account for system prompt when estimating new context size

src/core/condense/index.ts

@@ -62,6 +62,7 @@ export type SummarizeResponse = {
 export async function summarizeConversation(
 	messages: ApiMessage[],
 	apiHandler: ApiHandler,
+	systemPrompt?: string,
 ): Promise<SummarizeResponse> {
 	const response: SummarizeResponse = { messages, cost: 0, summary: "" }
 	const messagesToSummarize = getMessagesSinceLastSummary(messages.slice(0, -N_MESSAGES_TO_KEEP))
@@ -111,6 +112,9 @@ export async function summarizeConversation(
 	// Count the tokens in the context for the next API request
 	// We only estimate the tokens in summaryMesage if outputTokens is 0, otherwise we use outputTokens
 	const contextMessages = outputTokens ? [...keepMessages] : [summaryMessage, ...keepMessages]
+	if (systemPrompt) {
+		contextMessages.unshift({ role: "user", content: systemPrompt })
+	}
 	const contextBlocks = contextMessages.flatMap((message) =>
 		typeof message.content === "string" ? [{ text: message.content, type: "text" as const }] : message.content,
 	)

src/core/sliding-window/index.ts

@@ -53,6 +53,7 @@ export function truncateConversation(messages: ApiMessage[], fracToRemove: numbe
  * @param {number} maxTokens - The maximum number of tokens allowed.
  * @param {ApiHandler} apiHandler - The API handler to use for token counting.
  * @param {boolean} autoCondenseContext - Whether to use LLM summarization or sliding window implementation
+ * @param {string} systemPrompt - The system prompt, used for estimating the new context size after summarizing.
  * @returns {ApiMessage[]} The original or truncated conversation messages.
  */
 
@@ -63,6 +64,7 @@ type TruncateOptions = {
 	maxTokens?: number | null
 	apiHandler: ApiHandler
 	autoCondenseContext?: boolean
+	systemPrompt?: string
 }
 
 type TruncateResponse = SummarizeResponse & { prevContextTokens: number }
@@ -81,6 +83,7 @@ export async function truncateConversationIfNeeded({
 	maxTokens,
 	apiHandler,
 	autoCondenseContext,
+	systemPrompt,
 }: TruncateOptions): Promise<TruncateResponse> {
 	// Calculate the maximum tokens reserved for response
 	const reservedTokens = maxTokens || contextWindow * 0.2
@@ -103,7 +106,7 @@ export async function truncateConversationIfNeeded({
 	if (effectiveTokens <= allowedTokens) {
 		return { messages, summary: "", cost: 0, prevContextTokens: effectiveTokens }
 	} else if (autoCondenseContext) {
-		const result = await summarizeConversation(messages, apiHandler)
+		const result = await summarizeConversation(messages, apiHandler, systemPrompt)
 		if (messages !== result.messages) {
 			return { ...result, prevContextTokens: effectiveTokens }
 		}

src/core/task/Task.ts

@@ -1479,6 +1479,7 @@ export class Task extends EventEmitter<ClineEvents> {
 				contextWindow,
 				apiHandler: this.api,
 				autoCondenseContext,
+				systemPrompt,
 			})
 			if (truncateResult.messages !== this.apiConversationHistory) {
 				await this.overwriteApiConversationHistory(truncateResult.messages)