From cb01593f7d958607c4f1793f990536d3489240e0 Mon Sep 17 00:00:00 2001 From: Jason Date: Tue, 9 Jun 2026 21:39:09 +0800 Subject: [PATCH] =?UTF-8?q?fix(proxy):=20exclude=20cache=5Fread=20and=20ca?= =?UTF-8?q?che=5Fcreation=20from=20input=20on=20Claude=E2=86=90OpenAI=20pa?= =?UTF-8?q?ths?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Builds on #2774 (which fixed cache_read for the streaming openai_chat path). Two gaps remained, both double-counting cache tokens when a Claude client meters as app_type="claude" (input_includes_cache_read=false): 1. cache_read was still added to input on the non-streaming openai_chat path (transform.rs openai_to_anthropic) and the whole openai_responses family (transform_responses.rs build_anthropic_usage_from_responses, covering the non-streaming call site and both streaming_responses call sites). 2. cache_creation was never subtracted on any converted path, including the streaming openai_chat path #2774 had already touched. Claude billing treats cache_creation as a separate bucket, so an inclusive upstream carrying a direct cache_creation_input_tokens field billed it twice. All four metering points now compute: input = prompt_tokens - cache_read - cache_creation restoring the invariant input + cache_read + cache_creation == prompt_tokens. Pure OpenAI upstreams are unaffected (no cache_creation concept/field). Tests: update direct-cache assertions (40->20), add a streaming conservation regression test, and pin prompt Value { - // OpenAI prompt_tokens 含缓存，Anthropic input_tokens 不含，需减去 + // OpenAI prompt_tokens 含缓存，Anthropic input_tokens 不含，需减去 cache_read 与 cache_creation + // （三桶互斥，恒等 input + cache_read + cache_creation == prompt_tokens）。 let cached = extract_cache_read_tokens(usage).unwrap_or(0); - let input_tokens = usage.prompt_tokens.saturating_sub(cached); + let cache_creation = usage.cache_creation_input_tokens.unwrap_or(0); + let input_tokens = usage + .prompt_tokens + .saturating_sub(cached) + .saturating_sub(cache_creation); let mut usage_json = json!({ "input_tokens": input_tokens, "output_tokens": usage.completion_tokens @@ -110,8 +115,8 @@ fn build_anthropic_usage_json(usage: &Usage) -> Value { if cached > 0 { usage_json["cache_read_input_tokens"] = json!(cached); } - if let Some(created) = usage.cache_creation_input_tokens { - usage_json["cache_creation_input_tokens"] = json!(created); + if cache_creation > 0 { + usage_json["cache_creation_input_tokens"] = json!(cache_creation); } usage_json } @@ -227,13 +232,19 @@ pub fn create_anthropic_sse_stream( }); if let Some(u) = &chunk.usage { let cached = extract_cache_read_tokens(u).unwrap_or(0); - let input = u.prompt_tokens.saturating_sub(cached); + let cache_creation = + u.cache_creation_input_tokens.unwrap_or(0); + let input = u + .prompt_tokens + .saturating_sub(cached) + .saturating_sub(cache_creation); start_usage["input_tokens"] = json!(input); if cached > 0 { start_usage["cache_read_input_tokens"] = json!(cached); } - if let Some(created) = u.cache_creation_input_tokens { - start_usage["cache_creation_input_tokens"] = json!(created); + if cache_creation > 0 { + start_usage["cache_creation_input_tokens"] = + json!(cache_creation); } } @@ -1043,6 +1054,81 @@ mod tests { ); } + #[tokio::test] + async fn test_usage_chunk_subtracts_cache_read_and_creation_from_input() { + // prompt_tokens(1000) 含 cache_read(600) 与 cache_creation(300)；转 Anthropic 后 + // input 应为 fresh，守恒：input(100) + cache_read(600) + cache_creation(300) == prompt(1000)。 + let input = concat!( + "data: {\"id\":\"chatcmpl_cc\",\"model\":\"glm-5.1\",\"choices\":[{\"delta\":{\"tool_calls\":[{\"index\":0,\"id\":\"tool-1\",\"type\":\"function\",\"function\":{\"name\":\"Bash\",\"arguments\":\"{\\\"command\\\":\\\"pwd\\\"}\"}}]}}]}\n\n", + "data: {\"id\":\"chatcmpl_cc\",\"model\":\"glm-5.1\",\"choices\":[{\"delta\":{},\"finish_reason\":\"tool_calls\"}]}\n\n", + "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":1000,\"completion_tokens\":50,\"prompt_tokens_details\":{\"cached_tokens\":600},\"cache_creation_input_tokens\":300}}\n\n", + "data: [DONE]\n\n" + ); + + let events = collect_anthropic_events(input).await; + let message_delta = events + .iter() + .find(|event| event_type(event) == Some("message_delta")) + .expect("should emit message_delta with usage"); + + // fresh input = 1000 - 600 - 300 = 100 + assert_eq!( + message_delta + .pointer("/usage/input_tokens") + .and_then(|v| v.as_u64()), + Some(100) + ); + assert_eq!( + message_delta + .pointer("/usage/cache_read_input_tokens") + .and_then(|v| v.as_u64()), + Some(600) + ); + assert_eq!( + message_delta + .pointer("/usage/cache_creation_input_tokens") + .and_then(|v| v.as_u64()), + Some(300) + ); + } + + #[tokio::test] + async fn test_usage_chunk_clamps_input_to_zero_when_cache_exceeds_prompt() { + // prompt(100) < cache_read(80)+cache_creation(50)=130：saturating 钳到 0，防下溢。 + // 钉桩：阻止未来把 saturating_sub 误改成普通减法(debug panic / release wrap)。 + let input = concat!( + "data: {\"id\":\"chatcmpl_uf\",\"model\":\"glm-5.1\",\"choices\":[{\"delta\":{\"tool_calls\":[{\"index\":0,\"id\":\"tool-1\",\"type\":\"function\",\"function\":{\"name\":\"Bash\",\"arguments\":\"{\\\"command\\\":\\\"pwd\\\"}\"}}]}}]}\n\n", + "data: {\"id\":\"chatcmpl_uf\",\"model\":\"glm-5.1\",\"choices\":[{\"delta\":{},\"finish_reason\":\"tool_calls\"}]}\n\n", + "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":100,\"completion_tokens\":50,\"prompt_tokens_details\":{\"cached_tokens\":80},\"cache_creation_input_tokens\":50}}\n\n", + "data: [DONE]\n\n" + ); + + let events = collect_anthropic_events(input).await; + let message_delta = events + .iter() + .find(|event| event_type(event) == Some("message_delta")) + .expect("should emit message_delta with usage"); + + assert_eq!( + message_delta + .pointer("/usage/input_tokens") + .and_then(|v| v.as_u64()), + Some(0) + ); + assert_eq!( + message_delta + .pointer("/usage/cache_read_input_tokens") + .and_then(|v| v.as_u64()), + Some(80) + ); + assert_eq!( + message_delta + .pointer("/usage/cache_creation_input_tokens") + .and_then(|v| v.as_u64()), + Some(50) + ); + } + #[tokio::test] async fn test_message_delta_includes_zero_usage_when_stream_has_no_usage() { let input = concat!( diff --git a/src-tauri/src/proxy/providers/transform.rs b/src-tauri/src/proxy/providers/transform.rs index 57dd7928b..9b0dd8f72 100644 --- a/src-tauri/src/proxy/providers/transform.rs +++ b/src-tauri/src/proxy/providers/transform.rs @@ -644,10 +644,31 @@ pub fn openai_to_anthropic(body: Value) -> Result { // usage — map cache tokens from OpenAI format to Anthropic format let usage = body.get("usage").cloned().unwrap_or(json!({})); + // OpenAI prompt_tokens 含缓存命中，Anthropic input_tokens 不含 → 减去 cache_read 与 + // cache_creation，使 input 成为 fresh input。本路径以 app_type="claude" 记账（calculator + // 不再扣减），若不减则缓存会被计入 input 与各 cache 桶两次。三桶互斥，恒等： + // input + cache_read + cache_creation == prompt_tokens（inclusive 上游）。 + // 与流式 build_anthropic_usage_json (#2774) 及 transform_gemini 的 saturating_sub 对称。 + // 最终 cache_read：直传字段优先于 nested；cache_creation 仅来自直传字段（OpenAI 无此概念）。 + let cached = usage + .get("cache_read_input_tokens") + .and_then(|v| v.as_u64()) + .or_else(|| { + usage + .pointer("/prompt_tokens_details/cached_tokens") + .and_then(|v| v.as_u64()) + }) + .unwrap_or(0); + let cache_creation = usage + .get("cache_creation_input_tokens") + .and_then(|v| v.as_u64()) + .unwrap_or(0); let input_tokens = usage .get("prompt_tokens") .and_then(|v| v.as_u64()) - .unwrap_or(0) as u32; + .unwrap_or(0) + .saturating_sub(cached) + .saturating_sub(cache_creation) as u32; let output_tokens = usage .get("completion_tokens") .and_then(|v| v.as_u64()) @@ -658,19 +679,11 @@ pub fn openai_to_anthropic(body: Value) -> Result { "output_tokens": output_tokens }); - // OpenAI standard: prompt_tokens_details.cached_tokens - if let Some(cached) = usage - .pointer("/prompt_tokens_details/cached_tokens") - .and_then(|v| v.as_u64()) - { + if cached > 0 { usage_json["cache_read_input_tokens"] = json!(cached); } - // Some compatible servers return these fields directly - if let Some(v) = usage.get("cache_read_input_tokens") { - usage_json["cache_read_input_tokens"] = v.clone(); - } - if let Some(v) = usage.get("cache_creation_input_tokens") { - usage_json["cache_creation_input_tokens"] = v.clone(); + if cache_creation > 0 { + usage_json["cache_creation_input_tokens"] = json!(cache_creation); } let result = json!({ @@ -1314,7 +1327,8 @@ mod tests { }); let result = openai_to_anthropic(input).unwrap(); - assert_eq!(result["usage"]["input_tokens"], 100); + // prompt_tokens(100) 含 cached(80)，转换后 input 应为 fresh = 100 - 80 = 20 + assert_eq!(result["usage"]["input_tokens"], 20); assert_eq!(result["usage"]["output_tokens"], 50); assert_eq!(result["usage"]["cache_read_input_tokens"], 80); } @@ -1338,10 +1352,38 @@ mod tests { }); let result = openai_to_anthropic(input).unwrap(); + // cache_read(60)+cache_creation(20) 均从 prompt(100) 扣除，fresh = 100 - 60 - 20 = 20 + // 守恒：input(20) + cache_read(60) + cache_creation(20) == prompt(100) + assert_eq!(result["usage"]["input_tokens"], 20); assert_eq!(result["usage"]["cache_read_input_tokens"], 60); assert_eq!(result["usage"]["cache_creation_input_tokens"], 20); } + #[test] + fn test_openai_to_anthropic_clamps_input_when_cache_exceeds_prompt() { + // prompt(100) < cache_read(60)+cache_creation(50)=110：saturating 钳到 0，防下溢。 + // 钉桩：阻止未来把 saturating_sub 误改成普通减法(debug panic / release wrap)。 + let input = json!({ + "id": "chatcmpl-uf", + "model": "gpt-4", + "choices": [{ + "index": 0, + "message": {"role": "assistant", "content": "x"}, + "finish_reason": "stop" + }], + "usage": { + "prompt_tokens": 100, + "completion_tokens": 10, + "cache_read_input_tokens": 60, + "cache_creation_input_tokens": 50 + } + }); + let result = openai_to_anthropic(input).unwrap(); + assert_eq!(result["usage"]["input_tokens"], 0); + assert_eq!(result["usage"]["cache_read_input_tokens"], 60); + assert_eq!(result["usage"]["cache_creation_input_tokens"], 50); + } + #[test] fn test_openai_to_anthropic_finish_reason_content_filter_maps_end_turn() { let input = json!({ diff --git a/src-tauri/src/proxy/providers/transform_responses.rs b/src-tauri/src/proxy/providers/transform_responses.rs index 537e0f5de..d7379cd6e 100644 --- a/src-tauri/src/proxy/providers/transform_responses.rs +++ b/src-tauri/src/proxy/providers/transform_responses.rs @@ -355,6 +355,23 @@ pub(crate) fn build_anthropic_usage_from_responses(usage: Option<&Value>) -> Val result["cache_creation_input_tokens"] = v.clone(); } + // OpenAI/Responses 的 input(prompt_tokens/input_tokens)含缓存命中，Anthropic input_tokens 不含 + // → 减去 cache_read 与 cache_creation，使其成为 fresh input。本函数在计量意义上是 claude 专属 + // （Codex Responses 透传走 from_codex_response_*，不调用本函数），故可安全在此扣减。三桶互斥， + // 恒等：input + cache_read + cache_creation == 上游 input(inclusive)。与 build_anthropic_usage_json + // (#2774) 及 transform_gemini 的 saturating_sub 对称；一处同时覆盖非流式与流式(streaming_responses)。 + let cached = result + .get("cache_read_input_tokens") + .and_then(|v| v.as_u64()) + .unwrap_or(0); + let cache_creation = result + .get("cache_creation_input_tokens") + .and_then(|v| v.as_u64()) + .unwrap_or(0); + if cached > 0 || cache_creation > 0 { + result["input_tokens"] = json!(input.saturating_sub(cached).saturating_sub(cache_creation)); + } + result } @@ -1156,7 +1173,8 @@ mod tests { }); let result = responses_to_anthropic(input).unwrap(); - assert_eq!(result["usage"]["input_tokens"], 100); + // input_tokens(100) 含 cached(80)，转换后 input 应为 fresh = 100 - 80 = 20 + assert_eq!(result["usage"]["input_tokens"], 20); assert_eq!(result["usage"]["output_tokens"], 50); assert_eq!(result["usage"]["cache_read_input_tokens"], 80); } @@ -1180,6 +1198,9 @@ mod tests { }); let result = responses_to_anthropic(input).unwrap(); + // cache_read(60)+cache_creation(20) 均从 input(100) 扣除，fresh = 100 - 60 - 20 = 20 + // 守恒：input(20) + cache_read(60) + cache_creation(20) == 上游 input(100) + assert_eq!(result["usage"]["input_tokens"], 20); assert_eq!(result["usage"]["cache_read_input_tokens"], 60); assert_eq!(result["usage"]["cache_creation_input_tokens"], 20); } @@ -1642,7 +1663,8 @@ mod tests { "cached_tokens": 80 } }))); - assert_eq!(result["input_tokens"], json!(100)); + // input_tokens(100) 含 nested cached(80)，转换后 input 应为 fresh = 100 - 80 = 20 + assert_eq!(result["input_tokens"], json!(20)); assert_eq!(result["output_tokens"], json!(50)); assert_eq!(result["cache_read_input_tokens"], json!(80)); } @@ -1657,9 +1679,26 @@ mod tests { }, "cache_read_input_tokens": 100 }))); + // 直传 cache_read(100) 优先于 nested(80)；input(100) - 100 = 0（fresh） + assert_eq!(result["input_tokens"], json!(0)); assert_eq!(result["cache_read_input_tokens"], json!(100)); // Direct field overrides nested } + #[test] + fn test_build_usage_clamps_input_when_cache_exceeds_input() { + // input(100) < cache_read(60)+cache_creation(50)=110：saturating 钳到 0，防下溢。 + // 钉桩：阻止未来把 saturating_sub 误改成普通减法(debug panic / release wrap)。 + let result = build_anthropic_usage_from_responses(Some(&json!({ + "input_tokens": 100, + "output_tokens": 10, + "cache_read_input_tokens": 60, + "cache_creation_input_tokens": 50 + }))); + assert_eq!(result["input_tokens"], json!(0)); + assert_eq!(result["cache_read_input_tokens"], json!(60)); + assert_eq!(result["cache_creation_input_tokens"], json!(50)); + } + #[test] fn test_build_usage_cache_tokens_without_input_output() { let result = build_anthropic_usage_from_responses(Some(&json!({