@code-yeongyu/senpi-ai 2026.9.29-5 → 2026.9.30

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

@@ -323,6 +323,38 @@ function hasOpenAIExtendedPromptCache(model) {
     }
     return OPENAI_EXTENDED_CACHE_MODEL_ID.test(model.id);
 }
+/** Env that puts Claude Code on API-key, gateway or cloud billing, where it caches the main conversation for 5 minutes. */
+const CLAUDE_CODE_API_BILLING_ENV = [
+    "ANTHROPIC_API_KEY",
+    "ANTHROPIC_AUTH_TOKEN",
+    "ANTHROPIC_BASE_URL",
+    "CLAUDE_CODE_USE_BEDROCK",
+    "CLAUDE_CODE_USE_VERTEX",
+    "CLAUDE_CODE_USE_FOUNDRY",
+];
+function isEnabledFlag(value) {
+    return value !== undefined && !/^(?:|0|false|no|off)$/i.test(value.trim());
+}
+/**
+ * The Claude SDK lane's cache TTL is chosen by Claude Code, not senpi: `CLAUDE_CODE_PROMPT_CACHE_TTL`
+ * (`5m` | `1h`) wins, then `FORCE_PROMPT_CACHING_5M` and `ENABLE_PROMPT_CACHING_1H`; otherwise a Claude
+ * subscription gets 1 hour and API-key, gateway or cloud billing gets 5 minutes. A subscription past its
+ * usage limits also drops to 5 minutes, which nothing here can observe.
+ */
+function claudeCodePromptCacheTtlSeconds(env) {
+    const explicit = getProviderEnvValue("CLAUDE_CODE_PROMPT_CACHE_TTL", env)?.trim().toLowerCase();
+    if (explicit === "5m")
+        return PROMPT_CACHE_TTL_SHORT_SECONDS;
+    if (explicit === "1h")
+        return PROMPT_CACHE_TTL_LONG_SECONDS;
+    if (isEnabledFlag(getProviderEnvValue("FORCE_PROMPT_CACHING_5M", env)))
+        return PROMPT_CACHE_TTL_SHORT_SECONDS;
+    if (isEnabledFlag(getProviderEnvValue("ENABLE_PROMPT_CACHING_1H", env)))
+        return PROMPT_CACHE_TTL_LONG_SECONDS;
+    return CLAUDE_CODE_API_BILLING_ENV.some((name) => getProviderEnvValue(name, env) !== undefined)
+        ? PROMPT_CACHE_TTL_SHORT_SECONDS
+        : PROMPT_CACHE_TTL_LONG_SECONDS;
+}
 /**
  * Classify the active model's prompt-cache lifetime from the provider's documented cache contract.
  * `cacheRetention: "none"` (or `PI_CACHE_RETENTION` resolving to it) always wins.
@@ -330,8 +362,7 @@ function hasOpenAIExtendedPromptCache(model) {
 export function resolvePromptCacheLifetime(model, env) {
     switch (model.api) {
         case "claude-sdk-oauth":
-            // The Claude SDK owns prompt caching for this lane and uses Anthropic's default 5m TTL.
-            return ttl(PROMPT_CACHE_TTL_SHORT_SECONDS);
+            return ttl(claudeCodePromptCacheTtlSeconds(env));
         case "anthropic-messages": {
             const anthropicModel = model;
             const retention = resolveAnthropicCacheRetention(anthropicModel.cacheRetention, env, "short");

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@code-yeongyu/senpi-ai",
-	"version": "2026.9.29-5",
+	"version": "2026.9.30",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"type": "module",
 	"main": "./dist/index.js",
@@ -85,7 +85,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "0.127.0",
 		"@aws-sdk/client-bedrock-runtime": "3.1136.0",
-		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.29-5",
+		"@earendil-works/pi-telemetry": "npm:@code-yeongyu/senpi-telemetry@2026.9.30",
 		"@google/genai": "2.23.0",
 		"@smithy/node-http-handler": "4.12.1",
 		"http-proxy-agent": "9.1.0",