auto-model-router 0.2.12 → 0.2.13

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": "0.2.12",
+    "version": "0.2.13",
     "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": "0.2.12",
+      "version": "0.2.13",
       "author": {
         "name": "drewappling",
         "email": "drewappling@gmail.com"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "auto-model-router",
-  "version": "0.2.12",
+  "version": "0.2.13",
   "private": false,
   "description": "Local cost/complexity-aware model router for Oh My Pi, backed by OpenRouter",
   "type": "module",

package/src/server/turn.ts

@@ -335,7 +335,15 @@ export async function runTurn(
 								if (ttftMs === null) ttftMs = Date.now() - startedAt;
 								if (doxActive) assistantText += ev.delta;
 								break;
+							// A tool call is output too. Only text/reasoning used to stamp
+							// TTFT, so a dispatch that emitted nothing BUT tool calls —
+							// the dominant shape in an agentic loop, 83% of dispatches —
+							// recorded ttft_ms NULL and was then discarded by
+							// LATENCY_SELECT, which requires it. Latency/throughput
+							// scoring was therefore measuring the minority of turns that
+							// happened to narrate, and steering all the rest with it.
 							case "reasoning":
+							case "tool_call":
 								if (ttftMs === null) ttftMs = Date.now() - startedAt;
 								break;
 							case "finish":

package/test/turn.test.ts

@@ -694,3 +694,36 @@ describe("spend reaches the conversation total however the dispatch ends", () =>
 		expect(accrued.get("conv-test")?.spentUsd).toBeCloseTo(0.002, 10);
 	});
 });
+
+describe("latency measurement covers the work the router actually does", () => {
+	test("a tool-call-only dispatch still records TTFT", async () => {
+		// 83% of dispatches in an agentic loop finish with `tool_calls`, and 64.5%
+		// of those recorded ttft_ms NULL because only text/reasoning stamped it.
+		// LATENCY_SELECT requires ttft_ms, so throughput scoring (weight 0.75) was
+		// measuring the narrating minority and steering everything else with it.
+		const { router } = mkRouter([mkDecision("simple", "cheap/model", { maxTokens: 1, escalateTo: null })]);
+		const { upstream } = mkUpstream([
+			{
+				kind: "chunks",
+				chunks: [
+					startChunk("cheap/model"),
+					chunk([{ type: "tool_call", index: 0, id: "c1", name: "read", argsDelta: '{"path":"a.ts"}' }]),
+					finishChunk("tool_calls"),
+					usageChunk({ promptTokens: 50_000, completionTokens: 160 }, 0.004),
+				],
+			},
+		]);
+		const { ledger, entries } = mkLedger();
+		const { store } = mkConversations();
+		const { sink, errors } = mkSink();
+
+		await runTurn(mkReq(), sink, { config: mkConfig(), router, upstream, ledger, conversations: store, catalog, context: createDisabledBridge() }, new AbortController().signal);
+
+		expect(errors).toHaveLength(0);
+		expect(entries).toHaveLength(1);
+		expect(entries[0]?.finishReason).toBe("tool_calls");
+		// The qualifying condition for latency stats: a real, positive TTFT.
+		expect(entries[0]?.ttftMs).not.toBeNull();
+		expect(entries[0]?.ttftMs ?? -1).toBeGreaterThanOrEqual(0);
+	});
+});