优化: 缓存命中字段解析与提示词前缀稳定化

- openai 兼容口径补解析嵌套 cached_tokens
- cache 显示区分无数据与真零,估算轮不渲染
- 轨迹展开折叠空过程记录行降噪
- token 统计行恒完整可见不截断
- openai_compat 纳入 system 稳定易变拆分
- 系统警示与主题标记改尾部注入不位移前缀
- 工具结果视图压缩按入库时刻冻结
- 会话内工具集冻结仅增不减
- anthropic 缓存补消息级断点,设置页加开关
- 补 intent 测试对 Agent domain 覆盖

src/i18n/en/settings.ts

@@ -203,6 +203,11 @@ export default {
     pathAuthBypassConfirmTitle: 'Enable "Trust all directories"?',
     pathAuthBypassConfirmBody: 'Once enabled, AI tools will no longer ask for per-access authorization and can directly reach any path, including sensitive directories such as .ssh and .aws. Sensitive credentials may end up in the conversation context; proceed at your own risk. Are you sure you want to enable it?',
 
+    // Anthropic prompt caching (Settings - AI Model; KV df-ai-anthropic-cache, default off, effective immediately)
+    panelAnthropicCache: 'AI Prompt Caching',
+    labelAnthropicCache: 'Enable prompt caching',
+    descAnthropicCache: 'When enabled, Anthropic-protocol endpoints get explicit cache breakpoints on the system prompt and conversation history, greatly reducing token cost for long sessions; OpenAI-compatible endpoints also get prefix-stable splitting (volatile content moves to the end of the message stream so automatic prefix caches can hit). Only enable for endpoints that support caching; gateways without support may reject the cache_control field.',
+
     // ===== ScriptNode command execution safety =====
     panelScriptSafety: '🛡️ Command Execution Safety',
     descScriptSafety: 'Configure ScriptNode command whitelist/blacklist. Command name is the first token (e.g. `rm -rf /` -> `rm`), compared trimmed and lowercase. Blacklist takes precedence over whitelist.',