jeopi-stats 16.2.13

package/src/db.ts

@@ -0,0 +1,1278 @@
+import { Database } from "bun:sqlite";
+import * as fs from "node:fs/promises";
+import type { Usage } from "jeopi-ai";
+import type { GeneratedProvider } from "jeopi-catalog/models";
+import { getBundledModel } from "jeopi-catalog/models";
+import { getConfigRootDir, getStatsDbPath } from "jeopi-utils";
+import { classifyAgentType } from "./parser";
+import type {
+	AgentType,
+	AgentTypeStats,
+	AggregatedStats,
+	BehaviorModelStats,
+	BehaviorOverallStats,
+	BehaviorTimeSeriesPoint,
+	CostTimeSeriesPoint,
+	FolderStats,
+	MessageStats,
+	ModelPerformancePoint,
+	ModelStats,
+	ModelTimeSeriesPoint,
+	TimeSeriesPoint,
+	UserMessageLink,
+	UserMessageStats,
+} from "./types";
+
+type ModelCost = { input: number; output: number; cacheRead: number; cacheWrite: number };
+type UsageCost = Usage["cost"];
+type CostTokens = Pick<Usage, "input" | "output" | "cacheRead" | "cacheWrite">;
+
+interface CostBackfillRow {
+	id: number;
+	provider: string;
+	model: string;
+	input_tokens: number;
+	output_tokens: number;
+	cache_read_tokens: number;
+	cache_write_tokens: number;
+}
+
+let db: Database | null = null;
+
+const BACKFILL_COMPLETE = "complete";
+const BACKFILL_PENDING = "pending";
+const USER_MESSAGES_BACKFILL_KEY = "user_messages_v6";
+const USER_MESSAGE_LINKS_REPAIR_KEY = "user_message_links_v1";
+const PRIORITY_PREMIUM_REQUESTS_BACKFILL_KEY = "premium_requests_priority_v1";
+const AGENT_TYPE_BACKFILL_KEY = "agent_type_v1";
+const FORK_DEDUPE_KEY = "fork_dedupe_v1";
+function shouldResetBackfill(value: string | undefined): boolean {
+	return value !== BACKFILL_COMPLETE && value !== BACKFILL_PENDING;
+}
+/**
+ * Initialize the database and create tables.
+ */
+export async function initDb(): Promise<Database> {
+	if (db) return db;
+
+	// Ensure directory exists
+	await fs.mkdir(getConfigRootDir(), { recursive: true });
+
+	db = new Database(getStatsDbPath());
+	// Install the busy handler BEFORE any lock-taking statement. See
+	// https://github.com/can1357/oh-my-pi/issues/2421.
+	db.run("PRAGMA busy_timeout = 5000");
+	db.run("PRAGMA journal_mode = WAL");
+
+	// Whether `messages` predates this init — drives the one-time agent_type
+	// backfill below, so it must be sampled before CREATE TABLE adds the table.
+	const messagesTableExisted =
+		db.prepare("SELECT 1 FROM sqlite_master WHERE type = 'table' AND name = 'messages'").get() !== undefined;
+
+	// Create tables
+	db.run(`
+		CREATE TABLE IF NOT EXISTS messages (
+			id INTEGER PRIMARY KEY AUTOINCREMENT,
+			session_file TEXT NOT NULL,
+			entry_id TEXT NOT NULL,
+			folder TEXT NOT NULL,
+			model TEXT NOT NULL,
+			provider TEXT NOT NULL,
+			api TEXT NOT NULL,
+			timestamp INTEGER NOT NULL,
+			duration INTEGER,
+			ttft INTEGER,
+			stop_reason TEXT NOT NULL,
+			error_message TEXT,
+			input_tokens INTEGER NOT NULL,
+			output_tokens INTEGER NOT NULL,
+			cache_read_tokens INTEGER NOT NULL,
+			cache_write_tokens INTEGER NOT NULL,
+			total_tokens INTEGER NOT NULL,
+			premium_requests REAL NOT NULL,
+			cost_input REAL NOT NULL,
+			cost_output REAL NOT NULL,
+			cost_cache_read REAL NOT NULL,
+			cost_cache_write REAL NOT NULL,
+			cost_total REAL NOT NULL,
+			agent_type TEXT NOT NULL DEFAULT 'main',
+			UNIQUE(session_file, entry_id)
+		);
+
+		CREATE INDEX IF NOT EXISTS idx_messages_timestamp ON messages(timestamp);
+		CREATE INDEX IF NOT EXISTS idx_messages_model ON messages(model);
+		CREATE INDEX IF NOT EXISTS idx_messages_folder ON messages(folder);
+		CREATE INDEX IF NOT EXISTS idx_messages_session ON messages(session_file);
+		CREATE INDEX IF NOT EXISTS idx_messages_timestamp_model_provider ON messages(timestamp, model, provider);
+		CREATE INDEX IF NOT EXISTS idx_messages_timestamp_folder ON messages(timestamp, folder);
+		CREATE INDEX IF NOT EXISTS idx_messages_stop_reason_timestamp ON messages(stop_reason, timestamp);
+
+		CREATE TABLE IF NOT EXISTS file_offsets (
+			session_file TEXT PRIMARY KEY,
+			offset INTEGER NOT NULL,
+			last_modified INTEGER NOT NULL
+		);
+
+		CREATE TABLE IF NOT EXISTS user_messages (
+			id INTEGER PRIMARY KEY AUTOINCREMENT,
+			session_file TEXT NOT NULL,
+			entry_id TEXT NOT NULL,
+			folder TEXT NOT NULL,
+			timestamp INTEGER NOT NULL,
+			model TEXT,
+			provider TEXT,
+			chars INTEGER NOT NULL,
+			words INTEGER NOT NULL,
+			yelling INTEGER NOT NULL,
+			profanity INTEGER NOT NULL,
+			anguish INTEGER NOT NULL,
+			negation INTEGER NOT NULL DEFAULT 0,
+			repetition INTEGER NOT NULL DEFAULT 0,
+			blame INTEGER NOT NULL DEFAULT 0,
+			UNIQUE(session_file, entry_id)
+		);
+
+		CREATE INDEX IF NOT EXISTS idx_user_messages_timestamp ON user_messages(timestamp);
+		CREATE INDEX IF NOT EXISTS idx_user_messages_timestamp_model ON user_messages(timestamp, model, provider);
+
+		CREATE TABLE IF NOT EXISTS meta (
+			key TEXT PRIMARY KEY,
+			value TEXT NOT NULL
+		);
+	`);
+
+	const messageColumns = db.prepare("PRAGMA table_info(messages)").all() as { name: string }[];
+	if (!messageColumns.some(column => column.name === "premium_requests")) {
+		db.run("ALTER TABLE messages ADD COLUMN premium_requests REAL NOT NULL DEFAULT 0");
+	}
+	db.run("UPDATE messages SET premium_requests = 0 WHERE premium_requests IS NULL");
+	// Token-usage-by-agent: each message is classified main / subagent / advisor
+	// from its transcript path. A brand-new table gets the column from CREATE
+	// TABLE and the parser labels rows at insert time; a pre-existing table gets
+	// the column here (defaulting every prior row to 'main') and enrolls the
+	// one-time path-based reclassification, gated by a meta sentinel.
+	const hasAgentTypeColumn = messageColumns.some(column => column.name === "agent_type");
+	if (!hasAgentTypeColumn) {
+		db.run("ALTER TABLE messages ADD COLUMN agent_type TEXT NOT NULL DEFAULT 'main'");
+	}
+	// For any pre-existing table, enroll the backfill PENDING unless a prior run
+	// already settled the sentinel — `OR IGNORE` leaves an existing
+	// COMPLETE/PENDING value intact, so an ALTER that committed before its
+	// sentinel write (process killed in between) still reclassifies on the next
+	// init instead of silently leaving every row as the 'main' default. A
+	// brand-new empty table has nothing to reclassify, so it settles COMPLETE.
+	db.prepare("INSERT OR IGNORE INTO meta (key, value) VALUES (?, ?)").run(
+		AGENT_TYPE_BACKFILL_KEY,
+		messagesTableExisted ? BACKFILL_PENDING : BACKFILL_COMPLETE,
+	);
+	db.run("CREATE INDEX IF NOT EXISTS idx_messages_timestamp_agent_type ON messages(timestamp, agent_type)");
+	// Each behavior-metric bump invalidates previously-ingested rows. We detect
+	// the stale schema by column name and drop the table; `IF NOT EXISTS` above
+	// already produced the new schema, but we want a clean wipe + re-ingest.
+	// `backfillUserMessages` then clears `file_offsets` so the next sync
+	// re-parses every session under the current metric definitions.
+	//   v1 -> v2: yelling sentences replace `caps_words`.
+	//   v2 -> v3: `drama_runs` folded into a single `anguish` signal that
+	//             also captures elongated interjections, `dude`, and dot runs,
+	//             gated on a stripped prose-line budget.
+	//   v3 -> v4: added `negation`, `repetition`, `blame` frustration signals
+	//             plus profanity dictionary expansion + word-boundary fix.
+	//   v4 -> v5: column `yelling_sentences` renamed to `yelling` to match
+	//             the other single-word signal columns.
+	const userMessageColumns = db.prepare("PRAGMA table_info(user_messages)").all() as {
+		name: string;
+	}[];
+	const hasStaleColumn =
+		userMessageColumns.length > 0 &&
+		(userMessageColumns.some(column => column.name === "caps_words") ||
+			userMessageColumns.some(column => column.name === "drama_runs") ||
+			userMessageColumns.some(column => column.name === "yelling_sentences"));
+	const hasV4Columns = userMessageColumns.some(column => column.name === "negation");
+	const hasOldUserMessages = userMessageColumns.length > 0;
+	if (hasStaleColumn || (hasOldUserMessages && !hasV4Columns)) {
+		db.run("DROP TABLE user_messages");
+		db.run(`
+			CREATE TABLE user_messages (
+				id INTEGER PRIMARY KEY AUTOINCREMENT,
+				session_file TEXT NOT NULL,
+				entry_id TEXT NOT NULL,
+				folder TEXT NOT NULL,
+				timestamp INTEGER NOT NULL,
+				model TEXT,
+				provider TEXT,
+				chars INTEGER NOT NULL,
+				words INTEGER NOT NULL,
+				yelling INTEGER NOT NULL,
+				profanity INTEGER NOT NULL,
+				anguish INTEGER NOT NULL,
+				negation INTEGER NOT NULL DEFAULT 0,
+				repetition INTEGER NOT NULL DEFAULT 0,
+				blame INTEGER NOT NULL DEFAULT 0,
+				UNIQUE(session_file, entry_id)
+			);
+			CREATE INDEX IF NOT EXISTS idx_user_messages_timestamp ON user_messages(timestamp);
+			CREATE INDEX IF NOT EXISTS idx_user_messages_timestamp_model ON user_messages(timestamp, model, provider);
+		`);
+	}
+	backfillUserMessages(db);
+	repairUserMessageLinks(db);
+	backfillPriorityPremiumRequests(db);
+	backfillAgentType(db);
+	backfillMissingCatalogCosts(db);
+	backfillForkDuplicates(db);
+	return db;
+}
+
+function hasBillableCost(cost: ModelCost): boolean {
+	return cost.input !== 0 || cost.output !== 0 || cost.cacheRead !== 0 || cost.cacheWrite !== 0;
+}
+
+function getBundledModelCost(provider: string, modelId: string): ModelCost | null {
+	const model = getBundledModel(provider as GeneratedProvider, modelId);
+	return model?.cost ?? null;
+}
+
+function getCatalogCost(provider: string, modelId: string): ModelCost | null {
+	const primaryCost = getBundledModelCost(provider, modelId);
+	if (primaryCost && hasBillableCost(primaryCost)) {
+		return primaryCost;
+	}
+
+	if (provider === "openai-codex") {
+		const openAICost = getBundledModelCost("openai", modelId);
+		if (openAICost && hasBillableCost(openAICost)) {
+			return openAICost;
+		}
+	}
+
+	return null;
+}
+
+function calculateCatalogCost(provider: string, modelId: string, tokens: CostTokens): UsageCost | null {
+	const cost = getCatalogCost(provider, modelId);
+	if (!cost) return null;
+
+	const input = (cost.input / 1_000_000) * tokens.input;
+	const output = (cost.output / 1_000_000) * tokens.output;
+	const cacheRead = (cost.cacheRead / 1_000_000) * tokens.cacheRead;
+	const cacheWrite = (cost.cacheWrite / 1_000_000) * tokens.cacheWrite;
+
+	return {
+		input,
+		output,
+		cacheRead,
+		cacheWrite,
+		total: input + output + cacheRead + cacheWrite,
+	};
+}
+
+function resolveStoredCost(stats: MessageStats): UsageCost {
+	if (stats.usage.cost.total !== 0) {
+		return stats.usage.cost;
+	}
+
+	return calculateCatalogCost(stats.provider, stats.model, stats.usage) ?? stats.usage.cost;
+}
+
+function backfillMissingCatalogCosts(database: Database): void {
+	const rows = database
+		.prepare(`
+			SELECT id, provider, model, input_tokens, output_tokens, cache_read_tokens, cache_write_tokens
+			FROM messages
+			WHERE cost_total = 0 AND total_tokens > 0
+		`)
+		.all() as CostBackfillRow[];
+
+	if (rows.length === 0) return;
+
+	const update = database.prepare(`
+		UPDATE messages
+		SET cost_input = ?, cost_output = ?, cost_cache_read = ?, cost_cache_write = ?, cost_total = ?
+		WHERE id = ?
+	`);
+
+	const applyBackfill = database.transaction(() => {
+		for (const row of rows) {
+			const cost = calculateCatalogCost(row.provider, row.model, {
+				input: row.input_tokens,
+				output: row.output_tokens,
+				cacheRead: row.cache_read_tokens,
+				cacheWrite: row.cache_write_tokens,
+			});
+
+			if (!cost || cost.total === 0) continue;
+
+			update.run(cost.input, cost.output, cost.cacheRead, cost.cacheWrite, cost.total, row.id);
+		}
+	});
+
+	applyBackfill();
+}
+
+/**
+ * Get the stored offset for a session file.
+ */
+export function getFileOffset(sessionFile: string): { offset: number; lastModified: number } | null {
+	if (!db) return null;
+
+	const stmt = db.prepare("SELECT offset, last_modified FROM file_offsets WHERE session_file = ?");
+	const row = stmt.get(sessionFile) as { offset: number; last_modified: number } | undefined;
+
+	return row ? { offset: row.offset, lastModified: row.last_modified } : null;
+}
+
+/**
+ * Update the stored offset for a session file.
+ */
+export function setFileOffset(sessionFile: string, offset: number, lastModified: number): void {
+	if (!db) return;
+
+	const stmt = db.prepare(`
+		INSERT OR REPLACE INTO file_offsets (session_file, offset, last_modified)
+		VALUES (?, ?, ?)
+	`);
+	stmt.run(sessionFile, offset, lastModified);
+}
+
+/**
+ * Insert message stats into the database.
+ *
+ * Forked / branched sessions (see `SessionManager.fork()` and
+ * `createBranchedSession()` in `jeopi`) deep-copy a parent
+ * session's entries into a new JSONL — same `entry_id`, `timestamp`, `model`,
+ * `provider`, token counts, and `responseId`. The `UNIQUE(session_file,
+ * entry_id)` constraint alone keys each row by file, so without the guard
+ * below the same provider request would land twice and inflate every
+ * aggregate. The `WHERE NOT EXISTS` clause skips inserts whose
+ * `(entry_id, timestamp)` already exists under a different `session_file` —
+ * first-write-wins across the lineage. Same-file re-syncs still hit the
+ * `ON CONFLICT(session_file, entry_id)` upsert below so historical
+ * `premium_requests` fix-ups continue to work.
+ */
+export function insertMessageStats(stats: MessageStats[]): number {
+	if (!db || stats.length === 0) return 0;
+
+	const stmt = db.prepare(`
+		INSERT INTO messages (
+			session_file, entry_id, folder, model, provider, api, timestamp,
+			duration, ttft, stop_reason, error_message,
+			input_tokens, output_tokens, cache_read_tokens, cache_write_tokens, total_tokens, premium_requests,
+			cost_input, cost_output, cost_cache_read, cost_cache_write, cost_total, agent_type
+		)
+		SELECT ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?
+		WHERE NOT EXISTS (
+			SELECT 1 FROM messages
+			WHERE entry_id = ? AND timestamp = ? AND session_file <> ?
+		)
+		ON CONFLICT(session_file, entry_id) DO UPDATE SET
+			premium_requests = excluded.premium_requests
+		WHERE messages.premium_requests < excluded.premium_requests
+	`);
+
+	let inserted = 0;
+	const insert = db.transaction(() => {
+		for (const s of stats) {
+			const cost = resolveStoredCost(s);
+			const result = stmt.run(
+				s.sessionFile,
+				s.entryId,
+				s.folder,
+				s.model,
+				s.provider,
+				s.api,
+				s.timestamp,
+				s.duration,
+				s.ttft,
+				s.stopReason,
+				s.errorMessage,
+				s.usage.input,
+				s.usage.output,
+				s.usage.cacheRead,
+				s.usage.cacheWrite,
+				s.usage.totalTokens,
+				s.usage.premiumRequests ?? 0,
+				cost.input,
+				cost.output,
+				cost.cacheRead,
+				cost.cacheWrite,
+				cost.total,
+				s.agentType,
+				// `WHERE NOT EXISTS` binds: skip when a different session_file
+				// already holds this (entry_id, timestamp).
+				s.entryId,
+				s.timestamp,
+				s.sessionFile,
+			);
+			if (result.changes > 0) inserted++;
+		}
+	});
+
+	insert();
+	return inserted;
+}
+
+/**
+ * Build aggregated stats from query results.
+ */
+function buildAggregatedStats(rows: any[]): AggregatedStats {
+	if (rows.length === 0) {
+		return {
+			totalRequests: 0,
+			successfulRequests: 0,
+			failedRequests: 0,
+			errorRate: 0,
+			totalInputTokens: 0,
+			totalOutputTokens: 0,
+			totalCacheReadTokens: 0,
+			totalCacheWriteTokens: 0,
+			cacheRate: 0,
+			totalCost: 0,
+			totalPremiumRequests: 0,
+			avgDuration: null,
+			avgTtft: null,
+			avgTokensPerSecond: null,
+			firstTimestamp: 0,
+			lastTimestamp: 0,
+		};
+	}
+
+	const row = rows[0];
+	const totalRequests = row.total_requests || 0;
+	const failedRequests = row.failed_requests || 0;
+	const successfulRequests = totalRequests - failedRequests;
+	const totalInputTokens = row.total_input_tokens || 0;
+	const totalCacheReadTokens = row.total_cache_read_tokens || 0;
+	const totalPremiumRequests = row.total_premium_requests || 0;
+
+	return {
+		totalRequests,
+		successfulRequests,
+		failedRequests,
+		errorRate: totalRequests > 0 ? failedRequests / totalRequests : 0,
+		totalInputTokens,
+		totalOutputTokens: row.total_output_tokens || 0,
+		totalCacheReadTokens,
+		totalCacheWriteTokens: row.total_cache_write_tokens || 0,
+		cacheRate:
+			totalInputTokens + totalCacheReadTokens > 0
+				? totalCacheReadTokens / (totalInputTokens + totalCacheReadTokens)
+				: 0,
+		totalCost: row.total_cost || 0,
+		totalPremiumRequests,
+		avgDuration: row.avg_duration,
+		avgTtft: row.avg_ttft,
+		avgTokensPerSecond: row.avg_tokens_per_second,
+		firstTimestamp: row.first_timestamp || 0,
+		lastTimestamp: row.last_timestamp || 0,
+	};
+}
+
+/**
+ * Get overall aggregated stats.
+ */
+export function getOverallStats(cutoff?: number): AggregatedStats {
+	if (!db) return buildAggregatedStats([]);
+
+	const hasCutoff = cutoff !== undefined && cutoff > 0;
+	const stmt = db.prepare(`
+		SELECT
+			COUNT(*) as total_requests,
+			SUM(CASE WHEN stop_reason = 'error' THEN 1 ELSE 0 END) as failed_requests,
+			SUM(input_tokens) as total_input_tokens,
+			SUM(output_tokens) as total_output_tokens,
+			SUM(cache_read_tokens) as total_cache_read_tokens,
+			SUM(cache_write_tokens) as total_cache_write_tokens,
+			SUM(premium_requests) as total_premium_requests,
+			SUM(cost_total) as total_cost,
+			AVG(duration) as avg_duration,
+			AVG(ttft) as avg_ttft,
+			AVG(CASE WHEN duration > 0 THEN output_tokens * 1000.0 / duration ELSE NULL END) as avg_tokens_per_second,
+			MIN(timestamp) as first_timestamp,
+			MAX(timestamp) as last_timestamp
+		FROM messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+	`);
+
+	const rows = hasCutoff ? stmt.all(cutoff) : stmt.all();
+	return buildAggregatedStats(rows as any[]);
+}
+/**
+ * Get stats grouped by model.
+ */
+export function getStatsByModel(cutoff?: number): ModelStats[] {
+	if (!db) return [];
+
+	const hasCutoff = cutoff !== undefined && cutoff > 0;
+	const stmt = db.prepare(`
+		SELECT
+			model,
+			provider,
+			COUNT(*) as total_requests,
+			SUM(CASE WHEN stop_reason = 'error' THEN 1 ELSE 0 END) as failed_requests,
+			SUM(input_tokens) as total_input_tokens,
+			SUM(output_tokens) as total_output_tokens,
+			SUM(cache_read_tokens) as total_cache_read_tokens,
+			SUM(cache_write_tokens) as total_cache_write_tokens,
+			SUM(premium_requests) as total_premium_requests,
+			SUM(cost_total) as total_cost,
+			AVG(duration) as avg_duration,
+			AVG(ttft) as avg_ttft,
+			AVG(CASE WHEN duration > 0 THEN output_tokens * 1000.0 / duration ELSE NULL END) as avg_tokens_per_second,
+			MIN(timestamp) as first_timestamp,
+			MAX(timestamp) as last_timestamp
+		FROM messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+		GROUP BY model, provider
+		ORDER BY total_requests DESC
+	`);
+
+	const rows = (hasCutoff ? stmt.all(cutoff) : stmt.all()) as any[];
+	return rows.map(row => ({
+		model: row.model,
+		provider: row.provider,
+		...buildAggregatedStats([row]),
+	}));
+}
+
+/**
+ * Get stats grouped by folder.
+ */
+export function getStatsByFolder(cutoff?: number): FolderStats[] {
+	if (!db) return [];
+
+	const hasCutoff = cutoff !== undefined && cutoff > 0;
+	const stmt = db.prepare(`
+		SELECT
+			folder,
+			COUNT(*) as total_requests,
+			SUM(CASE WHEN stop_reason = 'error' THEN 1 ELSE 0 END) as failed_requests,
+			SUM(input_tokens) as total_input_tokens,
+			SUM(output_tokens) as total_output_tokens,
+			SUM(cache_read_tokens) as total_cache_read_tokens,
+			SUM(cache_write_tokens) as total_cache_write_tokens,
+			SUM(premium_requests) as total_premium_requests,
+			SUM(cost_total) as total_cost,
+			AVG(duration) as avg_duration,
+			AVG(ttft) as avg_ttft,
+			AVG(CASE WHEN duration > 0 THEN output_tokens * 1000.0 / duration ELSE NULL END) as avg_tokens_per_second,
+			MIN(timestamp) as first_timestamp,
+			MAX(timestamp) as last_timestamp
+		FROM messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+		GROUP BY folder
+		ORDER BY total_requests DESC
+	`);
+
+	const rows = (hasCutoff ? stmt.all(cutoff) : stmt.all()) as any[];
+	return rows.map(row => ({
+		folder: row.folder,
+		...buildAggregatedStats([row]),
+	}));
+}
+
+/**
+ * Get token usage grouped by agent type (main agent, task subagents, advisor).
+ * Token columns are explicit so the dashboard's share denominator matches the
+ * counts it renders. Rows missing `agent_type` (defensive) fall back to "main".
+ */
+export function getStatsByAgentType(cutoff?: number): AgentTypeStats[] {
+	if (!db) return [];
+
+	const hasCutoff = cutoff !== undefined && cutoff > 0;
+	const stmt = db.prepare(`
+		SELECT
+			agent_type,
+			COUNT(*) as total_requests,
+			SUM(input_tokens) as total_input_tokens,
+			SUM(output_tokens) as total_output_tokens,
+			SUM(cache_read_tokens) as total_cache_read_tokens,
+			SUM(cache_write_tokens) as total_cache_write_tokens,
+			SUM(cost_total) as total_cost
+		FROM messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+		GROUP BY agent_type
+	`);
+
+	const rows = (hasCutoff ? stmt.all(cutoff) : stmt.all()) as any[];
+	return rows.map(row => ({
+		agentType: (row.agent_type as AgentType) ?? "main",
+		totalRequests: row.total_requests || 0,
+		totalInputTokens: row.total_input_tokens || 0,
+		totalOutputTokens: row.total_output_tokens || 0,
+		totalCacheReadTokens: row.total_cache_read_tokens || 0,
+		totalCacheWriteTokens: row.total_cache_write_tokens || 0,
+		totalCost: row.total_cost || 0,
+	}));
+}
+
+/**
+ * Get time series data.
+ */
+export function getTimeSeries(hours = 24, cutoff?: number | null, bucketMs = 60 * 60 * 1000): TimeSeriesPoint[] {
+	if (!db) return [];
+
+	const hasCutoff = cutoff !== null;
+	const seriesCutoff = hasCutoff ? (cutoff ?? Date.now() - hours * 60 * 60 * 1000) : 0;
+
+	const stmt = db.prepare(`
+		SELECT
+			(timestamp / ?) * ? as bucket,
+			COUNT(*) as requests,
+			SUM(CASE WHEN stop_reason = 'error' THEN 1 ELSE 0 END) as errors,
+			SUM(total_tokens) as tokens,
+			SUM(cost_total) as cost
+		FROM messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+		GROUP BY bucket
+		ORDER BY bucket ASC
+	`);
+
+	const rows = hasCutoff
+		? (stmt.all(bucketMs, bucketMs, seriesCutoff) as any[])
+		: (stmt.all(bucketMs, bucketMs) as any[]);
+	return rows.map(row => ({
+		timestamp: row.bucket,
+		requests: row.requests,
+		errors: row.errors,
+		tokens: row.tokens,
+		cost: row.cost,
+	}));
+}
+
+/**
+ * Get daily performance time series data for the last N days.
+ */
+/**
+ * Get daily model usage time series data for the last N days.
+ */
+export function getModelTimeSeries(
+	days = 14,
+	cutoff?: number | null,
+	bucketMs = 24 * 60 * 60 * 1000,
+): ModelTimeSeriesPoint[] {
+	if (!db) return [];
+
+	const hasCutoff = cutoff !== null;
+	const seriesCutoff = hasCutoff ? (cutoff ?? Date.now() - days * 24 * 60 * 60 * 1000) : 0;
+
+	const stmt = db.prepare(`
+		SELECT
+			(timestamp / ?) * ? as bucket,
+			model,
+			provider,
+			COUNT(*) as requests
+		FROM messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+		GROUP BY bucket, model, provider
+		ORDER BY bucket ASC
+	`);
+
+	const rowsRaw = hasCutoff ? stmt.all(bucketMs, bucketMs, seriesCutoff) : stmt.all(bucketMs, bucketMs);
+	const rows = rowsRaw as Array<{ bucket: number; model: string; provider: string; requests: number }>;
+	return rows.map(row => ({
+		timestamp: row.bucket,
+		model: row.model,
+		provider: row.provider,
+		requests: row.requests,
+	}));
+}
+
+/**
+ * Get daily model performance time series data for the last N days.
+ */
+export function getModelPerformanceSeries(
+	days = 14,
+	cutoff?: number | null,
+	bucketMs = 24 * 60 * 60 * 1000,
+): ModelPerformancePoint[] {
+	if (!db) return [];
+
+	const hasCutoff = cutoff !== null;
+	const seriesCutoff = hasCutoff ? (cutoff ?? Date.now() - days * 24 * 60 * 60 * 1000) : 0;
+
+	const stmt = db.prepare(`
+		SELECT
+			(timestamp / ?) * ? as bucket,
+			model,
+			provider,
+			COUNT(*) as requests,
+			AVG(ttft) as avg_ttft,
+			AVG(CASE WHEN duration > 0 THEN output_tokens * 1000.0 / duration ELSE NULL END) as avg_tokens_per_second
+		FROM messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+		GROUP BY bucket, model, provider
+		ORDER BY bucket ASC
+	`);
+
+	const rowsRaw = hasCutoff ? stmt.all(bucketMs, bucketMs, seriesCutoff) : stmt.all(bucketMs, bucketMs);
+	const rows = rowsRaw as Array<{
+		bucket: number;
+		model: string;
+		provider: string;
+		requests: number;
+		avg_ttft: number | null;
+		avg_tokens_per_second: number | null;
+	}>;
+	return rows.map(row => ({
+		timestamp: row.bucket,
+		model: row.model,
+		provider: row.provider,
+		requests: row.requests,
+		avgTtft: row.avg_ttft,
+		avgTokensPerSecond: row.avg_tokens_per_second,
+	}));
+}
+
+/**
+ * Get total message count.
+ */
+export function getMessageCount(): number {
+	if (!db) return 0;
+	const stmt = db.prepare("SELECT COUNT(*) as count FROM messages");
+	const row = stmt.get() as { count: number };
+	return row.count;
+}
+
+/**
+ * Close the database connection.
+ */
+export function closeDb(): void {
+	if (db) {
+		db.close();
+		db = null;
+	}
+}
+
+function rowToMessageStats(row: any): MessageStats {
+	return {
+		id: row.id,
+		sessionFile: row.session_file,
+		entryId: row.entry_id,
+		folder: row.folder,
+		model: row.model,
+		provider: row.provider,
+		api: row.api,
+		timestamp: row.timestamp,
+		duration: row.duration,
+		ttft: row.ttft,
+		stopReason: row.stop_reason as any,
+		errorMessage: row.error_message,
+		usage: {
+			input: row.input_tokens,
+			output: row.output_tokens,
+			cacheRead: row.cache_read_tokens,
+			cacheWrite: row.cache_write_tokens,
+			totalTokens: row.total_tokens,
+			premiumRequests: row.premium_requests ?? 0,
+			cost: {
+				input: row.cost_input,
+				output: row.cost_output,
+				cacheRead: row.cost_cache_read,
+				cacheWrite: row.cost_cache_write,
+				total: row.cost_total,
+			},
+		},
+		agentType: (row.agent_type as AgentType) ?? "main",
+	};
+}
+
+export function getRecentRequests(limit = 100): MessageStats[] {
+	if (!db) return [];
+	const stmt = db.prepare(`
+		SELECT * FROM messages 
+		ORDER BY timestamp DESC 
+		LIMIT ?
+	`);
+	return (stmt.all(limit) as any[]).map(rowToMessageStats);
+}
+
+export function getRecentErrors(limit = 100): MessageStats[] {
+	if (!db) return [];
+	const stmt = db.prepare(`
+		SELECT * FROM messages 
+		WHERE stop_reason = 'error'
+		ORDER BY timestamp DESC 
+		LIMIT ?
+	`);
+	return (stmt.all(limit) as any[]).map(rowToMessageStats);
+}
+
+export function getMessageById(id: number): MessageStats | null {
+	if (!db) return null;
+	const stmt = db.prepare("SELECT * FROM messages WHERE id = ?");
+	const row = stmt.get(id);
+	return row ? rowToMessageStats(row) : null;
+}
+
+/**
+ * Get daily cost time series data for the last N days, broken down by model.
+ */
+export function getCostTimeSeries(days = 90, cutoff?: number | null): CostTimeSeriesPoint[] {
+	if (!db) return [];
+
+	const hasCutoff = cutoff !== null;
+	const seriesCutoff = hasCutoff ? (cutoff ?? Date.now() - days * 24 * 60 * 60 * 1000) : 0;
+
+	const stmt = db.prepare(`
+		SELECT
+			(timestamp / 86400000) * 86400000 as bucket,
+			model,
+			provider,
+			SUM(cost_total) as cost,
+			SUM(cost_input) as cost_input,
+			SUM(cost_output) as cost_output,
+			SUM(cost_cache_read) as cost_cache_read,
+			SUM(cost_cache_write) as cost_cache_write,
+			COUNT(*) as requests
+		FROM messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+		GROUP BY bucket, model, provider
+		ORDER BY bucket ASC
+	`);
+
+	const rows = hasCutoff ? (stmt.all(seriesCutoff) as any[]) : (stmt.all() as any[]);
+	return rows.map(row => ({
+		timestamp: row.bucket,
+		model: row.model,
+		provider: row.provider,
+		cost: row.cost,
+		costInput: row.cost_input,
+		costOutput: row.cost_output,
+		costCacheRead: row.cost_cache_read,
+		costCacheWrite: row.cost_cache_write,
+		requests: row.requests,
+	}));
+}
+
+/**
+ * Reset `file_offsets` (and any existing `user_messages` rows) so the next
+ * successful sync re-parses every session and re-derives behavioral metrics.
+ * Run once per metric-definition bump; the meta sentinel is only marked
+ * complete after `syncAllSessions` finishes. Older timestamp sentinel values
+ * are treated as pending so a failed compiled-binary sync cannot permanently
+ * suppress the backfill.
+ *
+ * - v1: initial introduction of `user_messages`.
+ * - v2: yelling-sentence metric replaces caps-word counts; existing rows are
+ *   computed under the old definition and must be discarded.
+ * - v3: drama runs collapsed into `anguish` (drama + elongated interjections
+ *   + `dude` + dot runs), scored on a stripped prose body and gated on
+ *   line count. Existing rows used the narrower definition.
+ * - v4: added `negation` / `repetition` / `blame` signals and fixed a
+ *   latent word-boundary bug in the profanity / anguish regexes that had
+ *   left those metrics matching nothing in real prose.
+ * - v5: renamed `yelling_sentences` column to `yelling` to match the other
+ *   single-word signal columns (profanity, anguish, negation, ...).
+ * - v6: dropped `git` from the profanity word list - it collided with the
+ *   version-control tool name, so existing rows over-counted profanity.
+ *
+ * Existing `messages` rows are unaffected - `INSERT OR IGNORE` keeps them.
+ */
+function backfillUserMessages(database: Database): void {
+	const row = database.prepare("SELECT value FROM meta WHERE key = ?").get(USER_MESSAGES_BACKFILL_KEY) as
+		| { value: string }
+		| undefined;
+	if (!shouldResetBackfill(row?.value)) return;
+
+	database.run("DELETE FROM user_messages");
+	database.run("DELETE FROM file_offsets");
+	database
+		.prepare("INSERT OR REPLACE INTO meta (key, value) VALUES (?, ?)")
+		.run(USER_MESSAGES_BACKFILL_KEY, BACKFILL_PENDING);
+}
+
+/**
+ * Reclassify pre-existing `messages` rows by agent type once, after the
+ * `agent_type` column is added to an older database (every prior row defaulted
+ * to 'main' on the ALTER). Classification is purely path-based — derived from
+ * the stored `session_file` — so no session re-parse is needed. Idempotent and
+ * crash-safe: enrolled (PENDING) only at migration time in {@link initDb} and
+ * marked COMPLETE inside the same transaction that applies the updates, so an
+ * interrupted run rolls back and retries on the next init.
+ */
+function backfillAgentType(database: Database): void {
+	const row = database.prepare("SELECT value FROM meta WHERE key = ?").get(AGENT_TYPE_BACKFILL_KEY) as
+		| { value: string }
+		| undefined;
+	if (row?.value !== BACKFILL_PENDING) return;
+
+	const sessionFiles = database.prepare("SELECT DISTINCT session_file FROM messages").all() as {
+		session_file: string;
+	}[];
+	const update = database.prepare("UPDATE messages SET agent_type = ? WHERE session_file = ?");
+	const markComplete = database.prepare("INSERT OR REPLACE INTO meta (key, value) VALUES (?, ?)");
+	const apply = database.transaction(() => {
+		for (const { session_file } of sessionFiles) {
+			const agentType = classifyAgentType(session_file);
+			// Rows already default to 'main'; only the nested transcripts move.
+			if (agentType !== "main") update.run(agentType, session_file);
+		}
+		markComplete.run(AGENT_TYPE_BACKFILL_KEY, BACKFILL_COMPLETE);
+	});
+	apply();
+}
+
+/**
+ * One-shot collapse of forked-session duplicates that landed under the old
+ * `UNIQUE(session_file, entry_id)`-only invariant. `SessionManager.fork()`
+ * and `createBranchedSession()` deep-copy a parent's entries into the new
+ * JSONL — same `entry_id`, `timestamp`, `model`, `responseId`, token counts,
+ * cost — and the previous insert path counted both files toward request /
+ * token / cost totals. The migration keeps the lowest-`id` row per
+ * `(entry_id, timestamp)` group (almost always the parent — sessions are
+ * filename-timestamped and sync processes them in name order, so the
+ * originating file lands first) and drops every other copy. Same fix on
+ * `user_messages` since forks copy user entries too. Idempotent and
+ * crash-safe: enrolled at module-load via the `meta` sentinel, marked
+ * COMPLETE inside the same transaction so an aborted run rolls back and
+ * retries on the next init.
+ */
+function backfillForkDuplicates(database: Database): void {
+	const row = database.prepare("SELECT value FROM meta WHERE key = ?").get(FORK_DEDUPE_KEY) as
+		| { value: string }
+		| undefined;
+	if (row?.value === BACKFILL_COMPLETE) return;
+
+	const markComplete = database.prepare("INSERT OR REPLACE INTO meta (key, value) VALUES (?, ?)");
+	const apply = database.transaction(() => {
+		database.run(`
+			DELETE FROM messages
+			WHERE id NOT IN (
+				SELECT MIN(id) FROM messages GROUP BY entry_id, timestamp
+			)
+		`);
+		database.run(`
+			DELETE FROM user_messages
+			WHERE id NOT IN (
+				SELECT MIN(id) FROM user_messages GROUP BY entry_id, timestamp
+			)
+		`);
+		markComplete.run(FORK_DEDUPE_KEY, BACKFILL_COMPLETE);
+	});
+	apply();
+}
+
+/**
+ * One-shot wipe of `file_offsets` to force `parseSessionFile` to re-parse
+ * every session from byte zero. We don't touch `user_messages`; the parser
+ * now emits a `UserMessageLink` for every assistant->parent pair, and the
+ * guarded `updateUserMessageLinks` UPDATE fixes any row whose `model` was
+ * left NULL by the old in-pass-only linking logic. Idempotent: gated by a
+ * sentinel row in `meta`.
+ */
+function repairUserMessageLinks(database: Database): void {
+	const row = database.prepare("SELECT value FROM meta WHERE key = ?").get(USER_MESSAGE_LINKS_REPAIR_KEY) as
+		| { value: string }
+		| undefined;
+	if (!shouldResetBackfill(row?.value)) return;
+
+	database.run("DELETE FROM file_offsets");
+	database
+		.prepare("INSERT OR REPLACE INTO meta (key, value) VALUES (?, ?)")
+		.run(USER_MESSAGE_LINKS_REPAIR_KEY, BACKFILL_PENDING);
+}
+
+/**
+ * One-shot wipe of `file_offsets` so the next sync re-parses every session
+ * and re-derives `premium_requests` from recorded `service_tier_change`
+ * entries. Earlier ingestions captured priority OpenAI traffic with
+ * `premium_requests = 0` because the AI layer only set the field for GitHub
+ * Copilot traffic. The parser now folds priority requests into the same
+ * counter; combined with the UPSERT in `insertMessageStats`, a single sync
+ * pass brings the messages table up to date without touching any other
+ * column. Idempotent: gated by a sentinel row in `meta`.
+ */
+function backfillPriorityPremiumRequests(database: Database): void {
+	const row = database.prepare("SELECT value FROM meta WHERE key = ?").get(PRIORITY_PREMIUM_REQUESTS_BACKFILL_KEY) as
+		| { value: string }
+		| undefined;
+	if (!shouldResetBackfill(row?.value)) return;
+
+	database.run("DELETE FROM file_offsets");
+	database
+		.prepare("INSERT OR REPLACE INTO meta (key, value) VALUES (?, ?)")
+		.run(PRIORITY_PREMIUM_REQUESTS_BACKFILL_KEY, BACKFILL_PENDING);
+}
+
+export function markPriorityPremiumRequestsBackfillComplete(): void {
+	if (!db) return;
+	db.prepare("INSERT OR REPLACE INTO meta (key, value) VALUES (?, ?)").run(
+		PRIORITY_PREMIUM_REQUESTS_BACKFILL_KEY,
+		BACKFILL_COMPLETE,
+	);
+}
+
+export function markUserMessagesBackfillComplete(): void {
+	if (!db) return;
+	db.prepare("INSERT OR REPLACE INTO meta (key, value) VALUES (?, ?)").run(
+		USER_MESSAGES_BACKFILL_KEY,
+		BACKFILL_COMPLETE,
+	);
+}
+
+export function markUserMessageLinksRepairComplete(): void {
+	if (!db) return;
+	db.prepare("INSERT OR REPLACE INTO meta (key, value) VALUES (?, ?)").run(
+		USER_MESSAGE_LINKS_REPAIR_KEY,
+		BACKFILL_COMPLETE,
+	);
+}
+
+/**
+ * Insert user-message stats. Idempotent via UNIQUE(session_file, entry_id).
+ * The `WHERE NOT EXISTS` clause matches {@link insertMessageStats}: forks
+ * copy user entries verbatim into the child JSONL, so the same
+ * `(entry_id, timestamp)` must not land twice across different session files.
+ */
+export function insertUserMessageStats(stats: UserMessageStats[]): number {
+	if (!db || stats.length === 0) return 0;
+
+	const stmt = db.prepare(`
+		INSERT OR IGNORE INTO user_messages (
+			session_file, entry_id, folder, timestamp, model, provider,
+			chars, words, yelling, profanity, anguish,
+			negation, repetition, blame
+		)
+		SELECT ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?
+		WHERE NOT EXISTS (
+			SELECT 1 FROM user_messages
+			WHERE entry_id = ? AND timestamp = ? AND session_file <> ?
+		)
+	`);
+
+	let inserted = 0;
+	const insert = db.transaction(() => {
+		for (const s of stats) {
+			const result = stmt.run(
+				s.sessionFile,
+				s.entryId,
+				s.folder,
+				s.timestamp,
+				s.model,
+				s.provider,
+				s.chars,
+				s.words,
+				s.yelling,
+				s.profanity,
+				s.anguish,
+				s.negation,
+				s.repetition,
+				s.blame,
+				// `WHERE NOT EXISTS` binds: skip when a different session_file
+				// already holds this (entry_id, timestamp).
+				s.entryId,
+				s.timestamp,
+				s.sessionFile,
+			);
+			if (result.changes > 0) inserted++;
+		}
+	});
+	insert();
+	return inserted;
+}
+
+/**
+ * Backfill the responding `model`/`provider` on user-message rows that were
+ * persisted before their assistant reply was parsed (a side effect of
+ * incremental `fromOffset` syncing: the `userByEntryId` map in
+ * `parseSessionFile` only spans a single pass). Each row is updated at most
+ * once because the `model IS NULL` guard short-circuits subsequent passes.
+ *
+ * Returns the number of rows actually updated.
+ */
+export function updateUserMessageLinks(links: UserMessageLink[]): number {
+	if (!db || links.length === 0) return 0;
+
+	const stmt = db.prepare(`
+		UPDATE user_messages
+		   SET model = ?, provider = ?
+		 WHERE session_file = ? AND entry_id = ? AND model IS NULL
+	`);
+
+	let updated = 0;
+	const apply = db.transaction(() => {
+		for (const link of links) {
+			const result = stmt.run(link.model, link.provider, link.sessionFile, link.entryId);
+			if (result.changes > 0) updated++;
+		}
+	});
+	apply();
+	return updated;
+}
+
+const UNKNOWN_MODEL = "unknown";
+
+interface BehaviorSeriesRow {
+	bucket: number;
+	model: string;
+	provider: string;
+	messages: number;
+	yelling: number | null;
+	profanity: number | null;
+	anguish: number | null;
+	negation: number | null;
+	repetition: number | null;
+	blame: number | null;
+	chars: number | null;
+}
+
+/**
+ * Daily behavioral time series, grouped by responding model+provider.
+ */
+export function getBehaviorTimeSeries(cutoff?: number | null): BehaviorTimeSeriesPoint[] {
+	if (!db) return [];
+	const hasCutoff = cutoff !== null && cutoff !== undefined && cutoff > 0;
+	const stmt = db.prepare(`
+		SELECT
+			(timestamp / 86400000) * 86400000 as bucket,
+			COALESCE(model, ?) as model,
+			COALESCE(provider, ?) as provider,
+			COUNT(*) as messages,
+			SUM(yelling) as yelling,
+			SUM(profanity) as profanity,
+			SUM(anguish) as anguish,
+			SUM(negation) as negation,
+			SUM(repetition) as repetition,
+			SUM(blame) as blame,
+			SUM(chars) as chars
+		FROM user_messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+		GROUP BY bucket, model, provider
+		ORDER BY bucket ASC
+	`);
+	const rows = (
+		hasCutoff ? stmt.all(UNKNOWN_MODEL, UNKNOWN_MODEL, cutoff) : stmt.all(UNKNOWN_MODEL, UNKNOWN_MODEL)
+	) as BehaviorSeriesRow[];
+	return rows.map(row => ({
+		timestamp: row.bucket,
+		model: row.model,
+		provider: row.provider,
+		messages: row.messages,
+		yelling: row.yelling ?? 0,
+		profanity: row.profanity ?? 0,
+		anguish: row.anguish ?? 0,
+		negation: row.negation ?? 0,
+		repetition: row.repetition ?? 0,
+		blame: row.blame ?? 0,
+		chars: row.chars ?? 0,
+	}));
+}
+
+interface BehaviorOverallRow {
+	total_messages: number;
+	total_yelling: number | null;
+	total_profanity: number | null;
+	total_anguish: number | null;
+	total_negation: number | null;
+	total_repetition: number | null;
+	total_blame: number | null;
+	total_chars: number | null;
+	first_timestamp: number | null;
+	last_timestamp: number | null;
+}
+
+/**
+ * Overall behavioral totals across the cutoff window.
+ */
+export function getBehaviorOverall(cutoff?: number | null): BehaviorOverallStats {
+	const empty: BehaviorOverallStats = {
+		totalMessages: 0,
+		totalYelling: 0,
+		totalProfanity: 0,
+		totalAnguish: 0,
+		totalNegation: 0,
+		totalRepetition: 0,
+		totalBlame: 0,
+		totalChars: 0,
+		firstTimestamp: 0,
+		lastTimestamp: 0,
+	};
+	if (!db) return empty;
+	const hasCutoff = cutoff !== null && cutoff !== undefined && cutoff > 0;
+	const stmt = db.prepare(`
+		SELECT
+			COUNT(*) as total_messages,
+			SUM(yelling) as total_yelling,
+			SUM(profanity) as total_profanity,
+			SUM(anguish) as total_anguish,
+			SUM(negation) as total_negation,
+			SUM(repetition) as total_repetition,
+			SUM(blame) as total_blame,
+			SUM(chars) as total_chars,
+			MIN(timestamp) as first_timestamp,
+			MAX(timestamp) as last_timestamp
+		FROM user_messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+	`);
+	const row = (hasCutoff ? stmt.get(cutoff) : stmt.get()) as BehaviorOverallRow | undefined;
+	if (!row?.total_messages) return empty;
+	return {
+		totalMessages: row.total_messages,
+		totalYelling: row.total_yelling ?? 0,
+		totalProfanity: row.total_profanity ?? 0,
+		totalAnguish: row.total_anguish ?? 0,
+		totalNegation: row.total_negation ?? 0,
+		totalRepetition: row.total_repetition ?? 0,
+		totalBlame: row.total_blame ?? 0,
+		totalChars: row.total_chars ?? 0,
+		firstTimestamp: row.first_timestamp ?? 0,
+		lastTimestamp: row.last_timestamp ?? 0,
+	};
+}
+
+interface BehaviorByModelRow {
+	model: string;
+	provider: string;
+	total_messages: number;
+	total_yelling: number | null;
+	total_profanity: number | null;
+	total_anguish: number | null;
+	total_negation: number | null;
+	total_repetition: number | null;
+	total_blame: number | null;
+	total_chars: number | null;
+	last_timestamp: number | null;
+}
+
+/**
+ * Per-model behavioral totals over the cutoff window. "Unknown" represents
+ * user messages that never received an assistant reply.
+ */
+export function getBehaviorByModel(cutoff?: number | null): BehaviorModelStats[] {
+	if (!db) return [];
+	const hasCutoff = cutoff !== null && cutoff !== undefined && cutoff > 0;
+	const stmt = db.prepare(`
+		SELECT
+			COALESCE(model, ?) as model,
+			COALESCE(provider, ?) as provider,
+			COUNT(*) as total_messages,
+			SUM(yelling) as total_yelling,
+			SUM(profanity) as total_profanity,
+			SUM(anguish) as total_anguish,
+			SUM(negation) as total_negation,
+			SUM(repetition) as total_repetition,
+			SUM(blame) as total_blame,
+			SUM(chars) as total_chars,
+			MAX(timestamp) as last_timestamp
+		FROM user_messages
+		${hasCutoff ? "WHERE timestamp >= ?" : ""}
+		GROUP BY model, provider
+		ORDER BY total_messages DESC
+	`);
+	const rows = (
+		hasCutoff ? stmt.all(UNKNOWN_MODEL, UNKNOWN_MODEL, cutoff) : stmt.all(UNKNOWN_MODEL, UNKNOWN_MODEL)
+	) as BehaviorByModelRow[];
+	return rows.map(row => ({
+		model: row.model,
+		provider: row.provider,
+		totalMessages: row.total_messages,
+		totalYelling: row.total_yelling ?? 0,
+		totalProfanity: row.total_profanity ?? 0,
+		totalAnguish: row.total_anguish ?? 0,
+		totalNegation: row.total_negation ?? 0,
+		totalRepetition: row.total_repetition ?? 0,
+		totalBlame: row.total_blame ?? 0,
+		totalChars: row.total_chars ?? 0,
+		lastTimestamp: row.last_timestamp ?? 0,
+	}));
+}