调整记忆和压缩生成内容

Author: phantom5099

docs/context.md

@@ -29,52 +29,46 @@ Coding Code 采用两层压缩策略，在不同阈值下自动触发：
 |--------|-----|------|
 | 触发阈值 | `promptEstimate > modelMaxTokens * 0.9` | prompt 估算超过模型最大 token 90% 时触发 |
 | 保留最近 turn | 1 | 保留最近 1 个 turn 不压缩 |
-| 压缩方式 | 调用 LLM 生成摘要 | 输出 `<summary>...</summary>` 块 |
+| 压缩方式 | 调用 LLM 生成摘要 | 整段输出即摘要（全量替换，不做标签抽取） |
 | 增量压缩 | 是 | 找到已有 SummaryEvent，只压缩 `endTurnId` 之后的事件 |
 | 失败追踪 | 连续 3 次失败后停止 | 24 小时 TTL 后重置 |
 
 ---
 
 ## 压缩输出格式
 
-LLM 压缩的摘要包含 10 个固定小节：
+LLM 压缩要求模型按 10 个固定小节输出；**模型的整段输出即摘要文本**，不做标签抽取：
 
 ```
-<analysis>
-自由推理区域，分析对话内容和关键信息
-</analysis>
-
-<summary>
-### Primary Request
+## 1. Primary Request and Intent
 用户的核心请求
 
-### Key Technical Concepts
+## 2. Key Technical Concepts
 涉及的关键技术概念
 
-### Files and Code Sections
+## 3. Files and Code Sections
 相关文件和代码段
 
-### Errors and Fixes
+## 4. Errors and Fixes
 遇到的错误和修复
 
-### Problem Solving
+## 5. Problem Solving
 问题解决过程
 
-### Decision Rationale
-决策理由
+## 6. Decision Rationale and Rejected Approaches
+决策理由与被否决的方案
 
-### All User Messages
-所有用户消息摘要
+## 7. All User Messages
+所有用户消息
 
-### Pending Tasks
+## 8. Pending Tasks
 待处理任务
 
-### Current Work
+## 9. Current Work
 当前工作内容
 
-### Optional Next Step
+## 10. Optional Next Step
 可选的下一步
-</summary>
 ```
 
 ---

docs/memory.md

@@ -31,10 +31,10 @@ Coding Code 支持跨会话的长期记忆：自动从对话中提取关键信
 
 1. 读取记忆文件全文作为"已有记忆"
 2. 将会话记录（按 `[user]` / `[assistant]` / `[tool:名称]` 标注）与已有记忆一起发送给 LLM
-3. LLM 输出整份**最新版记忆**，放在 `<memory>...</memory>` 块中
+3. LLM 输出整份**最新版记忆**（整段输出即内容，不做标签或片段抽取）
 4. 直接用输出内容整体替换记忆文件（受字节上限约束）
 
-模型自行决定更新哪些内容：可以新增条目、修改过时信息、删除不再相关的内容，代码不做"模型只改动哪部分"的任何假设。若模型没有输出有效内容、或输出与当前文件一致，则不写入。
+模型自行决定更新哪些内容：可以新增条目、修改过时信息、删除不再相关的内容，代码不做"模型只改动哪部分"的任何假设。若模型输出为空、或输出与当前文件一致，则不写入。
 
 ### 提取提示词
 

packages/codingcode/src/context/compaction-prompt.ts

@@ -1,12 +1,5 @@
-export const COMPACTION_SYSTEM_PROMPT = `You analyze and then summarize an agent conversation transcript.
+export const COMPACTION_SYSTEM_PROMPT = `Summarize an agent conversation transcript into the sections below.
 
-Output exactly two top-level blocks:
-
-<analysis>
-Free-form notes about the conversation. Identify the user's goal, what was done, what was learned, what remains. This block is for your reasoning — be thorough.
-</analysis>
-
-<summary>
 ## 1. Primary Request and Intent
 The user's overall objective and concrete asks.
 
@@ -35,5 +28,4 @@ Work the user explicitly asked for that is not yet done.
 What was happening at the moment of compaction.
 
 ## 10. Optional Next Step
-A recommended next action consistent with the user's intent.
-</summary>`;
+A recommended next action consistent with the user's intent.`;

packages/codingcode/src/context/context.ts

@@ -380,12 +380,11 @@ export const ContextLayer = Layer.effect(ContextService, Effect.gen(function* ()
           .complete({ messages: [userMsg], system }, model)
           .pipe(Effect.either);
         if (result._tag === 'Left') return null;
-        return extractSummary(result.right.content.trim());
+        return normalizeSummary(result.right.content);
       }).pipe(Effect.catchAllCause(() => Effect.succeed(null)));
 
-    function extractSummary(raw: string): string {
-      const m = raw.match(/<summary>([\s\S]*?)<\/summary>/);
-      return (m?.[1] ?? raw).trim();
+    function normalizeSummary(raw: string): string {
+      return raw.trim();
     }
 
     const willCompact = (

packages/codingcode/src/memory/extractor.ts

@@ -1,25 +1,21 @@
 import { Effect } from 'effect';
 import type { LLMShape } from '../llm/port.js';
 
-const SYSTEM_PROMPT = `你是记忆整理器。基于"已有记忆"和"会话记录"，输出整份最新版长期记忆，放在 <memory>...</memory> 块中，不要输出其它内容。
+const SYSTEM_PROMPT = `你是记忆整理器。基于"已有记忆"和"会话记录"，输出整份最新版长期记忆。
 
 规则：
 - 只保留值得跨会话记住的信息：用户角色、偏好与对 Agent 的纠正，项目架构决策、技术选型与部署信息，外部资源与链接等。
 - 忽略临时内容：一次性任务、调试过程、报错堆栈、闲聊。
 - 更新哪些内容由你决定：在已有记忆基础上自行增、删、改，输出必须是一份完整、自洽的最新记忆，而不是只输出变动部分。
 - 旧记忆与对话新信息矛盾时以最新为准；同一会话前后不一致时以最后出现为准。
 - 不要编造对话中未出现的信息。
-- 若没有值得记住的新信息且已有记忆为空，输出 <memory></memory>。
 
 格式：
 - 纯 Markdown，用 "### 主题" 小节组织，小节下用 "- " 列要点。
-- 条目需具体、自包含，避免"上面提到的那个"这类指代。
-- <memory> 内不要带任何解释性文字。`;
+- 条目需具体、自包含，避免"上面提到的那个"这类指代。`;
 
-function extractFrom(content: string): string | null {
-  const memoryMatch = content.match(/<memory>([\s\S]*?)<\/memory>/);
-  if (!memoryMatch) return null;
-  return memoryMatch[1]!.trim() || null;
+function normalizeMemoryOutput(content: string): string | null {
+  return content.trim() || null;
 }
 
 export function extractMemory(opts: {
@@ -45,7 +41,7 @@ ${transcript || '（空）'}`;
       model
     )
     .pipe(
-      Effect.map((res) => extractFrom(res.content)),
+      Effect.map((res) => normalizeMemoryOutput(res.content)),
       Effect.catchAllCause(() => Effect.succeed(null))
     );
 }

packages/codingcode/test/context/budget-integration.test.ts

@@ -5,7 +5,6 @@ import { randomUUID } from 'crypto';
 import { Effect, Layer } from 'effect';
 import { ContextService } from '../../src/context/port.js';
 import type { ContextShape } from '../../src/context/port.js';
-import { SessionService } from '../../src/session/port.js';
 import { SessionLayer } from '../../src/session/session.js';
 import { LLMService } from '../../src/llm/port.js';
 import type { SessionRef } from '../../src/contracts/session.js';

packages/codingcode/test/context/compressor/behavior.test.ts

@@ -134,6 +134,8 @@ describe('compressor behavior', () => {
         const summaries = readSummaryEvents(fx.transcriptPath);
         expect(summaries.length).toBe(1);
         expect(summaries[0]!.summaryText).toContain('### Goal');
+        // 全量替换：模型原文即摘要，不做任何抽取或裁剪
+        expect(summaries[0]!.summaryText).toBe(summary);
         expect(summaries[0]!.startTurnId).toBeLessThanOrEqual(summaries[0]!.endTurnId);
         expect(summaries[0]!.endTurnId).toBeGreaterThan(0);
       } finally {

packages/codingcode/test/context/compressor/l5-prompt.test.ts

@@ -1,83 +1,35 @@
 import { describe, it, expect } from 'vitest';
-
-describe('L5 compaction prompt and extraction', () => {
-  it('should extract summary from dual-tag format', () => {
-    const raw = `<analysis>
-This is analysis section with reasoning.
-Multiple lines of thinking.
-</analysis>
-
-<summary>
-## 1. Primary Request and Intent
-User wants to add feature X.
-
-## 2. Key Technical Concepts
-- Concept A
-- Concept B
-
-## 9. Optional Next Step
-Consider optimizing Y next.
-</summary>
-
-Some trailing text that should be ignored.`;
-
-    const match = raw.match(/<summary>([\s\S]*?)<\/summary>/);
-    expect(match).toBeDefined();
-    expect(match?.[1]).toContain('Primary Request');
-    expect(match?.[1]).toContain('Optional Next Step');
+import { COMPACTION_SYSTEM_PROMPT } from '../../../src/context/compaction-prompt.js';
+
+// 断言直接读真实提示词常量，避免测试与实现各写一份而悄悄漂移。
+describe('L5 compaction prompt contract', () => {
+  const SECTIONS = [
+    '## 1. Primary Request and Intent',
+    '## 2. Key Technical Concepts',
+    '## 3. Files and Code Sections',
+    '## 4. Errors and Fixes',
+    '## 5. Problem Solving',
+    '## 6. Decision Rationale and Rejected Approaches',
+    '## 7. All User Messages',
+    '## 8. Pending Tasks',
+    '## 9. Current Work',
+    '## 10. Optional Next Step',
+  ];
+
+  it('requests all ten sections in order', () => {
+    let cursor = -1;
+    for (const section of SECTIONS) {
+      const at = COMPACTION_SYSTEM_PROMPT.indexOf(section);
+      expect(at, `missing or out-of-order section: ${section}`).toBeGreaterThan(cursor);
+      cursor = at;
+    }
   });
 
-  it('should fallback to raw content if no summary tags', () => {
-    const raw = 'Just raw content without tags';
-    const match = raw.match(/<summary>([\s\S]*?)<\/summary>/);
-    const result = match ? match[1] : raw;
-    expect(result).toBe(raw);
+  it('no longer asks for a separate analysis block', () => {
+    expect(COMPACTION_SYSTEM_PROMPT).not.toMatch(/<\/?analysis>/);
   });
 
-  it('should validate 9 sections are present in summary structure', () => {
-    const sections = [
-      '## 1. Primary Request and Intent',
-      '## 2. Key Technical Concepts',
-      '## 3. Files and Code Sections',
-      '## 4. Errors and Fixes',
-      '## 5. Problem Solving',
-      '## 6. All User Messages',
-      '## 7. Pending Tasks',
-      '## 8. Current Work',
-      '## 9. Optional Next Step',
-    ];
-
-    sections.forEach((section) => {
-      expect(section).toMatch(/^##\s+\d+\./);
-    });
-    expect(sections).toHaveLength(9);
-  });
-
-  it('should separate analysis from summary', () => {
-    const full = `<analysis>
-Reasoning and thinking process here.
-This is how I approached the problem.
-</analysis>
-
-<summary>
-The final structured output.
-</summary>`;
-
-    const analysisMatch = full.match(/<analysis>([\s\S]*?)<\/analysis>/);
-    const summaryMatch = full.match(/<summary>([\s\S]*?)<\/summary>/);
-
-    expect(analysisMatch?.[1]).toContain('Reasoning');
-    expect(summaryMatch?.[1]).toContain('final structured');
-    expect(analysisMatch?.[1]).not.toContain('final structured');
-  });
-
-  it('should handle empty sections gracefully', () => {
-    const raw = `<analysis></analysis>
-<summary>
-## 1. Primary Request and Intent
-</summary>`;
-
-    const match = raw.match(/<summary>([\s\S]*?)<\/summary>/);
-    expect(match?.[1]).toBeDefined();
+  it('no longer wraps the summary in tags', () => {
+    expect(COMPACTION_SYSTEM_PROMPT).not.toMatch(/<\/?summary>/);
   });
 });

packages/codingcode/test/memory/extractor.test.ts

@@ -20,24 +20,28 @@ function extract(llm: LLMShape, currentMemory: string, transcript: string) {
 }
 
 describe('Memory Extractor', () => {
-  it('returns memory inside <memory> tags', async () => {
-    const response = `<memory>### 主题
-- 用户是 TypeScript 开发者</memory>`;
+  it('takes the whole model output as the new memory', async () => {
+    const response = `### 主题
+- 用户是 TypeScript 开发者`;
 
     const result = await extract(createMockLlm(response), '', '[user] I like TypeScript');
 
-    expect(result).toContain('### 主题');
-    expect(result).toContain('用户是 TypeScript 开发者');
+    expect(result).toBe(response);
   });
 
-  it('returns null when memory tags are empty', async () => {
-    const result = await extract(createMockLlm('<memory></memory>'), '', '[user] Some text');
+  it('keeps a preamble instead of trying to strip it', async () => {
+    const response = `好的，我整理了一份记忆：
 
-    expect(result).toBeNull();
+### 主题
+- 用户偏好简洁回答`;
+
+    const result = await extract(createMockLlm(response), '', '[user] 简洁点');
+
+    expect(result).toBe(response);
   });
 
-  it('returns null when memory tags not found', async () => {
-    const result = await extract(createMockLlm('No memory tags here'), '', '[user] Some text');
+  it('returns null when the model returns blank output', async () => {
+    const result = await extract(createMockLlm('   \n  '), '', '[user] Some text');
 
     expect(result).toBeNull();
   });
@@ -56,7 +60,7 @@ describe('Memory Extractor', () => {
   });
 
   it('passes currentMemory to the model as existing memory', async () => {
-    const mockLlm = createMockLlm('<memory></memory>');
+    const mockLlm = createMockLlm('');
 
     await extract(mockLlm, '### project\n- 旧信息', '[user] 新对话');
 
@@ -67,7 +71,7 @@ describe('Memory Extractor', () => {
   });
 
   it('keeps instructions in system and transcript data in messages', async () => {
-    const mockLlm = createMockLlm('<memory></memory>');
+    const mockLlm = createMockLlm('');
 
     await extract(mockLlm, '### project\n- Likes TypeScript', '[user] I use Python');
 
@@ -80,7 +84,7 @@ describe('Memory Extractor', () => {
   });
 
   it('passes the target model to the non-streaming channel', async () => {
-    const mockLlm = createMockLlm('<memory></memory>');
+    const mockLlm = createMockLlm('');
 
     await extract(mockLlm, '', '[user] hi');
 

packages/codingcode/test/memory/index.test.ts

@@ -173,7 +173,7 @@ describe('flushSessionToMemory', () => {
       { type: 'user', content: '记住新架构决策' },
       { type: 'assistant', content: '好的' },
     ] as any);
-    setLlmResponse('<memory>### 项目\n- 新的架构决策</memory>');
+    setLlmResponse('### 项目\n- 新的架构决策');
 
     const result = await run(service.flushSessionToMemory('session', TEST_MODEL, tmpDir));
 
@@ -183,14 +183,14 @@ describe('flushSessionToMemory', () => {
     expect(fs.readFileSync(memFile, 'utf-8')).toBe('### 项目\n- 新的架构决策');
   });
 
-  it('keeps file unchanged when model returns empty memory', async () => {
+  it('keeps file unchanged when the model returns blank output', async () => {
     await enableConfig();
     writeMemory('### 旧主题\n- 旧内容');
     const { readTranscript } = await import('../../src/session/file-ops.js');
     vi.mocked(readTranscript).mockImplementation(() => [
       { type: 'user', content: 'hello' },
     ] as any);
-    setLlmResponse('<memory></memory>');
+    setLlmResponse('');
 
     const result = await run(service.flushSessionToMemory('session', TEST_MODEL, tmpDir));
 
@@ -205,7 +205,7 @@ describe('flushSessionToMemory', () => {
     vi.mocked(readTranscript).mockImplementation(() => [
       { type: 'user', content: '无新信息' },
     ] as any);
-    setLlmResponse('<memory>### 主题\n- 不变的内容</memory>');
+    setLlmResponse('### 主题\n- 不变的内容');
 
     const result = await run(service.flushSessionToMemory('session', TEST_MODEL, tmpDir));
 
@@ -219,7 +219,7 @@ describe('flushSessionToMemory', () => {
     vi.mocked(readTranscript).mockImplementation(() => [
       { type: 'user', content: 'hello' },
     ] as any);
-    setLlmResponse('<memory>### 自动\n- 新记忆</memory>', () => {
+    setLlmResponse('### 自动\n- 新记忆', () => {
       writeMemory('### 手动\n- 用户并发编辑');
     });