Context in context (RooCodeInc#2745)

* context in context

* keeping comments

* minimal context window line

* nit

Author: pashpashpash

.changeset/curvy-walls-cry.md

@@ -0,0 +1,5 @@
+---
+"claude-dev": minor
+---
+
+Added context window usage to env variables context for messages

src/core/context-management/ContextManager.ts

@@ -1,7 +1,7 @@
 import { Anthropic } from "@anthropic-ai/sdk"
 import { ClineApiReqInfo, ClineMessage } from "../../shared/ExtensionMessage"
 import { ApiHandler } from "../../api"
-import { OpenAiHandler } from "../../api/providers/openai"
+import { getContextWindowInfo } from "./context-window-utils"
 import { formatResponse } from "../prompts/responses"
 import { GlobalFileNames } from "../storage/disk"
 import { fileExistsAtPath } from "../../utils/fs"
@@ -125,25 +125,7 @@ export class ContextManager {
 				const timestamp = previousRequest.ts
 				const { tokensIn, tokensOut, cacheWrites, cacheReads }: ClineApiReqInfo = JSON.parse(previousRequest.text)
 				const totalTokens = (tokensIn || 0) + (tokensOut || 0) + (cacheWrites || 0) + (cacheReads || 0)
-				let contextWindow = api.getModel().info.contextWindow || 128_000
-				// FIXME: hack to get anyone using openai compatible with deepseek to have the proper context window instead of the default 128k. We need a way for the user to specify the context window for models they input through openai compatible
-				if (api instanceof OpenAiHandler && api.getModel().id.toLowerCase().includes("deepseek")) {
-					contextWindow = 64_000
-				}
-				let maxAllowedSize: number
-				switch (contextWindow) {
-					case 64_000: // deepseek models
-						maxAllowedSize = contextWindow - 27_000
-						break
-					case 128_000: // most models
-						maxAllowedSize = contextWindow - 30_000
-						break
-					case 200_000: // claude models
-						maxAllowedSize = contextWindow - 40_000
-						break
-					default:
-						maxAllowedSize = Math.max(contextWindow - 40_000, contextWindow * 0.8) // for deepseek, 80% of 64k meant only ~10k buffer which was too small and resulted in users getting context window errors.
-				}
+				const { maxAllowedSize } = getContextWindowInfo(api)
 
 				// This is the most reliable way to know when we're close to hitting the context window.
 				if (totalTokens >= maxAllowedSize) {

src/core/context-management/context-window-utils.ts

@@ -0,0 +1,35 @@
+import { ApiHandler } from "../../api"
+import { OpenAiHandler } from "../../api/providers/openai"
+
+/**
+ * Gets context window information for the given API handler
+ *
+ * @param api The API handler to get context window information for
+ * @returns An object containing the raw context window size and the effective max allowed size
+ */
+export function getContextWindowInfo(api: ApiHandler) {
+	let contextWindow = api.getModel().info.contextWindow || 128_000
+	// FIXME: hack to get anyone using openai compatible with deepseek to have the proper context window instead of the default 128k. We need a way for the user to specify the context window for models they input through openai compatible
+
+	// Handle special cases like DeepSeek
+	if (api instanceof OpenAiHandler && api.getModel().id.toLowerCase().includes("deepseek")) {
+		contextWindow = 64_000
+	}
+
+	let maxAllowedSize: number
+	switch (contextWindow) {
+		case 64_000: // deepseek models
+			maxAllowedSize = contextWindow - 27_000
+			break
+		case 128_000: // most models
+			maxAllowedSize = contextWindow - 30_000
+			break
+		case 200_000: // claude models
+			maxAllowedSize = contextWindow - 40_000
+			break
+		default:
+			maxAllowedSize = Math.max(contextWindow - 40_000, contextWindow * 0.8) // for deepseek, 80% of 64k meant only ~10k buffer which was too small and resulted in users getting context window errors.
+	}
+
+	return { contextWindow, maxAllowedSize }
+}

src/core/task/index.ts

@@ -13,6 +13,7 @@ import { ApiHandler, buildApiHandler } from "../../api"
 import { AnthropicHandler } from "../../api/providers/anthropic"
 import { ClineHandler } from "../../api/providers/cline"
 import { OpenRouterHandler } from "../../api/providers/openrouter"
+import { getContextWindowInfo } from "../context-management/context-window-utils"
 import { ApiStream } from "../../api/transform/stream"
 import CheckpointTracker from "../../integrations/checkpoints/CheckpointTracker"
 import { DIFF_VIEW_URI_SCHEME, DiffViewProvider } from "../../integrations/editor/DiffViewProvider"
@@ -3682,6 +3683,36 @@ export class Task {
 			}
 		}
 
+		// Add context window usage information
+		const { contextWindow, maxAllowedSize } = getContextWindowInfo(this.api)
+
+		// Get the token count from the most recent API request to accurately reflect context management
+		const getTotalTokensFromApiReqMessage = (msg: ClineMessage) => {
+			if (!msg.text) {
+				return 0
+			}
+			try {
+				const { tokensIn, tokensOut, cacheWrites, cacheReads } = JSON.parse(msg.text)
+				return (tokensIn || 0) + (tokensOut || 0) + (cacheWrites || 0) + (cacheReads || 0)
+			} catch (e) {
+				return 0
+			}
+		}
+
+		const modifiedMessages = combineApiRequests(combineCommandSequences(this.clineMessages.slice(1)))
+		const lastApiReqMessage = findLast(modifiedMessages, (msg) => {
+			if (msg.say !== "api_req_started") {
+				return false
+			}
+			return getTotalTokensFromApiReqMessage(msg) > 0
+		})
+
+		const lastApiReqTotalTokens = lastApiReqMessage ? getTotalTokensFromApiReqMessage(lastApiReqMessage) : 0
+		const usagePercentage = Math.round((lastApiReqTotalTokens / contextWindow) * 100)
+
+		details += "\n# Context Window Usage"
+		details += `\n${lastApiReqTotalTokens.toLocaleString()} / ${(contextWindow / 1000).toLocaleString()}K tokens used (${usagePercentage}%)`
+
 		details += "\n\n# Current Mode"
 		if (this.chatSettings.mode === "plan") {
 			details += "\nPLAN MODE\n" + formatResponse.planModeInstructions()