fix(insights): preserve business priorities through investigation selection

SPEC.md

@@ -58,7 +58,10 @@ regressions retain priority even when the model selects none. Original measureme
 constraints and the unverified planning rationale stay in the frozen objective.
 Scheduled runs investigate at most two; a deliberate manual full scan investigates at
 most five and covers a distinct eligible specialist family before taking extra work from
-one family. The portfolio is diversified across correlated subjects and survives a
+one family. This does not reintroduce optional general work excluded by business-aware
+selection. Candidate input is bounded by serialized size rather than a count cutoff;
+the complete saved brief and newest relevant correction survive source budgeting, or
+selection retains the conservative fallback. The portfolio is diversified across correlated subjects and survives a
 retry unchanged. Each selected signal still gets its own exact agent turn, durable
 observation, and investigation history; a model does not manufacture a broad report
 from ungrounded raw data.

apps/insights/src/business-aware-selection.test.ts

@@ -4,6 +4,7 @@ import type { BusinessContext } from "@databuddy/ai/lib/business-context";
 import { MockLanguageModelV3 } from "ai/test";
 import { chooseInvestigationSignals } from "./business-aware-selection";
 import { planCoveragePortfolio } from "./coverage-planner";
+import { organizationProfileContext } from "./business-context";
 import type { DetectedSignal } from "./detection";
 import {
 	investigateWebsitePortfolioWithSources,
@@ -317,11 +318,6 @@ describe("business-aware investigation selection", () => {
 			[],
 			[outcome],
 			[traffic, { ...outcome, definitionEvidence: "x".repeat(64_001) }],
-			Array.from({ length: 33 }, (_, index) => ({
-				...outcome,
-				metric: `goal:${index}`,
-				subjectKey: `goal:${index}`,
-			})),
 		]) {
 			await planInvestigationsWithBusinessContext(
 				input,
@@ -335,23 +331,50 @@ describe("business-aware investigation selection", () => {
 				scope
 			);
 		}
-		await chooseInvestigationSignals(
-			{
-				businessContext: context,
-				candidates: Array.from({ length: 9 }, (_, index) => ({
-					signal: prepareInvestigation(
-						{
-							...outcome,
-							metric: `goal:${index}`,
-							subjectKey: `goal:${index}`,
-						},
-						7
-					).signal,
-				})),
-				limit: 2,
-			},
-			model
-		);
+		expect(model.doGenerateCalls).toHaveLength(0);
+	});
+
+	it.each([9, 24, 33])("uses business context for %i bounded candidates", async (count) => {
+		const key = `goal:${count - 1}`;
+		const model = new MockLanguageModelV3({ doGenerate: async (request) => {
+			expect(JSON.stringify(request.prompt)).toContain(explanation);
+			return response({ selections: [{ ...choice, signalKey: key }] });
+		} });
+		const plan = await planInvestigationsWithBusinessContext(input,
+			Array.from({ length: count }, (_, index) => ({ ...outcome, metric: `goal:${index}`, subjectKey: `goal:${index}` })),
+			{ loadBusinessProfile: async () => context, selectCandidates: params => chooseInvestigationSignals(params, model) },
+			false, scope, { reason: "scheduled" });
+		expect(plan.map(candidate => candidate.signal.signalKey)).toEqual([key]);
+		expect(model.doGenerateCalls).toHaveLength(1);
+	});
+
+	it("keeps the complete maximum saved brief and a current correction before optional background", async () => {
+		const saved = organizationProfileContext({
+			content: "Business overview. ".padEnd(11_940, " Background.") + " Final exclusion: generic traffic is already explained.",
+			origin: "mixed", sources: [{ url: "https://example.com/", title: "Public overview" }],
+			revision: 4, updatedAt: "2026-07-11T00:00:00.000Z", updatedBy: "example-editor", sourceWebsiteId: null,
+			teamContext: { priority: "Prioritize delivery. ".padEnd(2000, " Priority."), successDefinition: "Delivery means accepted by the recipient. ".padEnd(2000, " Definition."), exclusions: "Exclude demos. ".padEnd(2000, " Exclusion.") },
+		}, input.organizationId, new Date(input.asOf));
+		const correction = { id: "current-correction", kind: "team_reply" as const, subjectKey: choice.signalKey,
+			content: "Current correction: use accepted delivery. ".padEnd(3950, " Team details.") + " Correction ends here.", observedAt: input.asOf };
+		saved.sources.push(correction);
+		const model = new MockLanguageModelV3({ doGenerate: async (request) => {
+			const sent = JSON.stringify(request.prompt);
+			expect(sent).toContain("Business overview.");
+			expect(sent).toContain("Final exclusion: generic traffic is already explained.");
+			expect(sent).toContain("Current correction: use accepted delivery.");
+			expect(sent).toContain("Correction ends here.");
+			expect(sent).toContain("Exclude demos.");
+			return response({ selections: [choice] });
+		} });
+		const result = await chooseInvestigationSignals({ businessContext: saved, candidates: [traffic, outcome].map(signal => ({ signal: prepareInvestigation(signal, 7).signal })), limit: 2 }, model);
+		expect(result?.output.selections).toEqual([choice]);
+	});
+
+	it("falls back instead of selecting from an incomplete over-budget saved document", async () => {
+		const model = new MockLanguageModelV3({ doGenerate: async () => { throw new Error("Selection should be skipped"); } });
+		const saved = organizationProfileContext({ content: "\u0000".repeat(12_000), sources: [], origin: "team", revision: 1, updatedAt: input.asOf, updatedBy: "example-editor", sourceWebsiteId: null }, input.organizationId, new Date(input.asOf));
+		expect(await chooseInvestigationSignals({ businessContext: saved, candidates: [traffic, outcome].map(signal => ({ signal: prepareInvestigation(signal, 7).signal })), limit: 2 }, model)).toBeNull();
 		expect(model.doGenerateCalls).toHaveLength(0);
 	});
 

apps/insights/src/business-aware-selection.ts

@@ -45,7 +45,6 @@ export async function chooseInvestigationSignals(
 	if (
 		!(model || isAiGatewayConfigured) ||
 		candidates.length <= 1 ||
-		candidates.length > input.limit * 4 ||
 		!businessContext.sources.length ||
 		!["ready", "partial"].includes(businessContext.status) ||
 		JSON.stringify(candidates).length > 48_000
@@ -69,54 +68,47 @@ export async function chooseInvestigationSignals(
 		),
 		...replies,
 		...businessContext.sources.filter((source) => source.kind === "website"),
-	];
-	const sources: Pick<
-		BusinessContext["sources"][number],
-		| "id"
-		| "kind"
-		| "content"
-		| "observedAt"
-		| "subjectKey"
-		| "author"
-		| "url"
-		| "references"
-		| "origin"
-	>[] = [];
-	let pageCharacters = 0;
-	let characters = 0;
-	for (const {
-		id,
-		kind,
-		content,
-		observedAt,
-		subjectKey,
-		author,
-		origin,
-		url,
-		references,
-	} of ordered) {
-		const source = {
+	].map(
+		({ id, kind, content, observedAt, subjectKey, author, origin, url }) => ({
 			id,
 			kind,
-			references,
 			content,
 			observedAt,
 			subjectKey,
 			author,
 			origin,
 			url,
-		};
+		})
+	);
+	// Keep the complete saved document and the newest relevant correction.
+	// Bibliography stays on the investigation snapshot; it is not needed to rank work.
+	const characterLimit = Math.max(
+		18_000,
+		ordered
+			.filter(
+				(source) =>
+					source.kind === "organization_profile" || source.id === replies[0]?.id
+			)
+			.reduce((total, source) => total + JSON.stringify(source).length, 0)
+	);
+	if (characterLimit > 32_000) {
+		return null;
+	}
+	const sources: typeof ordered = [];
+	let pageCharacters = 0;
+	let characters = 0;
+	for (const source of ordered) {
 		const size = JSON.stringify(source).length;
 		if (
-			sources.some((item) => item.id === id) ||
-			characters + size > 18_000 ||
-			(kind === "website" && pageCharacters + size > 8000)
+			sources.some((item) => item.id === source.id) ||
+			characters + size > characterLimit ||
+			(source.kind === "website" && pageCharacters + size > 8000)
 		) {
 			continue;
 		}
 		sources.push(source);
 		characters += size;
-		if (kind === "website") {
+		if (source.kind === "website") {
 			pageCharacters += size;
 		}
 	}

apps/insights/src/coverage-planner.test.ts

@@ -303,12 +303,19 @@ describe("business preference constraints", () => {
 		expect(plan).toHaveLength(5);
 		expect(plan).toContain(error);
 		expect(plan).toContain(funnel);
-		expect(plan).toContain(traffic);
+		expect(plan).not.toContain(traffic);
 		expect(plan.filter((item) => item.metric.startsWith("goal:"))).toHaveLength(
-			2
+			3
 		);
 	});
 
+	it.each(["manual", "scheduled"] as const)("preserves an explicit traffic exclusion in a %s scan", (reason) => {
+		const goal = signal({ metric: "goal:activation", subjectKey: "goal:activation" });
+		const traffic = signal({ metric: "visitors" });
+		expect(planCoveragePortfolio([traffic, goal], { reason, selectedSignalKeys: keys([goal]) })).toEqual([goal]);
+		expect(planCoveragePortfolio([traffic, goal], { reason, selectedSignalKeys: keys([traffic]) })).toContain(traffic);
+	});
+
 	it("keeps one correlated subject, the due case first, and the scheduled limit", () => {
 		const due = signal({ metric: "goal:due", subjectKey: "goal:due" });
 		const visitors = signal({ metric: "visitors" });

apps/insights/src/coverage-planner.ts

@@ -199,7 +199,9 @@ export function planCoveragePortfolio(
 					(candidate) =>
 						selection.has(candidate.key) ||
 						isCriticalReliabilitySignal(candidate.signal) ||
-						(options.reason === "manual" && !usedFamilies.has(candidate.family))
+						(options.reason === "manual" &&
+							candidate.family !== "general" &&
+							!usedFamilies.has(candidate.family))
 				)
 			: available;
 		const preferred =

apps/insights/src/evals/README.md

@@ -1,5 +1,13 @@
 # Investigation quality evals
 
+`context-selection.ts --out <fresh-directory> --runs 2` compares native absent/present
+context paths before running selected investigations through this evaluator. It uses
+synthetic 2-, 9- and 24-signal portfolios, a maximum-sized context correction case and
+manual exclusion coverage. `--reverse` reverses candidate order for holdouts; `--cases`
+selects scenario IDs. Alternate arms, preserve the copied source and fixtures, and
+review the complete outputs as well as final selections. Its zero exit status means
+the run completed; `results.json` retains quality failures for manual comparison.
+
 Run from the repository root with `AI_GATEWAY_API_KEY` configured:
 
 ```sh