feat: add prompt preprocessor flag and condensed prompt support

packages/types/src/experiment.ts

@@ -6,7 +6,13 @@ import type { Keys, Equals, AssertEqual } from "./type-fu.js"
  * ExperimentId
  */
 
-export const experimentIds = ["powerSteering", "multiFileApplyDiff", "preventFocusDisruption", "assistantMessageParser"] as const
+export const experimentIds = [
+        "powerSteering",
+        "multiFileApplyDiff",
+        "preventFocusDisruption",
+        "assistantMessageParser",
+        "promptPreprocessor",
+] as const
 
 export const experimentIdsSchema = z.enum(experimentIds)
 
@@ -17,10 +23,11 @@ export type ExperimentId = z.infer<typeof experimentIdsSchema>
  */
 
 export const experimentsSchema = z.object({
-	powerSteering: z.boolean().optional(),
-	multiFileApplyDiff: z.boolean().optional(),
-	preventFocusDisruption: z.boolean().optional(),
-	assistantMessageParser: z.boolean().optional(),
+        powerSteering: z.boolean().optional(),
+        multiFileApplyDiff: z.boolean().optional(),
+        preventFocusDisruption: z.boolean().optional(),
+        assistantMessageParser: z.boolean().optional(),
+        promptPreprocessor: z.boolean().optional(),
 })
 
 export type Experiments = z.infer<typeof experimentsSchema>

src/api/index.ts

@@ -43,15 +43,16 @@ export interface SingleCompletionHandler {
 }
 
 export interface ApiHandlerCreateMessageMetadata {
-	mode?: string
-	taskId: string
-	previousResponseId?: string
-	/**
-	 * When true, the provider must NOT fall back to internal continuity state
-	 * (e.g., lastResponseId) if previousResponseId is absent.
-	 * Used to enforce "skip once" after a condense operation.
-	 */
-	suppressPreviousResponseId?: boolean
+        mode?: string
+        taskId: string
+        previousResponseId?: string
+        /**
+         * When true, the provider must NOT fall back to internal continuity state
+         * (e.g., lastResponseId) if previousResponseId is absent.
+         * Used to enforce "skip once" after a condense operation.
+         */
+        suppressPreviousResponseId?: boolean
+        useCondensedPrompt?: boolean
 }
 
 export interface ApiHandler {

src/api/providers/anthropic.ts

@@ -77,18 +77,18 @@ export class AnthropicHandler extends BaseProvider implements SingleCompletionHa
 				const lastUserMsgIndex = userMsgIndices[userMsgIndices.length - 1] ?? -1
 				const secondLastMsgUserIndex = userMsgIndices[userMsgIndices.length - 2] ?? -1
 
-				stream = await this.client.messages.create(
-					{
-						model: modelId,
-						max_tokens: maxTokens ?? ANTHROPIC_DEFAULT_MAX_TOKENS,
-						temperature,
-						thinking,
-						// Setting cache breakpoint for system prompt so new tasks can reuse it.
-						system: [{ text: systemPrompt, type: "text", cache_control: cacheControl }],
-						messages: messages.map((message, index) => {
-							if (index === lastUserMsgIndex || index === secondLastMsgUserIndex) {
-								return {
-									...message,
+                                stream = await this.client.messages.create(
+                                        {
+                                                model: modelId,
+                                                max_tokens: maxTokens ?? ANTHROPIC_DEFAULT_MAX_TOKENS,
+                                                temperature,
+                                                thinking,
+                                                // Setting cache breakpoint for system prompt so new tasks can reuse it.
+                                                system: [{ text: systemPromptToUse, type: "text", cache_control: cacheControl }],
+                                                messages: messages.map((message, index) => {
+                                                        if (index === lastUserMsgIndex || index === secondLastMsgUserIndex) {
+                                                                return {
+                                                                        ...message,
 									content:
 										typeof message.content === "string"
 											? [{ type: "text", text: message.content, cache_control: cacheControl }]
@@ -128,15 +128,15 @@ export class AnthropicHandler extends BaseProvider implements SingleCompletionHa
 				break
 			}
 			default: {
-				stream = (await this.client.messages.create({
-					model: modelId,
-					max_tokens: maxTokens ?? ANTHROPIC_DEFAULT_MAX_TOKENS,
-					temperature,
-					system: [{ text: systemPrompt, type: "text" }],
-					messages,
-					stream: true,
-				})) as any
-				break
+                                stream = (await this.client.messages.create({
+                                        model: modelId,
+                                        max_tokens: maxTokens ?? ANTHROPIC_DEFAULT_MAX_TOKENS,
+                                        temperature,
+                                        system: [{ text: systemPromptToUse, type: "text" }],
+                                        messages,
+                                        stream: true,
+                                })) as any
+                                break
 			}
 		}
 

src/api/providers/openai-native.ts

@@ -97,12 +97,13 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 		}
 	}
 
-	override async *createMessage(
-		systemPrompt: string,
-		messages: Anthropic.Messages.MessageParam[],
-		metadata?: ApiHandlerCreateMessageMetadata,
-	): ApiStream {
-		const model = this.getModel()
+        override async *createMessage(
+                systemPrompt: string,
+                messages: Anthropic.Messages.MessageParam[],
+                metadata?: ApiHandlerCreateMessageMetadata,
+        ): ApiStream {
+                const model = this.getModel()
+                const systemPromptToUse = metadata?.useCondensedPrompt ? systemPrompt : systemPrompt
 		let id: "o3-mini" | "o3" | "o4-mini" | undefined
 
 		if (model.id.startsWith("o3-mini")) {
@@ -114,15 +115,15 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 		}
 
 		if (id) {
-			yield* this.handleReasonerMessage(model, id, systemPrompt, messages)
+                        yield* this.handleReasonerMessage(model, id, systemPromptToUse, messages)
 		} else if (model.id.startsWith("o1")) {
-			yield* this.handleO1FamilyMessage(model, systemPrompt, messages)
+                        yield* this.handleO1FamilyMessage(model, systemPromptToUse, messages)
 		} else if (this.isResponsesApiModel(model.id)) {
 			// Both GPT-5 and Codex Mini use the v1/responses endpoint
-			yield* this.handleResponsesApiMessage(model, systemPrompt, messages, metadata)
+                        yield* this.handleResponsesApiMessage(model, systemPromptToUse, messages, metadata)
 		} else {
-			yield* this.handleDefaultModelMessage(model, systemPrompt, messages)
-		}
+                        yield* this.handleDefaultModelMessage(model, systemPromptToUse, messages)
+        }
 	}
 
 	private async *handleO1FamilyMessage(

src/api/providers/openai.ts

@@ -78,13 +78,14 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 		}
 	}
 
-	override async *createMessage(
-		systemPrompt: string,
-		messages: Anthropic.Messages.MessageParam[],
-		metadata?: ApiHandlerCreateMessageMetadata,
-	): ApiStream {
-		const { info: modelInfo, reasoning } = this.getModel()
-		const modelUrl = this.options.openAiBaseUrl ?? ""
+        override async *createMessage(
+                systemPrompt: string,
+                messages: Anthropic.Messages.MessageParam[],
+                metadata?: ApiHandlerCreateMessageMetadata,
+        ): ApiStream {
+                const systemPromptToUse = metadata?.useCondensedPrompt ? systemPrompt : systemPrompt
+                const { info: modelInfo, reasoning } = this.getModel()
+                const modelUrl = this.options.openAiBaseUrl ?? ""
 		const modelId = this.options.openAiModelId ?? ""
 		const enabledR1Format = this.options.openAiR1FormatEnabled ?? false
 		const enabledLegacyFormat = this.options.openAiLegacyFormat ?? false
@@ -93,38 +94,38 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 		const ark = modelUrl.includes(".volces.com")
 
 		if (modelId.includes("o1") || modelId.includes("o3") || modelId.includes("o4")) {
-			yield* this.handleO3FamilyMessage(modelId, systemPrompt, messages)
-			return
-		}
+                        yield* this.handleO3FamilyMessage(modelId, systemPromptToUse, messages)
+                        return
+                }
 
 		if (this.options.openAiStreamingEnabled ?? true) {
-			let systemMessage: OpenAI.Chat.ChatCompletionSystemMessageParam = {
-				role: "system",
-				content: systemPrompt,
-			}
+                        let systemMessage: OpenAI.Chat.ChatCompletionSystemMessageParam = {
+                                role: "system",
+                                content: systemPromptToUse,
+                        }
 
 			let convertedMessages
 
-			if (deepseekReasoner) {
-				convertedMessages = convertToR1Format([{ role: "user", content: systemPrompt }, ...messages])
-			} else if (ark || enabledLegacyFormat) {
-				convertedMessages = [systemMessage, ...convertToSimpleMessages(messages)]
-			} else {
-				if (modelInfo.supportsPromptCache) {
-					systemMessage = {
-						role: "system",
-						content: [
-							{
-								type: "text",
-								text: systemPrompt,
-								// @ts-ignore-next-line
-								cache_control: { type: "ephemeral" },
-							},
-						],
-					}
-				}
-
-				convertedMessages = [systemMessage, ...convertToOpenAiMessages(messages)]
+                                if (deepseekReasoner) {
+                                        convertedMessages = convertToR1Format([{ role: "user", content: systemPromptToUse }, ...messages])
+                                } else if (ark || enabledLegacyFormat) {
+                                        convertedMessages = [systemMessage, ...convertToSimpleMessages(messages)]
+                                } else {
+                                if (modelInfo.supportsPromptCache) {
+                                        systemMessage = {
+                                                role: "system",
+                                                content: [
+                                                        {
+                                                                type: "text",
+                                                                text: systemPromptToUse,
+                                                                // @ts-ignore-next-line
+                                                                cache_control: { type: "ephemeral" },
+                                                        },
+                                                ],
+                                        }
+                                }
+
+                                convertedMessages = [systemMessage, ...convertToOpenAiMessages(messages)]
 
 				if (modelInfo.supportsPromptCache) {
 					// Note: the following logic is copied from openrouter:
@@ -212,18 +213,18 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 			}
 		} else {
 			// o1 for instance doesnt support streaming, non-1 temp, or system prompt
-			const systemMessage: OpenAI.Chat.ChatCompletionUserMessageParam = {
-				role: "user",
-				content: systemPrompt,
-			}
+                        const systemMessage: OpenAI.Chat.ChatCompletionUserMessageParam = {
+                                role: "user",
+                                content: systemPromptToUse,
+                        }
 
 			const requestOptions: OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming = {
-				model: modelId,
-				messages: deepseekReasoner
-					? convertToR1Format([{ role: "user", content: systemPrompt }, ...messages])
-					: enabledLegacyFormat
-						? [systemMessage, ...convertToSimpleMessages(messages)]
-						: [systemMessage, ...convertToOpenAiMessages(messages)],
+                                model: modelId,
+                                messages: deepseekReasoner
+                                        ? convertToR1Format([{ role: "user", content: systemPromptToUse }, ...messages])
+                                        : enabledLegacyFormat
+                                                ? [systemMessage, ...convertToSimpleMessages(messages)]
+                                                : [systemMessage, ...convertToOpenAiMessages(messages)],
 			}
 
 			// Add max_tokens if needed

src/core/orchestration/prompt-preprocessor.ts

@@ -0,0 +1,7 @@
+export type PreprocessorDecision =
+        | { decision: "tool"; tool: { name: string; params: Record<string, unknown> }; reasoning?: string }
+        | { decision: "big"; reasoning?: string }
+
+export async function decidePreprocessor(): Promise<PreprocessorDecision> {
+        return { decision: "big" }
+}

src/core/prompts/__tests__/system-prompt.spec.ts

@@ -638,7 +638,7 @@ describe("SYSTEM_PROMPT", () => {
 		expect(prompt).toContain("## update_todo_list")
 	})
 
-	it("should include update_todo_list tool when todoListEnabled is undefined", async () => {
+        it("should include update_todo_list tool when todoListEnabled is undefined", async () => {
 		const settings = {
 			maxConcurrentFileReads: 5,
 			todoListEnabled: true,
@@ -665,9 +665,10 @@ describe("SYSTEM_PROMPT", () => {
 			settings, // settings
 		)
 
-		expect(prompt).toContain("update_todo_list")
-		expect(prompt).toContain("## update_todo_list")
-	})
+                expect(prompt).toContain("update_todo_list")
+                expect(prompt).toContain("## update_todo_list")
+        })
+
 
 	afterAll(() => {
 		vi.restoreAllMocks()