Merge branch 'main' into issue_5602

commit 6d0b607f41b80583b2f4dc015c728d57fd0a3f6a
Author: nityanandagohain <nityanandagohain@gmail.com>
Date:   Wed Jul 15 10:30:43 2026 +0530

    fix: update tests

commit 20f42474b0901e49cc745155a2c70bb829988a94
Merge: 36dca5dab0 31efe177a4
Author: nityanandagohain <nityanandagohain@gmail.com>
Date:   Tue Jul 14 23:09:00 2026 +0530

    Merge branch 'issue_5601': scoped-trace builder package + port aggregations

    Merges the telemetryscopedtraces refactor (generic scoped-trace topology moved
    out of telemetryai, feature-flagged gen_ai key enrichment, tracefield. context
    classification) and ports the scalar/time-series work onto it:

    - trace_aggregation.go / trace_having.go now live in telemetryscopedtraces
    - trace-level HAVING resolution composes the canonical pkg/variables
      replacement (dynamic __all__ drops a condition for any operator) with the
      standard filter pipeline, so values still bind as args
    - buildPerTraceScan adopts the merged embedExpr placeholder/arg checking
    - ColumnProvider gains ActivityGateAlias; gen_ai provider gates trace-level
      aggregations on llm_call_count

commit 36dca5dab0c0abb08aeb2645fee07f28cc764eb8
Author: nityanandagohain <nityanandagohain@gmail.com>
Date:   Tue Jul 14 23:09:00 2026 +0530

    feat: trace-level aggregations for scalar/time-series and trace-scoped span list

    - trace.-prefixed aggregations (avg(trace.output_tokens), count(trace.trace_id),
      arithmetic) run over window-clipped per-trace values via a native CTE pipeline
    - a trace-level filter condition qualifies traces in scalar/time-series/raw via
      __qualified / __trace_scope, resource-fingerprint pruned like the trace list
    - trace-level filter conditions resolve through the standard filter pipeline
      (bound args, operators, query variables) against the per-trace column aliases
    - per-trace rows with no LLM span in the window are dropped from trace-level
      aggregations (LLM-activity gate)
    - targeted rejections for group-by/order-by on trace-level columns

frontend/src/assets/Logos/auth0.svg

@@ -1 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path fill="#EB5424" d="M21.98 7.448 19.62 0H4.347L2.02 7.448c-1.352 4.312.03 9.206 3.815 12.015L12.007 24l6.157-4.552c3.755-2.81 5.182-7.688 3.815-12.015l-6.16 4.58 2.343 7.45-6.157-4.597-6.158 4.58 2.358-7.433-6.188-4.55 7.63-.045L12.008 0l2.356 7.404 7.615.044z"/></svg>

frontend/src/assets/Logos/cohere.svg

@@ -1 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 20 20"><path fill="#355146" fill-rule="evenodd" d="M6.82 11.908c.525 0 1.57-.03 3.013-.639 1.682-.71 5.029-2 7.443-3.323 1.689-.926 2.429-2.151 2.429-3.8 0-2.29-1.81-4.146-4.043-4.146H6.307C3.1 0 .5 2.666.5 5.954s2.434 5.954 6.32 5.954" clip-rule="evenodd"/><path fill="#d18ee2" fill-rule="evenodd" d="M8.402 16.01c0-1.611.947-3.064 2.399-3.682l2.946-1.254c2.98-1.268 6.26.977 6.26 4.286 0 2.563-2.027 4.64-4.527 4.64l-3.19-.002c-2.147 0-3.888-1.785-3.888-3.987" clip-rule="evenodd"/><path fill="#ff7759" d="M3.848 12.691C1.998 12.691.5 14.228.5 16.124v.444C.5 18.464 1.999 20 3.848 20s3.347-1.536 3.347-3.432v-.444c0-1.896-1.499-3.433-3.347-3.433"/></svg>

frontend/src/assets/Logos/hcp-vault.svg

@@ -1 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path fill="#FFEC6E" d="m0 0 11.955 24L24 0zm13.366 4.827h1.393v1.38h-1.393zm-2.77 5.569H9.22V8.993h1.389zm0-2.087H9.22V6.906h1.389zm0-2.086H9.22V4.819h1.389zm2.087 6.263h-1.377V11.08h1.388zm0-2.09h-1.377V8.993h1.388zm0-2.087h-1.377V6.906h1.388zm0-2.086h-1.377V4.819h1.388zm.683.683h1.393v1.389h-1.393zm0 3.475V8.993h1.389v1.388Z"/></svg>

frontend/src/assets/Logos/langflow.svg

@@ -1 +0,0 @@
-<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path fill="#9CA3AF" d="M9.755 1.52h-.001c-.31 0-.608.124-.828.343L4.037 6.752a1.17 1.17 0 0 1-.827.343H1.17A1.17 1.17 0 0 0 0 8.295l.052 1.984a1.17 1.17 0 0 0 1.17 1.14h2.37c.31 0 .607-.124.827-.344l4.93-4.93c.22-.22.517-.343.827-.343h2.874a1.17 1.17 0 0 0 1.17-1.17V2.69a1.17 1.17 0 0 0-1.17-1.17zm9.78 2.503c-.31 0-.608.123-.828.343l-4.889 4.889a1.17 1.17 0 0 1-.827.342h-2.756c-.31 0-.608.124-.827.344L4.15 15.197a1.17 1.17 0 0 1-.827.343H1.32a1.17 1.17 0 0 0-1.17 1.17v1.996c0 .646.524 1.17 1.17 1.17h2.017c.302 0 .592-.116.81-.325l5.535-5.304a1.17 1.17 0 0 1 .81-.326h2.88c.31 0 .607-.123.827-.342l4.93-4.93c.22-.22.517-.344.827-.344h2.873A1.17 1.17 0 0 0 24 7.135V5.193a1.17 1.17 0 0 0-1.17-1.17h-3.294zm0 8.559c-.31 0-.608.123-.828.343l-4.889 4.889a1.17 1.17 0 0 1-.827.343h-2.04a1.17 1.17 0 0 0-1.17 1.2l.052 1.984a1.17 1.17 0 0 0 1.17 1.14h2.37c.31 0 .607-.124.827-.343l4.93-4.93c.22-.22.517-.343.827-.343h2.873a1.17 1.17 0 0 0 1.17-1.17v-1.943a1.17 1.17 0 0 0-1.17-1.17h-3.294Z"/></svg>

frontend/src/container/InfraMonitoringK8sV2/Base/K8sEmptyState.module.scss

@@ -51,5 +51,4 @@
 	gap: var(--spacing-3);
 	max-width: 500px;
 	padding: 24px;
-	max-height: 100%;
 }

frontend/src/container/OnboardingV2Container/onboarding-configs/onboarding-config-with-links.ts

@@ -8,7 +8,6 @@ import apacheDruidUrl from '@/assets/Logos/apache-druid.svg';
 import apiGatewayUrl from '@/assets/Logos/api-gateway.svg';
 import argocdUrl from '@/assets/Logos/argocd.svg';
 import aspnetUrl from '@/assets/Logos/aspnet.svg';
-import auth0Url from '@/assets/Logos/auth0.svg';
 import autogenUrl from '@/assets/Logos/autogen.svg';
 import awsAlbUrl from '@/assets/Logos/aws-alb.svg';
 import azureAppServiceUrl from '@/assets/Logos/azure-app-service.svg';
@@ -28,7 +27,6 @@ import claudeCodeUrl from '@/assets/Logos/claude-code.svg';
 import clickhouseUrl from '@/assets/Logos/clickhouse.svg';
 import cloudflareUrl from '@/assets/Logos/cloudflare.svg';
 import cloudwatchLogsUrl from '@/assets/Logos/cloudwatch-logs.svg';
-import cohereUrl from '@/assets/Logos/cohere.svg';
 import confluentKafkaUrl from '@/assets/Logos/confluent-kafka.svg';
 import convexLogoUrl from '@/assets/Logos/convex-logo.svg';
 import cppUrl from '@/assets/Logos/cpp.svg';
@@ -77,7 +75,6 @@ import grokUrl from '@/assets/Logos/grok.svg';
 import groqUrl from '@/assets/Logos/groq.svg';
 import hasuraUrl from '@/assets/Logos/hasura.svg';
 import haystackUrl from '@/assets/Logos/haystack.svg';
-import hcpVaultUrl from '@/assets/Logos/hcp-vault.svg';
 import herokuUrl from '@/assets/Logos/heroku.svg';
 import honeycombUrl from '@/assets/Logos/honeycomb.svg';
 import hostmetricsUrl from '@/assets/Logos/hostmetrics.svg';
@@ -95,7 +92,6 @@ import kafkaUrl from '@/assets/Logos/kafka.svg';
 import kubernetesUrl from '@/assets/Logos/kubernetes.svg';
 import lambdaUrl from '@/assets/Logos/lambda.svg';
 import langchainUrl from '@/assets/Logos/langchain.svg';
-import langflowUrl from '@/assets/Logos/langflow.svg';
 import langtraceUrl from '@/assets/Logos/langtrace.svg';
 import litellmUrl from '@/assets/Logos/litellm.svg';
 import livekitUrl from '@/assets/Logos/livekit.svg';
@@ -3941,58 +3937,6 @@ const onboardingConfigWithLinks = [
 		],
 		link: '/docs/claude-code-monitoring/',
 	},
-	{
-		dataSource: 'cohere',
-		label: 'Cohere',
-		imgUrl: cohereUrl,
-		tags: ['LLM Monitoring'],
-		module: 'apm',
-		relatedSearchKeywords: [
-			'cohere',
-			'cohere api',
-			'cohere logs',
-			'cohere metrics',
-			'cohere monitoring',
-			'cohere observability',
-			'cohere traces',
-			'llm',
-			'llm monitoring',
-			'logging',
-			'logs',
-			'metrics',
-			'monitoring',
-			'observability',
-			'otel cohere integration',
-			'telemetry',
-		],
-		link: '/docs/cohere-monitoring/',
-	},
-	{
-		dataSource: 'langflow',
-		label: 'Langflow',
-		imgUrl: langflowUrl,
-		tags: ['LLM Monitoring'],
-		module: 'apm',
-		relatedSearchKeywords: [
-			'langflow',
-			'langflow logs',
-			'langflow metrics',
-			'langflow monitoring',
-			'langflow observability',
-			'langflow traces',
-			'llm',
-			'llm monitoring',
-			'logging',
-			'logs',
-			'low code ai',
-			'metrics',
-			'monitoring',
-			'observability',
-			'otel langflow integration',
-			'telemetry',
-		],
-		link: '/docs/langflow-observability/',
-	},
 	{
 		dataSource: 'deepseek-api',
 		label: 'DeepSeek API',
@@ -6523,30 +6467,6 @@ const onboardingConfigWithLinks = [
 		id: 'cert-manager',
 		link: '/docs/infrastructure-monitoring/cert-manager/',
 	},
-	{
-		dataSource: 'pgbouncer',
-		label: 'PgBouncer',
-		imgUrl: postgresqlUrl,
-		tags: ['infrastructure monitoring', 'metrics'],
-		module: 'metrics',
-		relatedSearchKeywords: [
-			'connection pooler',
-			'connection pooling',
-			'database',
-			'metrics',
-			'monitoring',
-			'observability',
-			'opentelemetry pgbouncer',
-			'pgbouncer',
-			'pgbouncer metrics',
-			'pgbouncer monitoring',
-			'pgbouncer observability',
-			'postgres',
-			'postgresql',
-		],
-		id: 'pgbouncer',
-		link: '/docs/metrics-management/opentelemetry-pgbouncer/',
-	},
 	{
 		dataSource: 'graphql',
 		label: 'GraphQL',
@@ -6571,28 +6491,6 @@ const onboardingConfigWithLinks = [
 		id: 'graphql',
 		link: '/docs/instrumentation/javascript/opentelemetry-graphql/',
 	},
-	{
-		dataSource: 'opentelemetry-ebpf',
-		label: 'OpenTelemetry eBPF (OBI)',
-		imgUrl: opentelemetryUrl,
-		tags: ['apm/traces'],
-		module: 'apm',
-		relatedSearchKeywords: [
-			'auto instrumentation',
-			'ebpf',
-			'obi',
-			'opentelemetry ebpf',
-			'opentelemetry obi',
-			'otel ebpf',
-			'zero code instrumentation',
-			'monitoring',
-			'observability',
-			'traces',
-			'tracing',
-		],
-		id: 'opentelemetry-ebpf',
-		link: '/docs/instrumentation/opentelemetry-ebpf/',
-	},
 	{
 		dataSource: 'railway',
 		label: 'Railway',
@@ -6615,54 +6513,6 @@ const onboardingConfigWithLinks = [
 		id: 'railway',
 		link: '/docs/integrations/outposts/railway/',
 	},
-	{
-		dataSource: 'hcp-vault',
-		label: 'HCP Vault',
-		imgUrl: hcpVaultUrl,
-		tags: ['logs'],
-		module: 'logs',
-		relatedSearchKeywords: [
-			'hashicorp',
-			'hashicorp vault',
-			'hcp',
-			'hcp vault',
-			'hcp vault logs',
-			'hcp vault monitoring',
-			'hcp vault observability',
-			'log forwarding',
-			'logging',
-			'logs',
-			'monitoring',
-			'observability',
-			'secrets management',
-			'vault',
-		],
-		id: 'hcp-vault',
-		link: '/docs/integrations/outposts/hcp-vault/',
-	},
-	{
-		dataSource: 'auth0',
-		label: 'Auth0',
-		imgUrl: auth0Url,
-		tags: ['logs'],
-		module: 'logs',
-		relatedSearchKeywords: [
-			'auth0',
-			'auth0 logs',
-			'auth0 monitoring',
-			'auth0 observability',
-			'authentication',
-			'authorization',
-			'identity',
-			'log forwarding',
-			'logging',
-			'logs',
-			'monitoring',
-			'observability',
-		],
-		id: 'auth0',
-		link: '/docs/integrations/outposts/auth0/',
-	},
 	{
 		dataSource: 'aspnet-core-metrics',
 		label: 'ASP.NET Core Metrics',

frontend/src/utils/__tests__/queryContextUtils.test.ts

@@ -648,176 +648,3 @@ describe('getQueryContextAtCursor - trailing dot in key/value', () => {
 		expect(ctx.keyToken).toBe('k8s.namespace');
 	});
 });
-
-describe('getQueryContextAtCursor - partial operator', () => {
-	it('treats text after an incomplete key as an operator prefix', () => {
-		const q = 'service.name c';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.isInKey).toBe(false);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('c');
-		expect(ctx.currentPair).toStrictEqual(
-			expect.objectContaining({
-				key: 'service.name',
-				operator: 'c',
-				position: expect.objectContaining({
-					operatorStart: 13,
-					operatorEnd: 13,
-				}),
-			}),
-		);
-	});
-
-	it('keeps the operator context while completing contains', () => {
-		const q = 'service.name cont';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('cont');
-	});
-
-	it('treats cursor mid-token as operator context', () => {
-		const q = 'service.name cont';
-		// cursor sits between "con" and "t" — user still typing the operator
-		const ctx = getQueryContextAtCursor(q, 15);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.isInKey).toBe(false);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('cont');
-	});
-
-	it('keeps operator context when an AND conjunction precedes the pair', () => {
-		const q = 'a = 1 AND service.name c';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.isInKey).toBe(false);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('c');
-		expect(ctx.currentPair).toStrictEqual(
-			expect.objectContaining({
-				key: 'service.name',
-				operator: 'c',
-				position: expect.objectContaining({
-					operatorStart: 23,
-					operatorEnd: 23,
-				}),
-			}),
-		);
-	});
-
-	it('keeps operator context when an open parenthesis precedes the pair', () => {
-		const q = '(service.name c';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.isInKey).toBe(false);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('c');
-	});
-
-	it('re-glues a partial operator that follows a NOT negation', () => {
-		const q = 'service.name NOT c';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.isInKey).toBe(false);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('NOT c');
-		// operatorStart points at the partial operator (post-NOT), not at the
-		// negation — so suggestion selection only replaces the partial, never
-		// the user's typed NOT.
-		expect(ctx.currentPair).toStrictEqual(
-			expect.objectContaining({
-				key: 'service.name',
-				operator: 'NOT c',
-				hasNegation: true,
-				position: expect.objectContaining({
-					negationStart: 13,
-					negationEnd: 15,
-					operatorStart: 17,
-					operatorEnd: 17,
-				}),
-			}),
-		);
-	});
-
-	it('re-glues a multi-character partial operator after NOT', () => {
-		const q = 'service.name NOT lik';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('NOT lik');
-		expect(ctx.currentPair?.hasNegation).toBe(true);
-	});
-
-	it('re-glues an uppercase partial operator after NOT', () => {
-		const q = 'service.name NOT EXI';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('NOT EXI');
-		expect(ctx.currentPair?.hasNegation).toBe(true);
-	});
-
-	it('preserves original NOT casing in the operator text', () => {
-		const q = 'service.name not c';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('not c');
-		expect(ctx.currentPair?.hasNegation).toBe(true);
-	});
-
-	it('tolerates extra whitespace between NOT and the partial operator', () => {
-		const q = 'service.name NOT  c';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.keyToken).toBe('service.name');
-		// Display text uses a canonical single space between NOT and the
-		// partial, regardless of how many spaces the user typed.
-		expect(ctx.operatorToken).toBe('NOT c');
-		expect(ctx.currentPair?.hasNegation).toBe(true);
-	});
-
-	it('keeps operator context for NOT-prefixed partial inside parentheses', () => {
-		const q = '(service.name NOT c';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('NOT c');
-		expect(ctx.currentPair?.hasNegation).toBe(true);
-	});
-
-	it('keeps operator context for NOT-prefixed partial after an AND conjunction', () => {
-		const q = 'a = 1 AND service.name NOT c';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.keyToken).toBe('service.name');
-		expect(ctx.operatorToken).toBe('NOT c');
-		expect(ctx.currentPair?.hasNegation).toBe(true);
-	});
-
-	it('re-glues the most recent incomplete pair when three partial tokens are typed', () => {
-		// Pins documented behavior: with two trailing partial pairs (`c` and
-		// `k`), the heuristic pairs the most recent two — `c` becomes the
-		// key, `k` becomes the partial operator. The earlier `service.name`
-		// is dropped from the current pair view.
-		const q = 'service.name c k';
-		const ctx = getQueryContextAtCursor(q, q.length);
-
-		expect(ctx.isInOperator).toBe(true);
-		expect(ctx.keyToken).toBe('c');
-		expect(ctx.operatorToken).toBe('k');
-	});
-});

frontend/src/utils/queryContextUtils.ts

@@ -605,98 +605,6 @@ export function getQueryContextAtCursor(
 			queryPairs,
 		);
 
-		// Re-glue a partial operator that ANTLR has lexed as a second key.
-		//
-		// When the user types `service.name c` (or `service.name NOT c`), the
-		// lexer sees two KEY tokens (`service.name`, `c`) instead of a key +
-		// partial operator, so `extractQueryPairs` emits two consecutive
-		// key-only incomplete pairs. Downstream, that makes the dropdown
-		// suggest keys when it should be suggesting operators.
-		//
-		// Detect that pattern — a previous incomplete key-only pair followed
-		// by another incomplete key-only `currentPair`, separated only by
-		// whitespace (or by a negation token attached to the previous pair) —
-		// and rebuild a single synthetic pair where the previous pair's key
-		// is the key and the current pair's key is treated as the partial
-		// operator. The synthetic pair inherits the previous pair's negation
-		// flag and positions via the spread, so `NOT <partial>` propagates
-		// correctly to consumers.
-		const previousIncompletePair = queryPairs
-			.filter(
-				(pair) =>
-					!pair.isComplete &&
-					!!pair.key &&
-					!pair.operator &&
-					pair.position.keyEnd < (currentPair?.position.keyStart ?? cursorIndex),
-			)
-			.sort((a, b) => b.position.keyEnd - a.position.keyEnd)[0];
-
-		if (
-			previousIncompletePair &&
-			currentPair &&
-			currentPair !== previousIncompletePair &&
-			!currentPair.operator &&
-			currentPair.position.keyStart > previousIncompletePair.position.keyEnd
-		) {
-			const negationStart = previousIncompletePair.position.negationStart ?? 0;
-			const negationEnd = previousIncompletePair.position.negationEnd ?? 0;
-			const negationAfterKey =
-				previousIncompletePair.hasNegation &&
-				negationStart > previousIncompletePair.position.keyEnd;
-			const gapStart = negationAfterKey
-				? negationEnd + 1
-				: previousIncompletePair.position.keyEnd + 1;
-			const textBetweenPairs = query.slice(
-				gapStart,
-				currentPair.position.keyStart,
-			);
-
-			if (textBetweenPairs.trim() === '') {
-				// The replacement range (operatorStart/operatorEnd) must point
-				// at the partial operator only, NOT the leading negation.
-				// Consumers like QuerySearch use it to splice the chosen
-				// suggestion in-place, so including the negation would let a
-				// `NOT lik` -> `LIKE` selection erase the user's typed `NOT`.
-				// Matches the convention used for complete pairs in
-				// extractQueryPairs, where operatorStart starts after the
-				// negation token.
-				const operatorStart = currentPair.position.keyStart;
-				const operatorEnd = currentPair.position.keyEnd;
-				const partialOperator = query.slice(operatorStart, operatorEnd + 1);
-				const operatorText = negationAfterKey
-					? `${query.slice(negationStart, negationEnd + 1)} ${partialOperator}`
-					: partialOperator;
-
-				return {
-					tokenType: -1,
-					text: '',
-					start: cursorIndex,
-					stop: cursorIndex,
-					currentToken: operatorText,
-					isInKey: false,
-					isInNegation: false,
-					isInOperator: true,
-					isInValue: false,
-					isInConjunction: false,
-					isInFunction: false,
-					isInParenthesis: false,
-					isInBracketList: false,
-					keyToken: previousIncompletePair.key,
-					operatorToken: operatorText,
-					queryPairs,
-					currentPair: {
-						...previousIncompletePair,
-						operator: operatorText,
-						position: {
-							...previousIncompletePair.position,
-							operatorStart,
-							operatorEnd,
-						},
-					},
-				};
-			}
-		}
-
 		// Check if cursor is within any of the specific context boundaries
 		// FIXED: Include the case where the cursor is exactly at the end of a boundary
 		const isInKeyBoundary =

pkg/statementbuilder/aistatementbuilder/statement_builder_test.go

@@ -993,8 +993,8 @@ func TestBuild_TraceList_MultiVariantGateKey(t *testing.T) {
 	assert.Contains(t, got, "mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_number, 'gen_ai.tool.name')")
 }
 
-// `trace.` marks a trace-level aggregate; `tracefield.` routes trace-level too but is
-// not a rewritable alias, so the HAVING rewriter rejects it.
+// A `trace.`-prefixed aggregate in the filter box and the same condition in the
+// explicit Having box build the same query; output-only aggregates are rejected.
 func TestBuild_TraceList_TraceContextPrefix(t *testing.T) {
 	b := newTestBuilder(t)
 	build := func(q qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]) (*qbtypes.Statement, error) {
@@ -1002,19 +1002,14 @@ func TestBuild_TraceList_TraceContextPrefix(t *testing.T) {
 		return b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeTrace, q, nil)
 	}
 
-	_, err := build(qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+	viaTrace, err := build(qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
 		Filter: &qbtypes.Filter{Expression: "trace.output_tokens > 1000"}})
 	require.NoError(t, err)
 
-	_, err = build(qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
-		Filter: &qbtypes.Filter{Expression: "tracefield.output_tokens > 1000"}})
-	require.Error(t, err)
-	assert.Contains(t, err.Error(), "Invalid references in `Having` expression: [tracefield.output_tokens]")
-
-	_, err = build(qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
-		Having: &qbtypes.Having{Expression: "tracefield.output_tokens > 1000"}})
-	require.Error(t, err)
-	assert.Contains(t, err.Error(), "Invalid references in `Having` expression: [tracefield.output_tokens]")
+	viaHaving, err := build(qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+		Having: &qbtypes.Having{Expression: "trace.output_tokens > 1000"}})
+	require.NoError(t, err)
+	assert.Equal(t, viaTrace.Query, viaHaving.Query)
 
 	_, err = build(qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
 		Filter: &qbtypes.Filter{Expression: "trace.span_count > 3"}})
@@ -1022,7 +1017,8 @@ func TestBuild_TraceList_TraceContextPrefix(t *testing.T) {
 	assert.Contains(t, err.Error(), "cannot be used")
 }
 
-// Query variables in a trace-level condition are substituted into the HAVING.
+// Query variables in a trace-level condition resolve like span filters: bound args,
+// list/IN handling, dynamic __all__ dropping the condition.
 func TestBuild_TraceList_VariableInAggregateFilter(t *testing.T) {
 	b := newTestBuilder(t)
 	build := func(expr string, vars map[string]qbtypes.VariableItem) (*qbtypes.Statement, error) {
@@ -1034,17 +1030,18 @@ func TestBuild_TraceList_VariableInAggregateFilter(t *testing.T) {
 			}, vars)
 	}
 
-	// scalar variable -> literal in HAVING
+	// scalar variable -> bound arg via the filter pipeline
 	stmt, err := build("trace.output_tokens > $threshold",
 		map[string]qbtypes.VariableItem{"threshold": {Value: 700}})
 	require.NoError(t, err)
-	assert.Contains(t, stmt.Query, "HAVING output_tokens > 700")
+	assert.Contains(t, stmt.Query, "HAVING output_tokens > ?")
+	assert.Contains(t, stmt.Args, float64(700))
 
 	// list variable with IN
 	stmt, err = build("trace.llm_call_count IN $counts",
 		map[string]qbtypes.VariableItem{"counts": {Value: []any{1, 2}}})
 	require.NoError(t, err)
-	assert.Contains(t, stmt.Query, "HAVING llm_call_count IN")
+	assert.Contains(t, stmt.Query, "HAVING llm_call_count IN (?, ?)")
 
 	// dynamic __all__ -> condition dropped, no HAVING at all
 	stmt, err = build("trace.output_tokens > $threshold",
@@ -1052,7 +1049,7 @@ func TestBuild_TraceList_VariableInAggregateFilter(t *testing.T) {
 	require.NoError(t, err)
 	assert.NotContains(t, stmt.Query, "HAVING")
 
-	// unresolved variable -> rejected, not compared as a literal
+	// unresolved variable -> rejected, though only as an unknown aggregate today
 	_, err = build("trace.output_tokens > $missing", map[string]qbtypes.VariableItem{"other": {Value: 1}})
 	require.Error(t, err)
 }

pkg/statementbuilder/aistatementbuilder/trace_aggregation_test.go

@@ -0,0 +1,771 @@
+package aistatementbuilder
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	qbtypes "github.com/SigNoz/signoz/pkg/types/querybuildertypes/querybuildertypesv5"
+	"github.com/SigNoz/signoz/pkg/types/telemetrytypes"
+	"github.com/SigNoz/signoz/pkg/valuer"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// Build tests for scalar / time-series through the gen_ai scope; the
+// rewriteTraceAggregation unit tests live in scopedtracesstatementbuilder.
+// The goldens build up one dimension at a time: base → span filter → trace filter
+// → group by → everything combined; then the time-series variants.
+
+// The empty base case never reaches the builder: request validation rejects a
+// scalar / time-series builder_ai_query with no aggregations, so the builder
+// assumes at least one (internal calls are validated upstream).
+func TestBuild_Aggregation_NoAggregations_RejectedByRequestValidation(t *testing.T) {
+	for _, rt := range []qbtypes.RequestType{qbtypes.RequestTypeScalar, qbtypes.RequestTypeTimeSeries} {
+		req := qbtypes.QueryRangeRequest{
+			Start:       testStartMs,
+			End:         testEndMs,
+			RequestType: rt,
+			CompositeQuery: qbtypes.CompositeQuery{
+				Queries: []qbtypes.QueryEnvelope{{
+					Type: qbtypes.QueryTypeBuilderAI,
+					Spec: qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+						Name:         "A",
+						Signal:       telemetrytypes.SignalTraces,
+						StepInterval: qbtypes.Step{Duration: 60 * time.Second},
+					},
+				}},
+			},
+		}
+		require.ErrorContains(t, req.Validate(), "at least one aggregation is required", rt.StringValue())
+	}
+}
+
+// Base scalar over per-trace values: one window-clipped per-trace scan, outer avg
+// across traces. Traces without token spans yield NULL, which avg skips.
+func TestBuild_FullSQL_Scalar_TraceAgg(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __scoped_traces AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id
+)
+SELECT avg(output_tokens) AS __result_0
+FROM __scoped_traces
+ORDER BY __result_0 DESC
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// Span-level filter: resolved through the standard filter pipeline and ANDed into
+// the per-trace scan's WHERE, next to the gate mask.
+func TestBuild_FullSQL_Scalar_SpanFilter(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+			Filter:       &qbtypes.Filter{Expression: "gen_ai.request.model = 'gpt-4o-mini'"},
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __scoped_traces AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND (attributes_string['gen_ai.request.model'] = 'gpt-4o-mini' AND mapContains(attributes_string, 'gen_ai.request.model'))
+    GROUP BY trace_id
+)
+SELECT avg(output_tokens) AS __result_0
+FROM __scoped_traces
+ORDER BY __result_0 DESC
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// Trace-level filter: qualification first — __qualified selects the trace ids whose
+// whole-window value passes, then the per-trace scan is constrained to them.
+func TestBuild_FullSQL_Scalar_TraceFilter(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+			Filter:       &qbtypes.Filter{Expression: "trace.output_tokens > 1000"},
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __qualified AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id
+    HAVING output_tokens > 1000
+),
+__scoped_traces AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND trace_id GLOBAL IN (SELECT trace_id FROM __qualified)
+    GROUP BY trace_id
+)
+SELECT avg(output_tokens) AS __result_0
+FROM __scoped_traces
+ORDER BY __result_0 DESC
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// Group by a span attribute: selected (stringified) and grouped in the per-trace
+// scan, then grouped again in the outer aggregation. A trace spanning two models
+// contributes one per-trace row per model.
+func TestBuild_FullSQL_Scalar_GroupBy(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+			GroupBy:      []qbtypes.GroupByKey{{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "gen_ai.request.model"}}},
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __scoped_traces AS (
+    SELECT trace_id,
+        toString(multiIf(mapContains(attributes_string, 'gen_ai.request.model'), attributes_string['gen_ai.request.model'], NULL)) AS gen_ai.request.model,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id, gen_ai.request.model
+)
+SELECT gen_ai.request.model, avg(output_tokens) AS __result_0
+FROM __scoped_traces
+GROUP BY gen_ai.request.model
+ORDER BY __result_0 DESC
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// Everything combined: span + trace filter parts split (WHERE + __qualified), group
+// by, two aggregations, HAVING on the alias (rewritten to __result_0), explicit
+// order and limit.
+func TestBuild_FullSQL_Scalar_FullCombo(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal: telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{
+				{Expression: "avg(trace.output_tokens)", Alias: "avg_out"},
+				{Expression: "count(trace.trace_id)"},
+			},
+			Filter:  &qbtypes.Filter{Expression: "gen_ai.request.model = 'gpt-4o-mini' AND trace.total_tokens > 100"},
+			GroupBy: []qbtypes.GroupByKey{{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "gen_ai.request.model"}}},
+			Having:  &qbtypes.Having{Expression: "avg_out > 50"},
+			Order:   []qbtypes.OrderBy{{Key: qbtypes.OrderByKey{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "avg_out"}}, Direction: qbtypes.OrderDirectionDesc}},
+			Limit:   5,
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __qualified AS (
+    SELECT trace_id,
+        coalesce(sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.input_tokens'), toFloat64(attributes_number['gen_ai.usage.input_tokens']), NULL)), 0) + coalesce(sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)), 0) AS total_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id
+    HAVING total_tokens > 100
+),
+__scoped_traces AS (
+    SELECT trace_id,
+        toString(multiIf(mapContains(attributes_string, 'gen_ai.request.model'), attributes_string['gen_ai.request.model'], NULL)) AS gen_ai.request.model,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND (attributes_string['gen_ai.request.model'] = 'gpt-4o-mini' AND mapContains(attributes_string, 'gen_ai.request.model'))
+      AND trace_id GLOBAL IN (SELECT trace_id FROM __qualified)
+    GROUP BY trace_id, gen_ai.request.model
+)
+SELECT gen_ai.request.model, avg(output_tokens) AS __result_0, count(trace_id) AS __result_1
+FROM __scoped_traces
+GROUP BY gen_ai.request.model
+HAVING __result_0 > 50
+ORDER BY __result_0 desc
+LIMIT 5
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// Time series: the per-trace scan buckets by span time, the outer aggregation per bucket.
+func TestBuild_FullSQL_TimeSeries_TraceAgg(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeTimeSeries,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			StepInterval: qbtypes.Step{Duration: 60 * time.Second},
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __scoped_traces AS (
+    SELECT trace_id,
+        toStartOfInterval(timestamp, INTERVAL 60 SECOND) AS ts,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id, ts
+)
+SELECT ts, avg(output_tokens) AS __result_0
+FROM __scoped_traces
+GROUP BY ts
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// Grouped, limited time series: groups are ranked on whole-window per-trace values
+// (__scoped_traces_total, no ts bucketing → exact for non-composable aggregates like
+// avg), __limit_cte keeps the top-N by the requested order, and the bucketed main
+// scan is pruned to those groups before aggregating. The alias HAVING applies to the
+// outer aggregation; the aggregation order key ranks __limit_cte, the series
+// themselves order by ts.
+func TestBuild_FullSQL_TimeSeries_GroupLimit(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeTimeSeries,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			StepInterval: qbtypes.Step{Duration: 60 * time.Second},
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "sum(trace.output_tokens)", Alias: "total_out"}},
+			GroupBy:      []qbtypes.GroupByKey{{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "gen_ai.request.model"}}},
+			Having:       &qbtypes.Having{Expression: "total_out > 500"},
+			Order:        []qbtypes.OrderBy{{Key: qbtypes.OrderByKey{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "total_out"}}, Direction: qbtypes.OrderDirectionDesc}},
+			Limit:        3,
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __scoped_traces_total AS (
+    SELECT trace_id,
+        toString(multiIf(mapContains(attributes_string, 'gen_ai.request.model'), attributes_string['gen_ai.request.model'], NULL)) AS gen_ai.request.model,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id, gen_ai.request.model
+),
+__limit_cte AS (
+    SELECT gen_ai.request.model, sum(output_tokens) AS __result_0
+    FROM __scoped_traces_total
+    GROUP BY gen_ai.request.model
+    ORDER BY __result_0 desc
+    LIMIT 3
+),
+__scoped_traces AS (
+    SELECT trace_id,
+        toStartOfInterval(timestamp, INTERVAL 60 SECOND) AS ts,
+        toString(multiIf(mapContains(attributes_string, 'gen_ai.request.model'), attributes_string['gen_ai.request.model'], NULL)) AS gen_ai.request.model,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND (toString(multiIf(mapContains(attributes_string, 'gen_ai.request.model'), attributes_string['gen_ai.request.model'], NULL))) GLOBAL IN (SELECT gen_ai.request.model FROM __limit_cte)
+    GROUP BY trace_id, ts, gen_ai.request.model
+)
+SELECT ts, gen_ai.request.model, sum(output_tokens) AS __result_0
+FROM __scoped_traces
+GROUP BY ts, gen_ai.request.model
+HAVING __result_0 > 500
+ORDER BY ts desc
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// Span-level scalar with a trace-level filter: delegated to the trace builder,
+// constrained by the __trace_scope qualification (the delegate's own scalar shape,
+// hence no SETTINGS suffix).
+func TestBuild_FullSQL_Scalar_SpanAgg_TraceScoped(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "sum(gen_ai.usage.output_tokens)"}},
+			Filter:       &qbtypes.Filter{Expression: "trace.output_tokens > 1000"},
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __trace_scope AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id
+    HAVING output_tokens > 1000
+)
+SELECT sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS __result_0
+FROM signoz_traces.distributed_signoz_index_v3
+WHERE trace_id GLOBAL IN (SELECT trace_id FROM __trace_scope)
+  AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+  AND timestamp >= '1747947419000000000'
+  AND timestamp < '1747983448000000000'
+  AND ts_bucket_start >= 1747945619
+  AND ts_bucket_start <= 1747983448
+ORDER BY __result_0 DESC
+`, stmt)
+}
+
+// Grouped, limited time series with two group keys and a split filter: the top-N
+// prune is a 2-tuple GLOBAL IN, the qualification and span predicate apply to the
+// ranking scan and the main scan alike, and with no explicit order the ranking
+// defaults to __result_0 DESC.
+func TestBuild_FullSQL_TimeSeries_GroupLimit_MultiKey(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeTimeSeries,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			StepInterval: qbtypes.Step{Duration: 60 * time.Second},
+			Aggregations: []qbtypes.TraceAggregation{
+				{Expression: "sum(trace.output_tokens)"},
+				{Expression: "count(trace.trace_id)"},
+			},
+			Filter: &qbtypes.Filter{Expression: "gen_ai.request.model = 'gpt-4o-mini' AND trace.total_tokens > 100"},
+			GroupBy: []qbtypes.GroupByKey{
+				{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "gen_ai.request.model"}},
+				{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "gen_ai.user.id"}},
+			},
+			Limit: 2,
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __qualified AS (
+    SELECT trace_id,
+        coalesce(sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.input_tokens'), toFloat64(attributes_number['gen_ai.usage.input_tokens']), NULL)), 0) + coalesce(sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)), 0) AS total_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id
+    HAVING total_tokens > 100
+),
+__scoped_traces_total AS (
+    SELECT trace_id,
+        toString(multiIf(mapContains(attributes_string, 'gen_ai.request.model'), attributes_string['gen_ai.request.model'], NULL)) AS gen_ai.request.model,
+        toString(multiIf(mapContains(attributes_string, 'gen_ai.user.id'), attributes_string['gen_ai.user.id'], NULL)) AS gen_ai.user.id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND (attributes_string['gen_ai.request.model'] = 'gpt-4o-mini' AND mapContains(attributes_string, 'gen_ai.request.model'))
+      AND trace_id GLOBAL IN (SELECT trace_id FROM __qualified)
+    GROUP BY trace_id, gen_ai.request.model, gen_ai.user.id
+),
+__limit_cte AS (
+    SELECT gen_ai.request.model, gen_ai.user.id, sum(output_tokens) AS __result_0, count(trace_id) AS __result_1
+    FROM __scoped_traces_total
+    GROUP BY gen_ai.request.model, gen_ai.user.id
+    ORDER BY __result_0 DESC
+    LIMIT 2
+),
+__scoped_traces AS (
+    SELECT trace_id,
+        toStartOfInterval(timestamp, INTERVAL 60 SECOND) AS ts,
+        toString(multiIf(mapContains(attributes_string, 'gen_ai.request.model'), attributes_string['gen_ai.request.model'], NULL)) AS gen_ai.request.model,
+        toString(multiIf(mapContains(attributes_string, 'gen_ai.user.id'), attributes_string['gen_ai.user.id'], NULL)) AS gen_ai.user.id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND (attributes_string['gen_ai.request.model'] = 'gpt-4o-mini' AND mapContains(attributes_string, 'gen_ai.request.model'))
+      AND trace_id GLOBAL IN (SELECT trace_id FROM __qualified)
+      AND (toString(multiIf(mapContains(attributes_string, 'gen_ai.request.model'), attributes_string['gen_ai.request.model'], NULL)), toString(multiIf(mapContains(attributes_string, 'gen_ai.user.id'), attributes_string['gen_ai.user.id'], NULL))) GLOBAL IN (SELECT gen_ai.request.model, gen_ai.user.id FROM __limit_cte)
+    GROUP BY trace_id, ts, gen_ai.request.model, gen_ai.user.id
+)
+SELECT ts, gen_ai.request.model, gen_ai.user.id, sum(output_tokens) AS __result_0, count(trace_id) AS __result_1
+FROM __scoped_traces
+GROUP BY ts, gen_ai.request.model, gen_ai.user.id
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// A time-series limit without group-by has nothing to rank: it is ignored, matching
+// the trace builder — the query equals its unlimited form.
+func TestBuild_TimeSeries_LimitWithoutGroupByIgnored(t *testing.T) {
+	b := newTestBuilder(t)
+	build := func(limit int) *qbtypes.Statement {
+		stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeTimeSeries,
+			qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+				Signal:       telemetrytypes.SignalTraces,
+				StepInterval: qbtypes.Step{Duration: 60 * time.Second},
+				Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+				Limit:        limit,
+			}, nil)
+		require.NoError(t, err)
+		return stmt
+	}
+	assert.Equal(t, build(0).Query, build(5).Query)
+}
+
+// ---------------------------------------------------------------------------
+// Behavior / branch tests not covered by the goldens above
+// ---------------------------------------------------------------------------
+
+// Mixing span- and trace-level aggregations across one query is rejected.
+func TestBuild_Aggregation_MixedDomainsRejected(t *testing.T) {
+	b := newTestBuilder(t)
+	_, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal: telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{
+				{Expression: "avg(trace.output_tokens)"},
+				{Expression: "sum(gen_ai.usage.output_tokens)"},
+			},
+		}, nil)
+	require.ErrorContains(t, err, "cannot be mixed")
+}
+
+// Output-only aggregates are rejected in trace-level filters on the aggregation
+// path too (the raw and trace-list paths are covered elsewhere).
+func TestBuild_Aggregation_OutputOnlyFilterRejected(t *testing.T) {
+	b := newTestBuilder(t)
+	_, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "count()"}},
+			Filter:       &qbtypes.Filter{Expression: "trace.span_count > 3"},
+		}, nil)
+	require.ErrorContains(t, err, `aggregate "span_count" cannot be used`)
+}
+
+// Trace-level columns are rejected as group-by keys with a targeted builder error;
+// order keys never reach the builder — request validation only admits group keys and
+// aggregation aliases/expressions — and ordering by the alias stays valid.
+func TestBuild_Aggregation_GroupByOrderValidation(t *testing.T) {
+	b := newTestBuilder(t)
+	ctx := context.Background()
+
+	_, err := b.Build(ctx, valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+			GroupBy:      []qbtypes.GroupByKey{{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "trace.llm_call_count"}}},
+		}, nil)
+	require.ErrorContains(t, err, `grouping by trace-level aggregate "trace.llm_call_count" is not supported`)
+
+	req := qbtypes.QueryRangeRequest{
+		Start:       testStartMs,
+		End:         testEndMs,
+		RequestType: qbtypes.RequestTypeScalar,
+		CompositeQuery: qbtypes.CompositeQuery{
+			Queries: []qbtypes.QueryEnvelope{{
+				Type: qbtypes.QueryTypeBuilderAI,
+				Spec: qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+					Name:         "A",
+					Signal:       telemetrytypes.SignalTraces,
+					Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+					Order:        []qbtypes.OrderBy{{Key: qbtypes.OrderByKey{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "trace.total_tokens"}}, Direction: qbtypes.OrderDirectionDesc}},
+				},
+			}},
+		},
+	}
+	require.ErrorContains(t, req.Validate(), "invalid order by key")
+
+	_, err = b.Build(ctx, valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)", Alias: "avg_out"}},
+			Order:        []qbtypes.OrderBy{{Key: qbtypes.OrderByKey{TelemetryFieldKey: telemetrytypes.TelemetryFieldKey{Name: "avg_out"}}, Direction: qbtypes.OrderDirectionAsc}},
+		}, nil)
+	require.NoError(t, err)
+}
+
+// Variables in trace-level conditions resolve through the standard pipeline as bound
+// args; a dynamic __all__ drops the condition entirely; an unresolved $var is only
+// rejected as an unknown aggregate today (a targeted variable error is a separate
+// concern).
+func TestBuild_FullSQL_Aggregation_VariablesInTraceFilter(t *testing.T) {
+	b := newTestBuilder(t)
+	ctx := context.Background()
+
+	q := qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+		Signal:       telemetrytypes.SignalTraces,
+		Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+		Filter:       &qbtypes.Filter{Expression: "trace.output_tokens > $threshold"},
+	}
+	stmt, err := b.Build(ctx, valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar, q,
+		map[string]qbtypes.VariableItem{"threshold": {Type: qbtypes.TextBoxVariableType, Value: float64(1000)}})
+	require.NoError(t, err)
+	assertSQLEqual(t, `
+WITH __qualified AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id
+    HAVING output_tokens > 1000
+),
+__scoped_traces AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND trace_id GLOBAL IN (SELECT trace_id FROM __qualified)
+    GROUP BY trace_id
+)
+SELECT avg(output_tokens) AS __result_0
+FROM __scoped_traces
+ORDER BY __result_0 DESC
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+
+	// an unresolved $var is only rejected as an unknown aggregate today; a targeted
+	// "unknown variable" error is a separate concern
+	_, err = b.Build(ctx, valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar, q, nil)
+	require.ErrorContains(t, err, `aggregate "$threshold" cannot be used`)
+
+	// __all__ drops the condition: the query equals its unfiltered form
+	stmt, err = b.Build(ctx, valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar, q,
+		map[string]qbtypes.VariableItem{"threshold": {Type: qbtypes.DynamicVariableType, Value: "__all__"}})
+	require.NoError(t, err)
+	unfiltered := q
+	unfiltered.Filter = nil
+	want, err := b.Build(ctx, valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar, unfiltered, nil)
+	require.NoError(t, err)
+	assert.Equal(t, want.Query, stmt.Query)
+
+	// list variables render as IN with bound args; the scan selects only trace_id
+	// since no aggregation touches a per-trace column
+	stmt, err = b.Build(ctx, valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "count(trace.trace_id)"}},
+			Filter:       &qbtypes.Filter{Expression: "trace.llm_call_count IN $counts"},
+		}, map[string]qbtypes.VariableItem{
+			"counts": {Type: qbtypes.QueryVariableType, Value: []any{float64(1), float64(2)}},
+		})
+	require.NoError(t, err)
+	assertSQLEqual(t, `
+WITH __qualified AS (
+    SELECT trace_id,
+        countIf(mapContains(attributes_string, 'gen_ai.request.model')) AS llm_call_count
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+    GROUP BY trace_id
+    HAVING llm_call_count IN (1, 2)
+),
+__scoped_traces AS (
+    SELECT trace_id
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND trace_id GLOBAL IN (SELECT trace_id FROM __qualified)
+    GROUP BY trace_id
+)
+SELECT count(trace_id) AS __result_0
+FROM __scoped_traces
+ORDER BY __result_0 DESC
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// Resource conditions on the native path: the __resource_filter CTE prunes the
+// qualification scan and the per-trace scan by fingerprint.
+func TestBuild_FullSQL_Aggregation_ResourceFilter_Native(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "avg(trace.output_tokens)"}},
+			Filter:       &qbtypes.Filter{Expression: "service.name = 'api' AND trace.output_tokens > 1000"},
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __resource_filter AS (
+    SELECT fingerprint
+    FROM signoz_traces.distributed_traces_v3_resource
+    WHERE (simpleJSONExtractString(labels, 'service.name') = 'api' AND labels LIKE '%service.name%' AND labels LIKE '%service.name":"api%')
+      AND seen_at_ts_bucket_start >= 1747945619
+      AND seen_at_ts_bucket_start <= 1747983448
+    GROUP BY fingerprint
+),
+__qualified AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND resource_fingerprint GLOBAL IN (SELECT fingerprint FROM __resource_filter)
+    GROUP BY trace_id
+    HAVING output_tokens > 1000
+),
+__scoped_traces AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND resource_fingerprint GLOBAL IN (SELECT fingerprint FROM __resource_filter)
+      AND trace_id GLOBAL IN (SELECT trace_id FROM __qualified)
+    GROUP BY trace_id
+)
+SELECT avg(output_tokens) AS __result_0
+FROM __scoped_traces
+ORDER BY __result_0 DESC
+SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000
+`, stmt)
+}
+
+// Resource conditions on the delegated path: the standalone __trace_scope inlines its
+// fingerprint subquery (it is built without the delegate's CTEs), while the delegate
+// keeps its own __resource_filter CTE and inline resource predicate.
+func TestBuild_FullSQL_Aggregation_ResourceFilter_Delegated(t *testing.T) {
+	b := newTestBuilder(t)
+	stmt, err := b.Build(context.Background(), valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar,
+		qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+			Signal:       telemetrytypes.SignalTraces,
+			Aggregations: []qbtypes.TraceAggregation{{Expression: "sum(gen_ai.usage.output_tokens)"}},
+			Filter:       &qbtypes.Filter{Expression: "service.name = 'api' AND trace.output_tokens > 1000"},
+		}, nil)
+	require.NoError(t, err)
+
+	assertSQLEqual(t, `
+WITH __resource_filter AS (
+    SELECT fingerprint
+    FROM signoz_traces.distributed_traces_v3_resource
+    WHERE ((simpleJSONExtractString(labels, 'service.name') = 'api' AND labels LIKE '%service.name%' AND labels LIKE '%service.name":"api%'))
+      AND seen_at_ts_bucket_start >= 1747945619
+      AND seen_at_ts_bucket_start <= 1747983448
+    GROUP BY fingerprint
+),
+__trace_scope AS (
+    SELECT trace_id,
+        sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS output_tokens
+    FROM signoz_traces.distributed_signoz_index_v3
+    WHERE timestamp >= '1747947419000000000'
+      AND timestamp < '1747983448000000000'
+      AND ts_bucket_start >= 1747945619
+      AND ts_bucket_start <= 1747983448
+      AND (mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))
+      AND resource_fingerprint GLOBAL IN (SELECT fingerprint FROM (SELECT fingerprint FROM signoz_traces.distributed_traces_v3_resource WHERE (simpleJSONExtractString(labels, 'service.name') = 'api' AND labels LIKE '%service.name%' AND labels LIKE '%service.name":"api%') AND seen_at_ts_bucket_start >= 1747945619 AND seen_at_ts_bucket_start <= 1747983448 GROUP BY fingerprint))
+    GROUP BY trace_id
+    HAVING output_tokens > 1000
+)
+SELECT sum(multiIf(mapContains(attributes_number, 'gen_ai.usage.output_tokens'), toFloat64(attributes_number['gen_ai.usage.output_tokens']), NULL)) AS __result_0
+FROM signoz_traces.distributed_signoz_index_v3
+WHERE resource_fingerprint GLOBAL IN (SELECT fingerprint FROM __resource_filter)
+  AND trace_id GLOBAL IN (SELECT trace_id FROM __trace_scope)
+  AND (((mapContains(attributes_string, 'gen_ai.request.model') OR mapContains(attributes_string, 'gen_ai.tool.name') OR mapContains(attributes_string, 'gen_ai.agent.name'))) AND ((multiIf(resource.service.name IS NOT NULL, resource.service.name::String, mapContains(resources_string, 'service.name'), resources_string['service.name'], NULL) = 'api' AND multiIf(resource.service.name IS NOT NULL, resource.service.name::String, mapContains(resources_string, 'service.name'), resources_string['service.name'], NULL) IS NOT NULL)))
+  AND timestamp >= '1747947419000000000'
+  AND timestamp < '1747983448000000000'
+  AND ts_bucket_start >= 1747945619
+  AND ts_bucket_start <= 1747983448
+ORDER BY __result_0 DESC
+`, stmt)
+}
+
+// rate() over a trace-level column divides by the window (scalar) / step (series).
+// Note the AggreFuncMap semantics: it counts per-trace rows per second — it does not
+// sum the column.
+func TestBuild_Aggregation_RateDividesByInterval(t *testing.T) {
+	b := newTestBuilder(t)
+	ctx := context.Background()
+	q := qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]{
+		Signal:       telemetrytypes.SignalTraces,
+		Aggregations: []qbtypes.TraceAggregation{{Expression: "rate(trace.llm_call_count)"}},
+	}
+
+	stmt, err := b.Build(ctx, valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeScalar, q, nil)
+	require.NoError(t, err)
+	assert.Contains(t, stmt.Query, "count(llm_call_count)/36029 AS __result_0") // (end-start) seconds
+
+	q.StepInterval = qbtypes.Step{Duration: 60 * time.Second}
+	stmt, err = b.Build(ctx, valuer.UUID{}, testStartMs, testEndMs, qbtypes.RequestTypeTimeSeries, q, nil)
+	require.NoError(t, err)
+	assert.Contains(t, stmt.Query, "count(llm_call_count)/60 AS __result_0")
+}

pkg/statementbuilder/scopedtracesstatementbuilder/statement_builder.go

@@ -4,7 +4,6 @@ import (
 	"context"
 	"fmt"
 	"log/slog"
-	"sort"
 	"strings"
 
 	"github.com/SigNoz/signoz/pkg/errors"
@@ -19,7 +18,6 @@ import (
 	qbtypes "github.com/SigNoz/signoz/pkg/types/querybuildertypes/querybuildertypesv5"
 	"github.com/SigNoz/signoz/pkg/types/telemetrytypes"
 	"github.com/SigNoz/signoz/pkg/valuer"
-	qbvariables "github.com/SigNoz/signoz/pkg/variables"
 	"github.com/huandu/go-sqlbuilder"
 )
 
@@ -115,6 +113,8 @@ func (b *scopedTraceStatementBuilder) Build(
 		return b.buildTraceListQuery(ctx, orgID, querybuilder.ToNanoSecs(start), querybuilder.ToNanoSecs(end), query, variables)
 	case qbtypes.RequestTypeRaw:
 		return b.buildDelegated(ctx, orgID, start, end, requestType, query, variables)
+	case qbtypes.RequestTypeScalar, qbtypes.RequestTypeTimeSeries:
+		return b.buildAggregation(ctx, orgID, start, end, requestType, query, variables)
 	default:
 		return nil, ErrUnsupportedRequestType
 	}
@@ -143,6 +143,62 @@ func (b *scopedTraceStatementBuilder) buildDelegated(
 	return b.traceStmtBuilder.Build(ctx, orgID, start, end, requestType, gated, variables)
 }
 
+// traceScopedStatementBuilder is the delegate's optional capability of constraining a
+// query to a set of trace ids (implemented by the traces statement builder).
+type traceScopedStatementBuilder interface {
+	BuildTraceScoped(ctx context.Context, orgID valuer.UUID, start, end uint64, requestType qbtypes.RequestType, query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation], variables map[string]qbtypes.VariableItem, traceScope *qbtypes.Statement) (*qbtypes.Statement, error)
+}
+
+// buildDelegatedAggregation serves span-level scalar/time-series: the gate is ANDed
+// into the filter's span-level part and the query delegates to the standard trace
+// builder; a trace-level part becomes a qualification the delegate constrains
+// trace_id by.
+func (b *scopedTraceStatementBuilder) buildDelegatedAggregation(
+	ctx context.Context,
+	orgID valuer.UUID,
+	start, end uint64,
+	requestType qbtypes.RequestType,
+	query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation],
+	variables map[string]qbtypes.VariableItem,
+) (*qbtypes.Statement, error) {
+	var spanExpr, traceExpr string
+	var err error
+	if query.Filter != nil && strings.TrimSpace(query.Filter.Expression) != "" {
+		spanExpr, traceExpr, err = querybuilder.SplitFilterForAggregates(query.Filter.Expression, b.aggregateAliasSet())
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	gate := b.scope.FilterExpression
+	expr := gate
+	if strings.TrimSpace(spanExpr) != "" {
+		expr = fmt.Sprintf("(%s) AND (%s)", gate, spanExpr)
+	}
+
+	// shallow copy; only Filter is replaced, caller's query untouched
+	gated := query
+	gated.Filter = &qbtypes.Filter{Expression: expr}
+
+	if strings.TrimSpace(traceExpr) == "" {
+		return b.traceStmtBuilder.Build(ctx, orgID, start, end, requestType, gated, variables)
+	}
+
+	scoped, ok := b.traceStmtBuilder.(traceScopedStatementBuilder)
+	if !ok {
+		return nil, errors.NewInternalf(errors.CodeInternal, "trace statement builder does not support trace-scoped queries")
+	}
+	scope, err := b.buildQualifiedStatement(ctx, orgID, querybuilder.ToNanoSecs(start), querybuilder.ToNanoSecs(end), traceExpr, query, variables)
+	if err != nil {
+		return nil, err
+	}
+	if scope == nil {
+		// every trace-level condition was dropped by variable resolution
+		return b.traceStmtBuilder.Build(ctx, orgID, start, end, requestType, gated, variables)
+	}
+	return scoped.BuildTraceScoped(ctx, orgID, start, end, requestType, gated, variables, scope)
+}
+
 // buildTraceListQuery wires the CTE pipeline (start/end are nanoseconds):
 // matched (windowed, mask-pruned top-N trace_ids) → ranked (their [start,end] from
 // the summary table) → buckets (ts_bucket_start prune) → enrichment (every per-trace
@@ -184,22 +240,17 @@ func (b *scopedTraceStatementBuilder) buildTraceListQuery(
 	if err != nil {
 		return nil, err
 	}
-	orderableSet := orderableAliasSet(resolved)
-
 	resourceFrag, resourceArgs, resourcePred, err := b.maybeAttachResourceFilter(ctx, orgID, query, start, end, variables)
 	if err != nil {
 		return nil, err
 	}
 
-	fp, err := b.splitFilter(ctx, orgID, query, b.aggregateAliasSet(), orderableSet, start, end, variables, matchedSB)
+	fp, err := b.splitFilter(ctx, orgID, query, b.aggregateAliasSet(), start, end, variables, matchedSB)
 	if err != nil {
 		return nil, err
 	}
 
-	matchedFrag, matchedArgs, err := b.buildMatchedCTE(matchedSB, start, end, startBucket, endBucket, resolved, orders, orderableSet, maskExpr, fp, resourcePred, limit, query.Offset)
-	if err != nil {
-		return nil, err
-	}
+	matchedFrag, matchedArgs := b.buildMatchedCTE(matchedSB, start, end, startBucket, endBucket, resolved, orders, maskExpr, fp, resourcePred, limit, query.Offset)
 	rankedFrag, rankedArgs := b.buildRankedCTE(start, end)
 
 	adj := querybuilder.BucketAdjustment // 30-min bucket width in seconds
@@ -380,27 +431,27 @@ func (b *scopedTraceStatementBuilder) resolveListOrders(order []qbtypes.OrderBy,
 	return orders, nil
 }
 
-// filterParts is the user filter split into a span-level predicate and a trace-level
-// HAVING expression.
+// filterParts is the user filter split into a span-level predicate and the resolved
+// trace-level HAVING (nil when there is none).
 type filterParts struct {
 	spanPred      string
 	hasSpanFilter bool
-	havingExpr    string
+	having        *traceHaving
 	warnings      []string
 	warningsURL   string
 }
 
-// splitFilter splits query.Filter into a span-level predicate (args bound into sb)
-// and a trace-level HAVING (explicit query.Having ANDed on), then validates the
-// trace-level part against the matched-pass aggregates.
-func (b *scopedTraceStatementBuilder) splitFilter(ctx context.Context, orgID valuer.UUID, query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation], classifySet, orderableSet map[string]struct{}, start, end uint64, variables map[string]qbtypes.VariableItem, sb *sqlbuilder.SelectBuilder) (filterParts, error) {
+// splitFilter splits query.Filter into a span-level predicate and a trace-level
+// HAVING (explicit query.Having ANDed on before resolution); args bind into sb.
+func (b *scopedTraceStatementBuilder) splitFilter(ctx context.Context, orgID valuer.UUID, query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation], classifySet map[string]struct{}, start, end uint64, variables map[string]qbtypes.VariableItem, sb *sqlbuilder.SelectBuilder) (filterParts, error) {
 	var fp filterParts
+	havingExpr := ""
 	if query.Filter != nil && strings.TrimSpace(query.Filter.Expression) != "" {
 		spanExpr, traceExpr, err := querybuilder.SplitFilterForAggregates(query.Filter.Expression, classifySet)
 		if err != nil {
 			return fp, err
 		}
-		fp.havingExpr = traceExpr
+		havingExpr = traceExpr
 		if strings.TrimSpace(spanExpr) != "" {
 			pred, warnings, url, err := b.resolveSpanPredicate(ctx, orgID, start, end, spanExpr, variables, sb)
 			if err != nil {
@@ -415,23 +466,17 @@ func (b *scopedTraceStatementBuilder) splitFilter(ctx context.Context, orgID val
 		}
 	}
 	if query.Having != nil && strings.TrimSpace(query.Having.Expression) != "" {
-		if fp.havingExpr != "" {
-			fp.havingExpr = fmt.Sprintf("(%s) AND (%s)", fp.havingExpr, query.Having.Expression)
+		if havingExpr != "" {
+			havingExpr = fmt.Sprintf("(%s) AND (%s)", havingExpr, query.Having.Expression)
 		} else {
-			fp.havingExpr = query.Having.Expression
+			havingExpr = query.Having.Expression
 		}
 	}
-	// the HAVING is a plain text rewrite, so substitute variables here
-	if strings.TrimSpace(fp.havingExpr) != "" && len(variables) > 0 {
-		replaced, err := qbvariables.ReplaceVariablesInExpression(fp.havingExpr, variables)
-		if err != nil {
-			return fp, err
-		}
-		fp.havingExpr = replaced
-	}
-	if err := validateAggregateFilter(fp.havingExpr, orderableSet); err != nil {
+	having, err := b.resolveTraceHaving(ctx, havingExpr, variables, sb)
+	if err != nil {
 		return fp, err
 	}
+	fp.having = having
 	return fp, nil
 }
 
@@ -473,8 +518,8 @@ func (b *scopedTraceStatementBuilder) resolveSpanPredicate(ctx context.Context,
 // span filter + HAVING + ORDER BY + LIMIT/OFFSET, selecting only the aliases ORDER BY
 // / HAVING reference. Expressions carry $n markers bound to sb, so each can appear
 // several times and every occurrence resolves to the same arg.
-func (b *scopedTraceStatementBuilder) buildMatchedCTE(sb *sqlbuilder.SelectBuilder, start, end, startBucket, endBucket uint64, resolved []resolvedColumn, orders []listOrder, orderableSet map[string]struct{}, maskExpr string, fp filterParts, resourcePred string, limit, offset int) (string, []any, error) {
-	needed := neededMatchedAliases(orders, fp.havingExpr, orderableSet)
+func (b *scopedTraceStatementBuilder) buildMatchedCTE(sb *sqlbuilder.SelectBuilder, start, end, startBucket, endBucket uint64, resolved []resolvedColumn, orders []listOrder, maskExpr string, fp filterParts, resourcePred string, limit, offset int) (string, []any) {
+	needed := neededMatchedAliases(orders, fp.having)
 	selects := []string{"trace_id"}
 	for _, rc := range resolved {
 		if _, ok := needed[rc.alias]; !ok {
@@ -511,22 +556,8 @@ func (b *scopedTraceStatementBuilder) buildMatchedCTE(sb *sqlbuilder.SelectBuild
 		having = append(having, "countIf("+maskExpr+") > 0")
 		having = append(having, "countIf("+fp.spanPred+") > 0")
 	}
-	if strings.TrimSpace(fp.havingExpr) != "" {
-		// the rewriter matches raw key text, so map the trace. form alongside the bare name
-		columnMap := make(map[string]string, len(orderableSet)*2)
-		for a := range orderableSet {
-			columnMap[a] = quoteAlias(a)
-			columnMap[telemetrytypes.FieldContextTrace.StringValue()+"."+a] = quoteAlias(a)
-		}
-		hv, err := querybuilder.NewHavingExpressionRewriter().Rewrite(fp.havingExpr, columnMap)
-		if err != nil {
-			return "", nil, err
-		}
-		if hv != "" {
-			// escape user text so a literal $ isn't read as an arg marker; the countIf
-			// entries hold live $n markers and must stay unescaped
-			having = append(having, sqlbuilder.Escape(hv))
-		}
+	if fp.having != nil {
+		having = append(having, fp.having.pred)
 	}
 	if len(having) > 0 {
 		sb.Having(strings.Join(having, " AND "))
@@ -539,7 +570,7 @@ func (b *scopedTraceStatementBuilder) buildMatchedCTE(sb *sqlbuilder.SelectBuild
 	}
 
 	sql, args := sb.BuildWithFlavor(sqlbuilder.ClickHouse)
-	return fmt.Sprintf("matched AS (%s)", sql), args, nil
+	return fmt.Sprintf("matched AS (%s)", sql), args
 }
 
 // buildRankedCTE builds `ranked`: [start,end] bounds per matched trace from the
@@ -592,59 +623,36 @@ func (b *scopedTraceStatementBuilder) aggregateAliasSet() map[string]struct{} {
 	return set
 }
 
-// orderableAliasSet is the subset of aliases computable in the matched pass.
-func orderableAliasSet(resolved []resolvedColumn) map[string]struct{} {
-	set := make(map[string]struct{})
-	for _, rc := range resolved {
-		if rc.orderable {
-			set[rc.alias] = struct{}{}
-		}
-	}
-	return set
-}
-
 // neededMatchedAliases is the minimal alias set the matched pass must select: those
-// in ORDER BY plus those in the aggregate HAVING.
-func neededMatchedAliases(orders []listOrder, havingExpr string, orderableSet map[string]struct{}) map[string]struct{} {
+// in ORDER BY plus those the resolved trace-level HAVING touches.
+func neededMatchedAliases(orders []listOrder, having *traceHaving) map[string]struct{} {
 	needed := make(map[string]struct{})
 	for _, o := range orders {
 		needed[o.alias] = struct{}{}
 	}
-	for _, name := range traceAggregateNames(havingExpr) {
-		if _, ok := orderableSet[name]; ok {
+	if having != nil {
+		for name := range having.used {
 			needed[name] = struct{}{}
 		}
 	}
 	return needed
 }
 
-// traceAggregateNames extracts the aggregate names a trace-level HAVING references;
-// only unspecified- and trace-context selectors name aggregates.
-func traceAggregateNames(havingExpr string) []string {
-	var names []string
-	for _, sel := range querybuilder.QueryStringToKeysSelectors(havingExpr) {
-		if sel.FieldContext == telemetrytypes.FieldContextUnspecified || sel.FieldContext == telemetrytypes.FieldContextTrace {
-			names = append(names, sel.Name)
-		}
-	}
-	return names
-}
-
-// validateAggregateFilter rejects a trace-level filter referencing an aggregate not
-// computable in the matched pass.
+// validateAggregateFilter rejects trace-level filters on aggregates not computable in
+// the matched pass (e.g. span_count) with a targeted top-level error; inside the
+// where-clause visitor it would surface only as a detail of a combined error. Only
+// unspecified- and trace-context selectors name aggregates.
 func validateAggregateFilter(havingExpr string, orderableSet map[string]struct{}) error {
 	if strings.TrimSpace(havingExpr) == "" {
 		return nil
 	}
-	allowed := make([]string, 0, len(orderableSet))
-	for a := range orderableSet {
-		allowed = append(allowed, a)
-	}
-	sort.Strings(allowed)
-	for _, name := range traceAggregateNames(havingExpr) {
-		if _, ok := orderableSet[name]; !ok {
+	for _, sel := range querybuilder.QueryStringToKeysSelectors(havingExpr) {
+		if sel.FieldContext != telemetrytypes.FieldContextUnspecified && sel.FieldContext != telemetrytypes.FieldContextTrace {
+			continue
+		}
+		if _, ok := orderableSet[sel.Name]; !ok {
 			return errors.NewInvalidInputf(errors.CodeInvalidInput,
-				"aggregate %q cannot be used in the trace-list filter; filterable aggregates: %s", name, strings.Join(allowed, ", "))
+				"aggregate %q cannot be used in a trace-level filter; filterable aggregates: %s", sel.Name, strings.Join(sortedAliases(orderableSet), ", "))
 		}
 	}
 	return nil

pkg/statementbuilder/scopedtracesstatementbuilder/trace_aggregation.go

@@ -0,0 +1,759 @@
+package scopedtracesstatementbuilder
+
+import (
+	"context"
+	"fmt"
+	"sort"
+	"strings"
+
+	chparser "github.com/AfterShip/clickhouse-sql-parser/parser"
+	"github.com/SigNoz/signoz/pkg/errors"
+	"github.com/SigNoz/signoz/pkg/querybuilder"
+	"github.com/SigNoz/signoz/pkg/telemetryschema/tracestelemetryschema"
+	qbtypes "github.com/SigNoz/signoz/pkg/types/querybuildertypes/querybuildertypesv5"
+	"github.com/SigNoz/signoz/pkg/types/telemetrytypes"
+	"github.com/SigNoz/signoz/pkg/valuer"
+	"github.com/huandu/go-sqlbuilder"
+)
+
+// Scalar / time-series for scoped-trace queries. The `trace.` prefix picks the
+// domain per expression: span-level (bare keys) delegates to the standard trace
+// builder with the gate ANDed in; trace-level aggregates window-clipped per-trace
+// values through the native pipeline (buildTraceAggregationQuery):
+//
+//	__qualified       traces whose whole-window aggregates satisfy the trace-level
+//	   │              filter part; present only when the filter has one.
+//	   ▼
+//	__scoped_traces   per-trace values: windowed, mask-pruned GROUP BY trace_id
+//	   │              (+ ts bucket for time series, + group-by columns). Columns
+//	   ▼              off a trace slice are NULL and skipped by outer aggregates.
+//	main              outer aggregation over the per-trace rows → __result_i.
+
+// traceAggregation is one aggregation rewritten to run over the per-trace scan.
+type traceAggregation struct {
+	expr   string              // rewritten SQL over the per-trace column aliases
+	used   map[string]struct{} // per-trace aliases referenced
+	isRate bool
+}
+
+// buildAggregation routes scalar/time-series requests by aggregation domain.
+func (b *scopedTraceStatementBuilder) buildAggregation(
+	ctx context.Context,
+	orgID valuer.UUID,
+	start, end uint64,
+	requestType qbtypes.RequestType,
+	query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation],
+	variables map[string]qbtypes.VariableItem,
+) (*qbtypes.Statement, error) {
+	traceAggs, err := b.classifyAggregations(query.Aggregations)
+	if err != nil {
+		return nil, err
+	}
+	if err := b.validateGroupBy(query); err != nil {
+		return nil, err
+	}
+	if len(traceAggs) == 0 {
+		return b.buildDelegatedAggregation(ctx, orgID, start, end, requestType, query, variables)
+	}
+	return b.buildTraceAggregationQuery(ctx, orgID, querybuilder.ToNanoSecs(start), querybuilder.ToNanoSecs(end), requestType, query, variables, traceAggs)
+}
+
+// classifyAggregations returns the rewritten trace-domain aggregations, nil when all
+// are span-domain; mixing the two domains is rejected.
+func (b *scopedTraceStatementBuilder) classifyAggregations(aggs []qbtypes.TraceAggregation) ([]traceAggregation, error) {
+	traceCols := b.orderableColumnSet()
+	var out []traceAggregation
+	spanCount := 0
+	for _, agg := range aggs {
+		ta, isTrace, err := rewriteTraceAggregation(agg.Expression, traceCols)
+		if err != nil {
+			return nil, err
+		}
+		if isTrace {
+			out = append(out, *ta)
+		} else {
+			spanCount++
+		}
+	}
+	if len(out) > 0 && spanCount > 0 {
+		return nil, errors.NewInvalidInputf(errors.CodeInvalidInput,
+			"span-level and trace-level (trace.) aggregations cannot be mixed in one query")
+	}
+	return out, nil
+}
+
+// orderableColumnSet is the static per-trace column set usable in trace-level
+// aggregations and filters.
+func (b *scopedTraceStatementBuilder) orderableColumnSet() map[string]struct{} {
+	set := make(map[string]struct{})
+	for _, c := range b.scope.Columns {
+		if c.Orderable {
+			set[c.Alias] = struct{}{}
+		}
+	}
+	return set
+}
+
+// validateGroupBy rejects trace-level columns as group-by keys with a targeted error
+// (not the field mapper's generic "field not found"). Order keys need no check here:
+// request validation only admits group keys and aggregation aliases/expressions.
+func (b *scopedTraceStatementBuilder) validateGroupBy(query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]) error {
+	aliases := b.aggregateAliasSet()
+	for _, gb := range query.GroupBy {
+		if isTraceLevelKey(gb.Name, gb.FieldContext, aliases) {
+			return errors.NewInvalidInputf(errors.CodeInvalidInput,
+				"grouping by trace-level aggregate %q is not supported; group by span attributes instead (e.g. service.name)", gb.Name)
+		}
+	}
+	return nil
+}
+
+// isTraceLevelKey reports whether a group-by / order key explicitly (trace. prefix or
+// trace context) names a per-trace aggregate; bare names pass through since they may
+// be span columns sharing an alias name (duration_nano, timestamp).
+func isTraceLevelKey(name string, fieldContext telemetrytypes.FieldContext, aliases map[string]struct{}) bool {
+	stripped := strings.TrimPrefix(name, "trace.")
+	if _, ok := aliases[stripped]; !ok {
+		return false
+	}
+	return stripped != name || fieldContext == telemetrytypes.FieldContextTrace
+}
+
+// rewriteTraceAggregation rewrites an aggregation over trace.-prefixed columns to run
+// on the per-trace scan (trace.output_tokens → output_tokens, functions mapped via
+// AggreFuncMap); a pure span-level expression returns isTrace=false for the delegate.
+func rewriteTraceAggregation(expr string, traceCols map[string]struct{}) (*traceAggregation, bool, error) {
+	p := chparser.NewParser("SELECT " + expr)
+	stmts, err := p.ParseStmts()
+	if err != nil {
+		return nil, false, errors.WrapInvalidInputf(err, errors.CodeInvalidInput, "failed to parse aggregation expression %q", expr)
+	}
+	if len(stmts) == 0 {
+		return nil, false, errors.NewInvalidInputf(errors.CodeInvalidInput, "invalid aggregation expression %q", expr)
+	}
+	sel, ok := stmts[0].(*chparser.SelectQuery)
+	if !ok || len(sel.SelectItems) == 0 {
+		return nil, false, errors.NewInvalidInputf(errors.CodeInvalidInput, "invalid aggregation expression %q", expr)
+	}
+
+	v := &traceAggVisitor{traceCols: traceCols, used: make(map[string]struct{})}
+	if err := sel.SelectItems[0].Accept(v); err != nil {
+		return nil, false, err
+	}
+	if !v.hasTrace {
+		return nil, false, nil
+	}
+	if v.hasSpan {
+		return nil, false, errors.NewInvalidInputf(errors.CodeInvalidInput,
+			"aggregation %q mixes trace-level (trace.) and span-level columns; use one domain per aggregation", expr)
+	}
+	return &traceAggregation{expr: chparser.Format(sel.SelectItems[0]), used: v.used, isRate: v.isRate}, true, nil
+}
+
+// traceAggVisitor classifies column references and rewrites trace.-prefixed ones in
+// place; the ancestor stack tells a column identifier from a path segment, function
+// name, or alias, and rejects trace. columns inside *If combinators.
+type traceAggVisitor struct {
+	chparser.DefaultASTVisitor
+	traceCols map[string]struct{}
+	used      map[string]struct{}
+	stack     []chparser.Expr
+	hasTrace  bool
+	hasSpan   bool
+	isRate    bool
+}
+
+func (v *traceAggVisitor) Enter(expr chparser.Expr) { v.stack = append(v.stack, expr) }
+func (v *traceAggVisitor) Leave(expr chparser.Expr) { v.stack = v.stack[:len(v.stack)-1] }
+
+// parent is the node enclosing the one currently being visited (the visited node
+// itself is the stack top).
+func (v *traceAggVisitor) parent() chparser.Expr {
+	if len(v.stack) < 2 {
+		return nil
+	}
+	return v.stack[len(v.stack)-2]
+}
+
+// enclosingCombinator returns the name of a surrounding *If-combinator function, if any.
+func (v *traceAggVisitor) enclosingCombinator() (string, bool) {
+	for _, e := range v.stack {
+		fn, ok := e.(*chparser.FunctionExpr)
+		if !ok {
+			continue
+		}
+		if agg, known := querybuilder.AggreFuncMap[valuer.NewString(strings.ToLower(fn.Name.Name))]; known && agg.FuncCombinator {
+			return fn.Name.Name, true
+		}
+	}
+	return "", false
+}
+
+// VisitPath classifies a dotted reference (trace.output_tokens); trace-level ones are
+// rewritten in place to the bare per-trace alias.
+func (v *traceAggVisitor) VisitPath(p *chparser.Path) error {
+	ref := chparser.Format(p)
+	col, isTrace := traceColumnRef(ref)
+	if !isTrace {
+		v.hasSpan = true
+		return nil
+	}
+	if err := v.acceptTraceColumn(ref, col); err != nil {
+		return err
+	}
+	p.Fields = p.Fields[len(p.Fields)-1:]
+	p.Fields[0].Name = col
+	return nil
+}
+
+// VisitIdent classifies a plain identifier (a backquoted `trace.output_tokens` is
+// trace-level); path segments, function names, and aliases are structural, not columns.
+func (v *traceAggVisitor) VisitIdent(i *chparser.Ident) error {
+	switch parent := v.parent().(type) {
+	case *chparser.Path:
+		return nil // segments are classified whole by VisitPath
+	case *chparser.FunctionExpr:
+		if parent.Name == i {
+			return nil
+		}
+	case *chparser.ColumnExpr:
+		if parent.Alias == i {
+			return nil
+		}
+	}
+	col, isTrace := traceColumnRef(i.Name)
+	if !isTrace {
+		v.hasSpan = true
+		return nil
+	}
+	if err := v.acceptTraceColumn(i.Name, col); err != nil {
+		return err
+	}
+	i.Name = col
+	return nil
+}
+
+// acceptTraceColumn validates one trace-level column reference and records it.
+func (v *traceAggVisitor) acceptTraceColumn(ref, col string) error {
+	if name, in := v.enclosingCombinator(); in {
+		return errors.NewInvalidInputf(errors.CodeInvalidInput,
+			"%q over trace-level (trace.) columns is not supported; put the trace-level condition in the filter expression instead", name)
+	}
+	// trace_id is always selected by the per-trace scan (count(trace.trace_id)
+	// counts traces); everything else must be a scope column.
+	if col != "trace_id" {
+		if _, known := v.traceCols[col]; !known {
+			return errors.NewInvalidInputf(errors.CodeInvalidInput,
+				"unknown trace-level aggregation column %q; usable columns: %s", ref, strings.Join(sortedAliases(v.traceCols), ", "))
+		}
+		v.used[col] = struct{}{}
+	}
+	v.hasTrace = true
+	return nil
+}
+
+// VisitFunctionExpr validates and maps the function name. Children were already
+// visited (post-order), so classification is complete for this subtree.
+func (v *traceAggVisitor) VisitFunctionExpr(fn *chparser.FunctionExpr) error {
+	name := strings.ToLower(fn.Name.Name)
+	aggFunc, ok := querybuilder.AggreFuncMap[valuer.NewString(name)]
+	if !ok {
+		return errors.NewInvalidInputf(errors.CodeInvalidInput, "unrecognized function: %s", name)
+	}
+	if fn.Params != nil && fn.Params.Items != nil && len(fn.Params.Items.Items) > 0 && aggFunc.FuncCombinator {
+		// combinator predicates over span columns stay span-level (countIf(has_error=true))
+		v.hasSpan = true
+		return nil
+	}
+	fn.Name.Name = aggFunc.FuncName
+	if aggFunc.Rate {
+		v.isRate = true
+	}
+	return nil
+}
+
+// traceColumnRef reports whether text is a pure trace.-prefixed column reference
+// (trace.output_tokens) and returns the bare column name.
+func traceColumnRef(text string) (string, bool) {
+	text = strings.TrimSpace(text)
+	rest, ok := strings.CutPrefix(text, "trace.")
+	if !ok {
+		return "", false
+	}
+	if rest == "" || strings.ContainsAny(rest, " ()'\"`,+-*/<>=!") {
+		return "", false
+	}
+	return rest, true
+}
+
+func sortedAliases(set map[string]struct{}) []string {
+	out := make([]string, 0, len(set))
+	for a := range set {
+		out = append(out, a)
+	}
+	sort.Strings(out)
+	return out
+}
+
+// ---------------------------------------------------------------------------
+// Qualification + per-trace scan
+// ---------------------------------------------------------------------------
+
+// buildQualifiedStatement builds the delegate's __trace_scope: trace ids whose
+// window-clipped aggregates satisfy the trace-level filter, resource-pruned inline
+// (the caller embeds it standalone). start/end are ns; nil when every condition was
+// dropped by variable resolution.
+func (b *scopedTraceStatementBuilder) buildQualifiedStatement(
+	ctx context.Context,
+	orgID valuer.UUID,
+	start, end uint64,
+	traceExpr string,
+	query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation],
+	variables map[string]qbtypes.VariableItem,
+) (*qbtypes.Statement, error) {
+	keys, err := b.fetchKeys(ctx, orgID)
+	if err != nil {
+		return nil, err
+	}
+	sb := sqlbuilder.NewSelectBuilder()
+	maskExpr, resolved, err := b.resolveFor(ctx, orgID, start, end, keys, sb)
+	if err != nil {
+		return nil, err
+	}
+	having, err := b.resolveTraceHaving(ctx, traceExpr, variables, sb)
+	if err != nil {
+		return nil, err
+	}
+	if having == nil {
+		return nil, nil //nolint:nilnil
+	}
+	var resourcePred string
+	// nil when the filter has no resource-attribute conditions
+	if stmt, err := b.resourceFilterStmtBuilder.Build(ctx, orgID, start, end, qbtypes.RequestTypeRaw, query, variables); err != nil {
+		return nil, err
+	} else if stmt != nil {
+		inlined, err := embedExpr(sb, stmt.Query, stmt.Args)
+		if err != nil {
+			return nil, err
+		}
+		resourcePred = fmt.Sprintf("resource_fingerprint GLOBAL IN (SELECT fingerprint FROM (%s))", inlined)
+	}
+	sql, args := b.buildPerTraceScan(sb, start, end, resolved, maskExpr, perTraceScanOpts{
+		needed:       having.used,
+		havingPred:   having.pred,
+		resourcePred: resourcePred,
+	})
+	return &qbtypes.Statement{Query: sql, Args: args}, nil
+}
+
+// embedExpr inlines a pre-built statement into sb, replacing each `?` with a builder
+// Var; a count mismatch would silently shift args into the wrong slots, so error out.
+func embedExpr(sb *sqlbuilder.SelectBuilder, expr string, args []any) (string, error) {
+	if n := strings.Count(expr, "?"); n != len(args) {
+		return "", errors.NewInternalf(errors.CodeInternal,
+			"scoped trace builder: %d placeholders != %d args embedding %q", n, len(args), expr)
+	}
+	var out strings.Builder
+	ai := 0
+	for i := 0; i < len(expr); i++ {
+		if expr[i] == '?' {
+			out.WriteString(sb.Var(args[ai]))
+			ai++
+			continue
+		}
+		out.WriteByte(expr[i])
+	}
+	return out.String(), nil
+}
+
+// groupColumn is a resolved span-attribute group-by column (arg-free expression).
+type groupColumn struct {
+	name string
+	expr string
+}
+
+// perTraceScanOpts parametrize one windowed, mask-pruned GROUP BY trace_id scan.
+// All expressions are already resolved against the scan's builder.
+type perTraceScanOpts struct {
+	stepSeconds  int64 // >0 → bucket per-trace values by time (ts column)
+	groupCols    []groupColumn
+	needed       map[string]struct{} // per-trace aliases to select
+	spanPred     string              // resolved span-level filter, ANDed per span
+	resourcePred string              // resource-fingerprint prune (CTE reference or inline subquery)
+	qualified    bool                // constrain to __qualified
+	limitPred    string              // top-N group prune (GLOBAL IN __limit_cte)
+	havingPred   string              // resolved HAVING predicate over the selected aliases
+}
+
+// buildPerTraceScan renders the scan: window + gate mask (+ span filter, resource
+// prune, qualification), grouped by trace_id (+ ts bucket, group-by columns).
+func (b *scopedTraceStatementBuilder) buildPerTraceScan(sb *sqlbuilder.SelectBuilder, start, end uint64, resolved []resolvedColumn, maskExpr string, o perTraceScanOpts) (string, []any) {
+	startBucket := start/querybuilder.NsToSeconds - querybuilder.BucketAdjustment
+	endBucket := end / querybuilder.NsToSeconds
+
+	selects := []string{"trace_id"}
+	if o.stepSeconds > 0 {
+		selects = append(selects, fmt.Sprintf("toStartOfInterval(timestamp, INTERVAL %d SECOND) AS ts", o.stepSeconds))
+	}
+	for _, gc := range o.groupCols {
+		selects = append(selects, fmt.Sprintf("toString(%s) AS `%s`", gc.expr, gc.name))
+	}
+	for _, rc := range resolved {
+		if _, ok := o.needed[rc.alias]; !ok {
+			continue
+		}
+		selects = append(selects, rc.expr+" AS "+quoteAlias(rc.alias))
+	}
+	sb.Select(selects...)
+	sb.From(fmt.Sprintf("%s.%s", tracestelemetryschema.DBName, tracestelemetryschema.SpanIndexV3TableName))
+
+	where := []string{
+		sb.GE("timestamp", fmt.Sprintf("%d", start)),
+		sb.L("timestamp", fmt.Sprintf("%d", end)),
+		sb.GE("ts_bucket_start", startBucket),
+		sb.LE("ts_bucket_start", endBucket),
+		maskExpr,
+	}
+	if strings.TrimSpace(o.spanPred) != "" {
+		where = append(where, o.spanPred)
+	}
+	if o.resourcePred != "" {
+		where = append(where, o.resourcePred)
+	}
+	if o.qualified {
+		where = append(where, "trace_id GLOBAL IN (SELECT trace_id FROM __qualified)")
+	}
+	if o.limitPred != "" {
+		where = append(where, o.limitPred)
+	}
+	sb.Where(where...)
+
+	groupBy := []string{"trace_id"}
+	if o.stepSeconds > 0 {
+		groupBy = append(groupBy, "ts")
+	}
+	for _, gc := range o.groupCols {
+		groupBy = append(groupBy, "`"+gc.name+"`")
+	}
+	sb.GroupBy(groupBy...)
+	if strings.TrimSpace(o.havingPred) != "" {
+		sb.Having(o.havingPred)
+	}
+	return sb.BuildWithFlavor(sqlbuilder.ClickHouse)
+}
+
+// resolveGroupColumns resolves span-attribute group-by keys through the field mapper
+// (metadata-aware), for selection inside the per-trace scan.
+func (b *scopedTraceStatementBuilder) resolveGroupColumns(ctx context.Context, orgID valuer.UUID, start, end uint64, groupBy []qbtypes.GroupByKey) ([]groupColumn, error) {
+	if len(groupBy) == 0 {
+		return nil, nil
+	}
+	selectors := make([]*telemetrytypes.FieldKeySelector, 0, len(groupBy))
+	for i := range groupBy {
+		selectors = append(selectors, &telemetrytypes.FieldKeySelector{
+			Name:              groupBy[i].Name,
+			Signal:            telemetrytypes.SignalTraces,
+			FieldContext:      groupBy[i].FieldContext,
+			FieldDataType:     groupBy[i].FieldDataType,
+			SelectorMatchType: telemetrytypes.FieldSelectorMatchTypeExact,
+		})
+	}
+	keys, _, err := b.metadataStore.GetKeysMulti(ctx, orgID, selectors)
+	if err != nil {
+		return nil, err
+	}
+	out := make([]groupColumn, 0, len(groupBy))
+	for i := range groupBy {
+		expr, err := b.fm.ColumnExpressionFor(ctx, orgID, start, end, &groupBy[i].TelemetryFieldKey, telemetrytypes.FieldDataTypeString, keys)
+		if err != nil {
+			return nil, err
+		}
+		out = append(out, groupColumn{name: groupBy[i].Name, expr: sqlbuilder.Escape(expr)})
+	}
+	return out, nil
+}
+
+// ---------------------------------------------------------------------------
+// Native trace-domain aggregation query
+// ---------------------------------------------------------------------------
+
+// scanContext is one per-scan resolution: a fresh builder with the mask, columns,
+// span predicate, and optionally the trace-level HAVING resolved against it.
+type scanContext struct {
+	sb       *sqlbuilder.SelectBuilder
+	maskExpr string
+	resolved []resolvedColumn
+	spanPred string
+	having   *traceHaving
+	warnings []string
+	warnURL  string
+}
+
+// newScanContext resolves everything a per-trace scan embeds against a fresh builder.
+func (b *scopedTraceStatementBuilder) newScanContext(
+	ctx context.Context,
+	orgID valuer.UUID,
+	start, end uint64,
+	keys map[string][]*telemetrytypes.TelemetryFieldKey,
+	spanExpr, traceExpr string,
+	variables map[string]qbtypes.VariableItem,
+) (*scanContext, error) {
+	sc := &scanContext{sb: sqlbuilder.NewSelectBuilder()}
+	var err error
+	sc.maskExpr, sc.resolved, err = b.resolveFor(ctx, orgID, start, end, keys, sc.sb)
+	if err != nil {
+		return nil, err
+	}
+	if strings.TrimSpace(spanExpr) != "" {
+		pred, warns, url, err := b.resolveSpanPredicate(ctx, orgID, start, end, spanExpr, variables, sc.sb)
+		if err != nil {
+			return nil, err
+		}
+		sc.spanPred, sc.warnings, sc.warnURL = pred, warns, url
+	}
+	if strings.TrimSpace(traceExpr) != "" {
+		sc.having, err = b.resolveTraceHaving(ctx, traceExpr, variables, sc.sb)
+		if err != nil {
+			return nil, err
+		}
+	}
+	return sc, nil
+}
+
+// buildTraceAggregationQuery builds the native pipeline (see the file comment).
+// start/end are ns.
+func (b *scopedTraceStatementBuilder) buildTraceAggregationQuery(
+	ctx context.Context,
+	orgID valuer.UUID,
+	start, end uint64,
+	requestType qbtypes.RequestType,
+	query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation],
+	variables map[string]qbtypes.VariableItem,
+	traceAggs []traceAggregation,
+) (*qbtypes.Statement, error) {
+	keys, err := b.fetchKeys(ctx, orgID)
+	if err != nil {
+		return nil, err
+	}
+
+	var spanExpr, traceExpr string
+	if query.Filter != nil && strings.TrimSpace(query.Filter.Expression) != "" {
+		spanExpr, traceExpr, err = querybuilder.SplitFilterForAggregates(query.Filter.Expression, b.aggregateAliasSet())
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	resourceFrag, resourceArgs, resourcePred, err := b.maybeAttachResourceFilter(ctx, orgID, query, start, end, variables)
+	if err != nil {
+		return nil, err
+	}
+
+	var cteFragments []string
+	var cteArgs [][]any
+	if resourceFrag != "" {
+		cteFragments = append(cteFragments, resourceFrag)
+		cteArgs = append(cteArgs, resourceArgs)
+	}
+
+	// __qualified: its own scan resolution, HAVING = the trace-level filter part
+	qualified := false
+	if strings.TrimSpace(traceExpr) != "" {
+		qsc, err := b.newScanContext(ctx, orgID, start, end, keys, "", traceExpr, variables)
+		if err != nil {
+			return nil, err
+		}
+		if qsc.having != nil {
+			qsql, qargs := b.buildPerTraceScan(qsc.sb, start, end, qsc.resolved, qsc.maskExpr, perTraceScanOpts{
+				needed:       qsc.having.used,
+				havingPred:   qsc.having.pred,
+				resourcePred: resourcePred,
+			})
+			cteFragments = append(cteFragments, fmt.Sprintf("__qualified AS (%s)", qsql))
+			cteArgs = append(cteArgs, qargs)
+			qualified = true
+		}
+	}
+
+	groupCols, err := b.resolveGroupColumns(ctx, orgID, start, end, query.GroupBy)
+	if err != nil {
+		return nil, err
+	}
+	groupNames := make([]string, 0, len(groupCols))
+	for _, gc := range groupCols {
+		groupNames = append(groupNames, "`"+gc.name+"`")
+	}
+
+	needed := make(map[string]struct{})
+	for _, ta := range traceAggs {
+		for a := range ta.used {
+			needed[a] = struct{}{}
+		}
+	}
+
+	stepSeconds := int64(0)
+	rateInterval := (end - start) / querybuilder.NsToSeconds
+	if requestType == qbtypes.RequestTypeTimeSeries {
+		stepSeconds = int64(query.StepInterval.Seconds())
+		rateInterval = uint64(stepSeconds)
+	}
+
+	// outer aggregation over the per-trace rows
+	sb := sqlbuilder.NewSelectBuilder()
+	selects := []string{}
+	if stepSeconds > 0 {
+		selects = append(selects, "ts")
+	}
+	selects = append(selects, groupNames...)
+	for i, ta := range traceAggs {
+		selects = append(selects, fmt.Sprintf("%s AS __result_%d", ta.rendered(rateInterval), i))
+	}
+	sb.Select(selects...)
+	sb.From("__scoped_traces")
+
+	// grouped, limited time series → rank groups on whole-window per-trace values
+	// (exact for non-composable aggregates) and prune the main scan to the top-N.
+	limitPred := ""
+	if requestType == qbtypes.RequestTypeTimeSeries && query.Limit > 0 && len(groupCols) > 0 {
+		tsc, err := b.newScanContext(ctx, orgID, start, end, keys, spanExpr, "", variables)
+		if err != nil {
+			return nil, err
+		}
+		totalSQL, totalArgs := b.buildPerTraceScan(tsc.sb, start, end, tsc.resolved, tsc.maskExpr, perTraceScanOpts{
+			groupCols:    groupCols,
+			needed:       needed,
+			spanPred:     tsc.spanPred,
+			resourcePred: resourcePred,
+			qualified:    qualified,
+		})
+		cteFragments = append(cteFragments, fmt.Sprintf("__scoped_traces_total AS (%s)", totalSQL))
+		cteArgs = append(cteArgs, totalArgs)
+
+		limitSQL, limitArgs := outerLimitSQL(query, traceAggs, groupNames, (end-start)/querybuilder.NsToSeconds)
+		cteFragments = append(cteFragments, fmt.Sprintf("__limit_cte AS (%s)", limitSQL))
+		cteArgs = append(cteArgs, limitArgs)
+
+		exprs := make([]string, 0, len(groupCols))
+		for _, gc := range groupCols {
+			exprs = append(exprs, "toString("+gc.expr+")")
+		}
+		limitPred = fmt.Sprintf("(%s) GLOBAL IN (SELECT %s FROM __limit_cte)",
+			strings.Join(exprs, ", "), strings.Join(groupNames, ", "))
+	}
+
+	msc, err := b.newScanContext(ctx, orgID, start, end, keys, spanExpr, "", variables)
+	if err != nil {
+		return nil, err
+	}
+	perTraceSQL, perTraceArgs := b.buildPerTraceScan(msc.sb, start, end, msc.resolved, msc.maskExpr, perTraceScanOpts{
+		stepSeconds:  stepSeconds,
+		groupCols:    groupCols,
+		needed:       needed,
+		spanPred:     msc.spanPred,
+		resourcePred: resourcePred,
+		qualified:    qualified,
+		limitPred:    limitPred,
+	})
+	cteFragments = append(cteFragments, fmt.Sprintf("__scoped_traces AS (%s)", perTraceSQL))
+	cteArgs = append(cteArgs, perTraceArgs)
+
+	groupBys := []string{}
+	if stepSeconds > 0 {
+		groupBys = append(groupBys, "ts")
+	}
+	groupBys = append(groupBys, groupNames...)
+	if len(groupBys) > 0 {
+		sb.GroupBy(groupBys...)
+	}
+
+	if query.Having != nil && strings.TrimSpace(query.Having.Expression) != "" {
+		rewritten, err := querybuilder.NewHavingExpressionRewriter().RewriteForTraces(query.Having.Expression, query.Aggregations)
+		if err != nil {
+			return nil, err
+		}
+		sb.Having(sqlbuilder.Escape(rewritten))
+	}
+
+	if requestType == qbtypes.RequestTypeTimeSeries {
+		if len(query.Order) != 0 {
+			for _, orderBy := range query.Order {
+				if _, ok := traceAggOrderIndex(orderBy, query); !ok {
+					sb.OrderBy(fmt.Sprintf("`%s` %s", orderBy.Key.Name, orderBy.Direction.StringValue()))
+				}
+			}
+			sb.OrderBy("ts desc")
+		}
+	} else {
+		for _, orderBy := range query.Order {
+			if idx, ok := traceAggOrderIndex(orderBy, query); ok {
+				sb.OrderBy(fmt.Sprintf("__result_%d %s", idx, orderBy.Direction.StringValue()))
+			} else {
+				sb.OrderBy(fmt.Sprintf("`%s` %s", orderBy.Key.Name, orderBy.Direction.StringValue()))
+			}
+		}
+		if len(query.Order) == 0 {
+			sb.OrderBy("__result_0 DESC")
+		}
+		if query.Limit > 0 {
+			sb.Limit(query.Limit)
+		}
+	}
+
+	mainSQL, mainArgs := sb.BuildWithFlavor(sqlbuilder.ClickHouse)
+	finalSQL := querybuilder.CombineCTEs(cteFragments) + mainSQL + " SETTINGS distributed_product_mode='allow', max_memory_usage=10000000000"
+	finalArgs := querybuilder.PrependArgs(cteArgs, mainArgs)
+
+	return &qbtypes.Statement{
+		Query:          finalSQL,
+		Args:           finalArgs,
+		Warnings:       msc.warnings,
+		WarningsDocURL: msc.warnURL,
+	}, nil
+}
+
+// rendered returns the outer aggregation SQL, dividing rate aggregations by the
+// interval (step for time series, window length for scalar).
+func (ta traceAggregation) rendered(rateInterval uint64) string {
+	if ta.isRate {
+		return fmt.Sprintf("%s/%d", ta.expr, rateInterval)
+	}
+	return ta.expr
+}
+
+// outerLimitSQL renders the top-N group selection for a grouped, limited time series:
+// outer aggregations over whole-window per-trace values, ranked and limited.
+func outerLimitSQL(query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation], traceAggs []traceAggregation, groupNames []string, windowSeconds uint64) (string, []any) {
+	sb := sqlbuilder.NewSelectBuilder()
+	selects := append([]string{}, groupNames...)
+	for i, ta := range traceAggs {
+		selects = append(selects, fmt.Sprintf("%s AS __result_%d", ta.rendered(windowSeconds), i))
+	}
+	sb.Select(selects...)
+	sb.From("__scoped_traces_total")
+	sb.GroupBy(groupNames...)
+	for _, orderBy := range query.Order {
+		if idx, ok := traceAggOrderIndex(orderBy, query); ok {
+			sb.OrderBy(fmt.Sprintf("__result_%d %s", idx, orderBy.Direction.StringValue()))
+		} else {
+			sb.OrderBy(fmt.Sprintf("`%s` %s", orderBy.Key.Name, orderBy.Direction.StringValue()))
+		}
+	}
+	if len(query.Order) == 0 {
+		sb.OrderBy("__result_0 DESC")
+	}
+	sb.Limit(query.Limit)
+	return sb.BuildWithFlavor(sqlbuilder.ClickHouse)
+}
+
+// traceAggOrderIndex reports whether an order key refers to the i-th aggregation
+// (by alias, expression, or index), mirroring the trace builder.
+func traceAggOrderIndex(k qbtypes.OrderBy, q qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation]) (int, bool) {
+	for i, agg := range q.Aggregations {
+		if k.Key.Name == agg.Alias ||
+			k.Key.Name == agg.Expression ||
+			k.Key.Name == fmt.Sprintf("%d", i) {
+			return i, true
+		}
+	}
+	return 0, false
+}

pkg/statementbuilder/scopedtracesstatementbuilder/trace_aggregation_test.go

@@ -0,0 +1,60 @@
+package scopedtracesstatementbuilder
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestRewriteTraceAggregation(t *testing.T) {
+	cols := map[string]struct{}{
+		"input_tokens": {}, "output_tokens": {}, "total_tokens": {}, "llm_call_count": {}, "max_llm_latency_ns": {},
+	}
+
+	cases := []struct {
+		name    string
+		expr    string
+		isTrace bool
+		want    string // rewritten expr, only checked when isTrace
+		used    []string
+		wantErr string
+	}{
+		{name: "avg trace col", expr: "avg(trace.output_tokens)", isTrace: true, want: "avg(output_tokens)", used: []string{"output_tokens"}},
+		{name: "sum trace col", expr: "sum(trace.total_tokens)", isTrace: true, want: "sum(total_tokens)", used: []string{"total_tokens"}},
+		{name: "count traces", expr: "count(trace.trace_id)", isTrace: true, want: "count(trace_id)"},
+		{name: "p90 trace col", expr: "p90(trace.max_llm_latency_ns)", isTrace: true, want: "quantile(0.90)(max_llm_latency_ns)", used: []string{"max_llm_latency_ns"}},
+		{name: "arithmetic between trace cols", expr: "avg(trace.output_tokens + trace.input_tokens)", isTrace: true, want: "avg(output_tokens + input_tokens)", used: []string{"output_tokens", "input_tokens"}},
+		{name: "arithmetic with constant", expr: "sum(trace.output_tokens * 1.5)", isTrace: true, want: "sum(output_tokens * 1.5)", used: []string{"output_tokens"}},
+		{name: "ratio of two aggregations", expr: "sum(trace.output_tokens)/count(trace.trace_id)", isTrace: true, want: "sum(output_tokens) / count(trace_id)", used: []string{"output_tokens"}},
+		{name: "backquoted trace col", expr: "avg(`trace.output_tokens`)", isTrace: true, want: "avg(`output_tokens`)", used: []string{"output_tokens"}},
+		{name: "bare count is span-level", expr: "count()", isTrace: false},
+		{name: "span attribute is span-level", expr: "sum(gen_ai.usage.output_tokens)", isTrace: false},
+		{name: "countIf span predicate is span-level", expr: "countIf(has_error = true)", isTrace: false},
+		{name: "mixed domains in one expression", expr: "sum(trace.output_tokens) + sum(gen_ai.usage.input_tokens)", wantErr: "mixes trace-level"},
+		{name: "mixed domains in one function", expr: "sum(trace.output_tokens + gen_ai.usage.input_tokens)", wantErr: "mixes trace-level"},
+		{name: "output-only column rejected", expr: "avg(trace.span_count)", wantErr: "unknown trace-level aggregation column"},
+		{name: "unknown column rejected", expr: "avg(trace.bogus)", wantErr: "unknown trace-level aggregation column"},
+		{name: "countIf over trace col rejected", expr: "countIf(trace.output_tokens > 1000)", wantErr: "not supported"},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			ta, isTrace, err := rewriteTraceAggregation(tc.expr, cols)
+			if tc.wantErr != "" {
+				require.ErrorContains(t, err, tc.wantErr)
+				return
+			}
+			require.NoError(t, err)
+			assert.Equal(t, tc.isTrace, isTrace)
+			if !tc.isTrace {
+				return
+			}
+			assert.Equal(t, tc.want, ta.expr)
+			for _, u := range tc.used {
+				assert.Contains(t, ta.used, u)
+			}
+			assert.Len(t, ta.used, len(tc.used))
+		})
+	}
+}

pkg/statementbuilder/scopedtracesstatementbuilder/trace_having.go

@@ -0,0 +1,146 @@
+package scopedtracesstatementbuilder
+
+import (
+	"context"
+	"strings"
+
+	"github.com/SigNoz/signoz/pkg/errors"
+	"github.com/SigNoz/signoz/pkg/querybuilder"
+	qbtypes "github.com/SigNoz/signoz/pkg/types/querybuildertypes/querybuildertypesv5"
+	"github.com/SigNoz/signoz/pkg/types/telemetrytypes"
+	"github.com/SigNoz/signoz/pkg/valuer"
+	qbvariables "github.com/SigNoz/signoz/pkg/variables"
+	"github.com/huandu/go-sqlbuilder"
+)
+
+// traceHaving is the resolved trace-level filter part: a HAVING predicate over the
+// per-trace aliases plus the aliases it references (so scans select only those).
+type traceHaving struct {
+	pred string
+	used map[string]struct{}
+}
+
+// resolveTraceHaving resolves a trace-level filter through the standard pipeline
+// (variable replacement, then PrepareWhereClause against the per-trace aliases), so
+// operators, bound args, and __all__ behave exactly as in span filters. Returns nil
+// when the expression is empty or every condition was dropped; args bind into sb.
+func (b *scopedTraceStatementBuilder) resolveTraceHaving(ctx context.Context, expr string, variables map[string]qbtypes.VariableItem, sb *sqlbuilder.SelectBuilder) (*traceHaving, error) {
+	if strings.TrimSpace(expr) == "" {
+		return nil, nil //nolint:nilnil
+	}
+	// variables are replaced before validation so their literals are not mistaken for
+	// aggregate names; an unresolved $var is left in place and fails validation below
+	// (as an unknown aggregate — targeted variable errors are a separate concern)
+	if len(variables) > 0 {
+		replaced, err := qbvariables.ReplaceVariablesInExpression(expr, variables)
+		if err != nil {
+			return nil, err
+		}
+		expr = replaced
+		if strings.TrimSpace(expr) == "" {
+			return nil, nil //nolint:nilnil
+		}
+	}
+	allowed := b.orderableColumnSet()
+	// upfront targeted errors; the visitor folds them into a combined "Found N errors"
+	if err := validateAggregateFilter(expr, allowed); err != nil {
+		return nil, err
+	}
+
+	// both spellings resolve here: the key parser strips the trace. prefix into
+	// FieldContextTrace, which matches this entry's context
+	fieldKeys := make(map[string][]*telemetrytypes.TelemetryFieldKey, len(allowed))
+	for alias := range allowed {
+		key := &telemetrytypes.TelemetryFieldKey{Name: alias, FieldContext: telemetrytypes.FieldContextTrace}
+		fieldKeys[alias] = []*telemetrytypes.TelemetryFieldKey{key}
+	}
+
+	cb := &aliasConditionBuilder{allowed: allowed, used: make(map[string]struct{})}
+	prepared, err := querybuilder.PrepareWhereClause(expr, querybuilder.FilterExprVisitorOpts{
+		Context:          ctx,
+		Logger:           b.logger,
+		ConditionBuilder: cb,
+		FieldKeys:        fieldKeys,
+		Variables:        variables,
+		Builder:          sb,
+	})
+	if err != nil {
+		return nil, err
+	}
+	if prepared.IsEmpty() {
+		return nil, nil //nolint:nilnil
+	}
+	return &traceHaving{pred: prepared.Expr, used: cb.used}, nil
+}
+
+// aliasConditionBuilder renders filter conditions directly against the per-trace
+// aliases, recording the ones it touches; a key resolving to no alias is an error.
+type aliasConditionBuilder struct {
+	allowed map[string]struct{}
+	used    map[string]struct{}
+}
+
+var _ qbtypes.ConditionBuilder = (*aliasConditionBuilder)(nil)
+
+func (c *aliasConditionBuilder) ConditionFor(
+	_ context.Context,
+	_ valuer.UUID,
+	_, _ uint64,
+	key *telemetrytypes.TelemetryFieldKey,
+	keys map[string][]*telemetrytypes.TelemetryFieldKey,
+	_ qbtypes.ConditionBuilderOptions,
+	op qbtypes.FilterOperator,
+	value any,
+	sb *sqlbuilder.SelectBuilder,
+) ([]string, []string, error) {
+	matching := keys[key.Name]
+	if len(matching) == 0 {
+		return nil, nil, errors.NewInvalidInputf(errors.CodeInvalidInput,
+			"aggregate %q cannot be used in a trace-level filter; filterable aggregates: %s",
+			key.Name, strings.Join(sortedAliases(c.allowed), ", "))
+	}
+	alias := matching[0].Name
+	c.used[alias] = struct{}{}
+	col := quoteAlias(alias)
+
+	var cond string
+	switch op {
+	case qbtypes.FilterOperatorEqual:
+		cond = sb.E(col, value)
+	case qbtypes.FilterOperatorNotEqual:
+		cond = sb.NE(col, value)
+	case qbtypes.FilterOperatorGreaterThan:
+		cond = sb.G(col, value)
+	case qbtypes.FilterOperatorGreaterThanOrEq:
+		cond = sb.GE(col, value)
+	case qbtypes.FilterOperatorLessThan:
+		cond = sb.L(col, value)
+	case qbtypes.FilterOperatorLessThanOrEq:
+		cond = sb.LE(col, value)
+	case qbtypes.FilterOperatorIn, qbtypes.FilterOperatorNotIn:
+		values, ok := value.([]any)
+		if !ok {
+			values = []any{value}
+		}
+		if op == qbtypes.FilterOperatorIn {
+			cond = sb.In(col, values...)
+		} else {
+			cond = sb.NotIn(col, values...)
+		}
+	case qbtypes.FilterOperatorBetween, qbtypes.FilterOperatorNotBetween:
+		values, ok := value.([]any)
+		if !ok || len(values) != 2 {
+			return nil, nil, errors.NewInvalidInputf(errors.CodeInvalidInput,
+				"between on trace-level aggregate %q requires exactly two values", alias)
+		}
+		if op == qbtypes.FilterOperatorBetween {
+			cond = sb.Between(col, values[0], values[1])
+		} else {
+			cond = sb.NotBetween(col, values[0], values[1])
+		}
+	default:
+		return nil, nil, errors.NewInvalidInputf(errors.CodeInvalidInput,
+			"trace-level aggregate %q supports only comparison operators (=, !=, <, <=, >, >=, in, between)", alias)
+	}
+	return []string{cond}, nil, nil
+}

pkg/statementbuilder/tracesstatementbuilder/statement_builder.go

@@ -32,6 +32,9 @@ type traceQueryStatementBuilder struct {
 	resourceFilterResolver         *resourcefilter.ResourceFingerprintResolver[qbtypes.TraceAggregation]
 	aggExprRewriter                qbtypes.AggExprRewriter
 	skipResourceFingerprintEnabled bool
+	// traceScope, set only on the per-call copy made by BuildTraceScoped, constrains
+	// queries to spans whose trace_id is in the __trace_scope CTE.
+	traceScope *qbtypes.Statement
 }
 
 var _ qbtypes.StatementBuilder[qbtypes.TraceAggregation] = (*traceQueryStatementBuilder)(nil)
@@ -95,6 +98,33 @@ func NewTraceQueryStatementBuilder(
 	}
 }
 
+// BuildTraceScoped is Build constrained to trace_ids selected by traceScope; the
+// receiver is copied so the shared builder stays stateless.
+func (b *traceQueryStatementBuilder) BuildTraceScoped(
+	ctx context.Context,
+	orgID valuer.UUID,
+	start uint64,
+	end uint64,
+	requestType qbtypes.RequestType,
+	query qbtypes.QueryBuilderQuery[qbtypes.TraceAggregation],
+	variables map[string]qbtypes.VariableItem,
+	traceScope *qbtypes.Statement,
+) (*qbtypes.Statement, error) {
+	scoped := *b
+	scoped.traceScope = traceScope
+	return scoped.Build(ctx, orgID, start, end, requestType, query, variables)
+}
+
+// attachTraceScope adds the trace-scope condition to sb and returns the CTE fragment
+// + args to prepend; both empty when no scope is set.
+func (b *traceQueryStatementBuilder) attachTraceScope(sb *sqlbuilder.SelectBuilder) (string, []any) {
+	if b.traceScope == nil {
+		return "", nil
+	}
+	sb.Where("trace_id GLOBAL IN (SELECT trace_id FROM __trace_scope)")
+	return fmt.Sprintf("__trace_scope AS (%s)", b.traceScope.Query), b.traceScope.Args
+}
+
 // Build builds a SQL query for traces based on the given parameters.
 func (b *traceQueryStatementBuilder) Build(
 	ctx context.Context,
@@ -519,6 +549,11 @@ func (b *traceQueryStatementBuilder) buildTimeSeriesQuery(
 		cteArgs = append(cteArgs, args)
 	}
 
+	if scopeFrag, scopeArgs := b.attachTraceScope(sb); scopeFrag != "" {
+		cteFragments = append(cteFragments, scopeFrag)
+		cteArgs = append(cteArgs, scopeArgs)
+	}
+
 	sb.SelectMore(fmt.Sprintf(
 		"toStartOfInterval(timestamp, INTERVAL %d SECOND) AS ts",
 		int64(query.StepInterval.Seconds()),
@@ -679,6 +714,13 @@ func (b *traceQueryStatementBuilder) buildScalarQuery(
 		cteArgs = append(cteArgs, args)
 	}
 
+	// skipResourceCTE means this scalar is embedded as a CTE of a time-series query,
+	// which has already emitted the __trace_scope fragment — add only the condition.
+	if scopeFrag, scopeArgs := b.attachTraceScope(sb); scopeFrag != "" && !skipResourceCTE {
+		cteFragments = append(cteFragments, scopeFrag)
+		cteArgs = append(cteArgs, scopeArgs)
+	}
+
 	allAggChArgs := []any{}
 
 	fieldNames := make([]string, 0, len(query.GroupBy))

tests/fixtures/querierai.py

@@ -37,10 +37,10 @@ def ai_trace(
     *,
     now: datetime,
     service: str,
-    user: str,
     in_tokens: int | None,
     out_tokens: int,
-    cost: float,
+    user: str = "user",
+    cost: float = 0.1,
     model: str = "gpt-4o-mini",
     environment: str = "production",
 ) -> list[Traces]:
@@ -79,6 +79,28 @@ def ai_trace(
     ]
 
 
+def tool_only_trace(*, now: datetime, service: str) -> list[Traces]:
+    """Root + one tool span: passes the gen_ai gate but has NO LLM span."""
+    trace_id = TraceIdGenerator.trace_id()
+    root_id = TraceIdGenerator.span_id()
+    resources = {"service.name": service}
+    return [
+        root_span(now=now, trace_id=trace_id, span_id=root_id, resources=resources, duration_s=2),
+        Traces(
+            timestamp=now - timedelta(seconds=4),
+            duration=timedelta(seconds=0.5),
+            trace_id=trace_id,
+            span_id=TraceIdGenerator.span_id(),
+            parent_span_id=root_id,
+            name="execute_tool",
+            kind=TracesKind.SPAN_KIND_INTERNAL,
+            status_code=TracesStatusCode.STATUS_CODE_OK,
+            resources=resources,
+            attributes={"gen_ai.tool.name": "get_weather", "gen_ai.tool.type": "function"},
+        ),
+    ]
+
+
 def ai_trace_mixed_spans(*, now: datetime, service: str, user: str) -> list[Traces]:
     """Root + LLM + tool + agent spans; only the LLM span carries gen_ai.request.model."""
     trace_id = TraceIdGenerator.trace_id()

tests/integration/tests/querierai/01_ai_traces.py

@@ -77,8 +77,8 @@ def test_ai_list_having_aggregate_filter(
     get_token: Callable[[str, str], str],
     insert_traces: Callable[[list[Traces]], None],
 ) -> None:
-    """Span + aggregate condition in one filter box splits into WHERE + HAVING; bare
-    and `trace.` spellings behave identically; an output-only aggregate is rejected."""
+    """One filter box splits into WHERE + HAVING; bare and `trace.` spellings behave
+    identically; an output-only aggregate is rejected."""
     now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
     service = "ai-it-having"
 
@@ -327,9 +327,8 @@ def test_ai_list_nested_group_span_or_and_aggregate(
     get_token: Callable[[str, str], str],
     insert_traces: Callable[[list[Traces]], None],
 ) -> None:
-    """service.name = X AND (has_error = true OR gen_ai.request.model = 'gpt-4o') AND
-    total_tokens > 100: the nested OR group must not flatten, span predicates go to
-    WHERE, the aggregate to HAVING."""
+    """A nested (span OR span) group ANDed with an aggregate must not flatten: span
+    predicates go to WHERE, the aggregate to HAVING."""
     now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
     service = "ai-it-nested"
 

tests/integration/tests/querierai/02_ai_aggregations.py

@@ -0,0 +1,539 @@
+"""
+builder_ai_query scalar / time-series aggregations. The `trace.` prefix picks the
+domain per expression: bare keys aggregate over gen_ai spans, trace.* over
+window-clipped per-trace values; a trace-level filter condition qualifies whole
+traces on both domains. Tests isolate their data via unique service.name.
+"""
+
+from collections.abc import Callable
+from datetime import UTC, datetime, timedelta
+from http import HTTPStatus
+
+import pytest
+
+from fixtures import types
+from fixtures.auth import USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD
+from fixtures.querier import (
+    Aggregation,
+    BuilderQuery,
+    OrderBy,
+    RequestType,
+    TelemetryFieldKey,
+    get_scalar_table_data,
+    make_query_request,
+)
+from fixtures.querierai import ai_trace, query_window, tool_only_trace
+from fixtures.traces import TraceIdGenerator, Traces, TracesKind, TracesStatusCode
+
+
+def scalar_query(
+    service: str,
+    expression: str,
+    *,
+    filter_extra: str = "",
+    group_by: list[TelemetryFieldKey] | None = None,
+    alias: str | None = None,
+    having: str | None = None,
+    order: list[OrderBy] | None = None,
+    limit: int | None = None,
+) -> dict:
+    filter_expression = f"service.name = '{service}'"
+    if filter_extra:
+        filter_expression += f" AND {filter_extra}"
+    return BuilderQuery(
+        signal="traces",
+        query_type="builder_ai_query",
+        name="A",
+        filter_expression=filter_expression,
+        aggregations=[Aggregation(expression=expression, alias=alias)],
+        group_by=group_by,
+        having_expression=having,
+        order=order,
+        limit=limit,
+    ).to_dict()
+
+
+def scalar_value(signoz: types.SigNoz, token: str, start_ms: int, end_ms: int, service: str, expression: str, filter_extra: str = "") -> float:
+    """Run one single-aggregation scalar query and return its value."""
+    resp = make_query_request(
+        signoz,
+        token,
+        start_ms,
+        end_ms,
+        [scalar_query(service, expression, filter_extra=filter_extra)],
+        request_type=RequestType.SCALAR,
+    )
+    assert resp.status_code == HTTPStatus.OK, f"{expression}: {resp.text}"
+    data = get_scalar_table_data(resp.json())
+    assert len(data) == 1, f"{expression}: expected one row, got {data}"
+    return float(data[0][-1])
+
+
+def series_values(response_json: dict) -> list[list[float]]:
+    """Per-series lists of bucket values (bucket order as returned)."""
+    series = response_json["data"]["data"]["results"][0]["aggregations"][0]["series"]
+    return [[v["value"] for v in ser["values"]] for ser in series]
+
+
+def test_ai_scalar_trace_level_aggregations(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """Trace-level scalars over per-trace values: out-tokens 100/300 give avg=200 and
+    count=2, while the span-level count() sees the two LLM spans (roots gated out)."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-scalar"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=100) + ai_trace(now=now, service=service, in_tokens=30, out_tokens=300))
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    def value(expression: str) -> float:
+        return scalar_value(signoz, token, start_ms, end_ms, service, expression)
+
+    assert value("avg(trace.output_tokens)") == pytest.approx(200)
+    assert value("count(trace.trace_id)") == 2
+    assert value("max(trace.total_tokens)") == pytest.approx(330)
+    assert value("p50(trace.output_tokens)") == pytest.approx(200)  # AggreFuncMap -> quantile(0.50)
+    # arithmetic inside one function and between functions
+    assert value("avg(trace.output_tokens + trace.input_tokens)") == pytest.approx(220)
+    assert value("sum(trace.output_tokens)/count(trace.trace_id)") == pytest.approx(200)
+    # span-level domain still works through the same request type
+    assert value("count()") == 2  # the two LLM spans; roots are not gen_ai
+    assert value("sum(gen_ai.usage.output_tokens)") == pytest.approx(400)
+
+    # multiple trace-level aggregations in one query -> one column per aggregation
+    multi = BuilderQuery(
+        signal="traces",
+        query_type="builder_ai_query",
+        name="A",
+        filter_expression=f"service.name = '{service}'",
+        aggregations=[Aggregation(expression="avg(trace.output_tokens)"), Aggregation(expression="count(trace.trace_id)")],
+    )
+    resp = make_query_request(signoz, token, start_ms, end_ms, [multi.to_dict()], request_type=RequestType.SCALAR)
+    assert resp.status_code == HTTPStatus.OK, resp.text
+    data = get_scalar_table_data(resp.json())
+    assert len(data) == 1 and [float(v) for v in data[0]] == [pytest.approx(200), 2], data
+
+
+def test_ai_scalar_trace_level_filter_qualifies_traces(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """`trace.output_tokens > 100` qualifies whole traces before aggregation: with
+    out-tokens 100/300 only the 300 trace survives, on both aggregation domains."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-qualify"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=100) + ai_trace(now=now, service=service, in_tokens=30, out_tokens=300))
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    for expression in (
+        "sum(trace.output_tokens)",  # native trace-domain path
+        "sum(gen_ai.usage.output_tokens)",  # delegated span-domain path (__trace_scope)
+    ):
+        got = scalar_value(signoz, token, start_ms, end_ms, service, expression, filter_extra="trace.output_tokens > 100")
+        assert got == pytest.approx(300), expression
+
+    # the qualification also constrains delegated (span-domain) time series
+    ts = BuilderQuery(
+        signal="traces",
+        query_type="builder_ai_query",
+        name="A",
+        filter_expression=f"service.name = '{service}' AND trace.output_tokens > 100",
+        aggregations=[Aggregation(expression="sum(gen_ai.usage.output_tokens)")],
+        step_interval=60,
+    )
+    resp = make_query_request(signoz, token, start_ms, end_ms, [ts.to_dict()], request_type=RequestType.TIME_SERIES)
+    assert resp.status_code == HTTPStatus.OK, resp.text
+    assert series_values(resp.json()) == [[pytest.approx(300)]]
+
+
+def test_ai_scalar_group_by_model(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """Trace-level aggregation grouped by a span attribute: per-model avg of per-trace tokens."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-groupby"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=100, model="gpt-4o") + ai_trace(now=now, service=service, in_tokens=10, out_tokens=300, model="gpt-4o") + ai_trace(now=now, service=service, in_tokens=10, out_tokens=50, model="gpt-4o-mini"))
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    resp = make_query_request(
+        signoz,
+        token,
+        start_ms,
+        end_ms,
+        [scalar_query(service, "avg(trace.output_tokens)", group_by=[TelemetryFieldKey(name="gen_ai.request.model")])],
+        request_type=RequestType.SCALAR,
+    )
+    assert resp.status_code == HTTPStatus.OK, resp.text
+    data = get_scalar_table_data(resp.json())
+    by_model = {row[0]: float(row[-1]) for row in data}
+    assert by_model == {"gpt-4o": pytest.approx(200), "gpt-4o-mini": pytest.approx(50)}, data
+
+
+def test_ai_timeseries_trace_level_aggregation(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """Time-series over per-trace values: all spans fall in one step bucket, avg=200."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-ts"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=100) + ai_trace(now=now, service=service, in_tokens=30, out_tokens=300))
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    query = BuilderQuery(
+        signal="traces",
+        query_type="builder_ai_query",
+        name="A",
+        filter_expression=f"service.name = '{service}'",
+        aggregations=[Aggregation(expression="avg(trace.output_tokens)")],
+        step_interval=60,
+    )
+    resp = make_query_request(signoz, token, start_ms, end_ms, [query.to_dict()], request_type=RequestType.TIME_SERIES)
+    assert resp.status_code == HTTPStatus.OK, resp.text
+    assert series_values(resp.json()) == [[pytest.approx(200)]]
+
+
+def test_ai_timeseries_top_n_groups(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """Grouped, limited time series ranks groups on whole-window per-trace values in
+    the requested order: gpt-4o sums to 400 vs gpt-4o-mini's 50, so limit=1 keeps
+    gpt-4o for the default/desc ranking and gpt-4o-mini when ranking asc."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-topn"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=300, model="gpt-4o") + ai_trace(now=now, service=service, in_tokens=10, out_tokens=100, model="gpt-4o") + ai_trace(now=now, service=service, in_tokens=10, out_tokens=50, model="gpt-4o-mini"))
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    def top_series(order: list[OrderBy] | None) -> dict:
+        query = BuilderQuery(
+            signal="traces",
+            query_type="builder_ai_query",
+            name="A",
+            filter_expression=f"service.name = '{service}'",
+            aggregations=[Aggregation(expression="sum(trace.output_tokens)", alias="total_out")],
+            group_by=[TelemetryFieldKey(name="gen_ai.request.model")],
+            order=order,
+            step_interval=60,
+            limit=1,
+        )
+        resp = make_query_request(signoz, token, start_ms, end_ms, [query.to_dict()], request_type=RequestType.TIME_SERIES)
+        assert resp.status_code == HTTPStatus.OK, resp.text
+        series = resp.json()["data"]["data"]["results"][0]["aggregations"][0]["series"]
+        assert len(series) == 1, f"limit=1 must keep exactly one group, got {len(series)} series"
+        return series[0]
+
+    top = top_series(None)  # default ranking: first aggregation desc
+    assert top["labels"][0]["value"] == "gpt-4o", top["labels"]
+    assert [v["value"] for v in top["values"]] == [pytest.approx(400)]
+
+    bottom = top_series([OrderBy(key=TelemetryFieldKey(name="total_out"), direction="asc")])
+    assert bottom["labels"][0]["value"] == "gpt-4o-mini", bottom["labels"]
+    assert [v["value"] for v in bottom["values"]] == [pytest.approx(50)]
+
+
+def test_ai_timeseries_limit_without_group_by(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """A time-series limit without group-by has nothing to rank and is ignored:
+    the single series comes back complete."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-limit-nogroup"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=100) + ai_trace(now=now, service=service, in_tokens=10, out_tokens=300))
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    query = BuilderQuery(
+        signal="traces",
+        query_type="builder_ai_query",
+        name="A",
+        filter_expression=f"service.name = '{service}'",
+        aggregations=[Aggregation(expression="avg(trace.output_tokens)")],
+        step_interval=60,
+        limit=1,
+    )
+    resp = make_query_request(signoz, token, start_ms, end_ms, [query.to_dict()], request_type=RequestType.TIME_SERIES)
+    assert resp.status_code == HTTPStatus.OK, resp.text
+    assert series_values(resp.json()) == [[pytest.approx(200)]]
+
+
+def test_ai_scalar_group_order_limit(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """Scalar limit is a plain top-N over the grouped rows: sums 400/50/10 with
+    order by the aggregation alias desc and limit=2 keep the two largest models."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-scalar-limit"
+    insert_traces(
+        ai_trace(now=now, service=service, in_tokens=10, out_tokens=300, model="gpt-4o")
+        + ai_trace(now=now, service=service, in_tokens=10, out_tokens=100, model="gpt-4o")
+        + ai_trace(now=now, service=service, in_tokens=10, out_tokens=50, model="gpt-4o-mini")
+        + ai_trace(now=now, service=service, in_tokens=10, out_tokens=10, model="gpt-4")
+    )
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    resp = make_query_request(
+        signoz,
+        token,
+        start_ms,
+        end_ms,
+        [
+            scalar_query(
+                service,
+                "sum(trace.output_tokens)",
+                group_by=[TelemetryFieldKey(name="gen_ai.request.model")],
+                alias="total_out",
+                order=[OrderBy(key=TelemetryFieldKey(name="total_out"), direction="desc")],
+                limit=2,
+            )
+        ],
+        request_type=RequestType.SCALAR,
+    )
+    assert resp.status_code == HTTPStatus.OK, resp.text
+    data = get_scalar_table_data(resp.json())
+    assert [(row[0], float(row[-1])) for row in data] == [("gpt-4o", pytest.approx(400)), ("gpt-4o-mini", pytest.approx(50))], data
+
+
+def test_ai_timeseries_span_time_bucketing(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """Per-trace values are clipped per (bucket, trace): two LLM calls two minutes
+    apart contribute each call's tokens to its own bucket, not the total to both."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-buckets"
+
+    trace_id = TraceIdGenerator.trace_id()
+    root_id = TraceIdGenerator.span_id()
+    resources = {"service.name": service}
+
+    def llm(offset_s: float, out_tokens: int) -> Traces:
+        return Traces(
+            timestamp=now - timedelta(seconds=offset_s),
+            duration=timedelta(seconds=1),
+            trace_id=trace_id,
+            span_id=TraceIdGenerator.span_id(),
+            parent_span_id=root_id,
+            name="chat",
+            kind=TracesKind.SPAN_KIND_CLIENT,
+            status_code=TracesStatusCode.STATUS_CODE_OK,
+            resources=resources,
+            attributes={"gen_ai.request.model": "gpt-4o-mini", "gen_ai.usage.output_tokens": out_tokens},
+        )
+
+    root = Traces(
+        timestamp=now - timedelta(seconds=130),
+        duration=timedelta(seconds=130),
+        trace_id=trace_id,
+        span_id=root_id,
+        parent_span_id="",
+        name="POST /api/chat",
+        kind=TracesKind.SPAN_KIND_SERVER,
+        status_code=TracesStatusCode.STATUS_CODE_OK,
+        resources=resources,
+        attributes={"http.request.method": "POST"},
+    )
+    insert_traces([root, llm(124, 100), llm(4, 300)])
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    query = BuilderQuery(
+        signal="traces",
+        query_type="builder_ai_query",
+        name="A",
+        filter_expression=f"service.name = '{service}'",
+        aggregations=[Aggregation(expression="avg(trace.output_tokens)")],
+        step_interval=60,
+    )
+    resp = make_query_request(signoz, token, start_ms, end_ms, [query.to_dict()], request_type=RequestType.TIME_SERIES)
+    assert resp.status_code == HTTPStatus.OK, resp.text
+
+    values = series_values(resp.json())
+    assert len(values) == 1, values
+    assert sorted(values[0]) == [pytest.approx(100), pytest.approx(300)], f"each call's tokens in its own bucket: {values}"
+
+
+def test_ai_scalar_variables_in_trace_level_filter(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """Variables resolve inside trace-level conditions with span-filter semantics;
+    an unresolvable $var is a 400 (today via aggregate validation — a targeted
+    unknown-variable error is a separate concern)."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-vars"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=100) + ai_trace(now=now, service=service, in_tokens=30, out_tokens=300))
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+    query = scalar_query(service, "sum(trace.output_tokens)", filter_extra="trace.output_tokens > $threshold")
+
+    resp = make_query_request(
+        signoz,
+        token,
+        start_ms,
+        end_ms,
+        [query],
+        request_type=RequestType.SCALAR,
+        variables={"threshold": {"type": "text", "value": 100}},
+    )
+    assert resp.status_code == HTTPStatus.OK, resp.text
+    data = get_scalar_table_data(resp.json())
+    assert len(data) == 1 and float(data[0][-1]) == pytest.approx(300), data
+
+    resp = make_query_request(signoz, token, start_ms, end_ms, [query], request_type=RequestType.SCALAR)
+    assert resp.status_code == HTTPStatus.BAD_REQUEST, resp.text
+    # quotes in the message are JSON-escaped, so match the halves separately
+    assert "$threshold" in resp.text and "cannot be used in a trace-level filter" in resp.text, resp.text
+
+    # a dynamic variable resolved to __all__ drops the condition (both traces count)
+    resp = make_query_request(
+        signoz,
+        token,
+        start_ms,
+        end_ms,
+        [query],
+        request_type=RequestType.SCALAR,
+        variables={"threshold": {"type": "dynamic", "value": "__all__"}},
+    )
+    assert resp.status_code == HTTPStatus.OK, resp.text
+    data = get_scalar_table_data(resp.json())
+    assert len(data) == 1 and float(data[0][-1]) == pytest.approx(400), data
+
+
+def test_ai_scalar_tool_only_trace_null_semantics(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """A tool-only trace (in the gate, no LLM span) follows plain SQL NULL semantics:
+    count(trace.trace_id) counts it (consistent with the trace list), avg over its
+    NULL tokens skips it, and `trace.llm_call_count > 0` is the explicit opt-out."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-toolonly"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=100) + tool_only_trace(now=now, service=service))
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    def value(expression: str, filter_extra: str = "") -> float:
+        return scalar_value(signoz, token, start_ms, end_ms, service, expression, filter_extra)
+
+    assert value("count(trace.trace_id)") == 2, "tool-only trace is an AI trace and must be counted"
+    assert value("avg(trace.output_tokens)") == pytest.approx(100), "NULL tokens are skipped by avg"
+    assert value("avg(trace.tool_call_count)") == pytest.approx(0.5), "tool-only trace feeds tool aggregates (1 and 0 calls)"
+    assert value("count()") == 2, "span-level count sees the LLM and the tool span"
+
+    # filtering on LLM activity is explicit, not implicit
+    assert value("count(trace.trace_id)", filter_extra="trace.llm_call_count > 0") == 1
+
+
+def test_ai_scalar_having_on_aggregation(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """The outer having filters aggregation results per group (by alias)."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-having"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=300, model="gpt-4o") + ai_trace(now=now, service=service, in_tokens=10, out_tokens=50, model="gpt-4o-mini"))
+
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    resp = make_query_request(
+        signoz,
+        token,
+        start_ms,
+        end_ms,
+        [
+            scalar_query(
+                service,
+                "avg(trace.output_tokens)",
+                group_by=[TelemetryFieldKey(name="gen_ai.request.model")],
+                alias="avg_out",
+                having="avg_out > 100",
+            )
+        ],
+        request_type=RequestType.SCALAR,
+    )
+    assert resp.status_code == HTTPStatus.OK, resp.text
+    data = get_scalar_table_data(resp.json())
+    assert len(data) == 1 and data[0][0] == "gpt-4o", data
+
+
+def test_ai_aggregation_rejections(
+    signoz: types.SigNoz,
+    create_user_admin: None,  # pylint: disable=unused-argument
+    get_token: Callable[[str, str], str],
+    insert_traces: Callable[[list[Traces]], None],
+) -> None:
+    """Targeted 400s: mixed domains and group-by on a trace column come from the
+    builder; order-by is stopped earlier by request validation (only group keys and
+    aggregation aliases/expressions are admitted)."""
+    now = datetime.now(tz=UTC).replace(second=0, microsecond=0)
+    service = "ai-it-agg-reject"
+    insert_traces(ai_trace(now=now, service=service, in_tokens=10, out_tokens=100))
+    token = get_token(USER_ADMIN_EMAIL, USER_ADMIN_PASSWORD)
+    start_ms, end_ms = query_window(now)
+
+    def expect_bad_request(query: dict, message: str) -> None:
+        resp = make_query_request(signoz, token, start_ms, end_ms, [query], request_type=RequestType.SCALAR)
+        assert resp.status_code == HTTPStatus.BAD_REQUEST, resp.text
+        assert message in resp.text, resp.text
+
+    # span-level and trace-level aggregations cannot be mixed in one query
+    mixed = BuilderQuery(
+        signal="traces",
+        query_type="builder_ai_query",
+        name="A",
+        filter_expression=f"service.name = '{service}'",
+        aggregations=[Aggregation(expression="avg(trace.output_tokens)"), Aggregation(expression="count()")],
+    )
+    expect_bad_request(mixed.to_dict(), "cannot be mixed")
+
+    expect_bad_request(
+        scalar_query(service, "avg(trace.output_tokens)", group_by=[TelemetryFieldKey(name="trace.llm_call_count")]),
+        "grouping by trace-level aggregate",
+    )
+
+    expect_bad_request(
+        scalar_query(service, "avg(trace.output_tokens)", order=[OrderBy(key=TelemetryFieldKey(name="trace.total_tokens"), direction="desc")]),
+        "invalid order by key",
+    )

tests/integration/tests/querierai/conftest.py

@@ -0,0 +1,34 @@
+import pytest
+from testcontainers.core.container import Network
+
+from fixtures import types
+from fixtures.signoz import create_signoz
+
+
+@pytest.fixture(name="signoz", scope="package")
+def signoz_ai_observability(
+    network: Network,
+    migrator: types.Operation,  # pylint: disable=unused-argument
+    zeus: types.TestContainerDocker,
+    gateway: types.TestContainerDocker,
+    sqlstore: types.TestContainerSQL,
+    clickhouse: types.TestContainerClickhouse,
+    request: pytest.FixtureRequest,
+    pytestconfig: pytest.Config,
+) -> types.SigNoz:
+    """Package-scoped SigNoz with AI observability enabled: the flag gates the static
+    gen_ai key definitions (enrichWithGenAIKeys) — without it the gate keys only
+    resolve once a span carrying them has been ingested."""
+    return create_signoz(
+        network=network,
+        zeus=zeus,
+        gateway=gateway,
+        sqlstore=sqlstore,
+        clickhouse=clickhouse,
+        request=request,
+        pytestconfig=pytestconfig,
+        cache_key="signoz-ai-observability",
+        env_overrides={
+            "SIGNOZ_FLAGGER_CONFIG_BOOLEAN_ENABLE__AI__OBSERVABILITY": True,
+        },
+    )