@oh-my-pi/pi-catalog 17.1.7 → 17.1.8

package/CHANGELOG.md

@@ -2,6 +2,16 @@
 
 ## [Unreleased]
 
+## [17.1.8] - 2026-07-28
+
+### Added
+
+- Added `resolveVertexEndpointHost(location)` utility to resolve the correct Vertex AI API endpoint hostnames for global, multi-region, and regional locations.
+
+### Fixed
+
+- Fixed an issue where `calculateCost` under-reported Anthropic cache-write costs by honoring the `usage.cttl` breakdown to correctly price 1-hour retention writes at 2x the base input rate.
+
 ## [17.1.7] - 2026-07-27
 
 ### Added

package/dist/types/hosts.d.ts

@@ -136,6 +136,19 @@ export declare function modelMatchesHost(model: {
     provider: string;
     baseUrl: string;
 }, host: KnownHost): boolean;
+/**
+ * Hostname for a Vertex AI GenerateContent / rawPredict / OpenAI-compat
+ * request for the given location.
+ *
+ * - `global` → global endpoint (`aiplatform.googleapis.com`)
+ * - `eu` / `us` multi-regions → REP endpoints (`aiplatform.{eu|us}.rep.googleapis.com`)
+ * - every other location → regional (`{location}-aiplatform.googleapis.com`)
+ *
+ * Multi-region codes do NOT follow the regional `{location}-aiplatform` pattern;
+ * interpolating them that way yields hosts like `eu-aiplatform.googleapis.com`
+ * that 404.
+ */
+export declare function resolveVertexEndpointHost(location: string): string;
 /** Vertex AI express-mode OpenAI-compatible endpoint (`…/endpoints/openapi`). */
 export declare function isVertexExpressOpenAIUrl(baseUrl: string): boolean;
 /** Vertex AI Anthropic raw-predict endpoints (`:streamRawPredict` / `:rawPredict`). */

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "17.1.7",
+	"version": "17.1.8",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -34,12 +34,12 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.1",
-		"@oh-my-pi/pi-utils": "17.1.7",
+		"@oh-my-pi/pi-utils": "17.1.8",
 		"arktype": "2.2.3",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "17.1.7",
+		"@oh-my-pi/pi-ai": "17.1.8",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/hosts.ts

@@ -111,6 +111,24 @@ function includesAsciiCaseInsensitive(value: string, lowerNeedle: string): boole
 
 // --- Endpoint-shape predicates (URL path/verb shapes, not vendor hosts) ---
 
+/**
+ * Hostname for a Vertex AI GenerateContent / rawPredict / OpenAI-compat
+ * request for the given location.
+ *
+ * - `global` → global endpoint (`aiplatform.googleapis.com`)
+ * - `eu` / `us` multi-regions → REP endpoints (`aiplatform.{eu|us}.rep.googleapis.com`)
+ * - every other location → regional (`{location}-aiplatform.googleapis.com`)
+ *
+ * Multi-region codes do NOT follow the regional `{location}-aiplatform` pattern;
+ * interpolating them that way yields hosts like `eu-aiplatform.googleapis.com`
+ * that 404.
+ */
+export function resolveVertexEndpointHost(location: string): string {
+	if (location === "global") return "aiplatform.googleapis.com";
+	if (location === "eu" || location === "us") return `aiplatform.${location}.rep.googleapis.com`;
+	return `${location}-aiplatform.googleapis.com`;
+}
+
 /** Vertex AI express-mode OpenAI-compatible endpoint (`…/endpoints/openapi`). */
 export function isVertexExpressOpenAIUrl(baseUrl: string): boolean {
 	return baseUrl.includes("/endpoints/openapi");

package/src/models.ts

@@ -48,10 +48,38 @@ export function calculateCost<TApi extends Api>(model: Model<TApi>, usage: Usage
 	usage.cost.input = (model.cost.input / 1000000) * (usage.input + (orchestration?.input ?? 0));
 	usage.cost.output = (model.cost.output / 1000000) * (usage.output + (orchestration?.output ?? 0));
 	usage.cost.cacheRead = (model.cost.cacheRead / 1000000) * (usage.cacheRead + (orchestration?.cacheRead ?? 0));
-	usage.cost.cacheWrite = (model.cost.cacheWrite / 1000000) * usage.cacheWrite;
+	usage.cost.cacheWrite = cacheWriteCost(model, usage);
 	usage.cost.total = usage.cost.input + usage.cost.output + usage.cost.cacheRead + usage.cost.cacheWrite;
 	return usage.cost;
 }
+
+/**
+ * Price cache-write tokens, honoring the TTL breakdown when the provider reports one.
+ *
+ * `model.cost.cacheWrite` is the 5-minute write rate (Anthropic bills 5m writes at
+ * 1.25x base input). When `usage.cttl` is present the write mixes 5m and 1h
+ * breakpoints — omp defaults to 1h retention on first-party Anthropic, and 1h writes
+ * bill at 2x base input — so each component is priced at its own rate instead of the
+ * flat 5m rate. Deriving 1h from `input * 2` (Anthropic's published multiplier) is
+ * model-independent and stays correct even for legacy entries whose stored
+ * `cacheWrite` scalar drifts from 1.25x input. Providers that omit `cttl`
+ * (everyone but Anthropic) keep the flat-rate calculation.
+ *
+ * The breakdown is documented to sum to `usage.cacheWrite`, but the two are written
+ * from independent wire fields (`cache_creation` vs `cache_creation_input_tokens`),
+ * so any unattributed remainder is priced at the flat rate instead of being dropped:
+ * a partial or stale breakdown must never make write tokens free.
+ */
+function cacheWriteCost<TApi extends Api>(model: Model<TApi>, usage: Usage): number {
+	const rate5m = model.cost.cacheWrite / 1000000;
+	const cttl = usage.cttl;
+	if (!cttl) return rate5m * usage.cacheWrite;
+	const fiveMinute = cttl.ephemeral5m ?? 0;
+	const oneHour = cttl.ephemeral1h ?? 0;
+	const residual = Math.max(0, usage.cacheWrite - fiveMinute - oneHour);
+	return rate5m * (fiveMinute + residual) + ((model.cost.input * 2) / 1000000) * oneHour;
+}
+
 /**
  * Check if two models are equal by comparing both their id and provider.
  * Returns false if either model is null or undefined.