auto-model-router 0.24.0 → 0.25.0

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": "0.24.0",
+    "version": "0.25.0",
     "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": "0.24.0",
+      "version": "0.25.0",
       "author": {
         "name": "drewappling",
         "email": "drewappling@gmail.com"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "auto-model-router",
-  "version": "0.24.0",
+  "version": "0.25.0",
   "private": false,
   "description": "Local cost/complexity-aware model router for Oh My Pi, backed by OpenRouter",
   "type": "module",

package/src/router/candidates.ts

@@ -254,20 +254,33 @@ export function buildCandidates(args: BuildCandidatesArgs): { candidates: Candid
 		// Price ceilings at the ACTUAL prompt size: long-context overrides can
 		// push a model over the ceiling exactly when conversations get long.
 		// Catalog prices are per-token; ceilings are per million tokens.
+		//
+		// The ceiling is compared against the BIASED price, because that is what the turn
+		// actually costs this deployment: capacity already paid for — an Ollama plan's
+		// included credits, a Claude Pro/Max subscription — carries the list price of the
+		// twin it is priced from, and a subscription model at $2-5/Mtok list would be
+		// thrown out here on every cheap tier before `costBias` was ever consulted. That
+		// made the bias silently inert: setting it to 0.00001 changed no decision at all.
 		const price = priceAt(model, Math.max(1, features.promptTokens));
-		if (!relaxPrice && priceCeiling !== undefined && price.prompt * 1e6 > priceCeiling) {
+		const providerBias =
+			snapshot.providerBias?.[model.provider] ??
+			(model.provider === "ollama" ? cfg.ollama.costBias : (cfg.upstreams.find((u) => u.id === model.provider)?.costBias ?? 1));
+		const biasedPrompt = price.prompt * providerBias;
+		const biasedCompletion = price.completion * providerBias;
+		const biasNote = providerBias === 1 ? "" : ` (×${providerBias} bias on $${(price.prompt * 1e6).toFixed(2)} list)`;
+		if (!relaxPrice && priceCeiling !== undefined && biasedPrompt * 1e6 > priceCeiling) {
 			rejected.push({
 				slug,
 				reason: "over_price_ceiling",
-				detail: `input $${(price.prompt * 1e6).toFixed(2)}/Mtok > ceiling $${priceCeiling.toFixed(2)}`,
+				detail: `input $${(biasedPrompt * 1e6).toFixed(2)}/Mtok > ceiling $${priceCeiling.toFixed(2)}${biasNote}`,
 			});
 			continue;
 		}
-		if (!relaxPrice && tierCfg.maxOutputPerMtok !== undefined && price.completion * 1e6 > tierCfg.maxOutputPerMtok) {
+		if (!relaxPrice && tierCfg.maxOutputPerMtok !== undefined && biasedCompletion * 1e6 > tierCfg.maxOutputPerMtok) {
 			rejected.push({
 				slug,
 				reason: "over_price_ceiling",
-				detail: `output $${(price.completion * 1e6).toFixed(2)}/Mtok > ceiling $${tierCfg.maxOutputPerMtok}`,
+				detail: `output $${(biasedCompletion * 1e6).toFixed(2)}/Mtok > ceiling $${tierCfg.maxOutputPerMtok}${biasNote}`,
 			});
 			continue;
 		}
@@ -357,14 +370,9 @@ export function buildCandidates(args: BuildCandidatesArgs): { candidates: Candid
 			filters.escalationCostWeight > 0 && args.escalationUsdPerPromptToken !== undefined
 				? filters.escalationCostWeight * escalationRate * args.escalationUsdPerPromptToken * features.promptTokens
 				: 0;
-		// Provider bias: capacity already paid for — an Ollama plan's included credits, a
-		// Claude Pro/Max subscription — is money already spent, so an operator may value it
-		// below list price in ranking. The ledger still records list price.
-		// The snapshot carries the LIVE bias (credit-aware, and each named upstream's own);
-		// the static config value is the fallback for snapshots built without one.
-		const providerBias =
-			snapshot.providerBias?.[model.provider] ??
-			(model.provider === "ollama" ? cfg.ollama.costBias : (cfg.upstreams.find((u) => u.id === model.provider)?.costBias ?? 1));
+		// `providerBias` is computed with the price ceilings above — capacity already paid for
+		// is money already spent, so it is valued below list in ranking AND against the
+		// ceilings. The ledger still records list price either way.
 		const effectiveUsd = (fc.expectedUsd / Math.max(trustScore, 0.5) + escalationUsd) * latencyMult * providerBias;
 		// Score is assigned in a SECOND PASS below: both qualityNormalization and
 		// capabilityFloorUsd are properties of the candidate SET, not of one

package/test/tier-plan.test.ts

@@ -358,6 +358,22 @@ describe("adaptive price ceilings", () => {
 		expect(on.candidates.map((c) => c.model.slug)).not.toContain("a/4");
 		expect(on.rejected.some((r) => r.slug === "a/4" && r.reason === "over_price_ceiling")).toBe(true);
 	});
+
+	test("a price ceiling judges the BIASED price, so prepaid capacity is not thrown out on list", () => {
+		// A subscription upstream inherits its OpenRouter twin's list price ($4/Mtok here) and is
+		// discounted by `costBias` because the capacity is already paid for. Judging the ceiling on
+		// list threw it out before the bias was ever read, which made the bias entirely inert.
+		const snap = snapshot(priced);
+		const biased = { ...snap, providerBias: { [snap.models[0]!.provider]: 0.1 } };
+		const run = (s: typeof snap) =>
+			buildCandidates({ req, features, tier: "moderate", task: "coding", snapshot: s, ledger: null, cfg: { ...BASE, adaptivePriceCeilings: true }, expectedCompletionTokens: 512, warmSlug: null });
+		// Unbiased: the band tightens moderate to $3 and a/4 is over it.
+		expect(run(snap).rejected.some((r) => r.slug === "a/4" && r.reason === "over_price_ceiling")).toBe(true);
+		// Biased ×0.1: $4 list is $0.40 to this deployment, so it clears the same ceiling.
+		const on = run(biased);
+		expect(on.candidates.map((c) => c.model.slug)).toContain("a/4");
+		expect(on.rejected.some((r) => r.slug === "a/4")).toBe(false);
+	});
 });
 
 describe("quality normalization and capability floor (benchmark findings 4/6)", () => {