admin/CC-Switch
1
0
Fork 0
mirror of https://github.com/farion1231/cc-switch.git synced 2026-07-28 00:35:32 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(usage): account for Anthropic cache write TTLs

Browse Source 
Parse and retain Anthropic's ephemeral 5-minute and 1-hour cache-creation token buckets while preserving the existing aggregate cache-write metric for compatibility.

Price 1-hour writes at the documented premium relative to the configured 5-minute write rate, clamp inconsistent provider details safely, and include TTL buckets in usage diagnostics.

Persist 1-hour cache-write tokens with schema version 14 so zero-cost backfills and later pricing updates retain the original TTL semantics. Keep session import paths compatible through zero-valued detail fields.
This commit is contained in:
Jason
2026-07-12 12:19:36 +08:00
parent b9263a8040
commit 13e7c1fcc4
11 changed files with 258 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src-tauri/src/proxy/usage/parser.rs
+84 -1
View File
@@ -27,6 +27,22 @@ fn openai_cache_write_tokens(usage: &Value) -> u32 {
.unwrap_or(0) as u32
}
fn cache_creation_ttl_tokens(usage: &Value) -> (u32, u32) {
let details = usage
.get("cache_creation")
.or_else(|| usage.pointer("/input_tokens_details/cache_creation"))
.or_else(|| usage.pointer("/prompt_tokens_details/cache_creation"));
let five_minutes = details
.and_then(|value| value.get("ephemeral_5m_input_tokens"))
.and_then(Value::as_u64)
.unwrap_or(0) as u32;
let one_hour = details
.and_then(|value| value.get("ephemeral_1h_input_tokens"))
.and_then(Value::as_u64)
.unwrap_or(0) as u32;
(five_minutes, one_hour)
}
/// Session 日志 request_id 前缀，与 `session_usage.rs` 中的格式保持一致
pub const SESSION_REQUEST_ID_PREFIX: &str = "session:";
@@ -37,6 +53,13 @@ pub struct TokenUsage {
pub output_tokens: u32,
pub cache_read_tokens: u32,
pub cache_creation_tokens: u32,
/// Anthropic cache-write TTL detail. The aggregate above remains the stable
/// public/storage metric; these buckets make 1-hour writes billable at 2x input
/// instead of the 5-minute 1.25x rate.
#[serde(default)]
pub cache_creation_5m_tokens: u32,
#[serde(default)]
pub cache_creation_1h_tokens: u32,
/// 从响应中提取的实际模型名称（如果可用）
pub model: Option<String>,
/// 从响应中提取的消息 ID（用于跨源去重）
@@ -47,6 +70,23 @@ pub struct TokenUsage {
}
impl TokenUsage {
/// Return mutually exclusive cache-write buckets, clamped to the aggregate.
/// Providers that do not expose TTL detail remain in `unspecified` and keep the
/// legacy cache-creation price.
pub fn normalized_cache_creation_buckets(&self) -> (u32, u32, u32) {
let one_hour = self
.cache_creation_1h_tokens
.min(self.cache_creation_tokens);
let five_minutes = self
.cache_creation_5m_tokens
.min(self.cache_creation_tokens.saturating_sub(one_hour));
let unspecified = self
.cache_creation_tokens
.saturating_sub(one_hour)
.saturating_sub(five_minutes);
(unspecified, five_minutes, one_hour)
}
/// 生成与 session 日志共享的 request_id，用于跨源去重。
/// 有 message_id 时返回 `session:{id}`，否则回退到随机 UUID。
pub fn dedup_request_id(&self) -> String {
@@ -83,6 +123,7 @@ impl TokenUsage {
/// 从 Claude API 非流式响应解析
pub fn from_claude_response(body: &Value) -> Option<Self> {
let usage = body.get("usage")?;
let (cache_creation_5m_tokens, cache_creation_1h_tokens) = cache_creation_ttl_tokens(usage);
// 提取响应中的模型名称
let model = body
.get("model")
@@ -104,6 +145,8 @@ impl TokenUsage {
.get("cache_creation_input_tokens")
.and_then(|v| v.as_u64())
.unwrap_or(0) as u32,
cache_creation_5m_tokens,
cache_creation_1h_tokens,
model,
message_id,
})
@@ -134,6 +177,8 @@ impl TokenUsage {
}
}
if let Some(msg_usage) = event.get("message").and_then(|m| m.get("usage")) {
let (cache_creation_5m, cache_creation_1h) =
cache_creation_ttl_tokens(msg_usage);
// 从 message_start 获取 input_tokens（原生 Claude API）
if let Some(input) =
msg_usage.get("input_tokens").and_then(|v| v.as_u64())
@@ -150,10 +195,14 @@ impl TokenUsage {
.and_then(|v| v.as_u64())
.unwrap_or(0)
as u32;
usage.cache_creation_5m_tokens = cache_creation_5m;
usage.cache_creation_1h_tokens = cache_creation_1h;
}
}
"message_delta" => {
if let Some(delta_usage) = event.get("usage") {
let (delta_cache_creation_5m, delta_cache_creation_1h) =
cache_creation_ttl_tokens(delta_usage);
// 从 message_delta 获取 output_tokens
if let Some(output) =
delta_usage.get("output_tokens").and_then(|v| v.as_u64())
@@ -192,6 +241,14 @@ impl TokenUsage {
}
if let Some(cache_creation) = delta_cache_creation {
usage.cache_creation_tokens = cache_creation;
if delta_cache_creation_5m > 0
|| delta_cache_creation_1h > 0
{
usage.cache_creation_5m_tokens =
delta_cache_creation_5m;
usage.cache_creation_1h_tokens =
delta_cache_creation_1h;
}
}
}
}
@@ -206,6 +263,10 @@ impl TokenUsage {
if usage.cache_creation_tokens == 0 {
if let Some(cache_creation) = delta_cache_creation {
usage.cache_creation_tokens = cache_creation;
if delta_cache_creation_5m > 0 || delta_cache_creation_1h > 0 {
usage.cache_creation_5m_tokens = delta_cache_creation_5m;
usage.cache_creation_1h_tokens = delta_cache_creation_1h;
}
}
}
}
@@ -237,6 +298,8 @@ impl TokenUsage {
output_tokens: usage.get("completion_tokens")?.as_u64()? as u32,
cache_read_tokens: 0,
cache_creation_tokens: 0,
cache_creation_5m_tokens: 0,
cache_creation_1h_tokens: 0,
model: None,
message_id: None,
})
@@ -270,12 +333,15 @@ impl TokenUsage {
let cached_tokens = openai_cache_read_tokens(usage);
let cache_write_tokens = openai_cache_write_tokens(usage);
let (cache_creation_5m_tokens, cache_creation_1h_tokens) = cache_creation_ttl_tokens(usage);
Some(Self {
input_tokens: input_tokens? as u32,
output_tokens: output_tokens? as u32,
cache_read_tokens: cached_tokens,
cache_creation_tokens: cache_write_tokens,
cache_creation_5m_tokens,
cache_creation_1h_tokens,
model,
message_id: None,
})
@@ -294,6 +360,7 @@ impl TokenUsage {
// 获取 cached_tokens (可能在 cache_read_input_tokens 或 input_tokens_details 中)
let cached_tokens = openai_cache_read_tokens(usage);
let cache_write_tokens = openai_cache_write_tokens(usage);
let (cache_creation_5m_tokens, cache_creation_1h_tokens) = cache_creation_ttl_tokens(usage);
// 调整 input_tokens: OpenAI total input 同时包含 cache read/write 两桶。
let adjusted_input = input_tokens
@@ -311,6 +378,8 @@ impl TokenUsage {
output_tokens,
cache_read_tokens: cached_tokens,
cache_creation_tokens: cache_write_tokens,
cache_creation_5m_tokens,
cache_creation_1h_tokens,
model,
message_id: None,
})
@@ -391,6 +460,7 @@ impl TokenUsage {
// 获取 cached_tokens (可能在 prompt_tokens_details 中)
let cached_tokens = openai_cache_read_tokens(usage);
let cache_write_tokens = openai_cache_write_tokens(usage);
let (cache_creation_5m_tokens, cache_creation_1h_tokens) = cache_creation_ttl_tokens(usage);
// 提取响应中的模型名称
let model = body
@@ -403,6 +473,8 @@ impl TokenUsage {
output_tokens: completion_tokens as u32,
cache_read_tokens: cached_tokens,
cache_creation_tokens: cache_write_tokens,
cache_creation_5m_tokens,
cache_creation_1h_tokens,
model,
message_id: None,
})
@@ -448,6 +520,8 @@ impl TokenUsage {
.and_then(|v| v.as_u64())
.unwrap_or(0) as u32,
cache_creation_tokens: 0,
cache_creation_5m_tokens: 0,
cache_creation_1h_tokens: 0,
model,
message_id: None,
})
@@ -499,6 +573,8 @@ impl TokenUsage {
output_tokens: total_output,
cache_read_tokens: total_cache_read,
cache_creation_tokens: 0,
cache_creation_5m_tokens: 0,
cache_creation_1h_tokens: 0,
model,
message_id: None,
})
@@ -521,7 +597,11 @@ mod tests {
"input_tokens": 100,
"output_tokens": 50,
"cache_read_input_tokens": 20,
"cache_creation_input_tokens": 10
"cache_creation_input_tokens": 10,
"cache_creation": {
"ephemeral_5m_input_tokens": 4,
"ephemeral_1h_input_tokens": 6
}
}
});
@@ -530,6 +610,9 @@ mod tests {
assert_eq!(usage.output_tokens, 50);
assert_eq!(usage.cache_read_tokens, 20);
assert_eq!(usage.cache_creation_tokens, 10);
assert_eq!(usage.cache_creation_5m_tokens, 4);
assert_eq!(usage.cache_creation_1h_tokens, 6);
assert_eq!(usage.normalized_cache_creation_buckets(), (0, 4, 6));
assert_eq!(usage.model, Some("claude-sonnet-4-20250514".to_string()));
}