auto-model-router 0.2.7 → 0.2.8

package/.omp-plugin/marketplace.json

@@ -7,14 +7,14 @@
   },
   "metadata": {
     "description": "auto-model-router: a local cost/complexity-aware model router for Oh My Pi, backed by OpenRouter",
-    "version": "0.2.7",
+    "version": "0.2.8",
     "pluginRoot": "."
   },
   "plugins": [
     {
       "name": "auto-model-router",
       "description": "Local cost/complexity-aware model router for Oh My Pi, backed by OpenRouter. Runs in-process, routes per turn by price and task complexity, with budget caps, mid-stream escalation, and cache-aware hysteresis.",
-      "version": "0.2.7",
+      "version": "0.2.8",
       "author": {
         "name": "drewappling",
         "email": "drewappling@gmail.com"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "auto-model-router",
-  "version": "0.2.7",
+  "version": "0.2.8",
   "private": false,
   "description": "Local cost/complexity-aware model router for Oh My Pi, backed by OpenRouter",
   "type": "module",

package/src/cost/ledger.ts

@@ -117,6 +117,13 @@ const TRUST_SELECT = `COUNT(*) AS attempts,
  * body streams once it starts. Errored/aborted and non-streaming rows (null
  * ttft) are excluded; throughput additionally requires a positive completion
  * count and elapsed time.
+ *
+ * Aggregated over a RECENT WINDOW (LATENCY_WINDOW_ROWS newest rows per slug),
+ * NOT all history: a model that degrades — e.g. deepseek-v4-flash collapsing
+ * from ~18 tok/s to ~7 — must move its score fast, or the penalty is drowned by
+ * hundreds of historical good rows and never demotes it (observed live: it kept
+ * 100% of coding at ~53s/turn despite latencyWeight=0.75). Trust (reliability)
+ * stays all-time; latency (volatile) is recency-weighted.
  */
 const LATENCY_SELECT = `COUNT(CASE WHEN ttft_ms IS NOT NULL AND ttft_ms > 0 AND error IS NULL THEN 1 END) AS samples,
 		AVG(CASE WHEN ttft_ms IS NOT NULL AND ttft_ms > 0 AND error IS NULL THEN ttft_ms END) AS ttft_ms,
@@ -127,6 +134,14 @@ const LATENCY_SELECT = `COUNT(CASE WHEN ttft_ms IS NOT NULL AND ttft_ms > 0 AND
 			AND json_extract(usage, '$.completionTokens') > 0
 			THEN latency_ms - ttft_ms END) AS elapsed_ms_sum`;
 
+/**
+ * Recent-rows window for latency stats: recent enough to react to a degrading
+ * model, wide enough to stay stable for a busy one. Rows are taken newest-first
+ * and then filtered by LATENCY_SELECT, so recent aborts naturally shrink the
+ * qualifying sample count (and can drop a model below latencyMinSamples).
+ */
+export const LATENCY_WINDOW_ROWS = 100;
+
 /**
  * Recovers the `UpstreamErrorKind` from the text turn.ts stored.
  *
@@ -231,8 +246,12 @@ export function createLedger(db: Database, cfg: RouterConfig): Ledger {
 	const trustStmt = db.query(`SELECT ${TRUST_SELECT} FROM ledger WHERE slug = ?`);
 	const trustHarnessStmt = db.query(`SELECT ${TRUST_SELECT} FROM ledger WHERE slug = ? AND harness_id = ?`);
 	const allTrustStmt = db.query(`SELECT slug, ${TRUST_SELECT} FROM ledger GROUP BY slug`);
-	const latencyStmt = db.query(`SELECT ${LATENCY_SELECT} FROM ledger WHERE slug = ?`);
-	const latencyHarnessStmt = db.query(`SELECT ${LATENCY_SELECT} FROM ledger WHERE slug = ? AND harness_id = ?`);
+	const latencyStmt = db.query(
+		`SELECT ${LATENCY_SELECT} FROM (SELECT * FROM ledger WHERE slug = ? ORDER BY created_at_ms DESC LIMIT ${LATENCY_WINDOW_ROWS})`,
+	);
+	const latencyHarnessStmt = db.query(
+		`SELECT ${LATENCY_SELECT} FROM (SELECT * FROM ledger WHERE slug = ? AND harness_id = ? ORDER BY created_at_ms DESC LIMIT ${LATENCY_WINDOW_ROWS})`,
+	);
 	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 cacheMetaStmt = db.query("SELECT fetched_at_ms FROM catalog_cache WHERE id = 1");

package/test/trust-attribution.test.ts

@@ -1,7 +1,7 @@
 import { describe, expect, test } from "bun:test";
 
 import { loadConfig } from "../src/config/load.ts";
-import { createLedger } from "../src/cost/ledger.ts";
+import { createLedger, LATENCY_WINDOW_ROWS } from "../src/cost/ledger.ts";
 import { EMPTY_USAGE, type LedgerEntry } from "../src/cost/types.ts";
 import { openDb } from "../src/util/sqlite.ts";
 
@@ -198,6 +198,22 @@ describe("latency signal", () => {
 		expect(l?.tokensPerSec).toBeCloseTo(100, 5);
 		expect(l?.samples).toBe(2);
 	});
+
+	test("throughput and ttft track a recent window, not the lifetime average", () => {
+		// Old rows are fast; the recent window is slow. Latency must reflect the
+		// recent (slow) behaviour so a degraded model is penalised, not masked by
+		// its history. Lifetime blend here would be ~57 tok/s; the window is 10.
+		const rows: Array<Partial<LedgerEntry>> = [];
+		let t = 1;
+		for (let i = 0; i < 50; i++)
+			rows.push({ createdAtMs: t++, ttftMs: 200, latencyMs: 1200, usage: { ...EMPTY_USAGE, completionTokens: 1000 } });
+		for (let i = 0; i < LATENCY_WINDOW_ROWS; i++)
+			rows.push({ createdAtMs: t++, ttftMs: 4000, latencyMs: 14000, usage: { ...EMPTY_USAGE, completionTokens: 100 } });
+		const l = latencyOf(rows);
+		expect(l?.samples).toBe(LATENCY_WINDOW_ROWS);
+		expect(l?.tokensPerSec).toBeCloseTo(10, 0);
+		expect(l?.ttftMs).toBeCloseTo(4000, 5);
+	});
 });
 
 describe("v4 migration", () => {