瀏覽代碼

fix: preserve reasoning_content in condense summary for DeepSeek-reasoner (#10292)

Hannes Rudolph 6 天之前
父節點
當前提交
89e9261367

+ 13 - 1
src/core/condense/__tests__/condense.spec.ts

@@ -86,7 +86,19 @@ describe("Condense", () => {
 			// Verify we have a summary message
 			const summaryMessage = result.messages.find((msg) => msg.isSummary)
 			expect(summaryMessage).toBeTruthy()
-			expect(summaryMessage?.content).toBe("Mock summary of the conversation")
+			// Summary content is now always an array with a synthetic reasoning block + text block
+			// for DeepSeek-reasoner compatibility
+			expect(Array.isArray(summaryMessage?.content)).toBe(true)
+			const contentArray = summaryMessage?.content as Anthropic.Messages.ContentBlockParam[]
+			expect(contentArray).toHaveLength(2)
+			expect(contentArray[0]).toEqual({
+				type: "reasoning",
+				text: "Condensing conversation context. The summary below captures the key information from the prior conversation.",
+			})
+			expect(contentArray[1]).toEqual({
+				type: "text",
+				text: "Mock summary of the conversation",
+			})
 
 			// With non-destructive condensing, all messages are retained (tagged but not deleted)
 			// Use getEffectiveApiHistory to verify the effective view matches the old behavior

+ 256 - 9
src/core/condense/__tests__/index.spec.ts

@@ -246,6 +246,94 @@ describe("getKeepMessagesWithToolBlocks", () => {
 		expect(result.keepMessages).toEqual(messages)
 		expect(result.toolUseBlocksToPreserve).toHaveLength(0)
 	})
+
+	it("should preserve reasoning blocks alongside tool_use blocks for DeepSeek/Z.ai interleaved thinking", () => {
+		const reasoningBlock = {
+			type: "reasoning" as const,
+			text: "Let me think about this step by step...",
+		}
+		const toolUseBlock = {
+			type: "tool_use" as const,
+			id: "toolu_deepseek_123",
+			name: "read_file",
+			input: { path: "test.txt" },
+		}
+		const toolResultBlock = {
+			type: "tool_result" as const,
+			tool_use_id: "toolu_deepseek_123",
+			content: "file contents",
+		}
+
+		const messages: ApiMessage[] = [
+			{ role: "user", content: "Hello", ts: 1 },
+			{ role: "assistant", content: "Let me help", ts: 2 },
+			{ role: "user", content: "Please read the file", ts: 3 },
+			{
+				role: "assistant",
+				// DeepSeek stores reasoning as content blocks alongside tool_use
+				content: [reasoningBlock as any, { type: "text" as const, text: "Reading file..." }, toolUseBlock],
+				ts: 4,
+			},
+			{
+				role: "user",
+				content: [toolResultBlock, { type: "text" as const, text: "Continue" }],
+				ts: 5,
+			},
+			{ role: "assistant", content: "Got it, the file says...", ts: 6 },
+			{ role: "user", content: "Thanks", ts: 7 },
+		]
+
+		const result = getKeepMessagesWithToolBlocks(messages, 3)
+
+		// keepMessages should be the last 3 messages
+		expect(result.keepMessages).toHaveLength(3)
+		expect(result.keepMessages[0].ts).toBe(5)
+
+		// Should preserve the tool_use block
+		expect(result.toolUseBlocksToPreserve).toHaveLength(1)
+		expect(result.toolUseBlocksToPreserve[0]).toEqual(toolUseBlock)
+
+		// Should preserve the reasoning block for DeepSeek/Z.ai interleaved thinking
+		expect(result.reasoningBlocksToPreserve).toHaveLength(1)
+		expect((result.reasoningBlocksToPreserve[0] as any).type).toBe("reasoning")
+		expect((result.reasoningBlocksToPreserve[0] as any).text).toBe("Let me think about this step by step...")
+	})
+
+	it("should return empty reasoningBlocksToPreserve when no reasoning blocks present", () => {
+		const toolUseBlock = {
+			type: "tool_use" as const,
+			id: "toolu_123",
+			name: "read_file",
+			input: { path: "test.txt" },
+		}
+		const toolResultBlock = {
+			type: "tool_result" as const,
+			tool_use_id: "toolu_123",
+			content: "file contents",
+		}
+
+		const messages: ApiMessage[] = [
+			{ role: "user", content: "Hello", ts: 1 },
+			{
+				role: "assistant",
+				// No reasoning block, just text and tool_use
+				content: [{ type: "text" as const, text: "Reading file..." }, toolUseBlock],
+				ts: 2,
+			},
+			{
+				role: "user",
+				content: [toolResultBlock],
+				ts: 3,
+			},
+			{ role: "assistant", content: "Done", ts: 4 },
+			{ role: "user", content: "Thanks", ts: 5 },
+		]
+
+		const result = getKeepMessagesWithToolBlocks(messages, 3)
+
+		expect(result.toolUseBlocksToPreserve).toHaveLength(1)
+		expect(result.reasoningBlocksToPreserve).toHaveLength(0)
+	})
 })
 
 describe("getMessagesSinceLastSummary", () => {
@@ -422,7 +510,14 @@ describe("summarizeConversation", () => {
 		const summaryMessage = result.messages.find((m) => m.isSummary)
 		expect(summaryMessage).toBeDefined()
 		expect(summaryMessage!.role).toBe("assistant")
-		expect(summaryMessage!.content).toBe("This is a summary")
+		// Summary content is now always an array with [synthetic reasoning, text]
+		// for DeepSeek-reasoner compatibility (requires reasoning_content on all assistant messages)
+		expect(Array.isArray(summaryMessage!.content)).toBe(true)
+		const content = summaryMessage!.content as any[]
+		expect(content).toHaveLength(2)
+		expect(content[0].type).toBe("reasoning")
+		expect(content[1].type).toBe("text")
+		expect(content[1].text).toBe("This is a summary")
 		expect(summaryMessage!.isSummary).toBe(true)
 
 		// Verify that the effective API history matches expected: first + summary + last N messages
@@ -827,14 +922,16 @@ describe("summarizeConversation", () => {
 		expect(summaryMessage!.isSummary).toBe(true)
 		expect(Array.isArray(summaryMessage!.content)).toBe(true)
 
-		// Content should be [text block, tool_use block]
+		// Content should be [synthetic reasoning, text block, tool_use block]
+		// The synthetic reasoning is always added for DeepSeek-reasoner compatibility
 		const content = summaryMessage!.content as Anthropic.Messages.ContentBlockParam[]
-		expect(content).toHaveLength(2)
-		expect(content[0].type).toBe("text")
-		expect((content[0] as Anthropic.Messages.TextBlockParam).text).toBe("Summary of conversation")
-		expect(content[1].type).toBe("tool_use")
-		expect((content[1] as Anthropic.Messages.ToolUseBlockParam).id).toBe("toolu_123")
-		expect((content[1] as Anthropic.Messages.ToolUseBlockParam).name).toBe("read_file")
+		expect(content).toHaveLength(3)
+		expect((content[0] as any).type).toBe("reasoning") // Synthetic reasoning for DeepSeek
+		expect(content[1].type).toBe("text")
+		expect((content[1] as Anthropic.Messages.TextBlockParam).text).toBe("Summary of conversation")
+		expect(content[2].type).toBe("tool_use")
+		expect((content[2] as Anthropic.Messages.ToolUseBlockParam).id).toBe("toolu_123")
+		expect((content[2] as Anthropic.Messages.ToolUseBlockParam).name).toBe("read_file")
 
 		// With non-destructive condensing, all messages are retained plus the summary
 		expect(result.messages.length).toBe(messages.length + 1) // all original + summary
@@ -981,7 +1078,10 @@ describe("summarizeConversation", () => {
 		expect(summaryMessage).toBeDefined()
 		expect(Array.isArray(summaryMessage!.content)).toBe(true)
 		const summaryContent = summaryMessage!.content as Anthropic.Messages.ContentBlockParam[]
-		expect(summaryContent[0]).toEqual({ type: "text", text: "This is a summary" })
+		// First block is synthetic reasoning for DeepSeek-reasoner compatibility
+		expect((summaryContent[0] as any).type).toBe("reasoning")
+		// Second block is the text summary
+		expect(summaryContent[1]).toEqual({ type: "text", text: "This is a summary" })
 
 		const preservedToolUses = summaryContent.filter(
 			(block): block is Anthropic.Messages.ToolUseBlockParam => block.type === "tool_use",
@@ -989,6 +1089,153 @@ describe("summarizeConversation", () => {
 		expect(preservedToolUses).toHaveLength(2)
 		expect(preservedToolUses.map((block) => block.id)).toEqual(["toolu_parallel_1", "toolu_parallel_2"])
 	})
+
+	it("should preserve reasoning blocks in summary message for DeepSeek/Z.ai interleaved thinking", async () => {
+		const reasoningBlock = {
+			type: "reasoning" as const,
+			text: "Let me think about this step by step...",
+		}
+		const toolUseBlock = {
+			type: "tool_use" as const,
+			id: "toolu_deepseek_reason",
+			name: "read_file",
+			input: { path: "test.txt" },
+		}
+		const toolResultBlock = {
+			type: "tool_result" as const,
+			tool_use_id: "toolu_deepseek_reason",
+			content: "file contents",
+		}
+
+		const messages: ApiMessage[] = [
+			{ role: "user", content: "Hello", ts: 1 },
+			{ role: "assistant", content: "Let me help", ts: 2 },
+			{ role: "user", content: "Please read the file", ts: 3 },
+			{
+				role: "assistant",
+				// DeepSeek stores reasoning as content blocks alongside tool_use
+				content: [reasoningBlock as any, { type: "text" as const, text: "Reading file..." }, toolUseBlock],
+				ts: 4,
+			},
+			{
+				role: "user",
+				content: [toolResultBlock, { type: "text" as const, text: "Continue" }],
+				ts: 5,
+			},
+			{ role: "assistant", content: "Got it, the file says...", ts: 6 },
+			{ role: "user", content: "Thanks", ts: 7 },
+		]
+
+		// Create a stream with usage information
+		const streamWithUsage = (async function* () {
+			yield { type: "text" as const, text: "Summary of conversation" }
+			yield { type: "usage" as const, totalCost: 0.05, outputTokens: 100 }
+		})()
+
+		mockApiHandler.createMessage = vi.fn().mockReturnValue(streamWithUsage) as any
+		mockApiHandler.countTokens = vi.fn().mockImplementation(() => Promise.resolve(50)) as any
+
+		const result = await summarizeConversation(
+			messages,
+			mockApiHandler,
+			defaultSystemPrompt,
+			taskId,
+			DEFAULT_PREV_CONTEXT_TOKENS,
+			false, // isAutomaticTrigger
+			undefined, // customCondensingPrompt
+			undefined, // condensingApiHandler
+			true, // useNativeTools - required for tool_use block preservation
+		)
+
+		// Find the summary message
+		const summaryMessage = result.messages.find((m) => m.isSummary)
+		expect(summaryMessage).toBeDefined()
+		expect(summaryMessage!.role).toBe("assistant")
+		expect(summaryMessage!.isSummary).toBe(true)
+		expect(Array.isArray(summaryMessage!.content)).toBe(true)
+
+		// Content should be [synthetic reasoning, preserved reasoning, text block, tool_use block]
+		// - Synthetic reasoning is always added for DeepSeek-reasoner compatibility
+		// - Preserved reasoning from the condensed assistant message
+		// This order ensures reasoning_content is always present for DeepSeek/Z.ai
+		const content = summaryMessage!.content as Anthropic.Messages.ContentBlockParam[]
+		expect(content).toHaveLength(4)
+
+		// First block should be synthetic reasoning
+		expect((content[0] as any).type).toBe("reasoning")
+		expect((content[0] as any).text).toContain("Condensing conversation context")
+
+		// Second block should be preserved reasoning from the condensed message
+		expect((content[1] as any).type).toBe("reasoning")
+		expect((content[1] as any).text).toBe("Let me think about this step by step...")
+
+		// Third block should be text (the summary)
+		expect(content[2].type).toBe("text")
+		expect((content[2] as Anthropic.Messages.TextBlockParam).text).toBe("Summary of conversation")
+
+		// Fourth block should be tool_use
+		expect(content[3].type).toBe("tool_use")
+		expect((content[3] as Anthropic.Messages.ToolUseBlockParam).id).toBe("toolu_deepseek_reason")
+
+		expect(result.error).toBeUndefined()
+	})
+
+	it("should include synthetic reasoning block in summary for DeepSeek-reasoner compatibility even without tool_use blocks", async () => {
+		// This test verifies the fix for the DeepSeek-reasoner 400 error:
+		// "Missing `reasoning_content` field in the assistant message at message index 1"
+		// DeepSeek-reasoner requires reasoning_content on ALL assistant messages, not just those with tool_calls.
+		// After condensation, the summary becomes an assistant message that needs reasoning_content.
+		const messages: ApiMessage[] = [
+			{ role: "user", content: "Tell me a joke", ts: 1 },
+			{ role: "assistant", content: "Why did the programmer quit?", ts: 2 },
+			{ role: "user", content: "I don't know, why?", ts: 3 },
+			{ role: "assistant", content: "He didn't get arrays!", ts: 4 },
+			{ role: "user", content: "Another one please", ts: 5 },
+			{ role: "assistant", content: "Why do programmers prefer dark mode?", ts: 6 },
+			{ role: "user", content: "Why?", ts: 7 },
+		]
+
+		// Create a stream with usage information (no tool calls in this conversation)
+		const streamWithUsage = (async function* () {
+			yield { type: "text" as const, text: "Summary: User requested jokes." }
+			yield { type: "usage" as const, totalCost: 0.05, outputTokens: 100 }
+		})()
+
+		mockApiHandler.createMessage = vi.fn().mockReturnValue(streamWithUsage) as any
+		mockApiHandler.countTokens = vi.fn().mockImplementation(() => Promise.resolve(50)) as any
+
+		const result = await summarizeConversation(
+			messages,
+			mockApiHandler,
+			defaultSystemPrompt,
+			taskId,
+			DEFAULT_PREV_CONTEXT_TOKENS,
+			false, // isAutomaticTrigger
+			undefined, // customCondensingPrompt
+			undefined, // condensingApiHandler
+			false, // useNativeTools - not using tools in this test
+		)
+
+		// Find the summary message
+		const summaryMessage = result.messages.find((m) => m.isSummary)
+		expect(summaryMessage).toBeDefined()
+		expect(summaryMessage!.role).toBe("assistant")
+		expect(summaryMessage!.isSummary).toBe(true)
+
+		// CRITICAL: Content must be an array with a synthetic reasoning block
+		// This is required for DeepSeek-reasoner which needs reasoning_content on all assistant messages
+		expect(Array.isArray(summaryMessage!.content)).toBe(true)
+		const content = summaryMessage!.content as any[]
+
+		// Should have [synthetic reasoning, text]
+		expect(content).toHaveLength(2)
+		expect(content[0].type).toBe("reasoning")
+		expect(content[0].text).toContain("Condensing conversation context")
+		expect(content[1].type).toBe("text")
+		expect(content[1].text).toBe("Summary: User requested jokes.")
+
+		expect(result.error).toBeUndefined()
+	})
 })
 
 describe("summarizeConversation with custom settings", () => {

+ 82 - 20
src/core/condense/index.ts

@@ -30,22 +30,49 @@ function getToolUseBlocks(message: ApiMessage): Anthropic.Messages.ToolUseBlock[
 	return message.content.filter((block) => block.type === "tool_use") as Anthropic.Messages.ToolUseBlock[]
 }
 
+/**
+ * Gets reasoning blocks from a message's content array.
+ * Task stores reasoning as {type: "reasoning", text: "..."} blocks,
+ * which convertToR1Format and convertToZAiFormat already know how to extract.
+ */
+function getReasoningBlocks(message: ApiMessage): Anthropic.Messages.ContentBlockParam[] {
+	if (message.role !== "assistant" || typeof message.content === "string") {
+		return []
+	}
+	// Filter for reasoning blocks and cast to ContentBlockParam (the type field is compatible)
+	return message.content.filter((block) => (block as any).type === "reasoning") as any[]
+}
+
+/**
+ * Result of getKeepMessagesWithToolBlocks
+ */
+export type KeepMessagesResult = {
+	keepMessages: ApiMessage[]
+	toolUseBlocksToPreserve: Anthropic.Messages.ToolUseBlock[]
+	// Reasoning blocks from the preceding assistant message, needed for DeepSeek/Z.ai
+	// when tool_use blocks are preserved. Task stores reasoning as {type: "reasoning", text: "..."}
+	// blocks, and convertToR1Format/convertToZAiFormat already extract these.
+	reasoningBlocksToPreserve: Anthropic.Messages.ContentBlockParam[]
+}
+
 /**
  * Extracts tool_use blocks that need to be preserved to match tool_result blocks in keepMessages.
  * When the first kept message is a user message with tool_result blocks,
  * we need to find the corresponding tool_use blocks from the preceding assistant message.
  * These tool_use blocks will be appended to the summary message to maintain proper pairing.
  *
+ * Also extracts reasoning blocks from the preceding assistant message, which are required
+ * by DeepSeek and Z.ai for interleaved thinking mode. Without these, the API returns a 400 error
+ * "Missing reasoning_content field in the assistant message".
+ * See: https://api-docs.deepseek.com/guides/thinking_mode#tool-calls
+ *
  * @param messages - The full conversation messages
  * @param keepCount - The number of messages to keep from the end
- * @returns Object containing keepMessages and any tool_use blocks to preserve
+ * @returns Object containing keepMessages, tool_use blocks, and reasoning blocks to preserve
  */
-export function getKeepMessagesWithToolBlocks(
-	messages: ApiMessage[],
-	keepCount: number,
-): { keepMessages: ApiMessage[]; toolUseBlocksToPreserve: Anthropic.Messages.ToolUseBlock[] } {
+export function getKeepMessagesWithToolBlocks(messages: ApiMessage[], keepCount: number): KeepMessagesResult {
 	if (messages.length <= keepCount) {
-		return { keepMessages: messages, toolUseBlocksToPreserve: [] }
+		return { keepMessages: messages, toolUseBlocksToPreserve: [], reasoningBlocksToPreserve: [] }
 	}
 
 	const startIndex = messages.length - keepCount
@@ -59,13 +86,20 @@ export function getKeepMessagesWithToolBlocks(
 			const precedingMessage = messages[precedingIndex]
 			const toolUseBlocks = getToolUseBlocks(precedingMessage)
 			if (toolUseBlocks.length > 0) {
-				// Return the tool_use blocks to be merged into the summary message
-				return { keepMessages, toolUseBlocksToPreserve: toolUseBlocks }
+				// Also extract reasoning blocks for DeepSeek/Z.ai interleaved thinking
+				// Task stores reasoning as {type: "reasoning", text: "..."} content blocks
+				const reasoningBlocks = getReasoningBlocks(precedingMessage)
+				// Return the tool_use blocks and reasoning blocks to be merged into the summary message
+				return {
+					keepMessages,
+					toolUseBlocksToPreserve: toolUseBlocks,
+					reasoningBlocksToPreserve: reasoningBlocks,
+				}
 			}
 		}
 	}
 
-	return { keepMessages, toolUseBlocksToPreserve: [] }
+	return { keepMessages, toolUseBlocksToPreserve: [], reasoningBlocksToPreserve: [] }
 }
 
 export const N_MESSAGES_TO_KEEP = 3
@@ -168,11 +202,15 @@ export async function summarizeConversation(
 	// Always preserve the first message (which may contain slash command content)
 	const firstMessage = messages[0]
 
-	// Get keepMessages and any tool_use blocks that need to be preserved for tool_result pairing
-	// Only preserve tool_use blocks when using native tools protocol (XML protocol doesn't need them)
-	const { keepMessages, toolUseBlocksToPreserve } = useNativeTools
+	// Get keepMessages and any tool_use/reasoning blocks that need to be preserved for tool_result pairing
+	// Only preserve these blocks when using native tools protocol (XML protocol doesn't need them)
+	const { keepMessages, toolUseBlocksToPreserve, reasoningBlocksToPreserve } = useNativeTools
 		? getKeepMessagesWithToolBlocks(messages, N_MESSAGES_TO_KEEP)
-		: { keepMessages: messages.slice(-N_MESSAGES_TO_KEEP), toolUseBlocksToPreserve: [] }
+		: {
+				keepMessages: messages.slice(-N_MESSAGES_TO_KEEP),
+				toolUseBlocksToPreserve: [],
+				reasoningBlocksToPreserve: [],
+			}
 
 	const keepStartIndex = Math.max(messages.length - N_MESSAGES_TO_KEEP, 0)
 	const includeFirstKeptMessageInSummary = toolUseBlocksToPreserve.length > 0
@@ -257,15 +295,39 @@ export async function summarizeConversation(
 	}
 
 	// Build the summary message content
-	// If there are tool_use blocks to preserve (for tool_result pairing), append them to the summary
-	let summaryContent: string | Anthropic.Messages.ContentBlockParam[]
+	// CRITICAL: Always include a reasoning block in the summary for DeepSeek-reasoner compatibility.
+	// DeepSeek-reasoner requires `reasoning_content` on ALL assistant messages, not just those with tool_calls.
+	// Without this, we get: "400 Missing `reasoning_content` field in the assistant message"
+	// See: https://api-docs.deepseek.com/guides/thinking_mode
+	//
+	// The summary content structure is:
+	// 1. Synthetic reasoning block (always present) - for DeepSeek-reasoner compatibility
+	// 2. Any preserved reasoning blocks from the condensed assistant message (if tool_use blocks are preserved)
+	// 3. Text block with the summary
+	// 4. Tool_use blocks (if any need to be preserved for tool_result pairing)
+
+	// Create a synthetic reasoning block that explains the summary
+	// This is minimal but satisfies DeepSeek's requirement for reasoning_content on all assistant messages
+	const syntheticReasoningBlock = {
+		type: "reasoning" as const,
+		text: "Condensing conversation context. The summary below captures the key information from the prior conversation.",
+	}
+
+	const textBlock: Anthropic.Messages.TextBlockParam = { type: "text", text: summary }
+
+	let summaryContent: Anthropic.Messages.ContentBlockParam[]
 	if (toolUseBlocksToPreserve.length > 0) {
-		// Create content array with text block followed by tool_use blocks
-		// Use TextBlockParam which doesn't require citations field
-		const textBlock: Anthropic.Messages.TextBlockParam = { type: "text", text: summary }
-		summaryContent = [textBlock, ...toolUseBlocksToPreserve]
+		// Include: synthetic reasoning, preserved reasoning (if any), summary text, and tool_use blocks
+		summaryContent = [
+			syntheticReasoningBlock as unknown as Anthropic.Messages.ContentBlockParam,
+			...reasoningBlocksToPreserve,
+			textBlock,
+			...toolUseBlocksToPreserve,
+		]
 	} else {
-		summaryContent = summary
+		// Include: synthetic reasoning and summary text
+		// This ensures the summary always has reasoning_content for DeepSeek-reasoner
+		summaryContent = [syntheticReasoningBlock as unknown as Anthropic.Messages.ContentBlockParam, textBlock]
 	}
 
 	// Generate a unique condenseId for this summary

+ 3 - 0
src/core/task-persistence/apiMessages.ts

@@ -20,6 +20,9 @@ export type ApiMessage = Anthropic.MessageParam & {
 	text?: string
 	// For OpenRouter reasoning_details array format (used by Gemini 3, etc.)
 	reasoning_details?: any[]
+	// For DeepSeek/Z.ai interleaved thinking: reasoning_content that must be preserved during tool call sequences
+	// See: https://api-docs.deepseek.com/guides/thinking_mode#tool-calls
+	reasoning_content?: string
 	// For non-destructive condense: unique identifier for summary messages
 	condenseId?: string
 	// For non-destructive condense: points to the condenseId of the summary that replaces this message