@code-yeongyu/senpi-ai 2026.11.2-1 → 2026.11.2-3

package/dist/providers/google-vertex.js

@@ -69,12 +69,12 @@ const vertexAuth = {
         const key = credential?.key ?? (await env("GOOGLE_CLOUD_API_KEY"));
         if (key)
             return { auth: { apiKey: key }, source: credential?.key ? "stored credential" : "GOOGLE_CLOUD_API_KEY" };
-        const adcPath = credential?.env?.GOOGLE_APPLICATION_CREDENTIALS ?? (await env("GOOGLE_APPLICATION_CREDENTIALS"));
+        const adcPath = credential?.env?.GOOGLE_APPLICATION_CREDENTIALS || (await env("GOOGLE_APPLICATION_CREDENTIALS"));
         signal.throwIfAborted();
         const hasCredentials = await ctx.fileExists(adcPath ?? VERTEX_ADC_PATH);
         signal.throwIfAborted();
-        const project = credential?.env?.GOOGLE_CLOUD_PROJECT ?? (await env("GOOGLE_CLOUD_PROJECT")) ?? (await env("GCLOUD_PROJECT"));
-        const location = credential?.env?.GOOGLE_CLOUD_LOCATION ?? (await env("GOOGLE_CLOUD_LOCATION"));
+        const project = credential?.env?.GOOGLE_CLOUD_PROJECT || (await env("GOOGLE_CLOUD_PROJECT")) || (await env("GCLOUD_PROJECT"));
+        const location = credential?.env?.GOOGLE_CLOUD_LOCATION || (await env("GOOGLE_CLOUD_LOCATION"));
         if (hasCredentials && project && location) {
             return credential
                 ? { auth: {}, env: credential.env, source: "stored credential" }

package/dist/utils/retry.js

@@ -88,11 +88,36 @@ const QUOTA_EXHAUSTION_PATTERNS = [
     escapeRegExp(ANTHROPIC_THIRD_PARTY_BILLING_MARKER),
 ];
 const QUOTA_EXHAUSTION_PATTERN = buildProviderErrorPattern(QUOTA_EXHAUSTION_PATTERNS);
+/**
+ * Quota markers that prove a spent account on their own. The bare `billing`
+ * substring is deliberately excluded: senpi#4632 shows a gateway 503 whose body
+ * merely mentions a billing lookup ("resolve Billing Scope: ... dial tcp: i/o
+ * timeout") — a dependency failure, not a spent account — and the substring
+ * also matches inside the code token `workspace_billing_scope_unavailable`.
+ */
+const STRONG_QUOTA_EXHAUSTION_PATTERN = buildProviderErrorPattern(QUOTA_EXHAUSTION_PATTERNS.filter((pattern) => pattern !== "billing"));
+/**
+ * Transient server-side evidence that gates bare billing prose (senpi#4632):
+ * a 5xx status token in the message, or the OpenAI-envelope `server_error`
+ * type. A genuine quota refusal carries an explicit code (insufficient_quota,
+ * usage_limit_reached, credits_required, ...) and keeps its verdict via the
+ * strong pattern above, whatever status wraps it.
+ */
+const TRANSIENT_SERVER_SIGNAL_PATTERN = /(?:^|[^0-9])(?:500|502|503|504|520|522|524)(?:[^0-9]|$)|"type"\s*:\s*"server_error"/i;
 export function isQuotaExhaustionMessage(errorMessage) {
-    return errorMessage !== undefined && QUOTA_EXHAUSTION_PATTERN.test(stripProviderRequestIds(errorMessage));
+    if (errorMessage === undefined)
+        return false;
+    const text = stripProviderRequestIds(errorMessage);
+    if (!QUOTA_EXHAUSTION_PATTERN.test(text))
+        return false;
+    // senpi#4632: bare "billing" prose on a transient 5xx is not a quota
+    // exhaustion unless a stronger quota marker is also present.
+    if (!STRONG_QUOTA_EXHAUSTION_PATTERN.test(text) && TRANSIENT_SERVER_SIGNAL_PATTERN.test(text)) {
+        return false;
+    }
+    return true;
 }
-const NON_RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
-    ...QUOTA_EXHAUSTION_PATTERNS,
+const NON_QUOTA_NON_RETRYABLE_PATTERNS = [
     // Request-shape rejections: the provider refused the payload we built, not the
     // work it describes. Gateways wrap these in whatever status they like — the
     // observed Apitopia/Kimi case arrives as `500 server_error: Invalid request:
@@ -109,6 +134,11 @@ const NON_RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     // Legacy text-only SIWC failures. Structured diagnostics take precedence below;
     // this code alone proves no reset time.
     "subscription_sharing_usage_limit_exceeded",
+];
+const NON_QUOTA_NON_RETRYABLE_PATTERN = buildProviderErrorPattern(NON_QUOTA_NON_RETRYABLE_PATTERNS);
+const NON_RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
+    ...QUOTA_EXHAUSTION_PATTERNS,
+    ...NON_QUOTA_NON_RETRYABLE_PATTERNS,
 ]);
 const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
     // Local credential/auth sidecar lock exhaustion is infrastructure contention,
@@ -534,8 +564,20 @@ export function classifyErrorMessage(errorMessage) {
     if (!errorMessage)
         return "unknown";
     const text = stripProviderRequestIds(errorMessage);
-    if (NON_RETRYABLE_PROVIDER_ERROR_PATTERN.test(text))
+    if (NON_RETRYABLE_PROVIDER_ERROR_PATTERN.test(text)) {
+        // senpi#4632: when the only non-retryable signal is bare "billing" prose
+        // on a transient 5xx, the failure is a server-side transient, not a
+        // spent account — let the retryable patterns (503, timeout, ...) decide.
+        // A stronger quota marker or a request-shape/subscription marker keeps
+        // the terminal verdict, whatever status wraps it.
+        if (isQuotaExhaustionMessage(text) || NON_QUOTA_NON_RETRYABLE_PATTERN.test(text)) {
+            return "non-retryable";
+        }
+        if (TRANSIENT_SERVER_SIGNAL_PATTERN.test(text)) {
+            return RETRYABLE_PROVIDER_ERROR_PATTERN.test(text) ? "retryable" : "unknown";
+        }
         return "non-retryable";
+    }
     if (RETRYABLE_PROVIDER_ERROR_PATTERN.test(text))
         return "retryable";
     return "unknown";

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@code-yeongyu/senpi-ai",
-	"version": "2026.11.02-1",
+	"version": "2026.11.2-3",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"type": "module",
 	"main": "./dist/index.js",
@@ -88,7 +88,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "0.127.0",
 		"@aws-sdk/client-bedrock-runtime": "3.1136.0",
-		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.11.02-1",
+		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.11.2-3",
 		"@google/genai": "2.23.0",
 		"@smithy/node-http-handler": "4.12.1",
 		"google-auth-library": "10.6.2",