n    // mirror below — never lost in the gap before `this.index` is assigned.\n    if (this.indexPromise) await this.indexPromise\n    // Chain this append behind any in-flight append for the same file. The tail\n    // promise never rejects (errors are isolated to each caller's `result`), so\n    // one failed append can't break the lock chain for the next one.\n    const prior = this.appendLocks.get(name) ?? Promise.resolve()\n    const result = prior.then(() => this.appendLocked(name, record))\n    const tail = result.then(\n      () => {},\n      () => {},\n    )\n    this.appendLocks.set(name, tail)\n    try {\n      await result\n    } finally {\n      // Drop the slot once this append is the current tail, so the map doesn't\n      // grow without bound across many appends to the same file.\n      if (this.appendLocks.get(name) === tail) this.appendLocks.delete(name)\n    }\n  }\n\n  private async appendLocked(name: string, record: unknown): Promise<void> {\n    await this.ensureDir()\n    const fs = await import('node:fs/promises')\n    const path = await import('node:path')\n    const active = path.join(this.dir, `${name}.ndjson`)\n    try {\n      const stat = await fs.stat(active)\n      if (stat.size >= this.maxBytes) {\n        // Strictly-increasing, collision-free even within one millisecond — a\n        // bare Date.now() would let two same-ms rollovers overwrite each other.\n        const stamp = Math.max(Date.now(), this.lastRolloverStamp + 1)\n        this.lastRolloverStamp = stamp\n        const rolled = path.join(this.dir, `${name}.${stamp}.ndjson`)\n        await fs.rename(active, rolled)\n      }\n    } catch {\n      /* file doesn't exist yet */\n    }\n    await fs.appendFile(active, `${JSON.stringify(record)}\\n`, 'utf8')\n    // Mirror genuinely-new rows into the lazy index. Update rows (marked\n    // with `_update: true` by updateRun/updateSpan) are applied by those\n    // methods directly via the index's update* APIs — re-inserting them\n    // here triggers a duplicate-id error once the first read populates\n    // the index. Await the mirror: a fire-and-forget would let an index\n    // rejection go unhandled and let disk + index diverge silently.\n    if (this.index && !(record as { _update?: boolean })?._update) {\n      await this.insertInto(name, record)\n    }\n  }\n\n  private async insertInto(name: string, record: unknown): Promise<void> {\n    if (!this.index) return\n    switch (name) {\n      case 'runs':\n        await this.index.appendRun(record as Run)\n        break\n      case 'spans':\n        await this.index.appendSpan(record as Span)\n        break\n      case 'events':\n        await this.index.appendEvent(record as TraceEvent)\n        break\n      case 'artifacts':\n        await this.index.appendArtifact(record as Artifact)\n        break\n      case 'budget':\n        await this.index.appendBudgetEntry(record as BudgetLedgerEntry)\n        break\n    }\n  }\n\n  private load(): Promise<InMemoryTraceStore> {\n    // Memoize: concurrent first reads share one build, and an append racing an\n    // in-flight load awaits this same promise (see append()).\n    this.indexPromise ??= this.buildIndex()\n    return this.indexPromise\n  }\n\n  private async buildIndex(): Promise<InMemoryTraceStore> {\n    const fs = await import('node:fs/promises')\n    const path = await import('node:path')\n    const store = new InMemoryTraceStore()\n    let entries: string[]\n    try {\n      entries = await fs.readdir(this.dir)\n    } catch {\n      // No dir yet (first run / empty corpus). Anything past this point — parse\n      // errors, missing patch bases — is real corruption and must fail loud.\n      this.index = store\n      return store\n    }\n    // Replay files in write order. For one entity, rolled files\n    // (`name.<ts>.ndjson`) are older than the active bare file (`name.ndjson`),\n    // and readdir order is not chronological — so without this sort a patch in\n    // an older rolled file could be applied AFTER (overwrite) a newer patch in\n    // the active file, silently restoring a stale field. Sort: per entity,\n    // rolled ascending by timestamp, then the bare active file last.\n    entries.sort((a, b) => {\n      // Group by the same entity base the loop uses (`split('.')[0]`); within an\n      // entity, real rollover files (`<base>.<ms>.ndjson`) sort by timestamp\n      // ascending and the bare active file (no numeric segment) sorts last.\n      // Non-rollover names tie (Infinity) and a stable sort preserves their\n      // readdir order.\n      const baseA = a.split('.')[0]!\n      const baseB = b.split('.')[0]!\n      if (baseA !== baseB) return baseA < baseB ? -1 : 1\n      const ta = a.match(/^[^.]+\\.(\\d+)\\.ndjson$/)\n      const tb = b.match(/^[^.]+\\.(\\d+)\\.ndjson$/)\n      const tsA = ta ? Number(ta[1]) : Number.POSITIVE_INFINITY\n      const tsB = tb ? Number(tb[1]) : Number.POSITIVE_INFINITY\n      return tsA - tsB\n    })\n    {\n      // Two-pass load. Pass 1 indexes all full (base) rows; pass 2 replays the\n      // `_update` patch rows that updateRun/updateSpan append. readdir order is\n      // not deterministic and rollover splits a stream across files, so a patch\n      // can be read before its base. Applying patches in a separate pass — after\n      // every base row is indexed — means a patch always finds its target.\n      // Doing it in one pass let a patch hit the catch and append a runId-less\n      // fragment, corrupting span/run counts cross-instance.\n      const runUpdates: Array<Record<string, unknown>> = []\n      const spanUpdates: Array<Record<string, unknown>> = []\n      for (const file of entries) {\n        if (!file.endsWith('.ndjson')) continue\n        const full = path.join(this.dir, file)\n        const content = await fs.readFile(full, 'utf8')\n        const base = file.split('.')[0]\n        const lines = content.split('\\n')\n        for (let ln = 0; ln < lines.length; ln++) {\n          const line = lines[ln]!\n          if (!line.trim()) continue\n          let record: ReturnType<typeof JSON.parse>\n          try {\n            record = JSON.parse(line)\n          } catch (err) {\n            // Fail loud WITH context — a bare SyntaxError loses which file/line\n            // is corrupt and which valid rows surround it.\n            throw new Error(\n              `FileSystemTraceStore: corrupt NDJSON in ${file} line ${ln + 1}: ${\n                err instanceof Error ? err.message : String(err)\n              } — ${line.slice(0, 120)}`,\n            )\n          }\n          if (base === 'runs') {\n            if (record?._update) {\n              runUpdates.push(record)\n              continue\n            }\n            // A duplicate full row (same dir loaded twice into one stream) is\n            // collapsed last-write-wins rather than thrown.\n            try {\n              await store.appendRun(record)\n            } catch {\n              await store.updateRun(record.runId, record)\n            }\n          } else if (base === 'spans') {\n            // `updateSpan` appends an `_update: true` patch row instead of\n            // rewriting the original span. Deferred to pass 2 so it collapses\n            // onto the original span — otherwise a fresh FileSystemTraceStore\n            // reading the same dir reports duplicate spans (one full, one\n            // fragment with no runId/kind/name), which breaks any downstream\n            // consumer that re-opens the store cross-process (e.g. the OTLP\n            // converter).\n            if (record?._update) {\n              spanUpdates.push(record)\n              continue\n            }\n            await store.appendSpan(record)\n          } else if (base === 'events') {\n            await store.appendEvent(record)\n          } else if (base === 'artifacts') {\n            await store.appendArtifact(record)\n          } else if (base === 'budget') {\n            await store.appendBudgetEntry(record)\n          }\n        }\n      }\n      // Pass 2: every base row is now indexed. A patch whose base is genuinely\n      // absent (truncated/partial corpus) throws inside update* — we let it\n      // propagate rather than silently appending a fragment, so a corrupt\n      // corpus fails loud instead of polluting counts.\n      for (const record of runUpdates) {\n        await store.updateRun(record.runId as string, record as Partial<Run>)\n      }\n      for (const record of spanUpdates) {\n        await store.updateSpan(record.spanId as string, record as Partial<Span>)\n      }\n    }\n    this.index = store\n    return store\n  }\n\n  async appendRun(run: Run): Promise<void> {\n    await this.append('runs', run)\n  }\n  async updateRun(runId: string, patch: Partial<Run>): Promise<void> {\n    // NDJSON is append-only; record updates as new rows with the same runId —\n    // readers collapse by last-write-wins on load.\n    await this.append('runs', { runId, ...patch, _update: true })\n    if (this.index) await this.index.updateRun(runId, patch)\n  }\n  async appendSpan(span: Span): Promise<void> {\n    await this.append('spans', span)\n  }\n  async updateSpan(spanId: string, patch: Partial<Span>): Promise<void> {\n    await this.append('spans', { spanId, ...patch, _update: true })\n    if (this.index) await this.index.updateSpan(spanId, patch)\n  }\n  async appendEvent(event: TraceEvent): Promise<void> {\n    await this.append('events', event)\n  }\n  async appendArtifact(artifact: Artifact): Promise<void> {\n    await this.append('artifacts', artifact)\n  }\n  async appendBudgetEntry(entry: BudgetLedgerEntry): Promise<void> {\n    await this.append('budget', entry)\n  }\n\n  async getRun(runId: string): Promise<Run | undefined> {\n    return (await this.load()).getRun(runId)\n  }\n  async listRuns(filter?: RunFilter): Promise<Run[]> {\n    return (await this.load()).listRuns(filter)\n  }\n  async spans(filter?: SpanFilter): Promise<Span[]> {\n    return (await this.load()).spans(filter)\n  }\n  async events(filter?: EventFilter): Promise<TraceEvent[]> {\n    return (await this.load()).events(filter)\n  }\n  async budget(runId: string): Promise<BudgetLedgerEntry[]> {\n    return (await this.load()).budget(runId)\n  }\n  async artifacts(runId: string): Promise<Artifact[]> {\n    return (await this.load()).artifacts(runId)\n  }\n}\n","/**\n * Run-completion integrity check — at end of run, verify the expected event\n * types were actually captured. The point is the launch-review failure mode:\n * a run *appears* successful but the raw provider events were never written,\n * so a downstream reviewer can't reconstruct what happened.\n *\n * Pattern:\n *\n *   const report = await assertRunCaptured(store, runId, {\n *     llmSpansMin: 1,\n *     judgeSpansMin: 1,\n *     rawSink: providerSink,                  // must have ≥ 1 event for this run\n *     requireRawCoverageOfLlmSpans: true,     // every llm span has matching raw events\n *   })\n *   if (!report.ok) throwIfRunIncomplete(report)  // or mark run failed and continue\n *\n * The function is read-only on the store and returns a structured report;\n * the caller chooses the failure mode (throw, mark run failed, log warning).\n * `throwIfRunIncomplete` is the convenient strict mode.\n */\n\nimport { CaptureIntegrityError } from '../errors'\nimport type { RawProviderSink } from './raw-provider-sink'\nimport type { TraceStore } from './store'\n\nexport interface RunIntegrityExpectations {\n  /** Minimum LLM span count. Default 0 (no requirement). */\n  llmSpansMin?: number\n  /** Minimum judge span count. Default 0. */\n  judgeSpansMin?: number\n  /** Minimum tool span count. Default 0. */\n  toolSpansMin?: number\n  /**\n   * Raw provider sink to consult for capture verification. When present,\n   * the check requires at least one raw event for the run.\n   */\n  rawSink?: RawProviderSink\n  /** Minimum raw provider event count. Default 0; ignored when `rawSink` absent. */\n  rawProviderEventsMin?: number\n  /**\n   * Every LLM span must have at least one matching raw `request` event\n   * (matched by spanId). Catches the common bug where the structured span\n   * was emitted but the raw HTTP capture was wired to a different sink.\n   */\n  requireRawCoverageOfLlmSpans?: boolean\n  /** Run outcome must be set (not null/undefined). Default false. */\n  requireOutcome?: boolean\n}\n\nexport type RunIntegrityIssueCode =\n  | 'no_run'\n  | 'missing_llm_spans'\n  | 'missing_judge_spans'\n  | 'missing_tool_spans'\n  | 'missing_raw_events'\n  | 'no_raw_sink'\n  | 'orphan_llm_span'\n  | 'missing_outcome'\n\nexport interface RunIntegrityIssue {\n  code: RunIntegrityIssueCode\n  message: string\n  detail?: Record<string, unknown>\n}\n\nexport interface RunIntegrityReport {\n  ok: boolean\n  runId: string\n  llmSpanCount: number\n  judgeSpanCount: number\n  toolSpanCount: number\n  rawProviderEventCount: number\n  /**\n   * Coverage of LLM spans by raw provider events keyed on spanId.\n   * `total` is the number of LLM spans; `covered` is the count with at\n   * least one matching `request` raw event.\n   */\n  rawSpanCoverage: { covered: number; total: number }\n  issues: RunIntegrityIssue[]\n}\n\nexport class RunIntegrityError extends CaptureIntegrityError {\n  constructor(public readonly report: RunIntegrityReport) {\n    super(\n      `Run ${report.runId} failed integrity check: ${report.issues.map((i) => i.code).join(', ')}`,\n    )\n  }\n}\n\nexport async function assertRunCaptured(\n  store: TraceStore,\n  runId: string,\n  expectations: RunIntegrityExpectations = {},\n): Promise<RunIntegrityReport> {\n  const issues: RunIntegrityIssue[] = []\n  const run = await store.getRun(runId)\n  if (!run) {\n    return {\n      ok: false,\n      runId,\n      llmSpanCount: 0,\n      judgeSpanCount: 0,\n      toolSpanCount: 0,\n      rawProviderEventCount: 0,\n      rawSpanCoverage: { covered: 0, total: 0 },\n      issues: [{ code: 'no_run', message: `Run ${runId} not found in store.` }],\n    }\n  }\n\n  const spans = await store.spans({ runId })\n  const llmSpans = spans.filter((s) => s.kind === 'llm')\n  const judgeSpans = spans.filter((s) => s.kind === 'judge')\n  const toolSpans = spans.filter((s) => s.kind === 'tool')\n\n  const llmMin = expectations.llmSpansMin ?? 0\n  const judgeMin = expectations.judgeSpansMin ?? 0\n  const toolMin = expectations.toolSpansMin ?? 0\n\n  if (llmSpans.length < llmMin) {\n    issues.push({\n      code: 'missing_llm_spans',\n      message: `Expected ≥ ${llmMin} LLM spans, found ${llmSpans.length}.`,\n      detail: { expected: llmMin, found: llmSpans.length },\n    })\n  }\n  if (judgeSpans.length < judgeMin) {\n    issues.push({\n      code: 'missing_judge_spans',\n      message: `Expected ≥ ${judgeMin} judge spans, found ${judgeSpans.length}.`,\n      detail: { expected: judgeMin, found: judgeSpans.length },\n    })\n  }\n  if (toolSpans.length < toolMin) {\n    issues.push({\n      code: 'missing_tool_spans',\n      message: `Expected ≥ ${toolMin} tool spans, found ${toolSpans.length}.`,\n      detail: { expected: toolMin, found: toolSpans.length },\n    })\n  }\n\n  let rawEventCount = 0\n  let coverage = { covered: 0, total: llmSpans.length }\n\n  if (expectations.rawSink) {\n    if (!expectations.rawSink.list) {\n      issues.push({\n        code: 'no_raw_sink',\n        message: 'Provided rawSink does not implement list(); cannot verify capture.',\n      })\n    } else {\n      const events = await expectations.rawSink.list({ runId })\n      rawEventCount = events.length\n      const rawMin = expectations.rawProviderEventsMin ?? 1\n      if (rawEventCount < rawMin) {\n        issues.push({\n          code: 'missing_raw_events',\n          message: `Expected ≥ ${rawMin} raw provider events, found ${rawEventCount}.`,\n          detail: { expected: rawMin, found: rawEventCount },\n        })\n      }\n      if (expectations.requireRawCoverageOfLlmSpans) {\n        const requestEventsBySpan = new Set(\n          events.filter((e) => e.direction === 'request' && e.spanId).map((e) => e.spanId!),\n        )\n        const orphaned = llmSpans.filter((s) => !requestEventsBySpan.has(s.spanId))\n        coverage = { covered: llmSpans.length - orphaned.length, total: llmSpans.length }\n        if (orphaned.length > 0) {\n          issues.push({\n            code: 'orphan_llm_span',\n            message: `${orphaned.length} LLM span(s) have no matching raw provider request event.`,\n            detail: { orphanedSpanIds: orphaned.map((s) => s.spanId) },\n          })\n        }\n      }\n    }\n  } else if (expectations.requireRawCoverageOfLlmSpans || expectations.rawProviderEventsMin) {\n    issues.push({\n      code: 'no_raw_sink',\n      message: 'Raw coverage required but no rawSink supplied to the integrity check.',\n    })\n  }\n\n  if (expectations.requireOutcome && (run.outcome === undefined || run.outcome === null)) {\n    issues.push({\n      code: 'missing_outcome',\n      message: `Run ${runId} has no outcome recorded.`,\n    })\n  }\n\n  return {\n    ok: issues.length === 0,\n    runId,\n    llmSpanCount: llmSpans.length,\n    judgeSpanCount: judgeSpans.length,\n    toolSpanCount: toolSpans.length,\n    rawProviderEventCount: rawEventCount,\n    rawSpanCoverage: coverage,\n    issues,\n  }\n}\n\n/** Strict mode: throws `RunIntegrityError` when the report isn't ok. */\nexport function throwIfRunIncomplete(report: RunIntegrityReport): void {\n  if (!report.ok) throw new RunIntegrityError(report)\n}\n"],"mappings":";;AA0EA,IAAa,qBAAb,MAAsD;CACpD,uBAAe,IAAI,IAAiB;CACpC,WAA2B,CAAC;CAC5B,YAAkC,CAAC;CACnC,eAAmC,CAAC;CACpC,YAAyC,CAAC;CAE1C,MAAM,UAAU,KAAyB;EACvC,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,GAAG,MAAM,IAAI,MAAM,OAAO,IAAI,MAAM,gBAAgB;EAC/E,KAAK,KAAK,IAAI,IAAI,OAAO,EAAE,GAAG,IAAI,CAAC;CACrC;CAEA,MAAM,UAAU,OAAe,OAAoC;EACjE,MAAM,WAAW,KAAK,KAAK,IAAI,KAAK;EACpC,IAAI,CAAC,UAAU,MAAM,IAAI,MAAM,OAAO,MAAM,WAAW;EACvD,KAAK,KAAK,IAAI,OAAO;GAAE,GAAG;GAAU,GAAG;EAAM,CAAC;CAChD;CAEA,MAAM,WAAW,MAA2B;EAC1C,KAAK,SAAS,KAAK,EAAE,GAAG,KAAK,CAAC;CAChC;CAEA,MAAM,WAAW,QAAgB,OAAqC;EACpE,MAAM,MAAM,KAAK,SAAS,WAAW,MAAM,EAAE,WAAW,MAAM;EAC9D,IAAI,MAAM,GAAG,MAAM,IAAI,MAAM,QAAQ,OAAO,WAAW;EACvD,KAAK,SAAS,OAAO;GAAE,GAAG,KAAK,SAAS;GAAM,GAAG;EAAM;CACzD;CAEA,MAAM,YAAY,OAAkC;EAClD,KAAK,UAAU,KAAK,EAAE,GAAG,MAAM,CAAC;CAClC;CAEA,MAAM,eAAe,UAAmC;EACtD,KAAK,aAAa,KAAK,EAAE,GAAG,SAAS,CAAC;CACxC;CAEA,MAAM,kBAAkB,OAAyC;EAC/D,KAAK,UAAU,KAAK,EAAE,GAAG,MAAM,CAAC;CAClC;CAEA,MAAM,OAAO,OAAyC;EACpD,MAAM,IAAI,KAAK,KAAK,IAAI,KAAK;EAC7B,OAAO,IAAI,EAAE,GAAG,EAAE,IAAI,KAAA;CACxB;CAEA,MAAM,SAAS,SAAoB,CAAC,GAAmB;EACrD,OAAO,CAAC,GAAG,KAAK,KAAK,OAAO,CAAC,CAAC,CAAC,QAAQ,MAAM,WAAW,GAAG,MAAM,CAAC;CACpE;CAEA,MAAM,MAAM,SAAqB,CAAC,GAAoB;EACpD,OAAO,KAAK,SAAS,QAAQ,MAAM,YAAY,GAAG,MAAM,CAAC,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CAClF;CAEA,MAAM,OAAO,SAAsB,CAAC,GAA0B;EAC5D,OAAO,KAAK,UAAU,QAAQ,MAAM,aAAa,GAAG,MAAM,CAAC,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CACpF;CAEA,MAAM,OAAO,OAA6C;EACxD,OAAO,KAAK,UAAU,QAAQ,MAAM,EAAE,UAAU,KAAK,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CAC9E;CAEA,MAAM,UAAU,OAAoC;EAClD,OAAO,KAAK,aAAa,QAAQ,MAAM,EAAE,UAAU,KAAK,CAAC,CAAC,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE;CACjF;AACF;AAEA,SAAS,WAAW,GAAQ,GAAuB;CACjD,IAAI,EAAE,cAAc,EAAE,eAAe,EAAE,YAAY,OAAO;CAC1D,IAAI,EAAE,aAAa,EAAE,cAAc,EAAE,WAAW,OAAO;CACvD,IAAI,EAAE,UAAU,EAAE,WAAW,EAAE,QAAQ,OAAO;CAC9C,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,SAAS,EAAE,IAAI,OAAO,OAAO;CACzD,IAAI,EAAE,eAAe,EAAE,gBAAgB,EAAE,aAAa,OAAO;CAC7D,IAAI,EAAE,aAAa,EAAE,cAAc,EAAE,WAAW,OAAO;CACvD,IAAI,EAAE,UAAU,EAAE,WAAW,EAAE,QAAQ,OAAO;CAC9C,IAAI,EAAE,SAAS,EAAE,UAAU,EAAE,OAAO,OAAO;CAC3C,OAAO;AACT;AAEA,SAAS,YAAY,GAAS,GAAwB;CACpD,IAAI,EAAE,SAAS,EAAE,UAAU,EAAE,OAAO,OAAO;CAC3C,IAAI,EAAE,gBAAgB,EAAE,iBAAiB,EAAE,cAAc,OAAO;CAChE,IAAI,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,OAAO;CACxC,IAAI,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,OAAO;CACxC,IAAI,EAAE,aAAa,EAAE,SAAS,UAAU,EAAE,aAAa,EAAE,WAAW,OAAO;CAC3E,IAAI,EAAE,YAAY,EAAE,SAAS,WAAW,EAAE,YAAY,EAAE,UAAU,OAAO;CACzE,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,OAAO;AACT;AAEA,SAAS,aAAa,GAAe,GAAyB;CAC5D,IAAI,EAAE,SAAS,EAAE,UAAU,EAAE,OAAO,OAAO;CAC3C,IAAI,EAAE,UAAU,EAAE,WAAW,EAAE,QAAQ,OAAO;CAC9C,IAAI,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,OAAO;CACxC,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,IAAI,EAAE,UAAU,KAAA,KAAa,EAAE,YAAY,EAAE,OAAO,OAAO;CAC3D,OAAO;AACT;AAUA,IAAa,uBAAb,MAAwD;CACtD;CACA;;CAEA;;;CAGA;;;CAGA,oBAA4B;;;;;;;;CAQ5B,8BAAsB,IAAI,IAA2B;CAErD,YAAY,SAAsC;EAChD,KAAK,MAAM,QAAQ;EACnB,KAAK,WAAW,QAAQ,YAAY,KAAK,OAAO;CAClD;CAEA,MAAc,YAA2B;EAEvC,OAAM,MADW,OAAO,oBAAA,CACf,MAAM,KAAK,KAAK,EAAE,WAAW,KAAK,CAAC;CAC9C;CAEA,MAAc,OAAO,MAAc,QAAgC;EAIjE,IAAI,KAAK,cAAc,MAAM,KAAK;EAKlC,MAAM,UADQ,KAAK,YAAY,IAAI,IAAI,KAAK,QAAQ,QAAQ,EAAA,CACvC,WAAW,KAAK,aAAa,MAAM,MAAM,CAAC;EAC/D,MAAM,OAAO,OAAO,WACZ,CAAC,SACD,CAAC,CACT;EACA,KAAK,YAAY,IAAI,MAAM,IAAI;EAC/B,IAAI;GACF,MAAM;EACR,UAAU;GAGR,IAAI,KAAK,YAAY,IAAI,IAAI,MAAM,MAAM,KAAK,YAAY,OAAO,IAAI;EACvE;CACF;CAEA,MAAc,aAAa,MAAc,QAAgC;EACvE,MAAM,KAAK,UAAU;EACrB,MAAM,KAAK,MAAM,OAAO;EACxB,MAAM,OAAO,MAAM,OAAO;EAC1B,MAAM,SAAS,KAAK,KAAK,KAAK,KAAK,GAAG,KAAK,QAAQ;EACnD,IAAI;GAEF,KAAI,MADe,GAAG,KAAK,MAAM,EAAA,CACxB,QAAQ,KAAK,UAAU;IAG9B,MAAM,QAAQ,KAAK,IAAI,KAAK,IAAI,GAAG,KAAK,oBAAoB,CAAC;IAC7D,KAAK,oBAAoB;IACzB,MAAM,SAAS,KAAK,KAAK,KAAK,KAAK,GAAG,KAAK,GAAG,MAAM,QAAQ;IAC5D,MAAM,GAAG,OAAO,QAAQ,MAAM;GAChC;EACF,QAAQ,CAER;EACA,MAAM,GAAG,WAAW,QAAQ,GAAG,KAAK,UAAU,MAAM,EAAE,KAAK,MAAM;EAOjE,IAAI,KAAK,SAAS,CAAE,QAAkC,SACpD,MAAM,KAAK,WAAW,MAAM,MAAM;CAEtC;CAEA,MAAc,WAAW,MAAc,QAAgC;EACrE,IAAI,CAAC,KAAK,OAAO;EACjB,QAAQ,MAAR;GACE,KAAK;IACH,MAAM,KAAK,MAAM,UAAU,MAAa;IACxC;GACF,KAAK;IACH,MAAM,KAAK,MAAM,WAAW,MAAc;IAC1C;GACF,KAAK;IACH,MAAM,KAAK,MAAM,YAAY,MAAoB;IACjD;GACF,KAAK;IACH,MAAM,KAAK,MAAM,eAAe,MAAkB;IAClD;GACF,KAAK;IACH,MAAM,KAAK,MAAM,kBAAkB,MAA2B;IAC9D;EACJ;CACF;CAEA,OAA4C;EAG1C,KAAK,iBAAiB,KAAK,WAAW;EACtC,OAAO,KAAK;CACd;CAEA,MAAc,aAA0C;EACtD,MAAM,KAAK,MAAM,OAAO;EACxB,MAAM,OAAO,MAAM,OAAO;EAC1B,MAAM,QAAQ,IAAI,mBAAmB;EACrC,IAAI;EACJ,IAAI;GACF,UAAU,MAAM,GAAG,QAAQ,KAAK,GAAG;EACrC,QAAQ;GAGN,KAAK,QAAQ;GACb,OAAO;EACT;EAOA,QAAQ,MAAM,GAAG,MAAM;GAMrB,MAAM,QAAQ,EAAE,MAAM,GAAG,CAAC,CAAC;GAC3B,MAAM,QAAQ,EAAE,MAAM,GAAG,CAAC,CAAC;GAC3B,IAAI,UAAU,OAAO,OAAO,QAAQ,QAAQ,KAAK;GACjD,MAAM,KAAK,EAAE,MAAM,wBAAwB;GAC3C,MAAM,KAAK,EAAE,MAAM,wBAAwB;GAG3C,QAFY,KAAK,OAAO,GAAG,EAAE,IAAI,OAAO,sBAC5B,KAAK,OAAO,GAAG,EAAE,IAAI,OAAO;EAE1C,CAAC;EACD;GAQE,MAAM,aAA6C,CAAC;GACpD,MAAM,cAA8C,CAAC;GACrD,KAAK,MAAM,QAAQ,SAAS;IAC1B,IAAI,CAAC,KAAK,SAAS,SAAS,GAAG;IAC/B,MAAM,OAAO,KAAK,KAAK,KAAK,KAAK,IAAI;IACrC,MAAM,UAAU,MAAM,GAAG,SAAS,MAAM,MAAM;IAC9C,MAAM,OAAO,KAAK,MAAM,GAAG,CAAC,CAAC;IAC7B,MAAM,QAAQ,QAAQ,MAAM,IAAI;IAChC,KAAK,IAAI,KAAK,GAAG,KAAK,MAAM,QAAQ,MAAM;KACxC,MAAM,OAAO,MAAM;KACnB,IAAI,CAAC,KAAK,KAAK,GAAG;KAClB,IAAI;KACJ,IAAI;MACF,SAAS,KAAK,MAAM,IAAI;KAC1B,SAAS,KAAK;MAGZ,MAAM,IAAI,MACR,2CAA2C,KAAK,QAAQ,KAAK,EAAE,IAC7D,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,EAChD,KAAK,KAAK,MAAM,GAAG,GAAG,GACzB;KACF;KACA,IAAI,SAAS,QAAQ;MACnB,IAAI,QAAQ,SAAS;OACnB,WAAW,KAAK,MAAM;OACtB;MACF;MAGA,IAAI;OACF,MAAM,MAAM,UAAU,MAAM;MAC9B,QAAQ;OACN,MAAM,MAAM,UAAU,OAAO,OAAO,MAAM;MAC5C;KACF,OAAO,IAAI,SAAS,SAAS;MAQ3B,IAAI,QAAQ,SAAS;OACnB,YAAY,KAAK,MAAM;OACvB;MACF;MACA,MAAM,MAAM,WAAW,MAAM;KAC/B,OAAO,IAAI,SAAS,UAClB,MAAM,MAAM,YAAY,MAAM;UACzB,IAAI,SAAS,aAClB,MAAM,MAAM,eAAe,MAAM;UAC5B,IAAI,SAAS,UAClB,MAAM,MAAM,kBAAkB,MAAM;IAExC;GACF;GAKA,KAAK,MAAM,UAAU,YACnB,MAAM,MAAM,UAAU,OAAO,OAAiB,MAAsB;GAEtE,KAAK,MAAM,UAAU,aACnB,MAAM,MAAM,WAAW,OAAO,QAAkB,MAAuB;EAE3E;EACA,KAAK,QAAQ;EACb,OAAO;CACT;CAEA,MAAM,UAAU,KAAyB;EACvC,MAAM,KAAK,OAAO,QAAQ,GAAG;CAC/B;CACA,MAAM,UAAU,OAAe,OAAoC;EAGjE,MAAM,KAAK,OAAO,QAAQ;GAAE;GAAO,GAAG;GAAO,SAAS;EAAK,CAAC;EAC5D,IAAI,KAAK,OAAO,MAAM,KAAK,MAAM,UAAU,OAAO,KAAK;CACzD;CACA,MAAM,WAAW,MAA2B;EAC1C,MAAM,KAAK,OAAO,SAAS,IAAI;CACjC;CACA,MAAM,WAAW,QAAgB,OAAqC;EACpE,MAAM,KAAK,OAAO,SAAS;GAAE;GAAQ,GAAG;GAAO,SAAS;EAAK,CAAC;EAC9D,IAAI,KAAK,OAAO,MAAM,KAAK,MAAM,WAAW,QAAQ,KAAK;CAC3D;CACA,MAAM,YAAY,OAAkC;EAClD,MAAM,KAAK,OAAO,UAAU,KAAK;CACnC;CACA,MAAM,eAAe,UAAmC;EACtD,MAAM,KAAK,OAAO,aAAa,QAAQ;CACzC;CACA,MAAM,kBAAkB,OAAyC;EAC/D,MAAM,KAAK,OAAO,UAAU,KAAK;CACnC;CAEA,MAAM,OAAO,OAAyC;EACpD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,OAAO,KAAK;CACzC;CACA,MAAM,SAAS,QAAoC;EACjD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,SAAS,MAAM;CAC5C;CACA,MAAM,MAAM,QAAsC;EAChD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,MAAM,MAAM;CACzC;CACA,MAAM,OAAO,QAA6C;EACxD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,OAAO,MAAM;CAC1C;CACA,MAAM,OAAO,OAA6C;EACxD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,OAAO,KAAK;CACzC;CACA,MAAM,UAAU,OAAoC;EAClD,QAAQ,MAAM,KAAK,KAAK,EAAA,CAAG,UAAU,KAAK;CAC5C;AACF;;;;;;;;;;;;;;;;;;;;;;;ACpXA,IAAa,oBAAb,cAAuC,sBAAsB;CAC/B;CAA5B,YAAY,QAA4C;EACtD,MACE,OAAO,OAAO,MAAM,2BAA2B,OAAO,OAAO,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,IAAI,GAC3F;EAH0B,KAAA,SAAA;CAI5B;AACF;AAEA,eAAsB,kBACpB,OACA,OACA,eAAyC,CAAC,GACb;CAC7B,MAAM,SAA8B,CAAC;CACrC,MAAM,MAAM,MAAM,MAAM,OAAO,KAAK;CACpC,IAAI,CAAC,KACH,OAAO;EACL,IAAI;EACJ;EACA,cAAc;EACd,gBAAgB;EAChB,eAAe;EACf,uBAAuB;EACvB,iBAAiB;GAAE,SAAS;GAAG,OAAO;EAAE;EACxC,QAAQ,CAAC;GAAE,MAAM;GAAU,SAAS,OAAO,MAAM;EAAsB,CAAC;CAC1E;CAGF,MAAM,QAAQ,MAAM,MAAM,MAAM,EAAE,MAAM,CAAC;CACzC,MAAM,WAAW,MAAM,QAAQ,MAAM,EAAE,SAAS,KAAK;CACrD,MAAM,aAAa,MAAM,QAAQ,MAAM,EAAE,SAAS,OAAO;CACzD,MAAM,YAAY,MAAM,QAAQ,MAAM,EAAE,SAAS,MAAM;CAEvD,MAAM,SAAS,aAAa,eAAe;CAC3C,MAAM,WAAW,aAAa,iBAAiB;CAC/C,MAAM,UAAU,aAAa,gBAAgB;CAE7C,IAAI,SAAS,SAAS,QACpB,OAAO,KAAK;EACV,MAAM;EACN,SAAS,cAAc,OAAO,oBAAoB,SAAS,OAAO;EAClE,QAAQ;GAAE,UAAU;GAAQ,OAAO,SAAS;EAAO;CACrD,CAAC;CAEH,IAAI,WAAW,SAAS,UACtB,OAAO,KAAK;EACV,MAAM;EACN,SAAS,cAAc,SAAS,sBAAsB,WAAW,OAAO;EACxE,QAAQ;GAAE,UAAU;GAAU,OAAO,WAAW;EAAO;CACzD,CAAC;CAEH,IAAI,UAAU,SAAS,SACrB,OAAO,KAAK;EACV,MAAM;EACN,SAAS,cAAc,QAAQ,qBAAqB,UAAU,OAAO;EACrE,QAAQ;GAAE,UAAU;GAAS,OAAO,UAAU;EAAO;CACvD,CAAC;CAGH,IAAI,gBAAgB;CACpB,IAAI,WAAW;EAAE,SAAS;EAAG,OAAO,SAAS;CAAO;CAEpD,IAAI,aAAa,SACf,IAAI,CAAC,aAAa,QAAQ,MACxB,OAAO,KAAK;EACV,MAAM;EACN,SAAS;CACX,CAAC;MACI;EACL,MAAM,SAAS,MAAM,aAAa,QAAQ,KAAK,EAAE,MAAM,CAAC;EACxD,gBAAgB,OAAO;EACvB,MAAM,SAAS,aAAa,wBAAwB;EACpD,IAAI,gBAAgB,QAClB,OAAO,KAAK;GACV,MAAM;GACN,SAAS,cAAc,OAAO,8BAA8B,cAAc;GAC1E,QAAQ;IAAE,UAAU;IAAQ,OAAO;GAAc;EACnD,CAAC;EAEH,IAAI,aAAa,8BAA8B;GAC7C,MAAM,sBAAsB,IAAI,IAC9B,OAAO,QAAQ,MAAM,EAAE,cAAc,aAAa,EAAE,MAAM,CAAC,CAAC,KAAK,MAAM,EAAE,MAAO,CAClF;GACA,MAAM,WAAW,SAAS,QAAQ,MAAM,CAAC,oBAAoB,IAAI,EAAE,MAAM,CAAC;GAC1E,WAAW;IAAE,SAAS,SAAS,SAAS,SAAS;IAAQ,OAAO,SAAS;GAAO;GAChF,IAAI,SAAS,SAAS,GACpB,OAAO,KAAK;IACV,MAAM;IACN,SAAS,GAAG,SAAS,OAAO;IAC5B,QAAQ,EAAE,iBAAiB,SAAS,KAAK,MAAM,EAAE,MAAM,EAAE;GAC3D,CAAC;EAEL;CACF;MACK,IAAI,aAAa,gCAAgC,aAAa,sBACnE,OAAO,KAAK;EACV,MAAM;EACN,SAAS;CACX,CAAC;CAGH,IAAI,aAAa,mBAAmB,IAAI,YAAY,KAAA,KAAa,IAAI,YAAY,OAC/E,OAAO,KAAK;EACV,MAAM;EACN,SAAS,OAAO,MAAM;CACxB,CAAC;CAGH,OAAO;EACL,IAAI,OAAO,WAAW;EACtB;EACA,cAAc,SAAS;EACvB,gBAAgB,WAAW;EAC3B,eAAe,UAAU;EACzB,uBAAuB;EACvB,iBAAiB;EACjB;CACF;AACF;;AAGA,SAAgB,qBAAqB,QAAkC;CACrE,IAAI,CAAC,OAAO,IAAI,MAAM,IAAI,kBAAkB,MAAM;AACpD"}

package/dist/integrity-rmVhXWA7.d.ts

@@ -0,0 +1,61 @@
+import { r as CaptureIntegrityError } from "./errors-CEk209JS.js";
+import { l as RawProviderSink } from "./raw-provider-sink-BU29Sh8h.js";
+import { s as TraceStore } from "./store-CT9YIIve.js";
+//#region src/trace/integrity.d.ts
+interface RunIntegrityExpectations {
+  /** Minimum LLM span count. Default 0 (no requirement). */
+  llmSpansMin?: number;
+  /** Minimum judge span count. Default 0. */
+  judgeSpansMin?: number;
+  /** Minimum tool span count. Default 0. */
+  toolSpansMin?: number;
+  /**
+   * Raw provider sink to consult for capture verification. When present,
+   * the check requires at least one raw event for the run.
+   */
+  rawSink?: RawProviderSink;
+  /** Minimum raw provider event count. Default 0; ignored when `rawSink` absent. */
+  rawProviderEventsMin?: number;
+  /**
+   * Every LLM span must have at least one matching raw `request` event
+   * (matched by spanId). Catches the common bug where the structured span
+   * was emitted but the raw HTTP capture was wired to a different sink.
+   */
+  requireRawCoverageOfLlmSpans?: boolean;
+  /** Run outcome must be set (not null/undefined). Default false. */
+  requireOutcome?: boolean;
+}
+type RunIntegrityIssueCode = 'no_run' | 'missing_llm_spans' | 'missing_judge_spans' | 'missing_tool_spans' | 'missing_raw_events' | 'no_raw_sink' | 'orphan_llm_span' | 'missing_outcome';
+interface RunIntegrityIssue {
+  code: RunIntegrityIssueCode;
+  message: string;
+  detail?: Record<string, unknown>;
+}
+interface RunIntegrityReport {
+  ok: boolean;
+  runId: string;
+  llmSpanCount: number;
+  judgeSpanCount: number;
+  toolSpanCount: number;
+  rawProviderEventCount: number;
+  /**
+   * Coverage of LLM spans by raw provider events keyed on spanId.
+   * `total` is the number of LLM spans; `covered` is the count with at
+   * least one matching `request` raw event.
+   */
+  rawSpanCoverage: {
+    covered: number;
+    total: number;
+  };
+  issues: RunIntegrityIssue[];
+}
+declare class RunIntegrityError extends CaptureIntegrityError {
+  readonly report: RunIntegrityReport;
+  constructor(report: RunIntegrityReport);
+}
+declare function assertRunCaptured(store: TraceStore, runId: string, expectations?: RunIntegrityExpectations): Promise<RunIntegrityReport>;
+/** Strict mode: throws `RunIntegrityError` when the report isn't ok. */
+declare function throwIfRunIncomplete(report: RunIntegrityReport): void;
+//#endregion
+export { RunIntegrityReport as a, RunIntegrityIssueCode as i, RunIntegrityExpectations as n, assertRunCaptured as o, RunIntegrityIssue as r, throwIfRunIncomplete as s, RunIntegrityError as t };
+//# sourceMappingURL=integrity-rmVhXWA7.d.ts.map

package/dist/integrity-rmVhXWA7.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"integrity-rmVhXWA7.d.ts","names":[],"sources":["../src/trace/integrity.ts"],"mappings":";;;;UAyBiB;;EAEf;;EAEA;;EAEA;;;;;EAKA,UAAU;;EAEV;;;;;;EAMA;;EAEA;;KAGU;UAUK;EACf,MAAM;EACN;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;EAMA;IAAmB;IAAiB;;EACpC,QAAQ;;cAGG,0BAA0B;WACT,QAAQ;EAApC,YAA4B,QAAQ;;iBAOhB,kBACpB,OAAO,YACP,eACA,eAAc,2BACb,QAAQ;;iBA6GK,qBAAqB,QAAQ"}