From 505e7ea481141cd8580177485264e6760937d4cc Mon Sep 17 00:00:00 2001
From: Saoud Rizwan <7799382+saoudrizwan@users.noreply.github.com>
Date: Tue, 31 Dec 2024 13:10:17 -0800
Subject: [PATCH] Fix deepseek price reporting

---
 src/api/providers/deepseek.ts     | 6 +++---
 src/api/providers/openrouter.ts   | 4 ----
 src/core/webview/ClineProvider.ts | 5 -----
 src/shared/api.ts                 | 6 +++---
 4 files changed, 6 insertions(+), 15 deletions(-)

diff --git a/src/api/providers/deepseek.ts b/src/api/providers/deepseek.ts
index 48afab68ba..0cadd57200 100644
--- a/src/api/providers/deepseek.ts
+++ b/src/api/providers/deepseek.ts
@@ -39,12 +39,12 @@ export class DeepSeekHandler implements ApiHandler {
 			if (chunk.usage) {
 				yield {
 					type: "usage",
-					inputTokens: 0, //chunk.usage.prompt_tokens || 0, (deepseek reports total input AND cache reads/writes, see context caching: https://api-docs.deepseek.com/guides/kv_cache)
+					inputTokens: chunk.usage.prompt_tokens || 0, // (deepseek reports total input AND cache reads/writes, see context caching: https://api-docs.deepseek.com/guides/kv_cache) but we use this to do the truncation algo, so we can't report cache stats right now because of how deepseek api reports input AND the cache reads/writes, while anthropic reports them as separate tokens
 					outputTokens: chunk.usage.completion_tokens || 0,
 					// @ts-ignore-next-line
-					cacheReadTokens: chunk.usage.prompt_cache_hit_tokens || 0,
+					// cacheReadTokens: chunk.usage.prompt_cache_hit_tokens || 0,
 					// @ts-ignore-next-line
-					cacheWriteTokens: chunk.usage.prompt_cache_miss_tokens || 0,
+					// cacheWriteTokens: chunk.usage.prompt_cache_miss_tokens || 0,
 				}
 			}
 		}
diff --git a/src/api/providers/openrouter.ts b/src/api/providers/openrouter.ts
index ccdde6c378..8170d41afc 100644
--- a/src/api/providers/openrouter.ts
+++ b/src/api/providers/openrouter.ts
@@ -98,10 +98,6 @@ export class OpenRouterHandler implements ApiHandler {
 
 		// Removes messages in the middle when close to context window limit. Should not be applied to models that support prompt caching since it would continuously break the cache.
 		let shouldApplyMiddleOutTransform = !this.getModel().info.supportsPromptCache
-		// except for deepseek (which we set supportsPromptCache to true for), where because the context window is so small our truncation algo might miss and we should use openrouter's middle-out transform as a fallback to ensure we don't exceed the context window
-		if (this.getModel().id === "deepseek/deepseek-chat") {
-			shouldApplyMiddleOutTransform = true
-		}
 
 		// @ts-ignore-next-line
 		const stream = await this.client.chat.completions.create({
diff --git a/src/core/webview/ClineProvider.ts b/src/core/webview/ClineProvider.ts
index 4081e617dd..78cc71b460 100644
--- a/src/core/webview/ClineProvider.ts
+++ b/src/core/webview/ClineProvider.ts
@@ -726,11 +726,6 @@ export class ClineProvider implements vscode.WebviewViewProvider {
 							modelInfo.cacheWritesPrice = 0.3
 							modelInfo.cacheReadsPrice = 0.03
 							break
-						case "deepseek/deepseek-chat":
-							modelInfo.supportsPromptCache = true
-							modelInfo.cacheWritesPrice = 0.14
-							modelInfo.cacheReadsPrice = 0.014
-							break
 					}
 
 					models[rawModel.id] = modelInfo
diff --git a/src/shared/api.ts b/src/shared/api.ts
index 578b1083e1..5f2b42b416 100644
--- a/src/shared/api.ts
+++ b/src/shared/api.ts
@@ -359,10 +359,10 @@ export const deepSeekModels = {
 		maxTokens: 8_000,
 		contextWindow: 64_000,
 		supportsImages: false,
-		supportsPromptCache: true,
+		supportsPromptCache: false, // technically supports context caching, but not in the way anthropic does it (deepseek reports input tokens and reads/writes in the same usage report) FIXME: we need to show users cache stats how deepseek does it
 		inputPrice: 0.14,
 		outputPrice: 0.28,
-		cacheWritesPrice: 0.14,
-		cacheReadsPrice: 0.014,
+		// cacheWritesPrice: 0.14,
+		// cacheReadsPrice: 0.014,
 	},
 } as const satisfies Record<string, ModelInfo>