@tangle-network/agent-eval 0.173.3 → 0.175.0

package/dist/{integrity-BWywb34E.js → integrity-DsHWCebQ.js}

@@ -1,5 +1,6 @@
 import { c as unscreenedRewardFields } from "./reward-nw2xZGZG.js";
 import { f as validateRolloutLine, r as ROLLOUT_SCHEMA, u as isRolloutLine } from "./schema-C1aaAxTf.js";
+import { d as parseJson, p as parseSupervisorTree, r as readTerminalRecord, s as isUnavailable, u as asRecord } from "./terminal-record-Ce9_UjRz.js";
 //#region src/supervisor-run/integrity-issues.ts
 function evidence(path, value) {
 	return {
@@ -762,436 +763,6 @@ function treeIssues(tree, uncertainty = {}) {
 //#region src/supervisor-run/integrity-types.ts
 const SUPERVISOR_RUN_INTEGRITY_SCHEMA = "tangle.supervisor-run-integrity@1";
 //#endregion
-//#region src/supervisor-run/source-facts.ts
-function asRecord(v) {
-	return typeof v === "object" && v !== null ? v : {};
-}
-function num(v) {
-	return typeof v === "number" && Number.isFinite(v) ? v : 0;
-}
-function readSpend(v) {
-	const rec = asRecord(v);
-	const tok = asRecord(rec.tokens);
-	const cacheRead = tok.cacheRead ?? tok.cache_read;
-	const cacheWrite = tok.cacheWrite ?? tok.cache_write;
-	return {
-		tokens: {
-			input: num(tok.input),
-			output: num(tok.output),
-			cacheRead: num(cacheRead),
-			cacheWrite: num(cacheWrite),
-			hasCache: cacheRead !== void 0 || cacheWrite !== void 0,
-			breakdownKnown: tok.cacheBreakdownKnown !== false
-		},
-		tokensKnown: rec.tokensKnown !== false,
-		usd: num(rec.usd),
-		usdKnown: rec.usdKnown !== false
-	};
-}
-function parseJsonl(text) {
-	return parseJsonlWithDiagnostics(text).rows;
-}
-function parseJsonlWithDiagnostics(text) {
-	if (text === null) return {
-		rows: [],
-		invalidRows: 0
-	};
-	const out = [];
-	let invalidRows = 0;
-	for (const line of text.split("\n")) {
-		const trimmed = line.trim();
-		if (!trimmed) continue;
-		try {
-			const parsed = JSON.parse(trimmed);
-			if (typeof parsed === "object" && parsed !== null && !Array.isArray(parsed)) out.push(parsed);
-			else invalidRows += 1;
-		} catch {
-			invalidRows += 1;
-		}
-	}
-	return {
-		rows: out,
-		invalidRows
-	};
-}
-function parseJson(text) {
-	if (text === null) return null;
-	try {
-		const parsed = JSON.parse(text);
-		return typeof parsed === "object" && parsed !== null ? parsed : null;
-	} catch {
-		return null;
-	}
-}
-function ms(at) {
-	if (typeof at !== "string") return null;
-	const t = Date.parse(at);
-	return Number.isFinite(t) ? t : null;
-}
-function readVerdict(v) {
-	if (typeof v === "string") return {
-		label: v,
-		valid: null,
-		score: null,
-		raw: v
-	};
-	if (typeof v !== "object" || v === null) return {
-		label: null,
-		valid: null,
-		score: null,
-		raw: null
-	};
-	const rec = v;
-	const valid = typeof rec.valid === "boolean" ? rec.valid : null;
-	return {
-		label: valid === null ? null : valid ? "valid" : "invalid",
-		valid,
-		score: typeof rec.score === "number" && Number.isFinite(rec.score) ? rec.score : null,
-		raw: v
-	};
-}
-function workerSourceKey(worker) {
-	return worker.workerId ?? worker.label;
-}
-/** Event kinds this parser folds into the tree. */
-const TREE_EVENT_KINDS = [
-	"spawned",
-	"settled",
-	"cancelled",
-	"metered"
-];
-/**
-* Event kinds `agent-runtime` journals that this parser recognizes but does not model as
-* tree nodes: `waiting` arms a wait node and `woken` settles it. They are counted by kind
-* rather than dropped, so "understood, not modeled" can never be mistaken for "absent".
-*/
-const UNMODELLED_EVENT_KINDS = ["waiting", "woken"];
-function treeEventKind(value) {
-	return TREE_EVENT_KINDS.includes(value) ? value : null;
-}
-/**
-* Classify one parsed journal row. Every row lands in exactly one outcome — interpreted,
-* recognized-but-not-modeled, or unreadable — because a row that falls through silently is
-* how a run with events reads as a run with none.
-*/
-function readJournalRow(row) {
-	const kind = row.kind;
-	if (typeof kind !== "string") return { outcome: "unreadable" };
-	if (kind === "begin" && typeof row.root === "string") return {
-		outcome: "ignored",
-		dialect: "runtime-envelope",
-		kind: "begin"
-	};
-	if (kind === "event") {
-		const inner = row.event;
-		if (typeof inner !== "object" || inner === null || Array.isArray(inner)) return { outcome: "unreadable" };
-		const innerKind = inner.kind;
-		if (typeof innerKind !== "string") return { outcome: "unreadable" };
-		const tree = treeEventKind(innerKind);
-		if (tree !== null) return {
-			outcome: "event",
-			dialect: "runtime-envelope",
-			kind: tree,
-			event: inner
-		};
-		return {
-			outcome: "ignored",
-			dialect: "runtime-envelope",
-			kind: innerKind
-		};
-	}
-	const tree = treeEventKind(kind);
-	if (tree !== null) return {
-		outcome: "event",
-		dialect: "flat",
-		kind: tree,
-		event: row
-	};
-	if (UNMODELLED_EVENT_KINDS.includes(kind)) return {
-		outcome: "ignored",
-		dialect: "flat",
-		kind
-	};
-	return { outcome: "unreadable" };
-}
-function summarizeDialect(seen) {
-	if (seen.size === 0) return "none";
-	if (seen.size > 1) return "mixed";
-	return seen.has("runtime-envelope") ? "runtime-envelope" : "flat";
-}
-function parseSupervisorTree(src) {
-	const parsedJournal = parseJsonlWithDiagnostics(src.journal);
-	const events = parsedJournal.rows;
-	const state = parseJson(src.state);
-	const spawns = [];
-	const closes = [];
-	let brainIn = 0;
-	let brainOut = 0;
-	let brainCacheRead = 0;
-	let brainCacheWrite = 0;
-	let brainHasCache = false;
-	let brainCacheBreakdownKnown = true;
-	let brainTokensKnownCount = 0;
-	let brainTokensUnknownCount = 0;
-	let brainUsd = 0;
-	let brainUsdKnownCount = 0;
-	let brainUsdUnknownCount = 0;
-	let meteredCount = 0;
-	let rootId = null;
-	let unreadableRows = 0;
-	const ignoredByKind = /* @__PURE__ */ new Map();
-	const dialectsSeen = /* @__PURE__ */ new Set();
-	const meteredRows = [];
-	for (const [sourceRow, row] of events.entries()) {
-		const reading = readJournalRow(row);
-		if (reading.outcome === "unreadable") {
-			unreadableRows += 1;
-			continue;
-		}
-		dialectsSeen.add(reading.dialect);
-		if (reading.outcome === "ignored") {
-			ignoredByKind.set(reading.kind, (ignoredByKind.get(reading.kind) ?? 0) + 1);
-			continue;
-		}
-		const { kind, event: ev } = reading;
-		const id = typeof ev.id === "string" ? ev.id : "";
-		if (kind === "spawned") {
-			const parent = typeof ev.parent === "string" && ev.parent.length > 0 ? ev.parent : null;
-			const label = typeof ev.label === "string" ? ev.label : "";
-			const invalidFields = [];
-			if (id.length === 0) invalidFields.push("id");
-			const parentOmittedForFirstRoot = ev.parent === void 0 && rootId === null;
-			if (ev.parent !== null && !parentOmittedForFirstRoot && !(typeof ev.parent === "string" && ev.parent.length > 0)) invalidFields.push("parent");
-			if (label.length === 0) invalidFields.push("label");
-			if (ev.role !== void 0 && ev.role !== "supervisor" && ev.role !== "worker") invalidFields.push("role");
-			if (ev.profileDigest !== void 0 && !(typeof ev.profileDigest === "string" && ev.profileDigest.length > 0)) invalidFields.push("profileDigest");
-			if (ev.runtime !== void 0 && !(typeof ev.runtime === "string" && ev.runtime.length > 0)) invalidFields.push("runtime");
-			const role = ev.role === "supervisor" || ev.role === "worker" ? ev.role : null;
-			const valid = invalidFields.length === 0;
-			if (valid && parent === null && rootId === null) rootId = id;
-			spawns.push({
-				sourceRow,
-				id,
-				parent,
-				label,
-				role,
-				profileDigest: typeof ev.profileDigest === "string" && ev.profileDigest.length > 0 ? ev.profileDigest : null,
-				runtime: typeof ev.runtime === "string" && ev.runtime.length > 0 ? ev.runtime : null,
-				at: ms(ev.at),
-				valid,
-				invalidFields
-			});
-		} else if (kind === "settled") {
-			const verdict = readVerdict(ev.verdict);
-			closes.push({
-				id,
-				kind: "settled",
-				status: typeof ev.status === "string" ? ev.status : null,
-				verdict: verdict.label,
-				valid: verdict.valid,
-				score: verdict.score,
-				rawVerdict: verdict.raw,
-				reason: typeof ev.reason === "string" ? ev.reason : null,
-				infra: typeof ev.infra === "boolean" ? ev.infra : null,
-				at: ms(ev.at),
-				spend: readSpend(ev.spent),
-				hasSpend: asRecord(ev.spent).tokens !== void 0
-			});
-		} else if (kind === "cancelled") closes.push({
-			id,
-			kind: "cancelled",
-			status: "cancelled",
-			verdict: typeof ev.reason === "string" ? ev.reason : null,
-			valid: null,
-			score: null,
-			rawVerdict: typeof ev.reason === "string" ? ev.reason : null,
-			reason: typeof ev.reason === "string" ? ev.reason : null,
-			infra: null,
-			at: ms(ev.at),
-			spend: {
-				tokens: {
-					input: 0,
-					output: 0,
-					cacheRead: 0,
-					cacheWrite: 0,
-					hasCache: false,
-					breakdownKnown: true
-				},
-				tokensKnown: true,
-				usd: 0,
-				usdKnown: true
-			},
-			hasSpend: false
-		});
-		else if (kind === "metered") meteredRows.push({
-			id,
-			spend: readSpend(ev.spend),
-			at: ms(ev.at)
-		});
-		else throw new Error(`unhandled journal event kind ${JSON.stringify(kind)}`);
-	}
-	for (const row of meteredRows) {
-		if (rootId === null || row.id !== rootId) continue;
-		if (row.spend.tokensKnown) {
-			brainIn += row.spend.tokens.input;
-			brainOut += row.spend.tokens.output;
-			brainCacheRead += row.spend.tokens.cacheRead;
-			brainCacheWrite += row.spend.tokens.cacheWrite;
-			brainHasCache = brainHasCache || row.spend.tokens.hasCache;
-			brainCacheBreakdownKnown = brainCacheBreakdownKnown && row.spend.tokens.breakdownKnown;
-			brainTokensKnownCount += 1;
-		} else brainTokensUnknownCount += 1;
-		if (row.spend.usdKnown) {
-			brainUsd += row.spend.usd;
-			brainUsdKnownCount += 1;
-		} else brainUsdUnknownCount += 1;
-		meteredCount += 1;
-	}
-	const workerSpawns = spawns.filter((s) => s.valid && s.id !== rootId);
-	const workerIds = new Set(workerSpawns.map((s) => s.id));
-	const workerCloses = closes.filter((c) => workerIds.has(c.id));
-	const workerLogs = /* @__PURE__ */ new Map();
-	const workerLogRows = [];
-	for (const [sourceIndex, w] of (src.workers ?? []).entries()) {
-		const facts = {
-			sourceIndex,
-			workerId: w.workerId ?? null,
-			label: w.label,
-			inboxCaptured: w.inbox !== null,
-			eventsCaptured: w.events !== null,
-			inboxInvalidRows: 0,
-			eventsInvalidRows: 0,
-			started: null,
-			finished: false,
-			finishedAt: null,
-			passed: null,
-			score: null,
-			finishedPatchBytes: null,
-			evidenceBytes: 0,
-			steersQueued: null,
-			steersDelivered: null,
-			steersQueuedUnavailable: null,
-			steersDeliveredUnavailable: null,
-			questions: 0,
-			steerRequests: [],
-			steerAcknowledgements: []
-		};
-		const parsedInbox = parseJsonlWithDiagnostics(w.inbox);
-		const parsedEvents = parseJsonlWithDiagnostics(w.events);
-		facts.inboxInvalidRows = parsedInbox.invalidRows;
-		facts.eventsInvalidRows = parsedEvents.invalidRows;
-		for (const [row, req] of parsedInbox.rows.entries()) {
-			if (typeof req.message !== "string" || req.message.trim().length === 0) continue;
-			facts.steerRequests.push({
-				requestId: typeof req.id === "string" && req.id.length > 0 ? req.id : null,
-				row
-			});
-		}
-		for (const [row, ev] of parsedEvents.rows.entries()) {
-			const kind = ev.kind;
-			if (kind === "message") {
-				if (ev.direction === "up") {
-					facts.questions += 1;
-					continue;
-				}
-				facts.steerAcknowledgements.push({
-					requestId: typeof ev.requestId === "string" && ev.requestId.length > 0 ? ev.requestId : null,
-					delivered: typeof ev.delivered === "boolean" ? ev.delivered : null,
-					row
-				});
-			} else if (kind === "started") facts.started = ms(ev.at);
-			else if (kind === "finished") {
-				facts.finished = true;
-				facts.finishedAt = ms(ev.at);
-				facts.passed = typeof ev.passed === "boolean" ? ev.passed : null;
-				facts.score = typeof ev.score === "number" && Number.isFinite(ev.score) ? ev.score : null;
-				facts.finishedPatchBytes = typeof ev.patchBytes === "number" ? ev.patchBytes : null;
-				facts.evidenceBytes = typeof ev.evidence === "string" ? ev.evidence.length : 0;
-			}
-		}
-		const requestsById = /* @__PURE__ */ new Map();
-		for (const request of facts.steerRequests) {
-			if (request.requestId === null) continue;
-			requestsById.set(request.requestId, (requestsById.get(request.requestId) ?? 0) + 1);
-		}
-		const acknowledgementsById = /* @__PURE__ */ new Map();
-		for (const acknowledgement of facts.steerAcknowledgements) {
-			if (acknowledgement.requestId === null) continue;
-			const matches = acknowledgementsById.get(acknowledgement.requestId) ?? [];
-			matches.push(acknowledgement);
-			acknowledgementsById.set(acknowledgement.requestId, matches);
-		}
-		const queueProblems = [];
-		if (!facts.inboxCaptured) queueProblems.push("inbox absent");
-		if (!facts.eventsCaptured) queueProblems.push("events absent");
-		if (facts.inboxInvalidRows > 0) queueProblems.push("inbox contains malformed rows");
-		if (facts.eventsInvalidRows > 0) queueProblems.push("events contain malformed rows");
-		if (facts.steerRequests.some((request) => request.requestId === null)) queueProblems.push("queued request id missing");
-		if ([...requestsById.values()].some((count) => count > 1)) queueProblems.push("queued request id duplicated");
-		if (facts.steerAcknowledgements.some((acknowledgement) => acknowledgement.requestId === null)) queueProblems.push("acknowledgement request id missing");
-		if ([...acknowledgementsById.keys()].some((id) => !requestsById.has(id))) queueProblems.push("acknowledgement has no queued request");
-		if (queueProblems.length === 0) facts.steersQueued = requestsById.size;
-		else facts.steersQueuedUnavailable = [...new Set(queueProblems)].join("; ");
-		const deliveryProblems = [...queueProblems];
-		if ([...acknowledgementsById.values()].some((rows) => rows.length > 1)) deliveryProblems.push("acknowledgement request id duplicated");
-		if (facts.steerAcknowledgements.some((acknowledgement) => acknowledgement.delivered === null)) deliveryProblems.push("acknowledgement delivery status missing");
-		if (deliveryProblems.length === 0) facts.steersDelivered = [...requestsById.keys()].filter((id) => {
-			const acknowledgement = acknowledgementsById.get(id);
-			return acknowledgement?.length === 1 && acknowledgement[0]?.delivered === true;
-		}).length;
-		else facts.steersDeliveredUnavailable = [...new Set(deliveryProblems)].join("; ");
-		workerLogRows.push(facts);
-		workerLogs.set(workerSourceKey(w), facts);
-	}
-	const startedAt = ms(state?.startedAt) ?? spawns[0]?.at ?? null;
-	const rootClose = rootId === null ? null : closes.find((close) => close.id === rootId);
-	const completedAt = ms(state?.completedAt) ?? rootClose?.at ?? null;
-	let firstEventAt = null;
-	let lastEventAt = null;
-	const widenEventSpan = (at) => {
-		if (at === null) return;
-		if (firstEventAt === null || at < firstEventAt) firstEventAt = at;
-		if (lastEventAt === null || at > lastEventAt) lastEventAt = at;
-	};
-	for (const spawn of spawns) widenEventSpan(spawn.at);
-	for (const close of closes) widenEventSpan(close.at);
-	for (const metered of meteredRows) widenEventSpan(metered.at);
-	return {
-		rootId,
-		spawns,
-		closes,
-		workerSpawns,
-		workerCloses,
-		brain: {
-			tokensIn: brainIn,
-			tokensOut: brainOut,
-			cacheRead: brainCacheRead,
-			cacheWrite: brainCacheWrite,
-			hasCache: brainHasCache,
-			cacheBreakdownKnown: brainCacheBreakdownKnown,
-			tokensKnownCount: brainTokensKnownCount,
-			tokensUnknownCount: brainTokensUnknownCount,
-			usd: brainUsd,
-			usdKnownCount: brainUsdKnownCount,
-			usdUnknownCount: brainUsdUnknownCount,
-			meteredCount
-		},
-		workerLogs,
-		workerLogRows,
-		journalRows: events.length + parsedJournal.invalidRows,
-		journalInvalidRows: parsedJournal.invalidRows + unreadableRows,
-		journalMalformedJsonRows: parsedJournal.invalidRows,
-		journalIgnoredRowsByKind: Object.fromEntries(ignoredByKind),
-		journalMeteredRows: meteredRows.length,
-		journalDialect: summarizeDialect(dialectsSeen),
-		state,
-		startedAt,
-		completedAt,
-		firstEventAt,
-		lastEventAt
-	};
-}
-//#endregion
 //#region src/supervisor-run/rollout-nodes.ts
 /**
 * The supervision tree as `tangle.rollout.v1` rows.
@@ -1260,6 +831,11 @@ function supervisorRunRolloutLinesFromFacts(src, tree, opts = {}) {
 	const result = parseJson(src.result);
 	const judge = parseJson(src.judge);
 	const stateResult = asRecord(state?.result);
+	const terminal = readTerminalRecord({
+		state,
+		result,
+		failure: src.failure === void 0 ? void 0 : parseJson(src.failure)
+	});
 	const rootId = tree.rootId;
 	const runId = opts.runId ?? rootId ?? src.runRef;
 	const suite = opts.suite ?? "supervisor-run";
@@ -1314,7 +890,7 @@ function supervisorRunRolloutLinesFromFacts(src, tree, opts = {}) {
 				verdict: judge ?? null,
 				metrics: {
 					arm: src.arm,
-					sup_status: typeof state?.status === "string" ? state.status : null,
+					sup_status: isUnavailable(terminal.supStatus) ? null : terminal.supStatus,
 					sup_verdict: typeof state?.verdict === "string" ? state.verdict : null,
 					delivered: typeof stateResult.delivered === "boolean" ? stateResult.delivered : null,
 					verify_pass: typeof result?.verify_pass === "boolean" ? result.verify_pass : null,
@@ -1325,9 +901,9 @@ function supervisorRunRolloutLinesFromFacts(src, tree, opts = {}) {
 					runtime: rootSpawn?.runtime ?? null,
 					profile_digest: rootSpawn?.profileDigest ?? null
 				},
-				is_completed: state?.status === "completed",
+				is_completed: terminal.completed === true,
 				is_truncated: false,
-				error: result?.kind === "interrupted" && typeof result.reason === "string" ? result.reason : null,
+				error: terminal.completed !== false ? null : !isUnavailable(terminal.supReason) && terminal.supReason !== null ? terminal.supReason : !isUnavailable(terminal.failure) && terminal.failure !== null ? terminal.failure.message ?? terminal.failure.name : null,
 				realness_gated: false
 			},
 			cost: {
@@ -1513,6 +1089,6 @@ function analyzeSupervisorRunIntegrity(input, options = {}) {
 	};
 }
 //#endregion
-export { parseJsonl as a, SUPERVISOR_RUN_INTEGRITY_SCHEMA as c, parseJson as i, supervisorRunRolloutLines as n, parseSupervisorTree as o, asRecord as r, workerSourceKey as s, analyzeSupervisorRunIntegrity as t };
+export { supervisorRunRolloutLines as n, SUPERVISOR_RUN_INTEGRITY_SCHEMA as r, analyzeSupervisorRunIntegrity as t };
 
-//# sourceMappingURL=integrity-BWywb34E.js.map
+//# sourceMappingURL=integrity-DsHWCebQ.js.map