优化: aichat效率批(save无变化捷径/断路器尾窗/system缓存指纹/会话列表刷新收敛) + 新增 LLM prompt caching(cache_control开关默认关/system稳定段易变段分离治缓存命中率) + 销账

This commit is contained in:
lxy
2026-08-09 21:35:13 +08:00
parent 4df91ed155
commit 7eb2e25ed5
13 changed files with 698 additions and 180 deletions
+15 -1
View File
@@ -18,6 +18,20 @@ use crate::provider::{tool_call_id_or_fallback, StreamChunk, TokenUsage, ToolCal
// Anthropic API 请求/响应结构体
// ============================================================
/// Anthropic `system` 字段形态(受 prompt caching 开关 `ANTHROPIC_CACHE_ENABLED` 控制):
/// - `Plain`:纯字符串。开关关时保持原形态,兼容非官方网关(GLM 订阅端点 / 任意 Messages API
/// 代理——部分网关会拒收 cache_control 字段)。
/// - `Cached`:blocks 数组 `[{type:text, text, cache_control:{type:ephemeral}}]`。开关开时启用
/// Anthropic prompt caching,让 system 稳定段在支持缓存的端点上命中。
#[derive(Debug, Clone, Serialize)]
#[serde(untagged)]
pub(crate) enum SystemBlock {
/// 纯字符串形态(兼容网关,开关关)
Plain(String),
/// blocks 数组形态(含 cache_control,开关开)
Cached(Vec<serde_json::Value>),
}
/// Anthropic 请求体
#[derive(Debug, Clone, Serialize)]
pub(crate) struct AnthropicRequest {
@@ -25,7 +39,7 @@ pub(crate) struct AnthropicRequest {
pub messages: Vec<serde_json::Value>,
pub max_tokens: u32,
#[serde(skip_serializing_if = "Option::is_none")]
pub system: Option<String>,
pub system: Option<SystemBlock>,
#[serde(skip_serializing_if = "Option::is_none")]
pub temperature: Option<f32>,
pub stream: bool,