Fix DeepSeek cost reporting

src/api/providers/deepseek.ts

@@ -39,12 +39,12 @@ export class DeepSeekHandler implements ApiHandler {
 			if (chunk.usage) {
 				yield {
 					type: "usage",
-					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
+					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) where the input tokens is the sum of the cache hits/misses, while anthropic reports them as separate tokens. This is important to know for 1) context management truncation algorithm, and 2) cost calculation (NOTE: we report both input and cache stats but for now set input price to 0 since all the cost calculation will be done using cache hits/misses)
 					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,
 				}
 			}
 		}

src/api/providers/openrouter.ts

@@ -98,6 +98,10 @@ 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 (FIXME: once we have a more robust token estimator we should not rely on this)
+		if (this.getModel().id === "deepseek/deepseek-chat") {
+			shouldApplyMiddleOutTransform = true
+		}
 
 		// @ts-ignore-next-line
 		const stream = await this.client.chat.completions.create({

src/core/webview/ClineProvider.ts

@@ -726,6 +726,13 @@ export class ClineProvider implements vscode.WebviewViewProvider {
 							modelInfo.cacheWritesPrice = 0.3
 							modelInfo.cacheReadsPrice = 0.03
 							break
+						case "deepseek/deepseek-chat":
+							modelInfo.supportsPromptCache = true
+							// see api.ts/deepSeekModels for more info
+							modelInfo.inputPrice = 0
+							modelInfo.cacheWritesPrice = 0.14
+							modelInfo.cacheReadsPrice = 0.014
+							break
 					}
 
 					models[rawModel.id] = modelInfo

src/shared/api.ts

@@ -359,10 +359,10 @@ export const deepSeekModels = {
 		maxTokens: 8_000,
 		contextWindow: 64_000,
 		supportsImages: false,
-		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,
+		supportsPromptCache: true, // 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, // technically there is no input price, it's all either a cache hit or miss (ApiOptions will not show this)
 		outputPrice: 0.28,
-		// cacheWritesPrice: 0.14,
-		// cacheReadsPrice: 0.014,
+		cacheWritesPrice: 0.14,
+		cacheReadsPrice: 0.014,
 	},
 } as const satisfies Record<string, ModelInfo>