admin/CC-Switch
1
0
Fork 0
mirror of https://github.com/farion1231/cc-switch.git synced 2026-07-29 09:37:37 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(usage): account for cache-write tokens across schema versions

Browse Source 
Parse cache_write_tokens from OpenAI usage details and preserve cache creation data across Chat, Responses, and Anthropic conversion paths.

Add explicit input-token semantics to request logs and rollups so legacy rows subtract cache reads only while new total-inclusive rows subtract both cache reads and writes. Migrate v12 databases, normalize rollups to fresh input, and cover historical backfill behavior with regression tests.
This commit is contained in:
Jason
2026-07-11 12:27:39 +08:00
parent 06039540ff
commit f991726ff0
13 changed files with 443 additions and 102 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src-tauri/src/proxy/providers/transform_codex_anthropic.rs
+24 -23
View File
@@ -125,19 +125,14 @@ pub(crate) fn map_anthropic_stop_reason_to_status(
/// Builds Responses usage from Anthropic usage.
///
/// Anthropic's `input_tokens` is the "cache-excluded" fresh input; OpenAI/Responses'
/// `input_tokens` includes cache hits. To keep downstream metering correct, this
/// adds them (symmetric to the subtraction done for the Claude side in
/// `transform_responses`):
/// input_tokens = input + cache_read
/// Anthropic's `input_tokens` is the cache-excluded fresh input. OpenAI Responses
/// reports total input and exposes cache reads/writes as subsets:
/// input_tokens = fresh + cache_read + cache_creation
/// input_tokens_details.cached_tokens = cache_read
/// input_tokens_details.cache_write_tokens = cache_creation
///
/// Note: **do not** fold `cache_creation` into `input_tokens`. The Codex billing
/// calculator (usage/calculator.rs) only subtracts `cache_read` for codex
/// (`billable = input - cache_read`), and separately lists cache-creation cost via
/// `cache_creation_input_tokens`; if creation were also added into
/// `input_tokens`, it would be double-charged at both the input price and the
/// cache-creation price.
/// The internal billing parser subtracts both subsets before charging the normal
/// input rate, then prices cache reads and writes separately.
pub(crate) fn build_responses_usage_from_anthropic(usage: Option<&Value>) -> Value {
let u = match usage {
Some(v) if v.is_object() => v,
@@ -166,10 +161,10 @@ pub(crate) fn build_responses_usage_from_anthropic(usage: Option<&Value>) -> Val
.and_then(|v| v.as_u64())
.unwrap_or(0);
let input_tokens = fresh_input.saturating_add(cache_read);
let total_tokens = input_tokens
.saturating_add(cache_creation)
.saturating_add(output);
let input_tokens = fresh_input
.saturating_add(cache_read)
.saturating_add(cache_creation);
let total_tokens = input_tokens.saturating_add(output);
let mut result = json!({
"input_tokens": input_tokens,
@@ -177,10 +172,14 @@ pub(crate) fn build_responses_usage_from_anthropic(usage: Option<&Value>) -> Val
"total_tokens": total_tokens,
"output_tokens_details": { "reasoning_tokens": reasoning }
});
if cache_read > 0 {
result["input_tokens_details"] = json!({ "cached_tokens": cache_read });
if cache_read > 0 || cache_creation > 0 {
result["input_tokens_details"] = json!({
"cached_tokens": cache_read,
"cache_write_tokens": cache_creation
});
}
// Explicitly pass through cache_creation so the downstream usage parser (from_codex_response) attributes billing correctly.
// Keep the legacy top-level alias for one compatibility window. New code reads
// the official nested cache_write_tokens field first.
if cache_creation > 0 {
result["cache_creation_input_tokens"] = json!(cache_creation);
}
@@ -2078,18 +2077,20 @@ mod tests {
}
});
let result = anthropic_response_to_responses(input).unwrap();
// input_tokens = fresh + cache_read = 20 + 60 = 80 (excluding cache_creation).
// The Codex billing calculator only subtracts cache_read from input (→ billable=fresh=20),
// and separately lists cache-creation cost via cache_creation_input_tokens; folding creation into input would double-charge.
assert_eq!(result["usage"]["input_tokens"], 80);
// Responses input_tokens is the inclusive total: fresh + read + write.
assert_eq!(result["usage"]["input_tokens"], 100);
assert_eq!(result["usage"]["output_tokens"], 5);
assert_eq!(
result["usage"]["output_tokens_details"]["reasoning_tokens"],
3
);
// total still includes everything: 80 + cache_creation 20 + output 5 = 105
// total includes input total + output exactly once.
assert_eq!(result["usage"]["total_tokens"], 105);
assert_eq!(result["usage"]["input_tokens_details"]["cached_tokens"], 60);
assert_eq!(
result["usage"]["input_tokens_details"]["cache_write_tokens"],
20
);
// cache_creation is passed through explicitly for downstream billing attribution (counted only once)
assert_eq!(result["usage"]["cache_creation_input_tokens"], 20);
}