auto-model-router 0.4.2 → 0.4.4

package/src/cost/ledger.ts

@@ -27,6 +27,7 @@ import type {
 	ModelCacheReliability,
 	ModelLatency,
 	ModelTrust,
+	SoftFailureSpike,
 	UsageCounts,
 } from "./types.ts";
 
@@ -35,6 +36,20 @@ const MIN_CALIBRATION_SAMPLES = 20;
 /** Calibration samples outside this bytes-per-token band are provider accounting quirks, not tokenizer facts. */
 const MIN_SANE_BYTES_PER_TOKEN = 1.5;
 const MAX_SANE_BYTES_PER_TOKEN = 8;
+/**
+ * Soft-failure spike detection (visibility only). A model is spiking when, over
+ * the recent window, it has at least SPIKE_MIN_DISPATCHES dispatches, at least
+ * SPIKE_MIN_FAILURES of them failed, its failure rate is at least
+ * SPIKE_MIN_RATE, and that rate is at least SPIKE_RATIO × its own baseline
+ * rate over the preceding window (a model with no baseline failures spikes on
+ * the absolute floor alone).
+ */
+const SPIKE_RECENT_MS = 60 * 60_000;
+const SPIKE_BASELINE_MS = 7 * 24 * 60 * 60_000;
+const SPIKE_MIN_DISPATCHES = 5;
+const SPIKE_MIN_FAILURES = 3;
+const SPIKE_MIN_RATE = 0.25;
+const SPIKE_RATIO = 2;
 /** Escalated attempts needed before their measured cost is trusted. */
 const MIN_ESCALATION_SAMPLES = 10;
 /** The escalation-cost aggregate scans a window of rows; memoised for this long. */
@@ -334,11 +349,23 @@ export function createLedger(db: Database, cfg: RouterConfig): Ledger {
 		`SELECT ${FEEDBACK_SELECT} FROM feedback f JOIN ledger l ON l.id = f.ledger_id WHERE f.slug = ? AND l.harness_id = ? AND f.created_at_ms > ?`,
 	);
 	const allFeedbackStmt = db.query(`SELECT f.slug, ${FEEDBACK_SELECT} FROM feedback f WHERE f.created_at_ms > ? GROUP BY f.slug`);
-	const feedbackFor = (slug: string, harnessId: string | undefined, cutoff: number): FeedbackRow | null => {
+	// Task-scoped variants (filters.feedbackByTask): the judged turn's task
+	// must match, or be unrecorded (older rows, or a turn that never classified).
+	const feedbackTaskStmt = db.query(
+		`SELECT ${FEEDBACK_SELECT} FROM feedback f JOIN ledger l ON l.id = f.ledger_id WHERE f.slug = ? AND (l.task = ? OR l.task IS NULL) AND f.created_at_ms > ?`,
+	);
+	const feedbackHarnessTaskStmt = db.query(
+		`SELECT ${FEEDBACK_SELECT} FROM feedback f JOIN ledger l ON l.id = f.ledger_id WHERE f.slug = ? AND l.harness_id = ? AND (l.task = ? OR l.task IS NULL) AND f.created_at_ms > ?`,
+	);
+	const feedbackFor = (slug: string, harnessId: string | undefined, cutoff: number, task?: string): FeedbackRow | null => {
 		if (cfg.filters.feedbackWeight <= 0) return null;
-		return harnessId !== undefined && harnessId !== ""
-			? (feedbackHarnessStmt.get(slug, harnessId, cutoff) as FeedbackRow | null)
-			: (feedbackStmt.get(slug, cutoff) as FeedbackRow | null);
+		const byHarness = harnessId !== undefined && harnessId !== "";
+		if (cfg.filters.feedbackByTask && task !== undefined && task !== "") {
+			return byHarness
+				? (feedbackHarnessTaskStmt.get(slug, harnessId, task, cutoff) as FeedbackRow | null)
+				: (feedbackTaskStmt.get(slug, task, cutoff) as FeedbackRow | null);
+		}
+		return byHarness ? (feedbackHarnessStmt.get(slug, harnessId, cutoff) as FeedbackRow | null) : (feedbackStmt.get(slug, cutoff) as FeedbackRow | null);
 	};
 	const latencyStmt = db.query(
 		`SELECT ${LATENCY_SELECT} FROM (SELECT * FROM ledger WHERE slug = ? ORDER BY created_at_ms DESC LIMIT ${LATENCY_WINDOW_ROWS})`,
@@ -348,6 +375,8 @@ export function createLedger(db: Database, cfg: RouterConfig): Ledger {
 	);
 	const ratioStmt = db.query("SELECT est_bytes, actual_tokens, samples FROM token_calibration WHERE tokenizer = ?");
 	const recentStmt = db.query("SELECT * FROM ledger ORDER BY created_at_ms DESC LIMIT ?");
+	const pruneStmt = db.query("DELETE FROM ledger WHERE created_at_ms < ?");
+	const wasteStmt = db.query("UPDATE ledger SET wasted = 1 WHERE id = ?");
 	const providerSpendStmt = db.query(
 		"SELECT COALESCE(SUM(COALESCE(reported_usd, predicted_usd)), 0) AS total FROM ledger WHERE created_at_ms >= ? AND COALESCE(served_slug, slug) LIKE ?",
 	);
@@ -363,6 +392,19 @@ export function createLedger(db: Database, cfg: RouterConfig): Ledger {
 		 FROM ledger WHERE attempt > 0 AND error IS NULL AND created_at_ms >= ?`,
 	);
 	let escalationMemo: { atMs: number; windowDays: number; value: EscalationCost | null } | null = null;
+	// Per-model failure counts over two adjacent windows: [recentStart, now] and
+	// [baselineStart, recentStart). Wasted rows (the failed attempt a retry
+	// replaced) stay in: they ARE the soft failures being counted. Digest side
+	// calls are excluded: they are not the session's turns.
+	const softFailureStmt = db.query(
+		`SELECT COALESCE(served_slug, slug) AS slug,
+			SUM(CASE WHEN created_at_ms >= $recentStart THEN 1 ELSE 0 END) AS recent_n,
+			SUM(CASE WHEN created_at_ms >= $recentStart AND (escalation_signal IS NOT NULL OR (${ATTRIBUTABLE_ERROR})) THEN 1 ELSE 0 END) AS recent_f,
+			SUM(CASE WHEN created_at_ms < $recentStart THEN 1 ELSE 0 END) AS base_n,
+			SUM(CASE WHEN created_at_ms < $recentStart AND (escalation_signal IS NOT NULL OR (${ATTRIBUTABLE_ERROR})) THEN 1 ELSE 0 END) AS base_f
+		 FROM ledger WHERE created_at_ms >= $baselineStart AND created_at_ms <= $now AND requested_model <> 'digest'
+		 GROUP BY COALESCE(served_slug, slug)`,
+	);
 	const cacheMetaStmt = db.query("SELECT fetched_at_ms FROM catalog_cache WHERE id = 1");
 	const cachePayloadStmt = db.query("SELECT payload FROM catalog_cache WHERE id = 1");
 
@@ -466,7 +508,7 @@ export function createLedger(db: Database, cfg: RouterConfig): Ledger {
 			return computeBlendedRate(db, cfg, windowDays);
 		},
 
-		trust(slug: string, harnessId?: string): ModelTrust | null {
+		trust(slug: string, harnessId?: string, task?: string): ModelTrust | null {
 			// Read the window at CALL time, not at construction: hot reload mutates
 			// the shared config object in place, so a pinned value would ignore an
 			// edit until restart. 0 => cutoff 0 => every row qualifies.
@@ -476,7 +518,7 @@ export function createLedger(db: Database, cfg: RouterConfig): Ledger {
 					? (trustHarnessStmt.get(slug, harnessId, cutoff) as TrustRow | null)
 					: (trustStmt.get(slug, cutoff) as TrustRow | null);
 			if (row === null || row.attempts === 0) return null;
-			return toTrust(slug, row, feedbackFor(slug, harnessId, cutoff), cfg.filters.feedbackWeight);
+			return toTrust(slug, row, feedbackFor(slug, harnessId, cutoff, task), cfg.filters.feedbackWeight);
 		},
 
 		allTrust(): ModelTrust[] {
@@ -497,7 +539,7 @@ export function createLedger(db: Database, cfg: RouterConfig): Ledger {
 			if (row === null) return null;
 			return toLatency(slug, row);
 		},
-		signals(slugs: readonly string[], harnessId?: string): Map<string, LedgerSignals> {
+		signals(slugs: readonly string[], harnessId?: string, task?: string): Map<string, LedgerSignals> {
 			const cutoff = cfg.filters.trustWindowDays > 0 ? Date.now() - cfg.filters.trustWindowDays * DAY_MS : 0;
 			const hasHarness = harnessId !== undefined && harnessId !== "";
 			const out = new Map<string, LedgerSignals>();
@@ -509,7 +551,7 @@ export function createLedger(db: Database, cfg: RouterConfig): Ledger {
 					? (latencyHarnessStmt.get(slug, harnessId) as LatencyRow | null)
 					: (latencyStmt.get(slug) as LatencyRow | null);
 				out.set(slug, {
-					trust: trustRow === null || trustRow.attempts === 0 ? null : toTrust(slug, trustRow, feedbackFor(slug, harnessId, cutoff), cfg.filters.feedbackWeight),
+					trust: trustRow === null || trustRow.attempts === 0 ? null : toTrust(slug, trustRow, feedbackFor(slug, harnessId, cutoff, task), cfg.filters.feedbackWeight),
 					latency: latencyRow === null ? null : toLatency(slug, latencyRow),
 				});
 			}
@@ -552,10 +594,44 @@ export function createLedger(db: Database, cfg: RouterConfig): Ledger {
 			const rows = recentStmt.all(limit) as LedgerRow[];
 			return rows.map(toEntry);
 		},
+		softFailureSpikes(nowMs = Date.now(), recentMs = SPIKE_RECENT_MS, baselineMs = SPIKE_BASELINE_MS): SoftFailureSpike[] {
+			const rows = softFailureStmt.all({ $now: nowMs, $recentStart: nowMs - recentMs, $baselineStart: nowMs - recentMs - baselineMs }) as {
+				slug: string;
+				recent_n: number;
+				recent_f: number;
+				base_n: number;
+				base_f: number;
+			}[];
+			const spikes: SoftFailureSpike[] = [];
+			for (const r of rows) {
+				if (r.recent_n < SPIKE_MIN_DISPATCHES || r.recent_f < SPIKE_MIN_FAILURES) continue;
+				const recentRate = r.recent_f / r.recent_n;
+				const baselineRate = r.base_n > 0 ? r.base_f / r.base_n : 0;
+				if (recentRate < SPIKE_MIN_RATE || recentRate < SPIKE_RATIO * baselineRate) continue;
+				spikes.push({
+					slug: r.slug,
+					recentDispatches: r.recent_n,
+					recentFailures: r.recent_f,
+					recentRate,
+					baselineDispatches: r.base_n,
+					baselineFailures: r.base_f,
+					baselineRate,
+				});
+			}
+			spikes.sort((a, b) => b.recentRate - a.recentRate || b.recentFailures - a.recentFailures);
+			return spikes;
+		},
 		providerSpendSince(slugPrefix: string, sinceMs: number): number {
 			const row = providerSpendStmt.get(sinceMs, `${slugPrefix}%`) as { total: number } | null;
 			return row?.total ?? 0;
 		},
+		prune(retentionDays: number, nowMs = Date.now()): number {
+			if (retentionDays <= 0) return 0;
+			return pruneStmt.run(nowMs - retentionDays * DAY_MS).changes;
+		},
+		markWasted(id: string): void {
+			wasteStmt.run(id);
+		},
 		latestForSession(ompSessionId: string): LedgerEntry | null {
 			if (ompSessionId === "") return null;
 			const row = sessionStmt.get(ompSessionId, 1) as LedgerRow | null;

package/src/cost/report.ts

@@ -34,6 +34,12 @@ export interface ReportTotals {
 	digests: number;
 	digestSpendUsd: number;
 	digestInputTokens: number;
+	/** Digests the agent went back on: the same tool re-run with the same primary argument afterwards (row marked wasted). */
+	digestReruns: number;
+	/** Forecast accuracy over clean kept rows with a reported cost: mean |predicted − reported| ÷ reported, and the share over-predicted. */
+	forecastSamples: number;
+	forecastMeanError: number;
+	forecastOverShare: number;
 }
 
 export interface ReportRow {
@@ -134,6 +140,8 @@ const COMP = "json_extract(usage, '$.completionTokens')";
 const PROVIDER = "CASE WHEN slug LIKE 'ollama/%' THEN 'ollama' ELSE 'openrouter' END";
 const STREAMED = "ttft_ms IS NOT NULL AND ttft_ms > 0 AND error IS NULL";
 const EST = "json_extract(usage, '$.cachedEstimated') = 1";
+/** Rows a forecast can be judged on: a reported cost, a prediction, clean and kept, not a side call. */
+const FORECASTABLE = "reported_usd > 0 AND predicted_usd IS NOT NULL AND wasted = 0 AND error IS NULL AND requested_model <> 'digest'";
 
 const ROW_SELECT = `
 	COUNT(*) AS dispatches,
@@ -185,14 +193,19 @@ function toRow(r: RawRow, windowSpend: number): ReportRow {
  */
 export function buildUsageReport(
 	db: Database,
-	opts: { windowDays: number; harnessId?: string; nowMs?: number; baselines?: readonly BaselinePrice[] },
+	opts: { windowDays: number; harnessId?: string; nowMs?: number; baselines?: readonly BaselinePrice[]; /** Exclusive upper bound; default open-ended. */ untilMs?: number },
 ): UsageReport {
 	const nowMs = opts.nowMs ?? Date.now();
 	const windowDays = Math.max(1, opts.windowDays);
 	const sinceMs = nowMs - windowDays * 86_400_000;
 	const harnessId = opts.harnessId ?? "";
-	const where = harnessId === "" ? "created_at_ms >= $since" : "created_at_ms >= $since AND harness_id = $harness";
-	const bind = harnessId === "" ? { $since: sinceMs } : { $since: sinceMs, $harness: harnessId };
+	const untilMs = opts.untilMs;
+	const where = [
+		"created_at_ms >= $since",
+		...(untilMs === undefined ? [] : ["created_at_ms < $until"]),
+		...(harnessId === "" ? [] : ["harness_id = $harness"]),
+	].join(" AND ");
+	const bind = { $since: sinceMs, ...(untilMs === undefined ? {} : { $until: untilMs }), ...(harnessId === "" ? {} : { $harness: harnessId }) };
 
 	const t = db
 		.query(
@@ -208,6 +221,10 @@ export function buildUsageReport(
 				SUM(CASE WHEN requested_model = 'digest' THEN 1 ELSE 0 END) AS digests,
 				COALESCE(SUM(CASE WHEN requested_model = 'digest' THEN ${USD} ELSE 0 END), 0) AS digest_spend,
 				COALESCE(SUM(CASE WHEN requested_model = 'digest' THEN ${PT} ELSE 0 END), 0) AS digest_input,
+				SUM(CASE WHEN requested_model = 'digest' AND wasted = 1 THEN 1 ELSE 0 END) AS digest_reruns,
+				SUM(CASE WHEN ${FORECASTABLE} THEN 1 ELSE 0 END) AS fc_n,
+				COALESCE(SUM(CASE WHEN ${FORECASTABLE} THEN ABS(predicted_usd - reported_usd) / reported_usd END), 0) AS fc_err,
+				SUM(CASE WHEN ${FORECASTABLE} AND predicted_usd > reported_usd THEN 1 ELSE 0 END) AS fc_over,
 				SUM(CASE WHEN escalation_signal IS NOT NULL THEN 1 ELSE 0 END) AS escalations,
 				SUM(CASE WHEN instr(reasons, 'failover:') > 0 THEN 1 ELSE 0 END) AS failovers,
 				SUM(CASE WHEN error IS NOT NULL THEN 1 ELSE 0 END) AS errors,
@@ -227,6 +244,10 @@ export function buildUsageReport(
 		digests: number | null;
 		digest_spend: number;
 		digest_input: number;
+		digest_reruns: number | null;
+		fc_n: number | null;
+		fc_err: number;
+		fc_over: number | null;
 		escalations: number | null;
 		failovers: number | null;
 		errors: number | null;
@@ -349,6 +370,10 @@ export function buildUsageReport(
 			digests: t.digests ?? 0,
 			digestSpendUsd: t.digest_spend,
 			digestInputTokens: t.digest_input,
+			digestReruns: t.digest_reruns ?? 0,
+			forecastSamples: t.fc_n ?? 0,
+			forecastMeanError: (t.fc_n ?? 0) > 0 ? t.fc_err / (t.fc_n ?? 1) : 0,
+			forecastOverShare: (t.fc_n ?? 0) > 0 ? (t.fc_over ?? 0) / (t.fc_n ?? 1) : 0,
 		},
 		providers,
 		models,
@@ -414,7 +439,12 @@ export function reportView(r: UsageReport, opts: { maxModels?: number } = {}): R
 		);
 	}
 	if (t.digests > 0) {
-		summary.push(`digests: ${num(t.digests)} tool results condensed (${num(t.digestInputTokens)} tok read by a cheap model) for ${usd(t.digestSpendUsd)}`);
+		summary.push(
+			`digests: ${num(t.digests)} tool results condensed (${num(t.digestInputTokens)} tok read by a cheap model) for ${usd(t.digestSpendUsd)} · re-run rate ${pct(t.digestReruns / t.digests)} (${num(t.digestReruns)} fetched again in full)`,
+		);
+	}
+	if (t.forecastSamples > 0) {
+		summary.push(`forecast: mean error ${pct(t.forecastMeanError)} of reported cost over ${num(t.forecastSamples)} turns · ${pct(t.forecastOverShare)} over-predicted`);
 	}
 	if (t.subagentDispatches > 0) {
 		summary.push(`subagents: ${num(t.subagentDispatches)} dispatches, ${usd(t.subagentSpendUsd)} (${pct(t.spendUsd > 0 ? t.subagentSpendUsd / t.spendUsd : 0)} of spend)`);

package/src/cost/summary.ts

@@ -0,0 +1,231 @@
+/**
+ * The daily summary: what the router did in the last 24 hours, in a few
+ * lines. Posted into the transcript once a day at omp session start
+ * (`report.dailySummary`) and on demand via `/router summary` or
+ * `GET /v1/router/summary`.
+ *
+ * Built from the same `buildUsageReport` the report hub uses, over a 1-day
+ * window, with the preceding day for comparison, plus the two live signals the
+ * report cannot carry: soft-failure spikes (the last hour) and the Ollama
+ * meter. The once-a-day gate is a marker in `router_kv`, keyed per harness, so
+ * several omp windows on one router share it and a restart does not repeat it.
+ */
+
+import type { Database } from "bun:sqlite";
+import { TIER_ORDER } from "../router/types.ts";
+import { buildUsageReport, type BaselinePrice, type BaselineRow, type UsageReport } from "./report.ts";
+import type { SoftFailureSpike } from "./types.ts";
+
+/** One 24-hour window's headline numbers. */
+export interface SummaryWindow {
+	spendUsd: number;
+	dispatches: number;
+	conversations: number;
+	cacheHitRate: number;
+	cacheEstimated: boolean;
+	escalations: number;
+	errors: number;
+	modelSwitches: number;
+	digests: number;
+	digestSpendUsd: number;
+	subagentSpendUsd: number;
+}
+
+export interface SummaryModel {
+	slug: string;
+	spendUsd: number;
+	share: number;
+	dispatches: number;
+}
+
+export interface SummaryOllama {
+	plan: string | null;
+	usedUsd: number;
+	creditsUsd: number;
+	/** Days of credits left at the recent burn; null when the burn is zero. */
+	runwayDays: number | null;
+}
+
+export interface DailySummary {
+	generatedAtMs: number;
+	sinceMs: number;
+	/** Empty ⇒ every harness. */
+	harnessId: string;
+	current: SummaryWindow;
+	/** The 24 hours before `sinceMs`. */
+	previous: SummaryWindow;
+	/** Top models by spend in the current window. */
+	topModels: SummaryModel[];
+	/** Tier moves between consecutive kept turns of one conversation. */
+	tierChanges: { up: number; down: number };
+	/** The first configured baseline the catalog knew, when any. */
+	baseline: BaselineRow | null;
+	spikes: SoftFailureSpike[];
+	ollama: SummaryOllama | null;
+}
+
+const DAY_MS = 86_400_000;
+/** How many top models the summary names. */
+const TOP_MODELS = 3;
+
+function windowOf(r: UsageReport): SummaryWindow {
+	const t = r.totals;
+	return {
+		spendUsd: t.spendUsd,
+		dispatches: t.dispatches,
+		conversations: t.conversations,
+		cacheHitRate: t.cacheHitRate,
+		cacheEstimated: t.cacheEstimated,
+		escalations: t.escalations,
+		errors: t.errors,
+		modelSwitches: t.modelSwitches,
+		digests: t.digests,
+		digestSpendUsd: t.digestSpendUsd,
+		subagentSpendUsd: t.subagentSpendUsd,
+	};
+}
+
+/** Counts tier moves up and down between consecutive kept turns of each conversation since `sinceMs`. */
+export function countTierChanges(db: Database, sinceMs: number, harnessId: string): { up: number; down: number } {
+	const where = harnessId === "" ? "created_at_ms >= $since" : "created_at_ms >= $since AND harness_id = $harness";
+	const bind = harnessId === "" ? { $since: sinceMs } : { $since: sinceMs, $harness: harnessId };
+	const seq = db
+		.query(`SELECT conversation_key AS ck, tier FROM ledger WHERE ${where} AND wasted = 0 AND requested_model <> 'digest' ORDER BY conversation_key, created_at_ms`)
+		.all(bind) as { ck: string; tier: string }[];
+	let up = 0;
+	let down = 0;
+	for (let i = 1; i < seq.length; i++) {
+		const a = seq[i - 1]!;
+		const b = seq[i]!;
+		if (a.ck !== b.ck) continue;
+		const ra = TIER_ORDER.indexOf(a.tier as (typeof TIER_ORDER)[number]);
+		const rb = TIER_ORDER.indexOf(b.tier as (typeof TIER_ORDER)[number]);
+		if (ra < 0 || rb < 0 || ra === rb) continue;
+		if (rb > ra) up++;
+		else down++;
+	}
+	return { up, down };
+}
+
+export function buildDailySummary(
+	db: Database,
+	opts: {
+		harnessId?: string;
+		nowMs?: number;
+		baselines?: readonly BaselinePrice[];
+		spikes?: readonly SoftFailureSpike[];
+		ollama?: SummaryOllama | null;
+	} = {},
+): DailySummary {
+	const nowMs = opts.nowMs ?? Date.now();
+	const harnessId = opts.harnessId ?? "";
+	const baselines = opts.baselines ?? [];
+	const current = buildUsageReport(db, { windowDays: 1, harnessId, nowMs, baselines });
+	const previous = buildUsageReport(db, { windowDays: 1, harnessId, nowMs: nowMs - DAY_MS, untilMs: current.sinceMs });
+	return {
+		generatedAtMs: nowMs,
+		sinceMs: current.sinceMs,
+		harnessId,
+		current: windowOf(current),
+		previous: windowOf(previous),
+		topModels: current.models.slice(0, TOP_MODELS).map((m) => ({ slug: m.key, spendUsd: m.spendUsd, share: m.share, dispatches: m.dispatches })),
+		tierChanges: countTierChanges(db, current.sinceMs, harnessId),
+		baseline: current.baselines[0] ?? null,
+		spikes: [...(opts.spikes ?? [])],
+		ollama: opts.ollama ?? null,
+	};
+}
+
+/** Whether the once-a-day auto summary is worth posting: something happened, or something is wrong. */
+export function summaryHasNews(s: DailySummary): boolean {
+	return s.current.dispatches > 0 || s.spikes.length > 0;
+}
+
+const usd = (v: number): string => (v >= 100 ? `$${v.toFixed(0)}` : v >= 1 ? `$${v.toFixed(2)}` : `$${v.toFixed(3)}`);
+const pct = (v: number, estimated = false): string => `${estimated ? "~" : ""}${Math.round(v * 100)}%`;
+
+function delta(current: number, previous: number): string {
+	if (previous <= 0) return current > 0 ? " (prev 24h: none)" : "";
+	const change = (current - previous) / previous;
+	const sign = change >= 0 ? "+" : "−";
+	return ` (prev 24h ${usd(previous)}, ${sign}${Math.round(Math.abs(change) * 100)}%)`;
+}
+
+/** Renders the summary as a few plain lines for the transcript. */
+export function renderDailySummary(s: DailySummary): string {
+	const scope = s.harnessId === "" ? "all harnesses" : `harness ${s.harnessId}`;
+	const out: string[] = [`auto-model-router daily summary — last 24h (${scope})`];
+	const c = s.current;
+	if (c.dispatches === 0) {
+		out.push("no routed turns in the last 24h");
+	} else {
+		out.push(
+			`spend ${usd(c.spendUsd)}${delta(c.spendUsd, s.previous.spendUsd)} · ${c.dispatches} turns · ${c.conversations} conversations · ${usd(c.spendUsd / c.dispatches)}/turn`,
+		);
+		const moves = c.modelSwitches > 0 ? ` (${s.tierChanges.up} tier up, ${s.tierChanges.down} down)` : "";
+		out.push(`cache hit ${pct(c.cacheHitRate, c.cacheEstimated)} · ${c.escalations} escalations · ${c.errors} errors · ${c.modelSwitches} model switches${moves}`);
+		if (s.topModels.length > 0) {
+			out.push(`top models: ${s.topModels.map((m) => `${m.slug} ${usd(m.spendUsd)} (${pct(m.share)}, ${m.dispatches} turns)`).join(" · ")}`);
+		}
+		if (s.baseline !== null && s.baseline.usd > 0) {
+			const b = s.baseline;
+			out.push(b.savedShare >= 0 ? `saved ${pct(b.savedShare)} vs ${b.slug} (${usd(b.usd)} at list)` : `cost ${pct(-b.savedShare)} MORE than ${b.slug} (${usd(b.usd)} at list)`);
+		}
+		const extras: string[] = [];
+		if (c.digests > 0) extras.push(`${c.digests} digests for ${usd(c.digestSpendUsd)}`);
+		if (c.subagentSpendUsd > 0) extras.push(`subagents ${usd(c.subagentSpendUsd)}`);
+		if (extras.length > 0) out.push(extras.join(" · "));
+	}
+	if (s.spikes.length === 0) out.push("soft failures: no model spiking in the last hour");
+	else {
+		out.push(`soft failures SPIKING (${s.spikes.length}):`);
+		for (const sp of s.spikes) {
+			out.push(`  ${sp.slug}: ${pct(sp.recentRate)} of ${sp.recentDispatches} failed in the last 1h (7d baseline ${pct(sp.baselineRate)} of ${sp.baselineDispatches})`);
+		}
+	}
+	const o = s.ollama;
+	if (o !== null) {
+		const runway = o.runwayDays === null ? "" : ` · ~${Math.round(o.runwayDays)} days of credits left`;
+		out.push(`ollama: ${o.plan === null ? "plan" : `${o.plan} plan`} $${o.usedUsd.toFixed(2)} of $${o.creditsUsd}${runway}`);
+	}
+	return out.join("\n");
+}
+
+// ---------------------------------------------------------------------------
+// Once-a-day gate
+// ---------------------------------------------------------------------------
+
+/** A summary posted less than this long ago is not due again. */
+export const DAILY_SUMMARY_INTERVAL_MS = 20 * 3_600_000;
+
+export interface KeyValueStore {
+	get(key: string): string | null;
+	set(key: string, value: string): void;
+}
+
+/** A tiny durable key/value store over the `router_kv` table. */
+export function createKv(db: Database): KeyValueStore {
+	const getStmt = db.query("SELECT value FROM router_kv WHERE key = $key");
+	const setStmt = db.query("INSERT INTO router_kv (key, value, updated_at_ms) VALUES ($key, $value, $at) ON CONFLICT(key) DO UPDATE SET value = excluded.value, updated_at_ms = excluded.updated_at_ms");
+	return {
+		get(key) {
+			const row = getStmt.get({ $key: key }) as { value: string } | null;
+			return row === null ? null : row.value;
+		},
+		set(key, value) {
+			setStmt.run({ $key: key, $value: value, $at: Date.now() });
+		},
+	};
+}
+
+const shownKey = (harnessId: string): string => `daily_summary_shown:${harnessId}`;
+
+/** True when no auto summary has been posted for this harness within the interval. */
+export function summaryDue(kv: KeyValueStore, harnessId: string, nowMs = Date.now()): boolean {
+	const last = Number(kv.get(shownKey(harnessId)) ?? "0");
+	return !(Number.isFinite(last) && nowMs - last < DAILY_SUMMARY_INTERVAL_MS);
+}
+
+export function markSummaryShown(kv: KeyValueStore, harnessId: string, nowMs = Date.now()): void {
+	kv.set(shownKey(harnessId), String(nowMs));
+}

package/src/cost/types.ts

@@ -230,6 +230,25 @@ export interface EscalationCost {
 	windowDays: number;
 }
 
+/**
+ * A model whose recent failure rate (probe rejections OpenRouter counts as
+ * success, plus attributable transport errors) is well above its own
+ * baseline. Visibility only: the ledger data showed soft failures do not
+ * cluster tightly enough for a breaker to save money, so the router reports
+ * spikes (/health, /router status, the daily summary) rather than acting.
+ */
+export interface SoftFailureSpike {
+	slug: string;
+	/** Dispatches and failures in the recent window. */
+	recentDispatches: number;
+	recentFailures: number;
+	recentRate: number;
+	/** The same, over the baseline window (recent window excluded). */
+	baselineDispatches: number;
+	baselineFailures: number;
+	baselineRate: number;
+}
+
 export interface Ledger {
 	record(entry: LedgerEntry): void;
 	/** Total reported (or predicted, when reported is null) spend for a conversation. */
@@ -240,8 +259,12 @@ export interface Ledger {
 	 */
 	spendSince(sinceMs: number, harnessId?: string): number;
 	blendedRate(windowDays: number): BlendedRate | null;
-	/** Per-model reliability over the ledger, optionally scoped to a harness. */
-	trust(slug: string, harnessId?: string): ModelTrust | null;
+	/**
+	 * Per-model reliability over the ledger, optionally scoped to a harness.
+	 * `task` (with `filters.feedbackByTask`) counts only verdicts given on
+	 * turns of that task type, plus verdicts on turns with no recorded task.
+	 */
+	trust(slug: string, harnessId?: string, task?: string): ModelTrust | null;
 	allTrust(): ModelTrust[];
 	/**
 	 * Per-model responsiveness (mean TTFT + completion throughput), optionally
@@ -251,7 +274,7 @@ export interface Ledger {
 	 */
 	latency(slug: string, harnessId?: string): ModelLatency | null;
 	/** Batch trust and latency for one candidate set; one query per signal kind. Optional — callers can fall back to per-slug calls. */
-	signals?(slugs: readonly string[], harnessId?: string): Map<string, LedgerSignals>;
+	signals?(slugs: readonly string[], harnessId?: string, task?: string): Map<string, LedgerSignals>;
 	/**
 	 * What an escalated retry actually bills per prompt token, measured over
 	 * the last `windowDays` of attempt > 0 rows. Null until enough escalated
@@ -271,8 +294,17 @@ export interface Ledger {
 	recentEntries(limit: number): LedgerEntry[];
 	/** Spend since an instant on slugs with a prefix (`ollama/`), for provider-level reconciliation. Optional. */
 	providerSpendSince?(slugPrefix: string, sinceMs: number): number;
+	/**
+	 * Models whose soft-failure rate over the last `recentMs` is a spike against
+	 * their own rate over the preceding `baselineMs`. Optional; visibility only.
+	 */
+	softFailureSpikes?(nowMs?: number, recentMs?: number, baselineMs?: number): SoftFailureSpike[];
 	/** Newest kept (non-wasted) entry for an omp session, for /router why and feedback. Optional so fakes need not implement it. */
 	latestForSession?(ompSessionId: string): LedgerEntry | null;
 	/** Newest entries for an omp session, newest first. Optional. */
 	entriesForSession?(ompSessionId: string, limit: number): LedgerEntry[];
+	/** Deletes rows older than `retentionDays` (0 ⇒ none); returns how many. Optional. */
+	prune?(retentionDays: number, nowMs?: number): number;
+	/** Marks one row wasted after the fact (a digest the agent went back on). Optional. */
+	markWasted?(id: string): void;
 }

package/src/router/candidates.ts

@@ -263,7 +263,7 @@ export function buildCandidates(args: BuildCandidatesArgs): { candidates: Candid
 		const signals = args.signals;
 		const trust =
 			signals?.get(slug)?.trust ??
-			ledger?.trust(slug, filters.trustScopedByHarness ? req.harnessId : undefined) ??
+			ledger?.trust(slug, filters.trustScopedByHarness ? req.harnessId : undefined, filters.feedbackByTask ? task : undefined) ??
 			null;
 		if (!relaxTrust && trust !== null && trust.attempts >= filters.minTrustSamples && trust.successRate < filters.minTrust) {
 			rejected.push({

package/src/router/classify.ts

@@ -10,7 +10,7 @@ import type { Ledger } from "../cost/types.ts";
 import { estimateTokens } from "../tokens/estimate.ts";
 import type { UpstreamClient } from "../upstream/types.ts";
 import { sha256Hex } from "../util/hash.ts";
-import { loadLearnedModel, predictRisk } from "./learned.ts";
+import { learnedRiskName, loadLearnedModel, predictRisk } from "./learned.ts";
 import type { NormRequest, ReasoningLevel } from "../wire/types.ts";
 import type { Classification, Features, TaskType, Tier } from "./types.ts";
 
@@ -313,7 +313,7 @@ export async function classify(
 		if (model !== null) {
 			const risk = predictRisk(model, f);
 			heuristic.learnedRisk = risk;
-			heuristic.reasons.push(`learned: p(escalate)=${risk.toFixed(3)}`);
+			heuristic.reasons.push(`learned: p(${learnedRiskName(model)})=${risk.toFixed(3)}`);
 		}
 	}
 	if (cc.ambiguityThreshold <= 0 || heuristic.confidence >= cc.ambiguityThreshold) return heuristic;

package/src/router/compaction.ts

@@ -33,6 +33,7 @@ const BREADCRUMB_BYTES = 120;
  */
 export function compactedBytes(originalBytes: number, edit: CompactionEdit | undefined): number {
 	if (edit === undefined) return originalBytes;
+	if (edit.digest !== undefined) return Math.min(originalBytes, Buffer.byteLength(edit.digest));
 	const kept = edit.mode === "stub" ? BREADCRUMB_BYTES : edit.keepHead + edit.keepTail + BREADCRUMB_BYTES;
 	return Math.min(originalBytes, kept);
 }
@@ -63,7 +64,7 @@ export function validatePlan(
  * for the resource a call operates on (a `path`, `id`, `query`, ...). Used to
  * detect when a later call supersedes an earlier read of the same resource.
  */
-function primaryArg(argsJson: string): string | null {
+export function primaryArg(argsJson: string): string | null {
 	try {
 		const parsed: unknown = JSON.parse(argsJson);
 		if (parsed !== null && typeof parsed === "object") {

package/src/router/learned.ts

@@ -77,8 +77,13 @@ export function learnedVector(f: Partial<Features>): number[] {
 	];
 }
 
+/** What a learned model's positive class means. */
+export type LearnedLabel = "escalation" | "feedback";
+
 export interface LearnedModel {
 	version: number;
+	/** Positive class: the turn escalated (default), or the user judged it bad. */
+	label?: LearnedLabel;
 	trainedAtMs: number;
 	rows: number;
 	positives: number;
@@ -93,7 +98,12 @@ export interface LearnedModel {
 
 const sigmoid = (z: number): number => 1 / (1 + Math.exp(-z));
 
-/** P(escalate) for one turn under a model. */
+/** The positive-class name a decision reason should print for a model. */
+export function learnedRiskName(model: LearnedModel): string {
+	return model.label === "feedback" ? "bad" : "escalate";
+}
+
+/** P(positive class) for one turn under a model: p(escalate), or p(bad) for a feedback-labelled model. */
 export function predictRisk(model: LearnedModel, f: Partial<Features>): number {
 	const x = learnedVector(f);
 	let z = model.bias;