@code-yeongyu/senpi-ai 2026.11.5-1 → 2026.11.6-2

package/dist/utils/prompt-cache-ttl.d.ts

@@ -23,6 +23,15 @@ export type PromptCacheLifetime = {
     readonly kind: "none";
 };
 export declare function isAnthropicApiBaseUrl(baseUrl: string): boolean;
+export declare function allowsAnthropicLongCacheRetention(model: Model<"anthropic-messages">): boolean;
+/**
+ * Whether this Anthropic-messages host applies `cache_control` at all, and therefore whether warming its
+ * prompt cache can achieve anything. The native host always does. A proxy does only when its model entry
+ * asserts cache support with `compat.supportsLongCacheRetention: true` — the documented assertion for a
+ * gateway that honours `cache_control` — because an unset or `false` flag cannot tell a host that caches
+ * for five minutes without the one-hour option apart from one that ignores `cache_control`.
+ */
+export declare function supportsAnthropicCacheWarming(model: Model<"anthropic-messages">): boolean;
 export declare function getAnthropicCompat(model: Model<"anthropic-messages">): Required<Omit<AnthropicMessagesCompat, "forceAdaptiveThinking" | "supportsMidConvoEffort" | "sessionAffinityFormat">> & Pick<AnthropicMessagesCompat, "sessionAffinityFormat">;
 export type ResolvedOpenAICompletionsCompat = Omit<Required<OpenAICompletionsCompat>, "vllmPriority" | "cacheControlFormat" | "toolCallFormat" | "deferredToolsMode" | "toolSchemaFlavor" | "supportsPromptCacheKey" | "chatTemplateArgs" | "supportsThinkingTokenBudget" | "thinkingTokenBudgetField" | "veniceParameters" | "supportsForcedToolChoice" | "supportsWebSearch" | "mixedStrictTools"> & {
     mixedStrictTools?: OpenAICompletionsCompat["mixedStrictTools"];

package/dist/utils/prompt-cache-ttl.js

@@ -50,6 +50,20 @@ function defaultSupportsToolReferences(model) {
     const minor = version[2] && version[2].length < 8 ? Number(version[2]) : 0;
     return major > 4 || (major === 4 && minor >= 5);
 }
+export function allowsAnthropicLongCacheRetention(model) {
+    return (model.compat?.supportsLongCacheRetention ??
+        (isAnthropicApiBaseUrl(model.baseUrl) && getAnthropicCompat(model).supportsLongCacheRetention));
+}
+/**
+ * Whether this Anthropic-messages host applies `cache_control` at all, and therefore whether warming its
+ * prompt cache can achieve anything. The native host always does. A proxy does only when its model entry
+ * asserts cache support with `compat.supportsLongCacheRetention: true` — the documented assertion for a
+ * gateway that honours `cache_control` — because an unset or `false` flag cannot tell a host that caches
+ * for five minutes without the one-hour option apart from one that ignores `cache_control`.
+ */
+export function supportsAnthropicCacheWarming(model) {
+    return isAnthropicApiBaseUrl(model.baseUrl) || model.compat?.supportsLongCacheRetention === true;
+}
 export function getAnthropicCompat(model) {
     // Auto-detect session affinity and cache control support from provider
     const isFireworks = model.provider === "fireworks";
@@ -381,9 +395,7 @@ export function resolvePromptCacheLifetime(model, env) {
             const retention = resolveAnthropicCacheRetention(anthropicModel.cacheRetention, env, "short");
             if (retention === "none")
                 return NO_PROMPT_CACHE;
-            return retention === "long" &&
-                isAnthropicApiBaseUrl(anthropicModel.baseUrl) &&
-                getAnthropicCompat(anthropicModel).supportsLongCacheRetention
+            return retention === "long" && allowsAnthropicLongCacheRetention(anthropicModel)
                 ? ttl(PROMPT_CACHE_TTL_LONG_SECONDS)
                 : ttl(PROMPT_CACHE_TTL_SHORT_SECONDS);
         }

package/dist/utils/provider-error.js

@@ -2,9 +2,10 @@ import { redactCredentialShapes, redactSecrets } from "./secret-redaction.js";
 export { redactCredentialShapes, redactSecrets };
 // A route that does not serve a configured model id: the OpenAI-compatible `model_not_found` code and OpenAI's
 // "The model `<id>` does not exist" prose, a gateway's "unknown provider for model", Anthropic's not_found_error whose
-// message names the model ("model: <id>"; any other not_found_error is not one), and Google's "models/<id> is not found".
+// message names the model ("model: <id>"; any other not_found_error is not one), Google's "models/<id> is not found",
+// and an OpenAI-compatible gateway's image allowlist reply "Model <id> is not supported on /v1/images/generations".
 // Every quantifier is bounded, so a long run never makes a match attempt expensive.
-const MODEL_CONFIG_ERROR = /\bmodel_not_found\b|\bunknown provider for model\b|\bthe model\s{1,8}[`'"]?[\w.:/-]{1,256}[`'"]?\s{1,8}does not exist\b|\bnot_found_error\b[^}]{0,120}?\bmodel:\s{0,8}[\w.:/-]|\bmodels\/[\w.:-]{1,256} is not found\b/i;
+const MODEL_CONFIG_ERROR = /\bmodel_not_found\b|\bunknown provider for model\b|\bmodel\s{1,8}[\w.:/-]{1,256}\s{1,8}is not supported on \/v1\/images\b|\bthe model\s{1,8}[`'"]?[\w.:/-]{1,256}[`'"]?\s{1,8}does not exist\b|\bnot_found_error\b[^}]{0,120}?\bmodel:\s{0,8}[\w.:/-]|\bmodels\/[\w.:-]{1,256} is not found\b/i;
 /** True when provider text says the route does not serve the configured model id (senpi#4282). */
 export function isModelConfigError(errorMessage) {
     return errorMessage !== undefined && MODEL_CONFIG_ERROR.test(errorMessage);

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@code-yeongyu/senpi-ai",
-	"version": "2026.11.5-1",
+	"version": "2026.11.6-2",
 	"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@2026.11.5-1",
+		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.11.6-2",
 		"@google/genai": "2.23.0",
 		"@smithy/node-http-handler": "4.12.1",
 		"google-auth-library": "10.6.2",