Refactor out of utils

src/core/ClaudeDev.ts

@@ -23,14 +23,15 @@ import { getApiMetrics } from "../shared/getApiMetrics"
 import { HistoryItem } from "../shared/HistoryItem"
 import { Tool, ToolName } from "../shared/Tool"
 import { ClaudeAskResponse } from "../shared/WebviewMessage"
-import { findLast, findLastIndex, formatContentBlockToMarkdown } from "../utils"
-import { truncateHalfConversation } from "../utils/context-management"
+import { findLast, findLastIndex } from "../utils/array"
+import { formatContentBlockToMarkdown } from "../integrations/misc/export-markdown"
+import { truncateHalfConversation } from "./sliding-window"
 import { extractTextFromFile } from "../integrations/misc/extract-text"
 import { regexSearchFiles } from "../services/ripgrep"
 import { parseMentions } from "./mentions/context-mentions"
 import { UrlContentFetcher } from "../services/browser/UrlContentFetcher"
 import { diagnosticsToProblemsString, getNewDiagnostics } from "../integrations/diagnostics"
-import { arePathsEqual } from "../utils/path-helpers"
+import { arePathsEqual } from "../utils/path"
 
 const SYSTEM_PROMPT = async (
 	supportsImages: boolean

src/core/sliding-window/index.ts

@@ -0,0 +1,26 @@
+import { Anthropic } from "@anthropic-ai/sdk"
+
+/*
+We can't implement a dynamically updating sliding window as it would break prompt cache
+every time. To maintain the benefits of caching, we need to keep conversation history
+static. This operation should be performed as infrequently as possible. If a user reaches
+a 200k context, we can assume that the first half is likely irrelevant to their current task.
+Therefore, this function should only be called when absolutely necessary to fit within
+context limits, not as a continuous process.
+*/
+export function truncateHalfConversation(
+	messages: Anthropic.Messages.MessageParam[]
+): Anthropic.Messages.MessageParam[] {
+	// API expects messages to be in user-assistant order, and tool use messages must be followed by tool results. We need to maintain this structure while truncating.
+
+	// Always keep the first Task message (this includes the project's file structure in environment_details)
+	const truncatedMessages = [messages[0]]
+
+	// Remove half of user-assistant pairs
+	const messagesToRemove = Math.floor(messages.length / 4) * 2 // has to be even number
+
+	const remainingMessages = messages.slice(messagesToRemove + 1) // has to start with assistant message since tool result cannot follow assistant message with no tool use
+	truncatedMessages.push(...remainingMessages)
+
+	return truncatedMessages
+}