@tangle-network/agent-eval 0.135.3 → 0.136.0

package/dist/{supervisor-run-B7lUGoyZ.js → supervisor-run-Dr5HnTup.js}

@@ -18,6 +18,8 @@ function showMeasured(v) {
 }
 /** A source that carries every fact the analyzer can use. */
 const NO_SOURCE_LIMITS = {
+	managerTokens: null,
+	workerTokens: null,
 	spendUsd: null,
 	workerVerdicts: null,
 	deliverables: null
@@ -82,6 +84,31 @@ function ms(at) {
 	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;
+}
 function parseSupervisorTree(src) {
 	const events = parseJsonl(src.journal);
 	const state = parseJson(src.state);
@@ -102,26 +129,36 @@ function parseSupervisorTree(src) {
 			const parent = typeof ev.parent === "string" ? ev.parent : null;
 			const label = typeof ev.label === "string" ? ev.label : "";
 			if (parent === null && rootId === null) rootId = id;
+			const role = parent === null || ev.role === "supervisor" ? "supervisor" : "worker";
 			spawns.push({
 				id,
 				parent,
 				label,
+				role,
 				at: ms(ev.at)
 			});
-		} else if (kind === "settled") closes.push({
-			id,
-			kind: "settled",
-			status: typeof ev.status === "string" ? ev.status : null,
-			verdict: typeof ev.verdict === "string" ? ev.verdict : null,
-			at: ms(ev.at),
-			spend: readSpend(ev.spent),
-			hasSpend: asRecord(ev.spent).tokens !== void 0
-		});
-		else if (kind === "cancelled") closes.push({
+		} 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,
+				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,
 			at: ms(ev.at),
 			spend: {
 				tokens: {
@@ -156,6 +193,7 @@ function parseSupervisorTree(src) {
 			finished: false,
 			finishedAt: null,
 			passed: null,
+			score: null,
 			finishedPatchBytes: null,
 			evidenceBytes: 0,
 			steersQueued: 0,
@@ -177,11 +215,12 @@ function parseSupervisorTree(src) {
 				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;
 			}
 		}
-		workerLogs.set(w.label, facts);
+		workerLogs.set(workerSourceKey(w), facts);
 	}
 	const startedAt = ms(state?.startedAt) ?? spawns[0]?.at ?? null;
 	const completedAt = ms(state?.completedAt) ?? [...spawns.map((s) => s.at), ...closes.map((c) => c.at)].reduce((acc, t) => t === null ? acc : acc === null ? t : Math.max(acc, t), null);
@@ -222,17 +261,34 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 	const judge = parseJson(src.judge);
 	const tree = parseSupervisorTree(src);
 	const { rootId, workerSpawns, workerCloses, startedAt, completedAt } = tree;
+	const spawnById = new Map(workerSpawns.map((spawn) => [spawn.id, spawn]));
+	const spawnsByLabel = /* @__PURE__ */ new Map();
+	for (const spawn of workerSpawns) {
+		const matches = spawnsByLabel.get(spawn.label) ?? [];
+		matches.push(spawn);
+		spawnsByLabel.set(spawn.label, matches);
+	}
+	const spawnsForSource = (worker) => {
+		if (worker.workerId === void 0) return spawnsByLabel.get(worker.label) ?? [];
+		const spawn = spawnById.get(worker.workerId);
+		return spawn === void 0 ? [] : [spawn];
+	};
+	const spawnForSource = (worker) => {
+		const matches = spawnsForSource(worker);
+		return matches.length === 1 ? matches[0] ?? null : null;
+	};
 	const supervisorWallMs = startedAt !== null && completedAt !== null && completedAt >= startedAt ? completedAt - startedAt : !haveJournal ? gap("supervisorWallMs", journalMissing) : gap("supervisorWallMs", "no parseable start/complete timestamps in state.json or journal");
 	const steerRows = [];
 	let steerQueuedTotal = 0;
 	let steerDeliveredTotal = 0;
 	let upLegMessages = 0;
 	if (src.workers !== null) for (const w of src.workers) {
-		const facts = tree.workerLogs.get(w.label);
+		const facts = tree.workerLogs.get(workerSourceKey(w));
 		const queued = facts?.steersQueued ?? 0;
 		const delivered = facts?.steersDelivered ?? 0;
 		upLegMessages += facts?.questions ?? 0;
 		steerRows.push({
+			workerId: w.workerId ?? null,
 			worker: w.label,
 			queued,
 			delivered
@@ -287,11 +343,39 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 		sumWorkerWallMs += span * live;
 	}
 	const firstWorkerSpawnAt = workerSpawns.reduce((acc, s) => s.at === null ? acc : acc === null ? s.at : Math.min(acc, s.at), null);
-	const firstSettleAt = workerCloses.reduce((acc, c) => c.at === null ? acc : acc === null ? c.at : Math.min(acc, c.at), null);
-	const respawns = firstSettleAt === null ? 0 : workerSpawns.filter((s) => s.at !== null && s.at > firstSettleAt).length;
-	const labelCounts = /* @__PURE__ */ new Map();
-	for (const s of workerSpawns) labelCounts.set(s.label, (labelCounts.get(s.label) ?? 0) + 1);
-	const repeatedLabels = [...labelCounts.entries()].filter(([, n]) => n > 1).map(([l]) => l);
+	const closeById = new Map(workerCloses.map((close) => [close.id, close]));
+	const childSpawnsByParent = /* @__PURE__ */ new Map();
+	for (const spawn of workerSpawns) {
+		if (spawn.parent === null) continue;
+		const siblings = childSpawnsByParent.get(spawn.parent) ?? [];
+		siblings.push(spawn);
+		childSpawnsByParent.set(spawn.parent, siblings);
+	}
+	let respawns = 0;
+	let observeThenRespawn = 0;
+	let respawnWithoutEvidence = 0;
+	const repeatedLabelSet = /* @__PURE__ */ new Set();
+	for (const siblings of childSpawnsByParent.values()) {
+		const labelCounts = /* @__PURE__ */ new Map();
+		for (const spawn of siblings) labelCounts.set(spawn.label, (labelCounts.get(spawn.label) ?? 0) + 1);
+		for (const [label, count] of labelCounts) if (count > 1) repeatedLabelSet.add(label);
+		const orderedSpawns = siblings.map((spawn, index) => ({
+			spawn,
+			index
+		})).filter((row) => row.spawn.at !== null).sort((a, b) => a.spawn.at - b.spawn.at || a.index - b.index);
+		const directCloseTimes = siblings.map((spawn) => closeById.get(spawn.id)?.at ?? null).filter((at) => at !== null).sort((a, b) => a - b);
+		const firstDirectClose = directCloseTimes[0] ?? null;
+		for (let i = 1; i < orderedSpawns.length; i += 1) {
+			const previous = orderedSpawns[i - 1]?.spawn.at;
+			const current = orderedSpawns[i]?.spawn.at;
+			if (previous === void 0 || current === void 0) continue;
+			if (firstDirectClose === null || current <= firstDirectClose) continue;
+			respawns += 1;
+			if (hasNumberBetween(directCloseTimes, previous, current)) observeThenRespawn += 1;
+			else respawnWithoutEvidence += 1;
+		}
+	}
+	const repeatedLabels = [...repeatedLabelSet];
 	const parentOf = new Map(tree.spawns.map((s) => [s.id, s.parent]));
 	let delegationDepth = 0;
 	for (const s of workerSpawns) {
@@ -333,35 +417,32 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 		if (c.verdict !== null) settledVerdicts[c.verdict] = (settledVerdicts[c.verdict] ?? 0) + 1;
 	}
 	const verdictLimit = src.limits.workerVerdicts;
+	const acceptedLimit = verdictLimit ?? src.limits.deliverables;
 	let accepted = 0;
-	let rejected = 0;
 	let emptyPass = 0;
 	let evidenceBytes = 0;
+	const sourceVerdicts = [];
 	for (const w of src.workers ?? []) {
-		const f = tree.workerLogs.get(w.label);
-		if (f === void 0 || !f.finished) continue;
-		evidenceBytes += f.evidenceBytes;
-		if (f.passed === true) if ((f.finishedPatchBytes ?? 0) > 0) accepted += 1;
+		const f = tree.workerLogs.get(workerSourceKey(w));
+		if (f?.finished) evidenceBytes += f.evidenceBytes;
+		const spawn = spawnForSource(w);
+		const passed = (spawn === null ? null : closeById.get(spawn.id) ?? null)?.valid ?? f?.passed ?? null;
+		if (passed !== null) sourceVerdicts.push(passed);
+		if (passed === true) if ((w.patchBytes ?? f?.finishedPatchBytes ?? 0) > 0) accepted += 1;
 		else emptyPass += 1;
-		else if (f.passed === false) rejected += 1;
-	}
-	let observeThenRespawn = 0;
-	let respawnWithoutEvidence = 0;
-	const spawnTimes = workerSpawns.map((s) => s.at).filter((t) => t !== null).sort((a, b) => a - b);
-	const closeTimes = workerCloses.map((c) => c.at).filter((t) => t !== null).sort((a, b) => a - b);
-	for (let i = 1; i < spawnTimes.length; i += 1) {
-		const prevSpawn = spawnTimes[i - 1];
-		const thisSpawn = spawnTimes[i];
-		const sawEvidence = closeTimes.some((t) => t >= prevSpawn && t <= thisSpawn);
-		if (firstSettleAt !== null && thisSpawn > firstSettleAt) if (sawEvidence) observeThenRespawn += 1;
-		else respawnWithoutEvidence += 1;
 	}
+	const settledCloses = workerCloses.filter((close) => close.kind === "settled");
+	const structuredVerdicts = settledCloses.map((close) => close.valid).filter((valid) => valid !== null);
+	const journalVerdictsComplete = settledCloses.length > 0 && structuredVerdicts.length === settledCloses.length;
+	const sourceVerdictsComplete = sourceVerdicts.length > 0 && sourceVerdicts.length >= settledCloses.length;
+	const rejected = journalVerdictsComplete ? structuredVerdicts.filter((valid) => !valid).length : sourceVerdictsComplete ? sourceVerdicts.filter((valid) => !valid).length : 0;
+	const rejectedLimit = journalVerdictsComplete || sourceVerdictsComplete ? null : settledCloses.length > 0 ? verdictLimit ?? "a settled journal verdict has no validity and no matched worker log" : verdictLimit !== null ? verdictLimit : !haveJournal && src.workers === null ? workersGapReason : null;
 	const decision = {
 		settledByStatus: haveJournal ? settledByStatus : gap("settledByStatus", journalMissing),
 		settledVerdicts: verdictLimit !== null ? unavailable(verdictLimit) : haveJournal ? settledVerdicts : unavailable(journalMissing),
-		accepted: verdictLimit !== null ? gap("accepted", verdictLimit) : src.workers === null ? unavailable(workersGapReason) : accepted,
-		rejected: verdictLimit !== null ? unavailable(verdictLimit) : src.workers === null ? unavailable(workersGapReason) : rejected,
-		emptyPass: verdictLimit !== null || src.limits.deliverables !== null ? unavailable(verdictLimit ?? src.limits.deliverables) : src.workers === null ? unavailable(workersGapReason) : emptyPass,
+		accepted: acceptedLimit !== null ? gap("accepted", acceptedLimit) : src.workers === null ? unavailable(workersGapReason) : accepted,
+		rejected: rejectedLimit === null ? rejected : unavailable(rejectedLimit),
+		emptyPass: acceptedLimit !== null ? gap("emptyPass", acceptedLimit) : src.workers === null ? unavailable(workersGapReason) : emptyPass,
 		observeThenRespawn: haveJournal ? observeThenRespawn : unavailable(journalMissing),
 		respawnWithoutEvidence: haveJournal ? respawnWithoutEvidence : unavailable(journalMissing),
 		reviewActions: src.workers === null ? unavailable(workersGapReason) : steerQueuedTotal + upLegMessages,
@@ -370,52 +451,64 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 	const journalWorkerIn = workerCloses.reduce((a, c) => a + c.spend.tokens.input, 0);
 	const journalWorkerOut = workerCloses.reduce((a, c) => a + c.spend.tokens.output, 0);
 	const journalWorkerUsd = workerCloses.reduce((a, c) => a + c.spend.usd, 0);
-	const labelById = new Map(tree.workerSpawns.map((s) => [s.id, s.label]));
+	const workerUsdById = /* @__PURE__ */ new Map();
+	for (const c of workerCloses) workerUsdById.set(c.id, (workerUsdById.get(c.id) ?? 0) + c.spend.usd);
+	const labelById = new Map(workerSpawns.map((spawn) => [spawn.id, spawn.label]));
 	const workerUsdByLabel = /* @__PURE__ */ new Map();
-	for (const c of workerCloses) {
-		const label = labelById.get(c.id);
+	for (const close of workerCloses) {
+		const label = labelById.get(close.id);
 		if (label === void 0) continue;
-		workerUsdByLabel.set(label, (workerUsdByLabel.get(label) ?? 0) + c.spend.usd);
+		workerUsdByLabel.set(label, (workerUsdByLabel.get(label) ?? 0) + close.spend.usd);
 	}
 	const sq = src.harnessWorkerTokens;
 	const harnessGapReason = src.harnessMissingReason ?? "harness session store unavailable and journal settled spend is 0";
-	const workerIn = sq !== null ? journalWorkerIn + sq.input : journalWorkerIn > 0 ? journalWorkerIn : gap("workers.tokensIn", harnessGapReason);
-	const workerOut = sq !== null ? journalWorkerOut + sq.output : journalWorkerOut > 0 ? journalWorkerOut : unavailable(harnessGapReason);
+	const workerTokenLimit = src.limits.workerTokens;
+	const workerIn = workerTokenLimit !== null ? gap("workers.tokensIn", workerTokenLimit) : sq !== null ? journalWorkerIn + sq.input : haveJournal ? journalWorkerIn : gap("workers.tokensIn", harnessGapReason);
+	const workerOut = workerTokenLimit !== null ? unavailable(workerTokenLimit) : sq !== null ? journalWorkerOut + sq.output : haveJournal ? journalWorkerOut : unavailable(harnessGapReason);
 	const stateResult = asRecord(state?.result);
 	const stateUsd = typeof stateResult.spentUsd === "number" ? stateResult.spentUsd : null;
 	const usdLimit = src.limits.spendUsd;
 	const totalUsd = usdLimit !== null ? gap("totalUsd", usdLimit) : stateUsd !== null ? round(stateUsd, 6) : haveJournal ? round(tree.brain.usd + journalWorkerUsd, 6) : gap("totalUsd", journalMissing);
 	const perWorker = (src.workers ?? []).map((w) => {
-		const f = tree.workerLogs.get(w.label);
+		const f = tree.workerLogs.get(workerSourceKey(w));
+		const matchingSpawns = spawnsForSource(w);
+		const spawn = spawnForSource(w);
+		const close = spawn === null ? null : closeById.get(spawn.id) ?? null;
+		const passed = close?.valid ?? f?.passed ?? null;
+		const matchingRoles = new Set(matchingSpawns.map((candidate) => candidate.role));
 		return {
+			workerId: w.workerId ?? null,
 			worker: w.label,
+			role: matchingRoles.size === 1 ? matchingSpawns[0]?.role ?? null : null,
 			wallMs: f?.started != null && f.finishedAt != null ? f.finishedAt - f.started : null,
 			tokensIn: w.tokensIn ?? null,
 			tokensOut: w.tokensOut ?? null,
-			usd: usdLimit !== null ? null : workerUsdByLabel.get(w.label) ?? null,
+			usd: usdLimit !== null ? null : w.workerId === void 0 ? workerUsdByLabel.get(w.label) ?? null : workerUsdById.get(w.workerId) ?? null,
 			patchBytes: w.patchBytes ?? f?.finishedPatchBytes ?? null,
-			passed: f?.passed ?? null
+			passed,
+			score: close?.score ?? f?.score ?? null
 		};
 	});
 	const walls = perWorker.map((w) => w.wallMs).filter((w) => w !== null).sort((a, b) => a - b);
 	const brainCalls = parseJsonl(src.brainLog);
+	const managerTokenLimit = src.limits.managerTokens;
 	const economics = {
 		brain: {
-			tokensIn: haveJournal ? tree.brain.tokensIn : gap("brain.tokensIn", journalMissing),
-			tokensOut: haveJournal ? tree.brain.tokensOut : unavailable(journalMissing),
+			tokensIn: managerTokenLimit !== null ? gap("brain.tokensIn", managerTokenLimit) : haveJournal ? tree.brain.tokensIn : gap("brain.tokensIn", journalMissing),
+			tokensOut: managerTokenLimit !== null ? unavailable(managerTokenLimit) : haveJournal ? tree.brain.tokensOut : unavailable(journalMissing),
 			usd: usdLimit !== null ? unavailable(usdLimit) : haveJournal ? round(tree.brain.usd, 6) : unavailable(journalMissing),
-			cacheRead: !haveJournal ? unavailable(journalMissing) : tree.brain.hasCache ? tree.brain.cacheRead : unavailable(NO_CACHE_COUNTERS),
-			cacheWrite: !haveJournal ? unavailable(journalMissing) : tree.brain.hasCache ? tree.brain.cacheWrite : unavailable(NO_CACHE_COUNTERS),
-			source: haveJournal ? `journal metered events (n=${tree.brain.meteredCount})` : journalMissing
+			cacheRead: managerTokenLimit !== null ? unavailable(managerTokenLimit) : !haveJournal ? unavailable(journalMissing) : tree.brain.hasCache ? tree.brain.cacheRead : unavailable(NO_CACHE_COUNTERS),
+			cacheWrite: managerTokenLimit !== null ? unavailable(managerTokenLimit) : !haveJournal ? unavailable(journalMissing) : tree.brain.hasCache ? tree.brain.cacheWrite : unavailable(NO_CACHE_COUNTERS),
+			source: managerTokenLimit ?? (haveJournal ? `journal metered events (n=${tree.brain.meteredCount})` : journalMissing)
 		},
 		brainTruncations: src.brainLog === null ? gap("brain.brainTruncations", src.supRunDir === null ? "no supervisor run dir under <ws>/.loops/supervisor" : "brain.jsonl absent — loops predates the brain-call tap, so truncation cannot be ruled out") : brainCalls.filter((c) => c.finish_reason === "length").length,
 		workers: {
 			tokensIn: workerIn,
 			tokensOut: workerOut,
-			cacheRead: sq?.cacheRead !== void 0 ? sq.cacheRead : unavailable(NO_CACHE_COUNTERS),
-			cacheWrite: sq?.cacheWrite !== void 0 ? sq.cacheWrite : unavailable(NO_CACHE_COUNTERS),
+			cacheRead: workerTokenLimit !== null ? unavailable(workerTokenLimit) : sq?.cacheRead !== void 0 ? sq.cacheRead : unavailable(NO_CACHE_COUNTERS),
+			cacheWrite: workerTokenLimit !== null ? unavailable(workerTokenLimit) : sq?.cacheWrite !== void 0 ? sq.cacheWrite : unavailable(NO_CACHE_COUNTERS),
 			usd: usdLimit !== null ? unavailable(usdLimit) : haveJournal ? round(journalWorkerUsd, 6) : unavailable(journalMissing),
-			source: sq !== null ? `journal settled spend + ${sq.store} sessions (n=${sq.sessions})` : `journal settled spend only — ${src.harnessMissingReason ?? "harness session store unavailable"}`
+			source: workerTokenLimit !== null ? workerTokenLimit : sq !== null ? `journal settled spend + ${sq.store} sessions (n=${sq.sessions})` : `journal settled spend only — ${src.harnessMissingReason ?? "harness session store unavailable"}`
 		},
 		totalUsd,
 		totalUsdSource: usdLimit !== null ? usdLimit : stateUsd !== null ? `state.json result.spentUsd${journalWorkerUsd === 0 ? " — brain-priced only; worker CLI inference is unpriced (see worker token counts)" : ""}` : haveJournal ? "journal metered + settled usd" : journalMissing,
@@ -459,6 +552,17 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 		traceCommand: src.traceCommand ?? "npx --yes @tangle-network/traces@latest analyze --harness opencode --cwd <worker-clone-cwd>"
 	};
 }
+/** Whether sorted values contain one value in the inclusive interval. */
+function hasNumberBetween(sorted, low, high) {
+	let left = 0;
+	let right = sorted.length;
+	while (left < right) {
+		const middle = left + Math.floor((right - left) / 2);
+		if (sorted[middle] < low) left = middle + 1;
+		else right = middle;
+	}
+	return left < sorted.length && sorted[left] <= high;
+}
 /** `[driver] registered tools: …supervisor_steer…` is a banner, not an invocation. */
 function registrationMentions(driverLog) {
 	let n = 0;
@@ -604,7 +708,7 @@ function rollupSupervisorRuns(reports) {
 * | observeThenRespawn / respawnWithoutEvidence | full | full | ordering of spawn vs settle instants |
 * | workerEvidenceBytes | full | PARTIAL | the child's closing message; 0 for pruned transcripts |
 * | brain tokens in/out + cache | full | full | main-thread `message.usage` |
-* | worker tokens in/out + cache | via harness join | PARTIAL | only for retained subagent transcripts |
+* | worker tokens in/out + cache | via harness join | full or unavailable | totals are refused if any spawned transcript was pruned; retained per-worker rows remain available |
 * | perWorker wall | full | full | spawn → settle instants |
 * | accepted / rejected / emptyPass / settledVerdicts | full | NONE | no per-worker verify step exists |
 * | brain/worker/total usd, costPerAcceptedPatch | full | NONE | transcripts carry no price |
@@ -774,6 +878,8 @@ async function readClaudeCodeSupervisorRun(opts) {
 	const sessionId = basename(opts.transcriptPath).replace(/\.jsonl$/, "");
 	const runRef = opts.runRef ?? opts.transcriptPath;
 	const limits = {
+		managerTokens: null,
+		workerTokens: null,
 		spendUsd: SPEND_UNPRICED,
 		workerVerdicts: NO_VERDICTS,
 		deliverables: null
@@ -799,6 +905,8 @@ async function readClaudeCodeSupervisorRun(opts) {
 		harnessMissingReason: `session transcript unreadable at ${opts.transcriptPath}`,
 		limits: {
 			...limits,
+			managerTokens: `session transcript unreadable at ${opts.transcriptPath}`,
+			workerTokens: `session transcript unreadable at ${opts.transcriptPath}`,
 			deliverables: NO_DELIVERABLES
 		},
 		traceCommand
@@ -871,13 +979,16 @@ async function readClaudeCodeSupervisorRun(opts) {
 		id: sessionId,
 		parent: null,
 		label: `session:${sessionId}`,
+		role: "supervisor",
 		at: startedAt
 	})];
+	const supervisorIds = new Set(spawns.map((spawn) => spawn.parentId));
 	for (const s of spawns) journalLines.push(line({
 		kind: "spawned",
 		id: s.agentId,
 		parent: s.parentId,
 		label: s.label,
+		role: supervisorIds.has(s.agentId) ? "supervisor" : "worker",
 		at: s.at
 	}));
 	const lastNotification = /* @__PURE__ */ new Map();
@@ -919,73 +1030,55 @@ async function readClaudeCodeSupervisorRun(opts) {
 	}));
 	const settleAtByAgent = new Map([...lastNotification.entries()].map(([id, n]) => [id, n.at]));
 	const spawnAtByAgent = new Map(spawns.map((s) => [s.agentId, s.at]));
-	const byLabel = /* @__PURE__ */ new Map();
-	for (const s of spawns) {
-		const ids = byLabel.get(s.label) ?? [];
-		ids.push(s.agentId);
-		byLabel.set(s.label, ids);
-	}
 	const workers = [];
-	for (const [label, agentIds] of byLabel) {
+	for (const spawn of spawns) {
+		const { agentId, label } = spawn;
 		const events = [];
 		const inbox = [];
-		let tokensIn = null;
-		let tokensOut = null;
-		let cacheRead = null;
-		let cacheWrite = null;
-		let transcriptRef = null;
-		for (const agentId of agentIds) {
-			const child = childByAgentId.get(agentId) ?? null;
-			const startAt = child?.firstAt ?? spawnAtByAgent.get(agentId) ?? null;
-			const endAt = settleAtByAgent.get(agentId) ?? child?.lastAt ?? null;
-			if (startAt !== null) events.push(line({
-				kind: "started",
-				label,
-				at: startAt,
-				agentId
+		const child = childByAgentId.get(agentId) ?? null;
+		const startAt = child?.firstAt ?? spawnAtByAgent.get(agentId) ?? null;
+		const endAt = settleAtByAgent.get(agentId) ?? child?.lastAt ?? null;
+		if (startAt !== null) events.push(line({
+			kind: "started",
+			label,
+			at: startAt,
+			agentId
+		}));
+		for (const steer of steersByTarget.get(agentId) ?? []) {
+			inbox.push(line({
+				id: `${agentId}:${steer.at}`,
+				at: steer.at,
+				worker: label,
+				message: "steer"
 			}));
-			for (const steer of steersByTarget.get(agentId) ?? []) {
-				inbox.push(line({
-					id: `${agentId}:${steer.at}`,
-					at: steer.at,
-					worker: label,
-					message: "steer"
-				}));
-				events.push(line({
-					kind: "message",
-					label,
-					direction: "down",
-					at: steer.at,
-					requestId: `${agentId}:${steer.at}`,
-					delivered: steer.delivered
-				}));
-			}
-			if (endAt !== null) events.push(line({
-				kind: "finished",
+			events.push(line({
+				kind: "message",
 				label,
-				at: endAt,
-				agentId,
-				...child?.finalReport === null || child?.finalReport === void 0 ? {} : { evidence: child.finalReport }
+				direction: "down",
+				at: steer.at,
+				requestId: `${agentId}:${steer.at}`,
+				delivered: steer.delivered
 			}));
-			if (child !== null) {
-				tokensIn = (tokensIn ?? 0) + child.tokensIn;
-				tokensOut = (tokensOut ?? 0) + child.tokensOut;
-				cacheRead = (cacheRead ?? 0) + child.cacheRead;
-				cacheWrite = (cacheWrite ?? 0) + child.cacheWrite;
-				transcriptRef = child.path;
-			}
 		}
+		if (endAt !== null) events.push(line({
+			kind: "finished",
+			label,
+			at: endAt,
+			agentId,
+			...child?.finalReport === null || child?.finalReport === void 0 ? {} : { evidence: child.finalReport }
+		}));
 		workers.push({
+			workerId: agentId,
 			label,
 			events: events.length === 0 ? null : `${events.join("\n")}\n`,
 			inbox: inbox.length === 0 ? null : `${inbox.join("\n")}\n`,
 			patchBytes: null,
-			transcriptRef,
+			transcriptRef: child?.path ?? null,
 			patchPath: null,
-			tokensIn,
-			tokensOut,
-			cacheRead,
-			cacheWrite
+			tokensIn: child?.tokensIn ?? null,
+			tokensOut: child?.tokensOut ?? null,
+			cacheRead: child?.cacheRead ?? null,
+			cacheWrite: child?.cacheWrite ?? null
 		});
 	}
 	const joined = children.length;
@@ -1027,6 +1120,7 @@ async function readClaudeCodeSupervisorRun(opts) {
 		harnessMissingReason,
 		limits: {
 			...limits,
+			workerTokens: spawns.length === 0 ? null : harnessMissingReason,
 			deliverables: NO_DELIVERABLES
 		},
 		rootTranscriptRef: opts.transcriptPath,
@@ -1100,7 +1194,7 @@ function renderSupervisorRunMarkdown(r) {
 	out.push(`| Spawn waves | ${showMeasured(o.waves)} |`);
 	out.push(`| Wave sizes | ${isUnavailable(o.waveSizes) ? showMeasured(o.waveSizes) : `[${o.waveSizes.join(", ")}]`} |`);
 	out.push(`| Max concurrency | ${showMeasured(o.maxConcurrency)} |`);
-	out.push(`| Respawns (spawns after first settle) | ${showMeasured(o.respawns)} |`);
+	out.push(`| Respawns (after same parent's first settle) | ${showMeasured(o.respawns)} |`);
 	out.push(`| Repeated labels | ${isUnavailable(o.repeatedLabels) ? showMeasured(o.repeatedLabels) : o.repeatedLabels.length === 0 ? "none" : o.repeatedLabels.join(", ")} |`);
 	out.push(`| Delegation depth | ${showMeasured(o.delegationDepth)} |`);
 	out.push(`| Time to first spawn | ${fmtMs(o.timeToFirstSpawnMs)} |`);
@@ -1111,9 +1205,9 @@ function renderSupervisorRunMarkdown(r) {
 	if (!isUnavailable(o.steersByWorker) && o.steersByWorker.length > 0) {
 		out.push("### Steers per worker");
 		out.push("");
-		out.push("| Worker | Queued | Delivered |");
-		out.push("|---|---:|---:|");
-		for (const s of o.steersByWorker) out.push(`| \`${s.worker}\` | ${s.queued} | ${s.delivered} |`);
+		out.push("| Worker id | Label | Queued | Delivered |");
+		out.push("|---|---|---:|---:|");
+		for (const s of o.steersByWorker) out.push(`| ${s.workerId === null ? "unavailable — legacy label join" : `\`${s.workerId}\``} | \`${s.worker}\` | ${s.queued} | ${s.delivered} |`);
 		out.push("");
 	} else if (isUnavailable(o.steersByWorker)) out.push(`### Steers per worker\n\nunavailable — ${o.steersByWorker.unavailable}\n`);
 	out.push("## Decision quality");
@@ -1126,7 +1220,7 @@ function renderSupervisorRunMarkdown(r) {
 	out.push(`| Rejected (verify red) | ${showMeasured(d.rejected)} |`);
 	out.push(`| Empty pass (green, no patch) | ${showMeasured(d.emptyPass)} |`);
 	out.push(`| Evidence → respawn sequences | ${showMeasured(d.observeThenRespawn)} |`);
-	out.push(`| Respawn with no settled evidence in front | ${showMeasured(d.respawnWithoutEvidence)} |`);
+	out.push(`| Respawn with no same-parent settled evidence in front | ${showMeasured(d.respawnWithoutEvidence)} |`);
 	out.push(`| Review actions (steers + worker questions) | ${showMeasured(d.reviewActions)} |`);
 	out.push(`| Worker evidence returned | ${isUnavailable(d.workerEvidenceBytes) ? showMeasured(d.workerEvidenceBytes) : `${d.workerEvidenceBytes} bytes`} |`);
 	out.push("");
@@ -1148,9 +1242,9 @@ function renderSupervisorRunMarkdown(r) {
 	}
 	out.push("");
 	if (!isUnavailable(e.perWorker) && e.perWorker.length > 0) {
-		out.push("| Worker | Wall | Tokens in | Tokens out | Patch bytes | Verify passed |");
-		out.push("|---|---:|---:|---:|---:|---|");
-		for (const w of e.perWorker) out.push(`| \`${w.worker}\` | ${w.wallMs === null ? "unavailable — no start/finish pair" : fmtMs(w.wallMs)} | ${w.tokensIn ?? "unavailable — store does not attribute tokens per worker"} | ${w.tokensOut ?? "unavailable — store does not attribute tokens per worker"} | ${w.patchBytes ?? "unavailable — no worker patch file"} | ${w.passed === null ? "unavailable — no finished event" : String(w.passed)} |`);
+		out.push("| Worker id | Label | Role | Wall | Tokens in | Tokens out | Patch bytes | Verify passed | Score |");
+		out.push("|---|---|---|---:|---:|---:|---:|---|---:|");
+		for (const w of e.perWorker) out.push(`| ${w.workerId === null ? "unavailable — legacy label join" : `\`${w.workerId}\``} | \`${w.worker}\` | ${w.role ?? "unavailable — no journal join"} | ${w.wallMs === null ? "unavailable — no start/finish pair" : fmtMs(w.wallMs)} | ${w.tokensIn ?? "unavailable — store does not attribute tokens per worker"} | ${w.tokensOut ?? "unavailable — store does not attribute tokens per worker"} | ${w.patchBytes ?? "unavailable — no worker patch file"} | ${w.passed === null ? "unavailable — no verdict" : String(w.passed)} | ${w.score ?? "unavailable — no numeric score"} |`);
 		out.push("");
 	}
 	out.push("## Outcome");
@@ -1237,9 +1331,12 @@ async function readLoopsSupervisorRun(runDir, opts = {}) {
 	const supRunDir = await findSupervisorRunDirIn(ws);
 	const result = await readMaybe(join(runDir, "result.json"));
 	const resultObj = parseJson(result);
+	const journal = supRunDir === null ? null : await readMaybe(join(supRunDir, "journal.jsonl"));
+	const journalWorkerSpawns = parseJsonl(journal).filter((event) => event.kind === "spawned" && typeof event.parent === "string" && event.role !== "supervisor").length;
 	let workers = null;
 	let workersMissingReason = null;
 	const workerCwds = [];
+	let workerStarts = 0;
 	if (supRunDir === null) workersMissingReason = `no supervisor run dir under ${join(ws, ".loops", "supervisor")}`;
 	else {
 		const workersDir = join(supRunDir, "workers");
@@ -1252,18 +1349,25 @@ async function readLoopsSupervisorRun(runDir, opts = {}) {
 				const events = await readMaybe(join(workersDir, `${label}.ndjson`));
 				const inbox = await readMaybe(join(workersDir, `${label}.inbox.ndjson`));
 				const patch = await readMaybe(join(workersDir, `${label}.patch`));
+				const startedRows = parseJsonl(events).filter((event) => event.kind === "started");
+				workerStarts += startedRows.length;
+				const startedIds = startedRows.map((event) => typeof event.workerId === "string" ? event.workerId : typeof event.agentId === "string" ? event.agentId : null).filter((id) => id !== null);
+				const distinctStartedIds = new Set(startedIds);
+				const workerId = startedRows.length > 0 && startedIds.length === startedRows.length && distinctStartedIds.size === 1 ? startedIds[0] : void 0;
 				workers.push({
+					...workerId === void 0 ? {} : { workerId },
 					label,
 					events,
 					inbox,
 					patchBytes: patch === null ? null : Buffer.byteLength(patch)
 				});
-				for (const ev of parseJsonl(events)) if (ev.kind === "started" && typeof ev.cwd === "string") workerCwds.push(ev.cwd);
+				for (const ev of startedRows) if (ev.kind === "started" && typeof ev.cwd === "string") workerCwds.push(ev.cwd);
 			}
 		}
 	}
 	let harnessWorkerTokens = null;
 	let harnessMissingReason = null;
+	let workerCwdsWithoutSessions = 0;
 	if (opts.opencodeDb === null) harnessMissingReason = "opencode join disabled";
 	else if (workerCwds.length === 0) harnessMissingReason = "no worker clone cwds in workers/*.ndjson (nothing to join)";
 	else {
@@ -1274,12 +1378,17 @@ async function readLoopsSupervisorRun(runDir, opts = {}) {
 			let sessions = 0;
 			let input = 0;
 			let output = 0;
-			for (const cwd of new Set(workerCwds)) for (const row of findOpencodeSessionsByDirectory(db, cwd)) {
-				if (seen.has(row.id)) continue;
-				seen.add(row.id);
-				sessions += 1;
-				input += row.tokensInput;
-				output += row.tokensOutput + row.tokensReasoning;
+			const distinctWorkerCwds = new Set(workerCwds);
+			for (const cwd of distinctWorkerCwds) {
+				const rows = findOpencodeSessionsByDirectory(db, cwd);
+				if (rows.length === 0) workerCwdsWithoutSessions += 1;
+				for (const row of rows) {
+					if (seen.has(row.id)) continue;
+					seen.add(row.id);
+					sessions += 1;
+					input += row.tokensInput;
+					output += row.tokensOutput + row.tokensReasoning;
+				}
 			}
 			harnessWorkerTokens = {
 				store: "opencode",
@@ -1287,10 +1396,17 @@ async function readLoopsSupervisorRun(runDir, opts = {}) {
 				input,
 				output
 			};
+			if (workerCwdsWithoutSessions > 0) harnessMissingReason = `${workerCwdsWithoutSessions}/${distinctWorkerCwds.size} worker clone cwds have no opencode session`;
 		} finally {
 			db.close();
 		}
 	}
+	const workerInvocations = Math.max(journalWorkerSpawns, workerStarts);
+	const workerTokenGaps = [];
+	if (workerCwds.length < workerInvocations) workerTokenGaps.push(`${workerInvocations - workerCwds.length}/${workerInvocations} worker invocations have no clone cwd for the opencode token join`);
+	if (workerInvocations > 0 && harnessWorkerTokens === null) workerTokenGaps.push(harnessMissingReason ?? "worker harness token join unavailable");
+	else if (workerCwdsWithoutSessions > 0 && harnessMissingReason !== null) workerTokenGaps.push(harnessMissingReason);
+	const workerTokenLimit = workerTokenGaps.length === 0 ? null : workerTokenGaps.join("; ");
 	const patchPath = opts.patchPath ?? (typeof resultObj?.patchPath === "string" ? resultObj.patchPath : null);
 	let judge = await readMaybe(join(runDir, "judge.json"));
 	let judgeSource = judge === null ? null : join(runDir, "judge.json");
@@ -1306,7 +1422,7 @@ async function readLoopsSupervisorRun(runDir, opts = {}) {
 		instanceId: typeof resultObj?.iid === "string" ? resultObj.iid : instanceIdFromPath(runDir),
 		arm: typeof resultObj?.arm === "string" ? resultObj.arm : basename(runDir),
 		supRunDir,
-		journal: supRunDir === null ? null : await readMaybe(join(supRunDir, "journal.jsonl")),
+		journal,
 		brainLog: supRunDir === null ? null : await readMaybe(join(supRunDir, "brain.jsonl")),
 		state: supRunDir === null ? null : await readMaybe(join(supRunDir, "state.json")),
 		progress: supRunDir === null ? null : await readMaybe(join(supRunDir, "progress.ndjson")),
@@ -1319,7 +1435,10 @@ async function readLoopsSupervisorRun(runDir, opts = {}) {
 		driverLog: await readMaybe(join(runDir, "driver.log")),
 		harnessWorkerTokens,
 		harnessMissingReason,
-		limits: NO_SOURCE_LIMITS,
+		limits: {
+			...NO_SOURCE_LIMITS,
+			workerTokens: workerTokenLimit
+		},
 		traceCommand: null
 	};
 }
@@ -1465,8 +1584,8 @@ async function findSupervisorRunDirs(root) {
 * The supervision tree as `tangle.rollout.v1` rows.
 *
 * A supervisor run IS a tree of rollouts, so its nodes are not a new shape:
-* the root becomes one `RolloutLine` with `role: 'supervisor'`, every spawned
-* worker becomes a `RolloutLine` with `role: 'worker'` and
+* the root becomes one `RolloutLine` with `role: 'supervisor'`, and every
+* spawned invocation keeps its explicit supervisor/worker role with
 * `parent_rollout_id` pointing at its spawner. The rows append to the same
 * ledger as solo-agent rollouts and join to them with the same keys.
 *
@@ -1493,6 +1612,9 @@ const EMPTY_COST = {
 	cache_write: null,
 	wall_s: null
 };
+function hasWorkerId(worker) {
+	return worker.workerId !== void 0;
+}
 /**
 * Mint the supervision tree as rollout rows. Returns the rows plus the gaps
 * that made any of them incomplete — same unavailable-vs-zero discipline as
@@ -1578,10 +1700,10 @@ function supervisorRunRolloutLines(src, opts = {}) {
 			cost: {
 				...EMPTY_COST,
 				usd: src.limits.spendUsd !== null ? null : typeof stateResult.spentUsd === "number" ? stateResult.spentUsd : tree.brain.usd,
-				tokens_in: tree.brain.tokensIn,
-				tokens_out: tree.brain.tokensOut,
-				cache_read: tree.brain.hasCache ? tree.brain.cacheRead : null,
-				cache_write: tree.brain.hasCache ? tree.brain.cacheWrite : null,
+				tokens_in: src.limits.managerTokens === null ? tree.brain.tokensIn : null,
+				tokens_out: src.limits.managerTokens === null ? tree.brain.tokensOut : null,
+				cache_read: src.limits.managerTokens === null && tree.brain.hasCache ? tree.brain.cacheRead : null,
+				cache_write: src.limits.managerTokens === null && tree.brain.hasCache ? tree.brain.cacheWrite : null,
 				wall_s: wallMs === null ? null : wallMs / 1e3
 			},
 			artifacts: {
@@ -1597,34 +1719,40 @@ function supervisorRunRolloutLines(src, opts = {}) {
 		});
 	} else gaps.push("tree.root: no parentless `spawned` event in the journal");
 	const closeById = new Map(tree.closes.map((c) => [c.id, c]));
-	const sourceByLabel = new Map((src.workers ?? []).map((w) => [w.label, w]));
+	const sourceById = new Map((src.workers ?? []).filter(hasWorkerId).map((worker) => [worker.workerId, worker]));
+	const fallbackSourceByLabel = new Map((src.workers ?? []).filter((worker) => worker.workerId === void 0).map((worker) => [worker.label, worker]));
 	for (const spawn of tree.workerSpawns) {
 		const close = closeById.get(spawn.id) ?? null;
-		const facts = tree.workerLogs.get(spawn.label) ?? null;
-		const workerSource = sourceByLabel.get(spawn.label) ?? null;
+		const workerSource = sourceById.get(spawn.id) ?? fallbackSourceByLabel.get(spawn.label) ?? null;
+		const facts = tree.workerLogs.get(workerSource?.workerId ?? workerSource?.label ?? spawn.label) ?? null;
 		const wallMs = facts?.started != null && facts.finishedAt != null ? facts.finishedAt - facts.started : null;
-		const reward = facts?.passed === null || facts === null ? null : facts.passed ? 1 : 0;
-		if (reward === null) gaps.push(`worker ${spawn.label}: no verify verdict (worker logs absent or unfinished)`);
+		const score = close?.score ?? facts?.score ?? null;
+		const passed = close?.valid ?? facts?.passed ?? null;
+		const reward = score ?? (passed === null ? null : passed ? 1 : 0);
+		if (reward === null) gaps.push(`child ${spawn.label}: no verify verdict (child logs absent or unfinished)`);
+		const isSupervisor = spawn.role === "supervisor";
 		nodes.push({
 			...base,
 			rollout_id: spawn.id,
 			parent_rollout_id: spawn.parent,
-			role: "worker",
+			role: spawn.role,
 			policy: {
-				harness: opts.workerHarness ?? null,
+				harness: isSupervisor ? opts.supervisorHarness ?? null : opts.workerHarness ?? null,
 				harness_version: null,
-				model: opts.workerModel ?? null,
+				model: isSupervisor ? opts.supervisorModel ?? null : opts.workerModel ?? null,
 				provider: null,
 				profile_commit: null,
 				sampling: null
 			},
 			outcome: {
 				...unscreenedRewardFields(reward),
-				reward_source: reward === null ? null : "worker-self-verify",
+				reward_source: reward === null ? null : score === null ? "worker-self-verify" : "worker-verdict-score",
 				verdict: close === null ? null : {
 					kind: close.kind,
 					status: close.status,
-					verdict: close.verdict
+					verdict: close.rawVerdict,
+					valid: close.valid,
+					score: close.score
 				},
 				metrics: {
 					label: spawn.label,
@@ -1633,7 +1761,7 @@ function supervisorRunRolloutLines(src, opts = {}) {
 					started_at: facts?.started ?? null,
 					finished_at: facts?.finishedAt ?? null,
 					wall_ms: wallMs,
-					patch_bytes: facts?.finishedPatchBytes ?? null,
+					patch_bytes: workerSource?.patchBytes ?? facts?.finishedPatchBytes ?? null,
 					evidence_bytes: facts?.evidenceBytes ?? null,
 					steers_queued: facts?.steersQueued ?? null,
 					steers_delivered: facts?.steersDelivered ?? null,
@@ -1676,4 +1804,4 @@ function supervisorRunRolloutLines(src, opts = {}) {
 //#endregion
 export { NO_SOURCE_LIMITS as C, showMeasured as D, isUnavailable as E, unavailable as O, rollupSupervisorRuns as S, SUPERVISOR_RUN_SCHEMA as T, claudeCodeSupervisorRunReader as _, loopsSupervisorRunReader as a, parsePatch as b, supervisorReportStem as c, renderSupervisorRollupMarkdown as d, renderSupervisorRunHeadline as f, DEFAULT_STEER_TOOLS as g, DEFAULT_SPAWN_TOOLS as h, findSupervisorRunDirs as i, writeSupervisorRunReport as l, DEFAULT_CANCEL_TOOLS as m, analyzeSupervisorRun as n, readLoopsSupervisorRun as o, renderSupervisorRunMarkdown as p, findSupervisorRunDirIn as r, reportSupervisorRound as s, supervisorRunRolloutLines as t, writeSupervisorRunReportSafe as u, readClaudeCodeSupervisorRun as v, SUPERVISOR_RUN_ROLLUP_SCHEMA as w, parseSupervisorTree as x, analyzeSupervisorRunSources as y };
 
-//# sourceMappingURL=supervisor-run-B7lUGoyZ.js.map
+//# sourceMappingURL=supervisor-run-Dr5HnTup.js.map