auto-model-router 0.27.0 → 0.28.0

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": "0.27.0",
+    "version": "0.28.0",
     "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": "0.27.0",
+      "version": "0.28.0",
       "author": {
         "name": "drewappling",
         "email": "drewappling@gmail.com"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "auto-model-router",
-  "version": "0.27.0",
+  "version": "0.28.0",
   "private": false,
   "description": "Local cost/complexity-aware model router for Oh My Pi, backed by OpenRouter",
   "type": "module",

package/src/catalog/ollama-catalog.ts

@@ -125,16 +125,22 @@ export function buildOllamaModels(args: BuildOllamaArgs): CatalogModel[] {
 	for (const l of listings) {
 		if (!l.isCloud && !cfg.includeLocal) continue;
 		const priceName = l.remoteModel ?? l.id;
-		const rate = ollamaRateFor(priceName, cfg.prices);
-		if (rate === null) {
-			skipped.push(l.id);
-			continue;
-		}
 		// A pin may name the tagged cloud name, its base, or the listed id.
 		const bare = bareCloudName(priceName);
 		const base = bare.includes(":") ? bare.slice(0, bare.indexOf(":")) : bare;
 		const pinned = cfg.twins[bare] ?? cfg.twins[base] ?? cfg.twins[l.id];
 		const twin = (pinned !== undefined ? bySlug.get(pinned) : undefined) ?? twins.get(ollamaTwinKey(priceName)) ?? null;
+		// Ollama publishes no prices, so they come from a static table — which means a model
+		// Ollama ships today is INVISIBLE until that table gains an entry, however good it is.
+		// Measured: deepseek-v4.1-flash was listed by /api/tags and dropped here, so routing
+		// never saw it. The OpenRouter twin sells the same weights per token, so it is the
+		// honest stand-in until the table catches up; a model with neither is still skipped.
+		const rate = ollamaRateFor(priceName, cfg.prices);
+		const price = rate !== null ? toPrice(rate.rate) : twin !== null ? { ...twin.price } : null;
+		if (price === null) {
+			skipped.push(`${l.id} (no price and no twin)`);
+			continue;
+		}
 		const contextLength = l.contextLength ?? twin?.contextLength ?? null;
 		if (contextLength === null) {
 			skipped.push(`${l.id} (no context length)`);
@@ -156,7 +162,7 @@ export function buildOllamaModels(args: BuildOllamaArgs): CatalogModel[] {
 			// Ollama's OpenAI-compatible endpoint documents `tool_choice` as unsupported.
 			supportsToolChoice: false,
 			inputModalities: modalities,
-			price: toPrice(rate.rate),
+			price,
 			priceTiers: [],
 			quality: twin === null ? {} : { ...twin.quality },
 			tokenizer: twin?.tokenizer ?? "Other",
@@ -167,7 +173,10 @@ export function buildOllamaModels(args: BuildOllamaArgs): CatalogModel[] {
 		if (twin?.maxCompletionTokens !== undefined) model.maxCompletionTokens = twin.maxCompletionTokens;
 		out.push(model);
 	}
-	if (skipped.length > 0) log?.debug("ollama models skipped (no price or context)", { skipped: skipped.join(", ") });
+	// A model the provider offers and the router refuses to route is operationally
+	// significant, not a debug detail: at `debug` this was invisible, and a newly published
+	// model stayed unroutable with nothing in the log to say so.
+	if (skipped.length > 0) log?.warn("ollama models listed but not routable", { skipped: skipped.join(", ") });
 	return out;
 }
 

package/src/server/http.ts

@@ -723,14 +723,34 @@ export function startServer(cfg: RouterConfig): StartedServer {
 					const named = Array.isArray(body?.anchors) ? (body.anchors as unknown[]).filter((a): a is string => typeof a === "string") : [];
 					const anchors = named.length > 0 ? [...new Set(named)] : pickAnchors(models, slug);
 					if (anchors.length < MIN_ANCHORS) return wireErrorResponse({ status: 422, code: "invalid_request_error", message: `need at least ${MIN_ANCHORS} scored, tool-capable anchor models to calibrate against` });
+					// A rate-limited dispatch is not a failed task. Ollama Cloud answers "too many
+					// concurrent requests" well below four models in flight, and the runner folds a
+					// throwing completion in as grade 0 — which would score a provider's throttle as
+					// the model being wrong. Retry with backoff, and keep the default concurrency
+					// low enough that the throttle is rarely reached in the first place.
 					const complete: Completer = async (target, messages) => {
-						const out = await upstream.complete({ model: target, stream: false, temperature: 0, max_tokens: 1024, messages }, AbortSignal.timeout(120_000));
-						return out.text;
+						let last: unknown = null;
+						for (let attempt = 0; attempt < 4; attempt++) {
+							if (attempt > 0) {
+								const { promise, resolve } = Promise.withResolvers<void>();
+								setTimeout(resolve, attempt * 4000);
+								await promise;
+							}
+							try {
+								const out = await upstream.complete({ model: target, stream: false, temperature: 0, max_tokens: 1024, messages }, AbortSignal.timeout(120_000));
+								return out.text;
+							} catch (err) {
+								last = err;
+							}
+						}
+						throw last instanceof Error ? last : new Error(String(last));
 					};
 					const judgeSlug = typeof body?.judge === "string" && body.judge !== "" ? body.judge : "";
+					const asked = typeof body?.concurrency === "number" ? Math.floor(body.concurrency) : 2;
 					const results = await runEval({
 						slugs: [slug, ...anchors],
 						complete,
+						concurrency: Math.min(Math.max(1, asked), 8),
 						...(judgeSlug === "" ? {} : { judge: makeJudge(complete, judgeSlug) }),
 					});
 					const target = results[0]!;

package/test/ollama.test.ts

@@ -164,6 +164,22 @@ describe("buildOllamaModels", () => {
 		expect(ds.quality).toEqual({});
 	});
 
+	test("a model Ollama lists but the price table does not know is priced from its twin", () => {
+		// Ollama publishes no prices, so they come from a static table. deepseek-v4.1-flash was
+		// listed by /api/tags and dropped here for want of an entry, so routing never saw it at
+		// all — a model the provider was actively offering. The twin sells the same weights.
+		const unpriced = { id: "glm-5.3-flash:cloud", remoteModel: "not-in-the-price-table", isCloud: true, contextLength: 262_144, capabilities: ["tools"], modifiedAtMs: 0 };
+		const twinPriced = buildOllamaModels({ listings: [{ ...unpriced, remoteModel: "glm-5.3-flash" }], openrouter: OR_MODELS, cfg: OLLAMA, log });
+		const or = OR_MODELS.find((m) => m.slug === "z-ai/glm-5.3-flash")!;
+		expect(twinPriced).toHaveLength(1);
+		expect(twinPriced[0]!.price.prompt).toBeCloseTo(0.15 / 1e6, 12);
+
+		// Unknown to the table AND matching no twin: still skipped, never priced at zero.
+		const orphan = buildOllamaModels({ listings: [{ ...unpriced, id: "brand-new-thing:cloud", remoteModel: "brand-new-thing" }], openrouter: OR_MODELS, cfg: OLLAMA, log });
+		expect(orphan).toEqual([]);
+		expect(or.price.prompt).toBeGreaterThan(0);
+	});
+
 	test("a pinned twin beats the name match", () => {
 		const cfg: OllamaConfig = { ...OLLAMA, twins: { "deepseek-v4-pro": "moonshotai/kimi-k3" } };
 		const ds = buildOllamaModels({ listings: listings(), openrouter: OR_MODELS, cfg, log }).find((m) => m.slug.startsWith("ollama/deepseek"))!;