Make synthesis context/response length admin-configurable; give recaps their own writing style and let them run full-length

Adds a "Context window" panel to the Models tab: num_ctx and
num_predict sliders, bounded by the selected synthesis model's own
detected max context (via Ollama's /api/show, GET
/api/admin/model-context) when available, falling back to a generous
default otherwise. These now drive every synthesis/recap/tag-extraction
call instead of the old fixed 8192/700 constants — too low a
num_predict is exactly why a long recap or merge sometimes cut off
mid-sentence.

Recaps also get their own per-tracked-item writing style, independent
of the global Merge-tab style: a "More" collapsible section (minimized
by default) next to each tracked item's recap cadence, with the same
preset + free-text pattern as the Merge tab. The recap system prompt no
longer caps output at "3-5 short paragraphs" — it now asks for a full,
comprehensive article sized to the material, which only works well
together with a properly-sized num_predict.

backend/src/api/admin.ts

@@ -208,6 +208,20 @@ export async function registerAdminRoutes(app: FastifyInstance) {
 		return { connected, host: settings.aiServiceHost, port: settings.aiServicePort, ramGB: null, gpu: null };
 	});
 
+	// Detects the selected synthesis model's own max context length (when Ollama exposes
+	// it) so the Models tab's num_ctx/num_predict sliders can be bounded by what the
+	// model actually supports, instead of an arbitrary fixed cap. contextLength is null
+	// when undetectable (older Ollama version, unusual model format, unreachable) — the
+	// frontend falls back to a generous default range in that case rather than blocking.
+	app.get('/api/admin/model-context', async (req, reply) => {
+		const { model } = req.query as { model?: string };
+		if (!model) return reply.code(400).send({ error: 'model query param required' });
+		const settings = settingsDb.getSettings();
+		const provider = new OllamaProvider(settings.aiServiceHost, settings.aiServicePort);
+		const contextLength = await provider.getModelContextLength(model);
+		return { contextLength };
+	});
+
 	// --- Telegram account (Connections tab — see telegram/client.ts and credentials.ts.
 	// API ID/hash and the resulting login session are stored encrypted at rest; none of
 	// these routes ever echo them back to the client.) ---

backend/src/inference/ollama-provider.ts

@@ -135,4 +135,32 @@ export class OllamaProvider implements InferenceProvider {
 			return false;
 		}
 	}
+
+	/**
+	 * Ollama's /api/show returns a model_info object whose keys are prefixed by the
+	 * model's own architecture name (e.g. "qwen2.context_length", "llama.context_length")
+	 * rather than one fixed field — there's no single stable key across model families.
+	 * Scanning for whichever key ends in ".context_length" avoids hardcoding a list of
+	 * known architectures that will inevitably miss a future/uncommon one. Returns null
+	 * (rather than throwing) on any failure — the admin-facing slider falls back to a
+	 * generous default cap when this can't be determined, rather than blocking the whole
+	 * Models tab on one unreliable, best-effort lookup.
+	 */
+	async getModelContextLength(model: string): Promise<number | null> {
+		try {
+			const res = await fetch(`${this.base()}/api/show`, {
+				method: 'POST',
+				headers: { 'Content-Type': 'application/json' },
+				body: JSON.stringify({ model, name: model }),
+				signal: AbortSignal.timeout(5000)
+			});
+			if (!res.ok) return null;
+			const data = (await res.json()) as { model_info?: Record<string, unknown> };
+			const entry = Object.entries(data.model_info ?? {}).find(([key]) => key.endsWith('.context_length'));
+			const value = entry?.[1];
+			return typeof value === 'number' && value > 0 ? value : null;
+		} catch {
+			return null;
+		}
+	}
 }

backend/src/inference/provider.ts

@@ -6,4 +6,6 @@ export interface InferenceProvider {
 	embed(text: string, opts?: { model?: string }): Promise<number[]>;
 	listModels(): Promise<string[]>;
 	isReachable(): Promise<boolean>;
+	/** The model's own reported max context length (training/architecture limit), or null if the server doesn't expose it — used to bound the admin-facing num_ctx slider (Models tab) so it can't be set past what the model actually supports. */
+	getModelContextLength(model: string): Promise<number | null>;
 }

backend/src/pipeline/publish.ts

@@ -327,7 +327,7 @@ export async function publishDirect(
 	let tagIds: string[] = [];
 	if (opts.provider) {
 		try {
-			const tagLabels = await extractTags(opts.provider, settings.selectedModels.synthesis, item);
+			const tagLabels = await extractTags(opts.provider, settings.selectedModels.synthesis, item, settings);
 			tagIds = await resolveTagIds(opts.provider, tagLabels, settings, 'synthesis');
 		} catch (err) {
 			logger.error('synthesis', `Tag extraction failed for "${item.title}": ${(err as Error).message}`);
@@ -475,7 +475,7 @@ export async function publishEventRecap(
 	event: TrackedEvent,
 	constituents: MergedArticle[]
 ): Promise<MergedArticle> {
-	const { title, body, tagLabels } = await synthesizeRecap(provider, settings.selectedModels.synthesis, event.name, constituents, settings);
+	const { title, body, tagLabels } = await synthesizeRecap(provider, settings.selectedModels.synthesis, event, constituents, settings);
 	const tagIds = await resolveTagIds(provider, tagLabels, settings, 'events');
 
 	const category = [...new Set(constituents.flatMap((a) => a.category))];

backend/src/pipeline/synthesis.ts

@@ -1,6 +1,5 @@
 import type { InferenceProvider } from '../inference/provider.js';
-import type { ContentItem, GlobalSettings, MergedArticle } from '../storage/db/types.js';
+import type { ContentItem, GlobalSettings, MergedArticle, TrackedEvent } from '../storage/db/types.js';
-import { DEFAULT_NUM_CTX, DEFAULT_NUM_PREDICT } from '../inference/ollama-provider.js';
 import { logger } from '../storage/db/logs.js';
 
 const TITLE_DELIMITER = '---TITLE---';
@@ -17,17 +16,21 @@ const TAG_DELIMITER_RE = /-{2,}\s*TAGS\s*-{2,}/i;
 // Ollama truncates prompts that don't fit its context window by keeping a small prefix
 // and dropping everything else in the middle — silently, with no error, and with no
 // regard for which sources end up cut (see ollama-provider.ts for the incident that
-// prompted this). Rather than relying on that, prompts here are sized to fit
+// prompted this). Rather than relying on that, prompts here are sized to fit the
-// DEFAULT_NUM_CTX up front: each source/article gets an equal character budget, cut only
+// admin-configured num_ctx/num_predict (Models tab) up front: each source/article gets
-// when the whole prompt would otherwise overflow, so every source stays at least
+// an equal character budget, cut only when the whole prompt would otherwise overflow, so
-// partially represented (and attributable) instead of some being dropped outright.
+// every source stays at least partially represented (and attributable) instead of some
-// ~4 chars/token is a rough heuristic (no tokenizer available here) — good enough for a
+// being dropped outright. ~4 chars/token is a rough heuristic (no tokenizer available
-// safety margin, not meant to be exact.
+// here) — good enough for a safety margin, not meant to be exact.
 const CHARS_PER_TOKEN = 4;
 const RESERVED_OVERHEAD_TOKENS = 300; // system prompt + per-entry headers/formatting
-const MAX_INPUT_CHARS = (DEFAULT_NUM_CTX - DEFAULT_NUM_PREDICT - RESERVED_OVERHEAD_TOKENS) * CHARS_PER_TOKEN;
 const MIN_ENTRY_CHARS = 300; // floor so a huge cluster/recap doesn't shrink every entry to nothing
 
+/** Character budget for prompt *input* — leaves numPredict's worth of the context window free for the model's own response, per the admin's configured num_ctx/num_predict (GlobalSettings.synthesisNumCtx/synthesisNumPredict). */
+function maxInputChars(numCtx: number, numPredict: number): number {
+	return Math.max(0, (numCtx - numPredict - RESERVED_OVERHEAD_TOKENS) * CHARS_PER_TOKEN);
+}
+
 function capEntryText(text: string, budgetChars: number): string {
 	return text.length > budgetChars ? text.slice(0, budgetChars) + '…' : text;
 }
@@ -56,14 +59,15 @@ function parseTagLabels(raw: string): string[] {
 export async function extractTags(
 	provider: InferenceProvider,
 	model: string,
-	item: Pick<ContentItem, 'title' | 'summary' | 'body'>
+	item: Pick<ContentItem, 'title' | 'summary' | 'body'>,
+	settings: GlobalSettings
 ): Promise<string[]> {
-	const summary = capEntryText(item.body || item.summary, MAX_INPUT_CHARS);
+	const summary = capEntryText(item.body || item.summary, maxInputChars(settings.synthesisNumCtx, TAG_EXTRACTION_NUM_PREDICT));
 	const prompt = `Title: ${item.title}\nSummary: ${summary}`;
 	const raw = await provider.generate(prompt, {
 		model,
 		system: TAG_EXTRACTION_SYSTEM_PROMPT,
-		numCtx: DEFAULT_NUM_CTX,
+		numCtx: settings.synthesisNumCtx,
 		numPredict: TAG_EXTRACTION_NUM_PREDICT,
 		label: `Extracting tags: "${item.title.slice(0, 60)}"`
 	});
@@ -73,11 +77,11 @@ export async function extractTags(
 const RECAP_SYSTEM_PROMPT_BASE = `You are a neutral news synthesis assistant. Given a chronological list of articles already published about an ongoing tracked event, write your response in exactly three parts, in this order:
 
 1. A short, specific headline for this recap (a single line, ideally under 12 words, no surrounding quotation marks, no trailing period).
-2. On a new line, write exactly "${TITLE_DELIMITER}", then the recap article:
+2. On a new line, write exactly "${TITLE_DELIMITER}", then the recap:
-   - Summarizes what has happened across the period covered, in chronological order
+   - Write a full, comprehensive news article covering the period — not a short summary or a bare list of bullet points. Use as many paragraphs and as much length as the material actually warrants; do not artificially cut it short.
-   - Highlights the most significant developments rather than restating every article
+   - Organize it in chronological order, but group and connect related developments into a coherent narrative rather than restating each source article one at a time
+   - Give real weight and detail to the most significant developments; minor ones can be covered more briefly, but nothing significant should be dropped for the sake of brevity
    - Stays neutral and factual, without editorializing
-   - Is 3-5 short paragraphs
 3. On a new line after the recap, write exactly "${TAG_DELIMITER}" followed by 2-4 short comma-separated topic/entity tags (e.g. proper nouns, named events) that this recap is about. If nothing salient qualifies, leave the tag line empty.`;
 
 const SYSTEM_PROMPT_BASE = `You are a neutral news synthesis assistant. Given summaries from multiple news sources describing the same event, write your response in exactly three parts, in this order:
@@ -99,7 +103,7 @@ const STYLE_PRESETS: Record<GlobalSettings['synthesisStylePreset'], string> = {
 	formal: 'Write in a formal, measured register — precise language, no contractions, no colloquialisms.'
 };
 
-/** Admin-configurable tone: a preset plus optional free-text instructions, both from GlobalSettings — the only two knobs that affect HOW the model writes, as opposed to WHAT gets clustered/published. Appended to the base prompt, never replacing its structural rules (attribution, paragraph count, tag format). */
+/** Admin-configurable tone: a preset plus optional free-text instructions, both from GlobalSettings — the only two knobs that affect HOW the model writes, as opposed to WHAT gets clustered/published. Appended to the base prompt, never replacing its structural rules (attribution, paragraph count, tag format). Applies only to regular same-story merges — recaps have their own independent style knob, see recapStyleAddendum below. */
 function styleAddendum(settings: GlobalSettings): string {
 	const preset = STYLE_PRESETS[settings.synthesisStylePreset] ?? '';
 	const custom = settings.synthesisCustomInstructions.trim();
@@ -108,6 +112,21 @@ function styleAddendum(settings: GlobalSettings): string {
 	return `\n\nAdditional style instructions from the site admin (follow these without breaking the rules above):\n${lines.join('\n')}`;
 }
 
+/**
+ * Per-tracked-item recap style — deliberately independent of the global synthesisStylePreset
+ * above (set in the Merge tab), since a recap's tone/scope is a very different kind of
+ * knob: it's set once per tracked item (the "More" section on its own edit panel, next to
+ * its recap cadence), not globally for every merge on the site. Reuses the same preset
+ * strings for consistency, but reads from the event's own fields instead of GlobalSettings.
+ */
+function recapStyleAddendum(event: TrackedEvent): string {
+	const preset = STYLE_PRESETS[event.recapStylePreset] ?? '';
+	const custom = event.recapCustomInstructions.trim();
+	const lines = [preset, custom].filter(Boolean);
+	if (lines.length === 0) return '';
+	return `\n\nAdditional style instructions from the site admin for this recap (follow these without breaking the rules above):\n${lines.join('\n')}`;
+}
+
 export interface SynthesisResult {
 	title: string;
 	body: string;
@@ -120,8 +139,8 @@ function fallbackTitle(body: string): string {
 	return firstLine.length > 100 ? firstLine.slice(0, 97) + '…' : firstLine;
 }
 
-function buildPrompt(items: ContentItem[], sourceNames: Map<string, string>): string {
+function buildPrompt(items: ContentItem[], sourceNames: Map<string, string>, numCtx: number, numPredict: number): string {
-	const budgetPerItem = Math.max(MIN_ENTRY_CHARS, Math.floor(MAX_INPUT_CHARS / items.length));
+	const budgetPerItem = Math.max(MIN_ENTRY_CHARS, Math.floor(maxInputChars(numCtx, numPredict) / items.length));
 	let truncated = 0;
 	const entries = items.map((item, i) => {
 		// Same fallback publishDirect uses (publish.ts) — body is the full article text
@@ -187,15 +206,16 @@ export async function synthesizeArticle(
 	sourceNames: Map<string, string>,
 	settings: GlobalSettings
 ): Promise<SynthesisResult> {
-	const prompt = buildPrompt(items, sourceNames);
+	const { synthesisNumCtx: numCtx, synthesisNumPredict: numPredict } = settings;
+	const prompt = buildPrompt(items, sourceNames, numCtx, numPredict);
 	const system = SYSTEM_PROMPT_BASE + styleAddendum(settings);
 	const label = `Merging ${items.length} source${items.length === 1 ? '' : 's'}: "${items[0]?.title.slice(0, 60) ?? ''}"`;
-	const raw = await provider.generate(prompt, { model, system, numCtx: DEFAULT_NUM_CTX, numPredict: DEFAULT_NUM_PREDICT, label });
+	const raw = await provider.generate(prompt, { model, system, numCtx, numPredict, label });
 	return assertNonEmpty(parseResult(raw), `"${items[0]?.title.slice(0, 60) ?? ''}"`);
 }
 
-function buildRecapPrompt(eventName: string, articles: MergedArticle[]): string {
+function buildRecapPrompt(eventName: string, articles: MergedArticle[], numCtx: number, numPredict: number): string {
-	const budgetPerArticle = Math.max(MIN_ENTRY_CHARS, Math.floor(MAX_INPUT_CHARS / articles.length));
+	const budgetPerArticle = Math.max(MIN_ENTRY_CHARS, Math.floor(maxInputChars(numCtx, numPredict) / articles.length));
 	let truncated = 0;
 	const entries = articles.map((article, i) => {
 		const body = capEntryText(article.body, budgetPerArticle);
@@ -219,17 +239,18 @@ function buildRecapPrompt(eventName: string, articles: MergedArticle[]): string
 export async function synthesizeRecap(
 	provider: InferenceProvider,
 	model: string,
-	eventName: string,
+	event: TrackedEvent,
 	articles: MergedArticle[],
 	settings: GlobalSettings
 ): Promise<SynthesisResult> {
-	const prompt = buildRecapPrompt(eventName, articles);
+	const { synthesisNumCtx: numCtx, synthesisNumPredict: numPredict } = settings;
+	const prompt = buildRecapPrompt(event.name, articles, numCtx, numPredict);
 	const raw = await provider.generate(prompt, {
 		model,
-		system: RECAP_SYSTEM_PROMPT_BASE + styleAddendum(settings),
+		system: RECAP_SYSTEM_PROMPT_BASE + recapStyleAddendum(event),
-		numCtx: DEFAULT_NUM_CTX,
+		numCtx,
-		numPredict: DEFAULT_NUM_PREDICT,
+		numPredict,
-		label: `Recapping event: "${eventName.slice(0, 60)}"`
+		label: `Recapping event: "${event.name.slice(0, 60)}"`
 	});
-	return assertNonEmpty(parseResult(raw), `event recap "${eventName.slice(0, 60)}"`);
+	return assertNonEmpty(parseResult(raw), `event recap "${event.name.slice(0, 60)}"`);
 }

backend/src/storage/db/events.ts

@@ -14,6 +14,8 @@ function rowToEvent(row: any): TrackedEvent {
 		isSpillover: !!row.is_spillover,
 		retentionOverrideDays: row.retention_override_days,
 		lastRecapAt: row.last_recap_at,
+		recapStylePreset: row.recap_style_preset,
+		recapCustomInstructions: row.recap_custom_instructions,
 		createdAt: row.created_at
 	};
 }
@@ -52,8 +54,8 @@ export function createEvent(input: Partial<TrackedEvent>): TrackedEvent {
 	const id = `evt-${randomUUID()}`;
 	const now = new Date().toISOString();
 	db.prepare(
-		`INSERT INTO tracked_events (id, name, description, source_ids, keywords, recap_interval_hours, active, is_spillover, retention_override_days, last_recap_at, created_at)
+		`INSERT INTO tracked_events (id, name, description, source_ids, keywords, recap_interval_hours, active, is_spillover, retention_override_days, last_recap_at, recap_style_preset, recap_custom_instructions, created_at)
-		 VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, NULL, ?)`
+		 VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, NULL, ?, ?, ?)`
 	).run(
 		id,
 		input.name ?? 'Untitled event',
@@ -64,6 +66,8 @@ export function createEvent(input: Partial<TrackedEvent>): TrackedEvent {
 		input.active === false ? 0 : 1,
 		input.isSpillover ? 1 : 0,
 		input.retentionOverrideDays ?? null,
+		input.recapStylePreset ?? 'default',
+		input.recapCustomInstructions ?? '',
 		now
 	);
 	return getEvent(id)!;
@@ -74,7 +78,7 @@ export function updateEvent(id: string, patch: Partial<TrackedEvent>): TrackedEv
 	if (!existing) return null;
 	const merged = { ...existing, ...patch };
 	db.prepare(
-		`UPDATE tracked_events SET name=?, description=?, source_ids=?, keywords=?, recap_interval_hours=?, active=?, is_spillover=?, retention_override_days=?, last_recap_at=? WHERE id=?`
+		`UPDATE tracked_events SET name=?, description=?, source_ids=?, keywords=?, recap_interval_hours=?, active=?, is_spillover=?, retention_override_days=?, last_recap_at=?, recap_style_preset=?, recap_custom_instructions=? WHERE id=?`
 	).run(
 		merged.name,
 		merged.description,
@@ -85,6 +89,8 @@ export function updateEvent(id: string, patch: Partial<TrackedEvent>): TrackedEv
 		merged.isSpillover ? 1 : 0,
 		merged.retentionOverrideDays,
 		merged.lastRecapAt,
+		merged.recapStylePreset,
+		merged.recapCustomInstructions,
 		id
 	);
 	return getEvent(id);

backend/src/storage/db/index.ts

@@ -161,6 +161,8 @@ export function migrate() {
 			is_spillover INTEGER NOT NULL DEFAULT 0,
 			retention_override_days INTEGER,
 			last_recap_at TEXT,
+			recap_style_preset TEXT NOT NULL DEFAULT 'default', -- default | casual | formal — this item's own recap tone, independent of the global Merge-tab style (see pipeline/synthesis.ts)
+			recap_custom_instructions TEXT NOT NULL DEFAULT '', -- free-text addendum for this item's recap prompt specifically
 			created_at TEXT NOT NULL
 		);
 
@@ -214,6 +216,8 @@ export function migrate() {
 			telegram_media_mode TEXT NOT NULL DEFAULT 'self-host', -- self-host | proxy (no "direct" — Telegram has no public hotlinkable media URL)
 			synthesis_style_preset TEXT NOT NULL DEFAULT 'default', -- default | casual | formal — see pipeline/synthesis.ts's STYLE_PRESETS
 			synthesis_custom_instructions TEXT NOT NULL DEFAULT '', -- free-text addendum appended to the synthesis system prompt, on top of the preset
+			synthesis_num_ctx INTEGER NOT NULL DEFAULT 8192, -- admin-tunable context window (Models tab) — see inference/ollama-provider.ts's DEFAULT_NUM_CTX
+			synthesis_num_predict INTEGER NOT NULL DEFAULT 700, -- admin-tunable max response length — too low silently truncates output mid-sentence
 			widget_weather_enabled INTEGER NOT NULL DEFAULT 1,
 			widget_stocks_enabled INTEGER NOT NULL DEFAULT 1,
 			widget_bookmarks_enabled INTEGER NOT NULL DEFAULT 1,
@@ -343,6 +347,12 @@ export function migrate() {
 	if (!hasColumn('tracked_events', 'recap_interval_hours')) {
 		db.exec('ALTER TABLE tracked_events ADD COLUMN recap_interval_hours INTEGER');
 	}
+	if (!hasColumn('tracked_events', 'recap_style_preset')) {
+		db.exec("ALTER TABLE tracked_events ADD COLUMN recap_style_preset TEXT NOT NULL DEFAULT 'default'");
+	}
+	if (!hasColumn('tracked_events', 'recap_custom_instructions')) {
+		db.exec("ALTER TABLE tracked_events ADD COLUMN recap_custom_instructions TEXT NOT NULL DEFAULT ''");
+	}
 	if (!hasColumn('merged_articles', 'is_recap')) {
 		db.exec('ALTER TABLE merged_articles ADD COLUMN is_recap INTEGER NOT NULL DEFAULT 0');
 	}
@@ -386,6 +396,12 @@ export function migrate() {
 	if (!hasColumn('global_settings', 'synthesis_custom_instructions')) {
 		db.exec("ALTER TABLE global_settings ADD COLUMN synthesis_custom_instructions TEXT NOT NULL DEFAULT ''");
 	}
+	if (!hasColumn('global_settings', 'synthesis_num_ctx')) {
+		db.exec('ALTER TABLE global_settings ADD COLUMN synthesis_num_ctx INTEGER NOT NULL DEFAULT 8192');
+	}
+	if (!hasColumn('global_settings', 'synthesis_num_predict')) {
+		db.exec('ALTER TABLE global_settings ADD COLUMN synthesis_num_predict INTEGER NOT NULL DEFAULT 700');
+	}
 
 	// Seed default categories if none exist yet. "News" sits right under "Top stories" —
 	// general news sources belong here, not on "Top stories" itself, which isn't a real

backend/src/storage/db/settings.ts

@@ -45,6 +45,8 @@ function rowToSettings(row: any): GlobalSettings {
 		telegramMediaMode: row.telegram_media_mode,
 		synthesisStylePreset: row.synthesis_style_preset,
 		synthesisCustomInstructions: row.synthesis_custom_instructions,
+		synthesisNumCtx: row.synthesis_num_ctx,
+		synthesisNumPredict: row.synthesis_num_predict,
 		...widgetsAndOrder(),
 		retention: {
 			publishedArticleMaxAgeDays: row.published_article_max_age_days,
@@ -93,6 +95,7 @@ export function updateSettings(patch: Partial<GlobalSettings>): GlobalSettings {
 			nitter_media_mode=$nitter_media_mode, fxtwitter_base_url=$fxtwitter_base_url, nitter_instance_url=$nitter_instance_url,
 			telegram_media_mode=$telegram_media_mode,
 			synthesis_style_preset=$synthesis_style_preset, synthesis_custom_instructions=$synthesis_custom_instructions,
+			synthesis_num_ctx=$synthesis_num_ctx, synthesis_num_predict=$synthesis_num_predict,
 			published_article_max_age_days=$published_article_max_age_days, raw_item_max_age_days=$raw_item_max_age_days,
 			storage_cap_enabled=$storage_cap_enabled, storage_cap_value=$storage_cap_value, storage_cap_unit=$storage_cap_unit
 		 WHERE id = 1`
@@ -112,6 +115,8 @@ export function updateSettings(patch: Partial<GlobalSettings>): GlobalSettings {
 		$telegram_media_mode: merged.telegramMediaMode,
 		$synthesis_style_preset: merged.synthesisStylePreset,
 		$synthesis_custom_instructions: merged.synthesisCustomInstructions,
+		$synthesis_num_ctx: merged.synthesisNumCtx,
+		$synthesis_num_predict: merged.synthesisNumPredict,
 		$published_article_max_age_days: merged.retention.publishedArticleMaxAgeDays,
 		$raw_item_max_age_days: merged.retention.rawItemMaxAgeDays,
 		$storage_cap_enabled: merged.retention.storageCapEnabled ? 1 : 0,

backend/src/storage/db/types.ts

@@ -185,6 +185,10 @@ export interface TrackedEvent {
 	isSpillover: boolean;
 	retentionOverrideDays: number | null;
 	lastRecapAt: string | null;
+	/** Tone preset for this item's own recap, independent of the global Merge-tab synthesis style — see pipeline/synthesis.ts's STYLE_PRESETS. 'default' adds nothing on top of the base recap prompt. */
+	recapStylePreset: 'default' | 'casual' | 'formal';
+	/** Free-text instructions appended to the recap system prompt for this item specifically — e.g. "focus on military developments", "write as a full narrative, not bullet points". Empty string means no addendum. */
+	recapCustomInstructions: string;
 	createdAt: string;
 }
 
@@ -291,6 +295,10 @@ export interface GlobalSettings {
 	synthesisStylePreset: 'default' | 'casual' | 'formal';
 	/** Free-text instructions appended to the synthesis system prompt alongside the style preset — e.g. "keep it under 3 sentences per paragraph". Empty string means no addendum. */
 	synthesisCustomInstructions: string;
+	/** Total context window (prompt + response) requested from Ollama for every synthesis/recap/tag-extraction call — see inference/ollama-provider.ts's DEFAULT_NUM_CTX for why this is ever explicit at all, and the Models tab for the admin-facing slider (bounded by the selected synthesis model's own reported max, when Ollama exposes it). */
+	synthesisNumCtx: number;
+	/** Max tokens the model is allowed to generate per synthesis/recap call — too low silently truncates the output mid-sentence rather than erroring (this is what a "cut off" recap/article means). Recaps in particular need real headroom: they're asked to summarize many source articles into several paragraphs, unlike a same-story merge. */
+	synthesisNumPredict: number;
 	/** Per-widget enable flags — see admin/settings' consolidated "Widgets" tab. Weather/Stocks/PoE2's backend pollers (scheduler.ts) are gated on these too, not just sidebar visibility; Bookmarks has no poller so its flag only affects the sidebar. */
 	widgets: {
 		weather: boolean;

frontend/src/lib/adminApi.ts

@@ -17,7 +17,8 @@ import type {
 	AdminWeatherSettings,
 	InstalledWidget,
 	WidgetUploadManifest,
-	PipelineStats
+	PipelineStats,
+	ModelContextInfo
 } from './adminTypes';
 
 async function request<T>(path: string, options: RequestInit = {}, fetchFn: typeof fetch = fetch): Promise<T> {
@@ -142,6 +143,9 @@ export const getModels = (fetchFn?: typeof fetch) =>
 export const getAiStatus = (fetchFn?: typeof fetch) =>
 	request<AiStatus>('/api/admin/ai-status', {}, fetchFn);
 
+export const getModelContext = (model: string, fetchFn?: typeof fetch) =>
+	request<ModelContextInfo>(`/api/admin/model-context?model=${encodeURIComponent(model)}`, {}, fetchFn);
+
 // Telegram account (Connections tab) — API ID/hash and the resulting login session are
 // stored encrypted at rest server-side (see backend telegram/credentials.ts); none of
 // these ever come back from the server, only status flags.

frontend/src/lib/adminTypes.ts

@@ -157,6 +157,10 @@ export interface AdminSettings {
 	telegramMediaMode: 'self-host' | 'proxy';
 	synthesisStylePreset: 'default' | 'casual' | 'formal';
 	synthesisCustomInstructions: string;
+	/** Total context window (prompt + response) requested from Ollama for every synthesis/recap/tag-extraction call. */
+	synthesisNumCtx: number;
+	/** Max tokens the model may generate per call — too low silently truncates output mid-sentence. */
+	synthesisNumPredict: number;
 	widgets: AdminWidgetsEnabled;
 	widgetOrder: ('weather' | 'stocks' | 'bookmarks' | 'poe2')[];
 	retention: RetentionSettings;
@@ -189,6 +193,10 @@ export interface AdminTrackedEvent {
 	active: boolean;
 	isSpillover: boolean;
 	retentionOverrideDays: number | null;
+	/** This item's own recap tone, independent of the global Merge-tab synthesis style. */
+	recapStylePreset: 'default' | 'casual' | 'formal';
+	/** Free-text instructions appended to this item's recap prompt specifically. */
+	recapCustomInstructions: string;
 }
 
 export interface ModelCatalog {
@@ -197,6 +205,11 @@ export interface ModelCatalog {
 	synthesis: string[];
 }
 
+/** Response from GET /api/admin/model-context — the selected model's own reported max context length, or null if Ollama doesn't expose it for this model/version. */
+export interface ModelContextInfo {
+	contextLength: number | null;
+}
+
 export interface AiStatus {
 	connected: boolean;
 	host: string;

frontend/src/lib/components/admin/EventsTab.svelte

@@ -1,6 +1,7 @@
 <script lang="ts">
 	import type { AdminTrackedEvent, AdminSource } from '$lib/adminTypes';
 	import { addEvent, updateEvent, deleteEvent } from '$lib/adminApi';
+	import CollapsibleSection from './CollapsibleSection.svelte';
 
 	let { events: initial, sources }: { events: AdminTrackedEvent[]; sources: AdminSource[] } = $props();
 	let events = $state([...initial]);
@@ -18,7 +19,9 @@
 			keywordsText: '',
 			recapIntervalHours: null as AdminTrackedEvent['recapIntervalHours'],
 			isSpillover: false,
-			retentionOverrideDays: null as number | null
+			retentionOverrideDays: null as number | null,
+			recapStylePreset: 'default' as AdminTrackedEvent['recapStylePreset'],
+			recapCustomInstructions: ''
 		};
 	}
 	let editForm = $state(emptyEditForm());
@@ -58,7 +61,9 @@
 			keywordsText: event.keywords.join(', '),
 			recapIntervalHours: event.recapIntervalHours,
 			isSpillover: event.isSpillover,
-			retentionOverrideDays: event.retentionOverrideDays
+			retentionOverrideDays: event.retentionOverrideDays,
+			recapStylePreset: event.recapStylePreset,
+			recapCustomInstructions: event.recapCustomInstructions
 		};
 	}
 
@@ -86,7 +91,9 @@
 			keywords,
 			recapIntervalHours: editForm.recapIntervalHours,
 			isSpillover: editForm.isSpillover,
-			retentionOverrideDays: editForm.retentionOverrideDays
+			retentionOverrideDays: editForm.retentionOverrideDays,
+			recapStylePreset: editForm.recapStylePreset,
+			recapCustomInstructions: editForm.recapCustomInstructions
 		});
 		events = events.map((e) => (e.id === editingId ? updated : e));
 		editingId = null;
@@ -177,6 +184,30 @@
 					</p>
 				</div>
 
+				<div class="more-section">
+					<CollapsibleSection title="More">
+						<div class="field-label">Recap writing style</div>
+						<p class="hint">
+							Independent from the global "Writing style" in the Merge tab — that only applies to
+							regular AI-merged articles. This item's own recap uses only what's set here.
+						</p>
+						<select bind:value={editForm.recapStylePreset}>
+							<option value="default">Default (neutral, wire-service tone)</option>
+							<option value="casual">Casual</option>
+							<option value="formal">Formal</option>
+						</select>
+						<label class="field-label" for="recap-custom-instructions" style="margin-top: 10px;">
+							Additional instructions (optional)
+						</label>
+						<textarea
+							id="recap-custom-instructions"
+							rows="3"
+							placeholder={'e.g. "Focus on military developments", "Write as one continuous narrative, not a list"'}
+							bind:value={editForm.recapCustomInstructions}
+						></textarea>
+					</CollapsibleSection>
+				</div>
+
 				<label class="spillover-toggle edit-spillover">
 					<input type="checkbox" bind:checked={editForm.isSpillover} />
 					Show in "More »" instead of its own nav tab
@@ -339,6 +370,17 @@
 	.cadence-block .hint {
 		margin-top: 6px;
 	}
+	.more-section {
+		margin-top: 12px;
+	}
+	.more-section select,
+	.more-section textarea {
+		width: 100%;
+	}
+	.more-section textarea {
+		resize: vertical;
+		font: inherit;
+	}
 	.edit-spillover {
 		display: flex;
 		margin-top: 12px;

frontend/src/lib/components/admin/MergeTab.svelte

@@ -204,8 +204,9 @@
 <div class="panel">
 	<span class="panel-title">Writing style</span>
 	<p class="hint">
-		Applies to AI-merged articles and event recaps only — a story with just one source
+		Applies to AI-merged articles only — a story with just one source publishes with its
-		publishes with its original text untouched, no AI involved.
+		original text untouched, no AI involved. Event recaps have their own independent writing
+		style, set per tracked item under Tracked items → Edit → More.
 	</p>
 	<select bind:value={local.synthesisStylePreset} onchange={scheduleSave}>
 		<option value="default">Default (neutral, wire-service tone)</option>

frontend/src/lib/components/admin/ModelsTab.svelte

@@ -1,6 +1,6 @@
 <script lang="ts">
 	import type { AdminSettings, ModelCatalog, AiStatus } from '$lib/adminTypes';
-	import { updateSettings, getAiStatus } from '$lib/adminApi';
+	import { updateSettings, getAiStatus, getModelContext } from '$lib/adminApi';
 	import SaveStatus from './SaveStatus.svelte';
 
 	let { settings, models, aiStatus: initialStatus }: { settings: AdminSettings; models: ModelCatalog; aiStatus: AiStatus } =
@@ -30,6 +30,63 @@
 			testing = false;
 		}
 	}
+
+	// Context window / max response length — see backend/src/inference/ollama-provider.ts's
+	// DEFAULT_NUM_CTX/DEFAULT_NUM_PREDICT for why these are ever explicit at all: too low a
+	// num_predict silently truncates the model's output mid-sentence rather than erroring
+	// (this is what a "cut off" article/recap means), and num_ctx bounds how much source
+	// text can even be included in the prompt before it gets trimmed.
+	let numCtx = $state(settings.synthesisNumCtx);
+	let numPredict = $state(settings.synthesisNumPredict);
+	let contextStatus = $state<'idle' | 'saving' | 'saved' | 'error'>('idle');
+	let contextSaveTimer: ReturnType<typeof setTimeout>;
+
+	// The selected synthesis model's own reported max context (via Ollama's /api/show) —
+	// null when undetectable (older Ollama, unusual model format, unreachable), in which
+	// case the slider falls back to a generous cap rather than blocking on it.
+	let detectedMax = $state<number | null>(null);
+	let detecting = $state(false);
+	const FALLBACK_MAX_CTX = 32768;
+	let maxCtx = $derived(detectedMax ?? FALLBACK_MAX_CTX);
+	// num_predict counts against the same context window as the prompt — capping it well
+	// under num_ctx leaves room for the prompt itself to actually fit.
+	let maxPredict = $derived(Math.max(100, numCtx - 512));
+
+	async function detectContext(model: string) {
+		if (!model) return;
+		detecting = true;
+		try {
+			const info = await getModelContext(model);
+			detectedMax = info.contextLength;
+		} catch {
+			detectedMax = null;
+		} finally {
+			detecting = false;
+		}
+	}
+
+	$effect(() => {
+		detectContext(selected.synthesis);
+	});
+
+	function scheduleContextSave() {
+		contextStatus = 'saving';
+		clearTimeout(contextSaveTimer);
+		contextSaveTimer = setTimeout(async () => {
+			try {
+				await updateSettings({ synthesisNumCtx: numCtx, synthesisNumPredict: numPredict });
+				contextStatus = 'saved';
+				setTimeout(() => (contextStatus = 'idle'), 1500);
+			} catch {
+				contextStatus = 'error';
+			}
+		}, 500);
+	}
+
+	function onNumCtxChange() {
+		if (numPredict > numCtx - 512) numPredict = Math.max(100, numCtx - 512);
+		scheduleContextSave();
+	}
 </script>
 
 <div class="panel">
@@ -84,6 +141,57 @@
 	</select>
 </div>
 
+<div class="panel">
+	<div class="head">
+		<span class="panel-title">Context window</span>
+		<SaveStatus status={contextStatus} />
+	</div>
+	<p class="hint">
+		How much text the synthesis model can take in (context window) and how long its response
+		can be (max response length). Too low a response limit is why an article or event recap
+		sometimes cuts off mid-sentence instead of finishing.
+		{#if detecting}
+			Detecting {selected.synthesis}'s limit…
+		{:else if detectedMax}
+			Detected max for {selected.synthesis}: {detectedMax.toLocaleString()} tokens.
+		{:else}
+			Couldn't detect a limit for {selected.synthesis} — defaulting the slider's ceiling to
+			{FALLBACK_MAX_CTX.toLocaleString()}. Setting num_ctx above what the model actually
+			supports will make Ollama reject or silently degrade requests.
+		{/if}
+	</p>
+
+	<label class="field-label" for="num-ctx">
+		Context window (num_ctx) — {numCtx.toLocaleString()} tokens
+	</label>
+	<div class="slider-row">
+		<input
+			id="num-ctx"
+			type="range"
+			min="1024"
+			max={maxCtx}
+			step="512"
+			bind:value={numCtx}
+			oninput={onNumCtxChange}
+		/>
+	</div>
+
+	<label class="field-label" for="num-predict" style="margin-top: 12px;">
+		Max response length (num_predict) — {numPredict.toLocaleString()} tokens
+	</label>
+	<div class="slider-row">
+		<input
+			id="num-predict"
+			type="range"
+			min="100"
+			max={maxPredict}
+			step="50"
+			bind:value={numPredict}
+			oninput={scheduleContextSave}
+		/>
+	</div>
+</div>
+
 <style>
 	.panel {
 		background: var(--surface-1);
@@ -121,6 +229,23 @@
 	.disconnected {
 		color: var(--text-danger);
 	}
+	.field-label {
+		display: block;
+		font-size: 11px;
+		color: var(--text-muted);
+		margin-bottom: 6px;
+	}
+	.slider-row {
+		display: flex;
+		align-items: center;
+		gap: 12px;
+	}
+	.slider-row input[type='range'] {
+		flex: 1;
+		border: none;
+		padding: 0;
+		background: transparent;
+	}
 	select {
 		width: 100%;
 	}