靠这10个优化点，我们把Multi-Agent工作流成本降了50%以上
AI 前沿· 5 分钟阅读0 阅读
我们用 AI Agent 驱动前后端全流程开发——从需求分析到自动化测试，整个 harness 工作流由 1 个 TL + 6 个子 Agent 组成。跑起来之后第一个问题就是：钱烧得很快，但
<div class="detail-content-box has-mask large">
<section style="margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 10px;padding-right: 10px;background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;font-family: 'Helvetica Neue', Helvetica, 'Hiragino Sans GB', 'Microsoft YaHei', Arial;font-size: 16px;color: rgb(0, 0, 0);line-height: 1.5em;word-spacing: 0em;letter-spacing: 0em;word-break: break-word;overflow-wrap: break-word;text-align: justify;"><section style="text-align: center;"></section><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><br/></p><blockquote style="font-style: normal;line-height: 1.8;color: rgb(14, 136, 235);margin-top: 20px;margin-bottom: 20px;margin-left: 0px;margin-right: 0px;padding-top: 10px;padding-bottom: 10px;padding-left: 20px;padding-right: 10px;border-top-style: none;border-bottom-style: none;border-left-style: solid;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0.05);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;box-shadow: rgba(0, 0, 0, 0) 0px 0px 0px 0px;display: block;overflow-x: auto;overflow-y: auto;"><p style="text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;display: inline;color: rgb(0, 0, 0);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>我们用 AI Agent 驱动前后端全流程开发——从需求分析到自动化测试，整个 harness 工作流由 1 个 TL + 6 个子 Agent 组成。跑起来之后第一个问题就是：钱烧得很快，但完全不知道烧在哪。我们先用 AgentLens 把成本拆开看，发现系统提示词、工具返回信息、历史消息是大头。围绕"让 AI 只看到当前需要的上下文、减少无关的上下文、减少重复的上下文"这三个原则，我们逐一改造了架构拆分、稳定前缀、渐进式披露、代码图谱、CLI 替代 MCP、长期记忆按需索引、工具调用并行化等 10 个方向。整体全流程预估降本 50%~65%。</span></p></blockquote><h3 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;text-align: left;margin-top: 0px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-family: 'Microsoft YaHei';padding-left: 10px;border-left-width: 5px;border-left-style: solid;border-left-color: #0E88EB;font-size: 20px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>一、背景与挑战</span></strong></span></h3><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>1.1 业务背景</span></span></h4><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>过去半年，我们团队一直在探索用 AI Agent 驱动前后端协同开发——从需求分析、技术方案设计，到前后端并行编码、自动化测试、视觉还原验证，全流程由一个叫 tech-leader 的 harness Skill 统一调度：</span></p><pre style="border-radius: 5px;box-shadow: rgba(0, 0, 0, 0.55) 0px 2px 10px;text-align: left;margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;"><code style="overflow-x: auto;padding: 16px;color: #abb2bf;padding-top: 15px;background: #282c34;border-radius: 5px;display: -webkit-box;font-family: Consolas, Monaco, Menlo, monospace;font-size: 12px;"><span>TL（调度层）</span><span><br/></span><span>├── Wave 1（并行）：后端 Agent + 前端 Agent</span><span><br/></span><span>├── Wave 2：质量审查 Agent</span><span><br/></span><span>├── Wave 3：测试 Agent（用例生成 + 执行）</span><span><br/></span><span>├── Wave 4：视觉验证 Agent</span><span><br/></span><span>└── Wave 5：Agent 评测</span><span><br/></span></code></pre><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>跑一次完整的中等需求，往往需要经历 5-6 个 Wave、20+ 次子 Agent 调用、数百轮工具调用。规模一上来，token 成本就成了绕不开的问题。</span></p><figure style="margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: flex;flex-direction: column;justify-content: center;align-items: center;"><span><img src="https://pic.imgdb.cn/i/034BII3mfmBiZzjR5A3ixY.png" class="rich_pages wxw-img" style="display: block;margin-top: 10px;margin-right: auto;margin-bottom: 10px;margin-left: auto;max-width: 100%;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;object-fit: fill;box-shadow: rgba(0, 0, 0, 0) 0px 0px 0px 0px;"/></span></figure><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>1.2 痛点分析：六类消耗来源</span></span></h4><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>一次任务的 token 消耗可以归为六类来源：</span></p><section style="margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;overflow-x: auto;"><table style="display: table;text-align: left;"><thead><tr class="firstRow"><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>来源</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>说明</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>成本特点</span></section></th></tr></thead><tbody><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>① 系统提示词</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>系统固定提示词、Skill 描述、MCP 工具 Schema</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>每轮必带，随 Agent/MCP 数量倍增。一个 40 工具的 MCP Server 每轮增加约 10-15KB Schema 开销</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>② 工具返回的信息</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>MCP 调用返回的工具列表、返回的具体内容（需求 JSON、设计稿节点树、截图 base64）</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>体积不可控，塞进 context 后后续几十轮都要跟着重复计费</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>③ 读取的文件信息</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>AI 读取的代码文件、知识库文档内容</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>盲搜式探索往往要 3-5 轮才能定位，每轮匹配行、整段文件都会累积</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>④ 长期记忆</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>AI 加载的历史经验、技术方案沉淀文档</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>一旦加载往往常驻会话，即使后续用不上大部分内容</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>⑤ 历史消息</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>多轮会话后累积的历史</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>真正的大头，append-only 且滚雪球式增长，越往后轮次越贵</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>⑥ 用户提示词</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>用户输入的一句话/一段需求描述</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>每轮增量，只出现一次、体量小，六类中最便宜的一项</span></section></td></tr></tbody></table></section><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>这六类来源叠加，导致一个中等需求的 token 消耗远超预期。而且在没有按 Wave 粒度拆分消耗之前，完全不知道钱主要烧在哪里——这是我们做这次优化专项的直接动因。</span></p><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>1.3 先建度量：看清成本在哪里</span></span></h4><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>本人使用的IDE是codebuddy内网版，内网版是把每轮对话的工具链调用和token用量上报到内网的AgentLens平台，如果您使用的是其他IDE，需要查阅下官方文档看看每轮对话的token用量怎么获取。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>没有度量就没有优化。CodeBuddy 已将每轮对话的 token 消耗上报至 AgentLens，通过 agentlens-mcp 可以按 TraceId 追踪单次调用链路，也可以按 SessionId 聚合一个完整需求的消耗分布：</span></p><figure style="margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: flex;flex-direction: column;justify-content: center;align-items: center;"><span><img src="https://pic.imgdb.cn/i/034BII5CnZebU0aqkvh3iV.png" class="rich_pages wxw-img" style="display: block;margin-top: 10px;margin-right: auto;margin-bottom: 10px;margin-left: auto;max-width: 100%;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;object-fit: fill;box-shadow: rgba(0, 0, 0, 0) 0px 0px 0px 0px;"/></span></figure><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>看清成本分布之后，目标就明确了：系统性压缩六类来源里能压缩的部分，把中型需求的全流程 token 成本降下来。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span><br/></span></p><h3 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;text-align: left;margin-top: 0px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-family: 'Microsoft YaHei';padding-left: 10px;border-left-width: 5px;border-left-style: solid;border-left-color: #0E88EB;font-size: 20px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>二、整体方案</span></span></h3><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>2.1 思路与路径</span></span></h4><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>我们把可以下手的动作归纳为三个原则：</span></p><pre style="border-radius: 5px;box-shadow: rgba(0, 0, 0, 0.55) 0px 2px 10px;text-align: left;margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;"><code style="overflow-x: auto;padding: 16px;color: #abb2bf;padding-top: 15px;background: #282c34;border-radius: 5px;display: -webkit-box;font-family: Consolas, Monaco, Menlo, monospace;font-size: 12px;"><span>① 让 AI 只看到当前需要的上下文 —— 该加载的东西，按需加载，不要一次性全带</span><span><br/></span><span>② 减少无关的上下文 —— 不该看到的东西，一开始就别带进来</span><span><br/></span><span>③ 减少重复的上下文 —— 同一份内容，不要在多轮里反复计费</span><span><br/></span></code></pre><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>动手之前还有一个更前置的架构判断：是否需要拆多 Agent。拆分本身有成本（多份系统提示词并行计费），只有需求规模足够大，拆分撬动的后续收益才能覆盖这笔开销。所以我们先做规模预判（S/M/L 模式）：小需求走单 Agent 直接处理；只有中大型需求，才进入多 Agent 并行调度。</span></p><figure style="margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: flex;flex-direction: column;justify-content: center;align-items: center;"><span><img src="https://pic.imgdb.cn/i/034BII6iNHB0pfB5B3bKh7.png" class="rich_pages wxw-img" style="display: block;margin-top: 10px;margin-right: auto;margin-bottom: 10px;margin-left: auto;max-width: 100%;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;object-fit: fill;box-shadow: rgba(0, 0, 0, 0) 0px 0px 0px 0px;"/></span></figure><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>2.2 关键技术选型</span></span></h4><section style="margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;overflow-x: auto;"><table style="display: table;text-align: left;"><thead><tr class="firstRow"><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>原则</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>优化手段</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>解决的问题</span></section></th></tr></thead><tbody><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>只看到当前需要的上下文</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>渐进式披露（L2/L3 分层）</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>SKILL.md 里条件性内容常驻，用不上也要计费</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>确定性的操作由脚本执行</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>大模型擅长推理，不擅长精确执行。反复试错是隐藏成本</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span><br/></span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>MCP 数据获取子 Agent 化</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>TAPD/Figma 原始 payload 直接进最长生命周期 Agent 的 context</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span><br/></span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>长期记忆按需索引加载</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>历史经验/技术方案全量加载，用不上的文档也占位计费</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span><br/></span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少无关上下文</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>单 Agent 拆分为多 Agent</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>全能型 Agent 背负所有领域的知识和历史</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>Agent 专属配置（工具白名单+模型分层）</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>每个 Agent 都能看到所有 MCP 工具 Schema</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span><br/></span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>代码图谱替代盲搜</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>关键词盲搜带入大量无关文件内容</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span><br/></span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少重复上下文</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>稳定前缀设计（状态外化）</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>进度状态/执行状态混在会话里，破坏缓存前缀</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>避免重复加载 Skill</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>同一份长期记忆被下游 Agent 各自重新加载</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span><br/></span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>rtk 压缩 CLI 输出</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>命令行原始输出噪音在多轮循环中反复出现</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span><br/></span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>工具调用并行化</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>无依赖的多次调用串行等待，历史被重复打包的轮次翻倍</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span><br/></span></section></td></tr></tbody></table></section><h3 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;text-align: left;margin-top: 0px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-family: 'Microsoft YaHei';padding-left: 10px;border-left-width: 5px;border-left-style: solid;border-left-color: #0E88EB;font-size: 20px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>三、实践详解</span></span></h3><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.1 让 AI 只看到当前需要的上下文</span></span></h4><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.1.1 渐进式披露</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>Anthropic 在 Agent Skill 规范中提出了一个"渐进式披露"架构，核心思想是：不是所有内容都需要常驻 context，只有真正需要时才加载。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>三层结构：</span></p><figure style="margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: flex;flex-direction: column;justify-content: center;align-items: center;"><span><img src="https://pic.imgdb.cn/i/034BII98CkLjq4Bb7T3nF9.png" class="rich_pages wxw-img" style="display: block;margin-top: 10px;margin-right: auto;margin-bottom: 10px;margin-left: auto;max-width: 100%;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;object-fit: fill;box-shadow: rgba(0, 0, 0, 0) 0px 0px 0px 0px;"/></span></figure><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>这种机制的好处是，即使安装了 20 个 Skill，初始加载也仅 1000-2000 token。相比单体式提示词，上下文使用量减少约 90%。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>本次优化主要是把正文部分内容迁移到资源层。</span></p><ul style="list-style-type: disc;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>条件性内容外移到资源层</span></section></li></ul><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>我们之前的问题是：</span></strong><span>大量只在特定场景才需要的内容，混在 SKILL.md 正文里常驻 context。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>以 自动化测试Skill 为例，改造前有几块内容其实是条件性的：</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>① spec 模板代码（约 40 行 TypeScript）</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>这段代码只在 Phase A 生成 spec 文件时用到一次，之后完全是冗余占位。所以我们把模版代码外置到references中，按需提取。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>② DB 前置操作规则（约 38 行）</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>只有检测到 DB 依赖用例时才需要。一样把DB操作规则外置到references中，按需读取。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>改造后自动化测试Skill正文从 198 行降到 128 行（-35%）。</span></p><ul style="list-style-type: disc;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>步骤详情外移到资源层</span></section></li></ul><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>比条件性内容更彻底的做法，是把"所有步骤的详细内容"都从 SKILL 正文移到资源层，正文只留骨架。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>以 </span><code style="color: rgb(30, 107, 184);line-height: 1.8em;letter-spacing: 0em;background-attachment: scroll;background-clip: border-box;background-color: rgba(27, 31, 35, 0.05);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 2px;margin-right: 2px;padding-top: 2px;padding-bottom: 2px;padding-left: 4px;padding-right: 4px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgb(0, 0, 0);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 4px;border-top-right-radius: 4px;border-bottom-right-radius: 4px;border-bottom-left-radius: 4px;overflow-wrap: break-word;word-break: break-all;font-family: 'Microsoft YaHei';font-size: 16px;"><span>tech-leader</span></code><span>Skill （主调度Skill）为例：</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>优化前，S/M 两种模式（小需求/中等需求）的完整工作流、Wave 1~5 各阶段的派发 prompt、审查规则、测试调度逻辑等执行细节全写在 SKILL.md 正文里，Skill 一旦激活就全量常驻 context。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>但 TL 每次实际只走一条路径——判完规模后要么进 S 模式、要么进 M 模式，进 M 模式后又按阶段逐步推进，任一时刻真正用到的只是其中一小段，其余步骤的详细内容全程占位计费却用不上。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>改法是把各步骤的详细执行内容拆到 </span><code style="color: rgb(30, 107, 184);line-height: 1.8em;letter-spacing: 0em;background-attachment: scroll;background-clip: border-box;background-color: rgba(27, 31, 35, 0.05);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 2px;margin-right: 2px;padding-top: 2px;padding-bottom: 2px;padding-left: 4px;padding-right: 4px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgb(0, 0, 0);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 4px;border-top-right-radius: 4px;border-bottom-right-radius: 4px;border-bottom-left-radius: 4px;overflow-wrap: break-word;word-break: break-all;font-family: 'Microsoft YaHei';font-size: 16px;"><span>references/</span></code><span> 资源层，SKILL.md 只保留</span><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>骨架</span></strong><span>——核心职责、规模预判维度表、分流规则、进度追踪机制、容错机制。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>骨架负责"决定下一步走哪"，具体"怎么走"则按需 </span><code style="color: rgb(30, 107, 184);line-height: 1.8em;letter-spacing: 0em;background-attachment: scroll;background-clip: border-box;background-color: rgba(27, 31, 35, 0.05);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 2px;margin-right: 2px;padding-top: 2px;padding-bottom: 2px;padding-left: 4px;padding-right: 4px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgb(0, 0, 0);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 4px;border-top-right-radius: 4px;border-bottom-right-radius: 4px;border-bottom-left-radius: 4px;overflow-wrap: break-word;word-break: break-all;font-family: 'Microsoft YaHei';font-size: 16px;"><span>read_file</span></code><span> 对应的资源文件。</span></p><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.1.2 确定性的操作由脚本执行</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>这里有两处优化：</span></p><ul style="list-style-type: disc;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>CLI 驱动一切环境操作和校验操作</span></section></li></ul><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>一开始我的工作流里面，没有任何操作脚本和验证脚本。后来发现后端Agent写代码时，频繁出现拼错数据库连接参数、找错启动服务命令、绕过编译等等；因为反复查找各种参数、命令，为此消耗了大量token。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>后来写了一个脚本，里面包含数据库迁移、编译、服务启动、健康检查等确定性的命令，全部让AI通过 dev-env.sh 脚本执行。AI 不拼接 mysql -u root -p 这种命令，只提供参数给脚本。</span></p><ul style="list-style-type: disc;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>能用CLI就不用MCP</span></section></li></ul><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>MCP 工具调用的隐藏成本：</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>MCP 工具给了 Agent 强大的能力，但每次 MCP 工具调用，实际上是一次完整的 LLM 推理轮次：</span></p><pre style="border-radius: 5px;box-shadow: rgba(0, 0, 0, 0.55) 0px 2px 10px;text-align: left;margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;"><code style="overflow-x: auto;padding: 16px;color: #abb2bf;padding-top: 15px;background: #282c34;border-radius: 5px;display: -webkit-box;font-family: Consolas, Monaco, Menlo, monospace;font-size: 12px;"><span>每次 MCP 工具调用的真实成本：</span><span><br/></span><span> 1. LLM 决策：判断调用哪个工具、构造参数 ← 一次 API 请求</span><span><br/></span><span> 2. 工具 JSON Schema 常驻 context ← 每轮 10-15KB 额外 input</span><span><br/></span><span> 3. 工具返回结果进入 context ← 可能几千行</span><span><br/></span><span> 4. LLM 处理结果：分析输出、决定下一步 ← 又一次 API 请求</span><span><br/></span></code></pre><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>自动化测试Skill中，我们最初用的是 Playwright MCP——通过 MCP 协议让 Agent 实时控制浏览器（打开页面、点击、截图）。但实践中发现几个问题：</span></p><ul style="list-style-type: disc;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>每个操作都要消耗一次 AI 推理：点击一个按钮就是一次 MCP 调用 + 一次模型推理，10 步操作就是 10 轮对话，token 消耗大、速度慢</span></p></section></li><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>不可重跑：MCP 模式下操作是实时的"一次性"行为，测试失败后想重跑一遍，Agent 需要重新推理一遍所有步骤</span></p></section></li><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>无法并行：MCP 控制的是单个浏览器实例，多个用例只能串行执行</span></p></section></li></ul><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>后面把Playwright MCP换成了Playwright CLI，CLI的核心思路是把自然语言用例翻译为 Playwright spec 文件，再调 Playwright CLI 批量执行。大模型做推理（生成 spec），脚本做执行（跑 CLI）。把"理解用例并翻译为代码"和"实际跑测试"拆成两步，前者需要 AI，后者完全不需要。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>这次改造不仅节省了token，耗时也大大缩短。</span></p><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.1.3 MCP 数据获取子 Agent 化</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>问题：主agent自己调 MCP，原始数据永久卡在最长生命周期的 context 里</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>tech-leader 工作流里，需求分析阶段要读 TAPD 需求、要读 Figma 设计稿，这两步 MCP 调用最早是主Agent自己直接发起的：</span></p><pre style="border-radius: 5px;box-shadow: rgba(0, 0, 0, 0.55) 0px 2px 10px;text-align: left;margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;"><code style="overflow-x: auto;padding: 16px;color: #abb2bf;padding-top: 15px;background: #282c34;border-radius: 5px;display: -webkit-box;font-family: Consolas, Monaco, Menlo, monospace;font-size: 12px;"><span>● TAPD -> 返回mcp所有工具描述、返回完整需求描述、关联缺陷、任务列表</span><span><br/></span><span><br/></span><span>● Figma -> 返回mcp所有工具描述、返回设计稿完整节点树 JSON + 截图</span><span><br/></span></code></pre><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>问题在于 TL 是整个工作流生命周期最长的 Agent。这两次 MCP 调用如果发生在 TL 自己的会话里，原始 payload——动辄几千字的需求描述、上万行的设计稿节点树 JSON——会一直留在 TL 的 context 里，后续几十轮工具调用都要重新计费一遍。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>改法：</span></strong><span>给 TAPD/Figma 各建一个专属子 Agent，只返回结构化摘要</span></p><figure style="margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: flex;flex-direction: column;justify-content: center;align-items: center;"><span><img src="https://pic.imgdb.cn/i/034BIIBxNEr3XnEoEJ7JRX.png" class="rich_pages wxw-img" style="display: block;margin-top: 10px;margin-right: auto;margin-bottom: 10px;margin-left: auto;max-width: 100%;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;object-fit: fill;box-shadow: rgba(0, 0, 0, 0) 0px 0px 0px 0px;"/></span></figure><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>实测同一工作流改造前后，单轮会话 input token 从 1,030,000 降到 634,905，-38.4%。首次调用两种方式消耗基本一样，真正的收益在后续多轮会话不会带着原始 payload 越滚越大。</span></p><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.1.4 长期记忆按需索引加载</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>context-keeper（知识沉淀Skill）负责在方案设计前加载历史经验/技术方案沉淀文档，最早的做法是把匹配到的候选文档全量读入，即使一份文档里只有一两条经验跟当前任务相关，剩下几十行也会一起进 context 常驻到会话结束。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>改法是引入一层轻量的 INDEX.md 目录索引，把"查目录"和"读正文"拆成两步：</span></p><pre style="border-radius: 5px;box-shadow: rgba(0, 0, 0, 0.55) 0px 2px 10px;text-align: left;margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;"><code style="overflow-x: auto;padding: 16px;color: #abb2bf;padding-top: 15px;background: #282c34;border-radius: 5px;display: -webkit-box;font-family: Consolas, Monaco, Menlo, monospace;font-size: 12px;"><span>❌ 改前：直接扫描 team/project 下所有历史文档，命中关键词就整篇 read_file</span><span><br/></span><span>✅ 改后：</span><span><br/></span><span> 1. 先 read_file INDEX.md（几十行的标题+标签+摘要表格）</span><span><br/></span><span> 2. 按关键词匹配标题/标签/摘要，结合类别权重计算相关度</span><span><br/></span><span> 3. 取 Top 3 命中条目，才对这几篇 read_file 正文</span><span><br/></span><span> 4. INDEX 不存在时才回退到全文 search_content（兜底，非常态路径）</span><span><br/></span><span><br/></span></code></pre><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>INDEX.md 本身是所有历史文档的目录，体量通常只有几十到一两百行——用几十行的索引筛选出真正相关的 2-3 篇，比一次性全量加载几十篇文档的正文便宜得多。相关度评分低于阈值的文档从一开始就不会进入 read_file 候选列表，从源头减少了"加载了但用不上"的常驻 token。</span></p><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.2 减少无关的上下文</span></span></h4><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.2.1 单 Agent 拆分为多 Agent</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>最早的 tech-leader 是一个从头到尾自己干完所有事的单体 Agent，带来两个问题——前端/后端/测试/视觉规范混在同一段历史里，越跑越臃肿；全程一条会话没有重置时机，历史只会越滚越大。</span></p><figure style="margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: flex;flex-direction: column;justify-content: center;align-items: center;"><span><img src="https://pic.imgdb.cn/i/034BIIDQud4fSfnOf6EStU.png" class="rich_pages wxw-img" style="display: block;margin-top: 10px;margin-right: auto;margin-bottom: 10px;margin-left: auto;max-width: 100%;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;object-fit: fill;box-shadow: rgba(0, 0, 0, 0) 0px 0px 0px 0px;"/></span></figure><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>改法是引入专职调度的 TL，把编码/测试/视觉校验分发给按角色划分的子 Agent（backend-dev / frontend-dev / test-runner / visual-reviewer / code-reviewer），每个子 Agent 只携带自己领域的提示词和工具，跑完即销毁，Wave 1 的前后端还能并行派发。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>需要注意的是：拆分本身不是靠"减少历史堆叠"直接省钱的——6 个 Agent 并行跑，等于同时有 6 份系统提示词在计费。真正的收益是分散滚雪球效应（短生命周期子 Agent 跑完销毁，滚雪球被切段）和为后续优化打开空间（稳定前缀、工具裁剪、模型分层、子 Agent 化都建立在"已经拆分"的前提上）。这也是为什么要先做规模预判——小需求不做无谓拆分。</span></p><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.2.2 Agent 专属配置</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>以前每个子 Agent 默认带上所有已注册 MCP Server 的工具 Schema（即使用不到），是一个隐藏的成本漏洞。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>以前我们用的是CodeBuddy自带的TeamCreate工具，自动派生子 Agent来完成任务。但是这种模式就没办法给子 Agent设置mcp白名单、模型。后来我的改法是：给每个角色创建单独的自定义Agent，在创建团队时调用相应的自定义子 Agent执行任务。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>这种做法可以在 Agent 定义文件的 frontmatter 中通过 tools 字段指定工具白名单，未列出的工具对该 Agent 完全不可见，比如后端开发Agent，就不需要任何mcp工具，只需要读文件、写文件等工具；</span></p><pre style="border-radius: 5px;box-shadow: rgba(0, 0, 0, 0.55) 0px 2px 10px;text-align: left;margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;"><code style="overflow-x: auto;padding: 16px;color: #abb2bf;padding-top: 15px;background: #282c34;border-radius: 5px;display: -webkit-box;font-family: Consolas, Monaco, Menlo, monospace;font-size: 12px;"><span>---</span><span><br/></span><span>name: backend-dev</span><span><br/></span><span>tools: list_dir, search_file, search_content, read_file, replace_in_file, write_to_file, execute_command</span><span><br/></span><span style="color: #5c6370;font-style: italic;line-height: 26px;"><span># 没有 mcpServers 字段 → Figma/TAPD/iWiki 等 MCP 全部不暴露</span></span><span><br/></span><span>---</span><span><br/></span></code></pre><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>顺带的收益是把原来 主Agent派发子 Agent提示词里的大段稳定指令迁移到子 Agent 系统提示词里（更容易命中缓存），TL 的 派发提示词 从 10-15 行精简到 2 行动态内容。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>同时可以按角色做模型分层路由：自动化测试、视觉还原对比 这类规则性强、推理要求低但轮次最多（修复循环）的角色换成 GLM-5v（智谱 GLM-5V 多模态模型，成本约 Sonnet 的 36%），成本节省随修复轮次倍增，测试/视觉 Agent 成本 -64%。</span></p><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.2.3 代码图谱替代盲搜</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>Agent 在编码阶段做的第一件事，往往是"理解项目结构"。传统方式是 search_content 关键词搜索：</span></p><pre style="border-radius: 5px;box-shadow: rgba(0, 0, 0, 0.55) 0px 2px 10px;text-align: left;margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;"><code style="overflow-x: auto;padding: 16px;color: #abb2bf;padding-top: 15px;background: #282c34;border-radius: 5px;display: -webkit-box;font-family: Consolas, Monaco, Menlo, monospace;font-size: 12px;"><span>搜索 </span><span style="color: #98c379;line-height: 26px;"><span>"UserService"</span></span><span> → 返回 20 个文件的匹配行</span><span><br/></span><span>搜索 </span><span style="color: #98c379;line-height: 26px;"><span>"getUserById"</span></span><span> → 返回 8 个文件的匹配行</span><span><br/></span><span>read_file UserService.java → 300 行进 context</span><span><br/></span><span>read_file UserController.java → 200 行进 context</span><span><br/></span><span>...</span><span><br/></span><span><br/></span></code></pre><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>每一次搜索返回的结果都会进入 context，大量无关匹配行随之带入。更糟的是，agent 往往需要多轮搜索才能逐渐定位——探索过程本身就是 token 消耗。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>代码图谱：一次查询直接定位</span></strong></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>后面我们的工作流中使用了 graphify 代码图谱，代码图谱使用AST和语义来给项目创建文件索引和依赖关系，在搜代码文件前先搜这个目录，锁定文件范围，再去read_file文件，能节省几十倍的token（官方预估数据）。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>代码图谱的 token 节省不只体现在单次查询上，更重要的是减少了 Agent 的探索轮次。少一轮工具调用，就少一次 API 请求，就少一整个 context window 的 input token 计费。在修复循环场景下，这个收益会被轮次数放大。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>实测数据对比：</span></strong></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>我们在相同任务下分别测试了使用代码图谱和未使用代码图谱两种模式的 token 消耗：</span></p><section style="margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;overflow-x: auto;"><table style="display: table;text-align: left;"><thead><tr class="firstRow"><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>指标</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>使用代码图谱</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>未使用代码图谱</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>差异</span></section></th></tr></thead><tbody><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>总计</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>676,987</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>875,352</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>-22.7%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>输入 token</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>670,281</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>868,544</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>-22.8%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>缓存命中</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>609,903</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>820,175</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>-25.6%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>缓存未命中</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>60,378</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>48,369</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>+24.8%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>输出 token</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>6,706</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>6,808</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>-1.5%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>缓存命中率</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>91.0%</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>94.4%</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>-3.4pp</span></section></td></tr></tbody></table></section><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>关键发现：</span></p><ol style="list-style-type: decimal;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>总 token 节省 22.7%——从 87.5 万降到 67.7 万，节省约 19.8 万 token</span></section></li><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>输入 token 是主要节省来源（-22.8%）——代码图谱减少了探索过程中带入 context 的冗余文件内容</span></section></li></ol><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>代码图谱使用小tips：</span></strong></p><ol style="list-style-type: decimal;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>首先需要安装</span><code style="color: rgb(30, 107, 184);line-height: 1.8em;letter-spacing: 0em;background-attachment: scroll;background-clip: border-box;background-color: rgba(27, 31, 35, 0.05);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 2px;margin-right: 2px;padding-top: 2px;padding-bottom: 2px;padding-left: 4px;padding-right: 4px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgb(0, 0, 0);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 4px;border-top-right-radius: 4px;border-bottom-right-radius: 4px;border-bottom-left-radius: 4px;overflow-wrap: break-word;word-break: break-all;font-family: 'Microsoft YaHei';font-size: 16px;"><span>graphify</span></code></section></li></ol><pre style="border-radius: 5px;box-shadow: rgba(0, 0, 0, 0.55) 0px 2px 10px;text-align: left;margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;"><code style="overflow-x: auto;padding: 16px;color: #abb2bf;padding-top: 15px;background: #282c34;border-radius: 5px;display: -webkit-box;font-family: Consolas, Monaco, Menlo, monospace;font-size: 12px;"><span>uv tool install graphifyy</span><span><br/></span></code></pre><ol style="list-style-type: decimal;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>执行初始化</span></section></li></ol><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>让graphify给你的代码仓库建立索引，大概需要几分钟的时间。初始化成功后，会生成一个graphify目录，</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>这个目录下面你只要提交4个文件就可以了，其他可以不用提交。</span></p><figure style="margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: flex;flex-direction: column;justify-content: center;align-items: center;"><span><img src="https://pic.imgdb.cn/i/034BIIFE2PqbMZXo6gZRwD.png" class="rich_pages wxw-img" style="display:block;margin-top:10px;margin-right:auto;margin-bottom:10px;margin-left:auto;max-width:100%;border-top-style:none;border-bottom-style:none;border-left-style:none;border-right-style:none;border-top-width:3px;border-bottom-width:3px;border-left-width:3px;border-right-width:3px;border-top-color:rgba(0, 0, 0, 0.4);border-bottom-color:rgba(0, 0, 0, 0.4);border-left-color:rgba(0, 0, 0, 0.4);border-right-color:rgba(0, 0, 0, 0.4);border-top-left-radius:0px;border-top-right-radius:0px;border-bottom-right-radius:0px;border-bottom-left-radius:0px;object-fit:fill;box-shadow:rgba(0, 0, 0, 0) 0px 0px 0px 0px;width:385px;height:355px;"/></span></figure><ol style="list-style-type: decimal;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>增量更新</span></section></li></ol><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>细心的同学可能会问了，那如果我的代码更新了怎么办？</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>你可以绑定Git Hook，graphify hook install一键绑定 post-commit + post-checkout，commit 后自动增量更新图谱、checkout 后自动切换分支图谱。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>省时间技巧：直接丢github链接给AI，让AI帮你完成安装，git hook绑定等等。（ https://github.com/Graphify-Labs/graphify ）</span></p><ol style="list-style-type: decimal;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>怎么用</span></section></li></ol><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>直接告诉AI代码搜索请使用代码图谱graphify，就会自动触发</span></p><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.3 减少重复的上下文</span></span></h4><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.3.1 稳定前缀设计</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>KV Cache：为什么前缀稳定等于省钱？</span></strong></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>理解这个优化，需要先了解一点 Transformer 的底层机制。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>LLM 处理每个 token 时，需要对它之前所有 token 做注意力计算（Attention），这个过程会产生两个中间矩阵：K（Key）和 V（Value），合称 KV Cache。KV 矩阵的计算量是 O(n²)——这是推理成本的主要来源。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>Prompt Cache 的本质是：如果你本次请求的前缀和上次完全一致，服务商直接复用上次的 KV 矩阵，跳过重复计算，只收缓存读取的低价（约正常价格的 10%）。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>我们发现 「派发子 Agent的提示词」里稳定指令和动态内容（技术方案文档）交错排列，导致前缀无法命中缓存，改法是把动态内容统一后置。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>主Agent自身还有个隐藏的前缀破坏者：</span></strong><span>进度状态不断在会话里输出累积。原来每个阶段切换都要在对话里输出完整进度看板，堆积在历史里导致历史不能被压缩。改法是把进度状态外化到文件中：</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>主Agent每次被唤醒先 read_file 进度文件，而不是回放历史。阶段切换也改成单行输出（✅ Wave 1 完成 → Wave 2 🔄 已启动），额外收益是会话中断后可直接读文件恢复现场。</span></p><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.3.2 避免重复加载 Skill</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>排查 「前端Agent」 的 context 膨胀时发现，"调用 知识沉淀Skill 加载项目上下文"这一步其实没必要——主Agent 做方案设计时已经加载过历史经验并写入了技术方案文档，</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>前端Agent 直接读文档就能拿到结论，不需要重新触发 use_skill 把 200 行 SKILL.md 再次加载进 context。信息应该在最上游收集一次，通过文档传递给下游，而不是让每个 Agent 各自重复获取。</span></p><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.3.3 rtk 压缩 CLI 输出</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>git status、npm test、docker ps 这类命令的原始输出夹带大量噪音，在自动化测试的修复循环中反复出现。我们接入了 rtk——一个在命令执行前拦截、重写为压缩版本的开源 CLI 代理，官方实测降幅 60%-90%。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>接入时踩了两个坑：</span></strong></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>① 官方 --agent 不支持 CodeBuddy，改为用 CodeBuddy 的 PreToolUse Hook 机制自己实现等价拦截；</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>② rtk 输出字段是 updatedInput，CodeBuddy 要求的是 modifiedInput，字段名不一致会导致静默失效（不报错但不生效），需要写一个几行的转换脚本做字段名转换。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>评估 rtk 效果时也踩了方法论的坑：最初想用"同一需求跑两遍完整工作流，rtk 开/关各一次"对比，但大模型的执行路径本身不确定（探索轮次、有没有踩坑重试都会波动），噪声比 rtk 本身的效果还大。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>更可靠的方式是用 rtk gain 统计或直接命令行对比——因为 rtk 的压缩是纯文本过滤，跟大模型决策无关，100% 可复现：</span></p><figure style="margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: flex;flex-direction: column;justify-content: center;align-items: center;"><span><img src="https://pic.imgdb.cn/i/034BIIGchKTZJ7EFXbxouf.png" class="rich_pages wxw-img" style="display: block;margin-top: 10px;margin-right: auto;margin-bottom: 10px;margin-left: auto;max-width: 100%;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;object-fit: fill;box-shadow: rgba(0, 0, 0, 0) 0px 0px 0px 0px;"/></span></figure><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>幅度因命令而异（ps aux -98.9%，纯 git status 仅 -31%），不能拿"60%-90%"当固定值。全局配置一次（~/.codebuddy/settings.json），所有子 Agent 自动获得压缩效果。</span></p><h5 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 16px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span style="margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;font-size: 16px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>3.3.4 工具调用并行化</span></span></h5><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>无依赖关系的多次工具调用如果串行执行，每一次调用都是一轮独立的 LLM 推理，前面所有轮次的历史都要跟着重新打包计费一遍——调用次数越多，滚雪球轮次越多。我们排查了两类典型的"本可并行却串行"场景：</span></p><pre style="border-radius: 5px;box-shadow: rgba(0, 0, 0, 0.55) 0px 2px 10px;text-align: left;margin-top: 10px;margin-bottom: 10px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;"><code style="overflow-x: auto;padding: 16px;color: #abb2bf;padding-top: 15px;background: #282c34;border-radius: 5px;display: -webkit-box;font-family: Consolas, Monaco, Menlo, monospace;font-size: 12px;"><span>❌ 改前：TAPD 摘要获取 → 等结果 → 再发起 Figma 摘要获取 → 等结果</span><span><br/></span><span> 两次子 Agent 调用互不依赖，却占用了 2 轮历史累积</span><span><br/></span><span><br/></span><span>✅ 改后：TAPD/Figma 若同时存在，同一轮消息内并行发起</span><span><br/></span><span> Task 工具同时传入两个 subagent_name 调用，一轮内拿到两份摘要</span><span><br/></span></code></pre><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>TAPD 需求摘要和 Figma 设计稿摘要本身没有先后依赖，我们把 tapd-req-analyzer、figma-design-analyzer 的派发方式从"依次调用、等结果"改成"同一轮消息内并行发起"，两个子 Agent 各自跑完即销毁，主Agent 一轮就拿到两份摘要，比串行少一轮历史打包。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>测试用例执行也是同理：自动化测试阶段原来是一条 spec 执行完再执行下一条，改法是让 Playwright CLI 一次接收多个 spec 文件、内置多 worker 并行跑（详见"CLI 替代 MCP"一节的批量执行方式），LLM 侧只需 1 次调用 + 1 次读汇总结果，不随用例数线性增加推理轮次。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><strong style="color: rgb(62, 71, 83);background-attachment: scroll;background-clip: border-box;background-color: rgba(0, 0, 0, 0);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;border-top-style: none;border-bottom-style: none;border-left-style: none;border-right-style: none;border-top-width: 3px;border-bottom-width: 3px;border-left-width: 3px;border-right-width: 3px;border-top-color: rgba(0, 0, 0, 0.4);border-bottom-color: rgba(0, 0, 0, 0.4);border-left-color: rgba(0, 0, 0, 0.4);border-right-color: rgba(0, 0, 0, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><span>判断是否可并行的原则很简单：</span></strong><span>两次调用之间没有数据依赖（后一次不需要前一次的输出作为输入）就应该并行，这类"沉默的串行"往往藏在最初写 prompt 时"想清楚一步再写下一步"的顺序思维里，需要专门排查才能发现。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span><br/></span></p><h3 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;text-align: left;margin-top: 0px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-family: 'Microsoft YaHei';padding-left: 10px;border-left-width: 5px;border-left-style: solid;border-left-color: #0E88EB;font-size: 20px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>四、效果与数据</span></span></h3><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>4.1 核心指标</span></span></h4><section style="margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;overflow-x: auto;"><table style="display: table;text-align: left;"><thead><tr class="firstRow"><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>维度</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>优化前</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>优化后</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>降幅</span></section></th></tr></thead><tbody><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>主Agent端到端 token</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>708,783（17 轮）</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>315,266（9 轮）</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>-55.5%，轮次 -47%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>子 Agent 单轮固定开销</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>常见上几十万 token</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少约 20,000 token</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>视配置</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>子 Agent 命令行输出</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>原始 CLI 输出</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>rtk 压缩后</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>-60%~90%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>测试/视觉子 Agent 模型成本</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>claude Sonnet 计价</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>GLM-5v 计价</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>-64%</span></section></td></tr></tbody></table></section><h4 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;font-size: 17px;color: rgb(3, 124, 230);margin-top: 10px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-size: 18px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>4.2 分项收益汇总</span></span></h4><section style="margin-top: 0px;margin-bottom: 0px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;overflow-x: auto;"><table style="display: table;text-align: left;"><thead><tr class="firstRow"><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>优化项</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>所属原则</span></section></th><th style="color: rgb(0, 0, 0);font-size: 16px;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;background: none left top / auto no-repeat scroll padding-box border-box rgb(240, 240, 240);height: auto;border-style: solid;border-width: 1px;border-color: rgba(204, 204, 204, 0.4);border-radius: 0px;padding: 5px 10px;min-width: 85px;"><section><span>主要收益</span></section></th></tr></thead><tbody><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>渐进式披露</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>看到需要的</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>SKILL.md 体积 -35%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>确定性操作由脚本执行</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>看到需要的</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>消除 Schema 开销，减少 LLM 推理轮次</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>MCP 数据获取子 Agent 化</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>看到需要的</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>单轮 input token -38.4%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>长期记忆按需索引加载</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>看到需要的</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>INDEX 先行筛选，避免全量文档常驻</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>单 Agent 拆分为多 Agent</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少无关</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>打开后续优化空间，分散滚雪球效应</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>Agent 专属配置</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少无关</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>测试/视觉 Agent 成本 -64%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>代码图谱</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少无关</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>总 token -22.7%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>稳定前缀设计</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少重复</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>提升缓存命中率</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>避免重复加载 Skill</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少重复</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>消除重复常驻的 SKILL.md 开销</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(248, 248, 248);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>rtk 压缩 CLI 输出</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少重复</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>命令行输出 -60%~90%</span></section></td></tr><tr style="color: rgb(0, 0, 0);background-attachment: scroll;background-clip: border-box;background-color: rgb(255, 255, 255);background-image: none;background-origin: padding-box;background-position-x: left;background-position-y: top;background-repeat: no-repeat;background-size: auto;width: auto;height: auto;"><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>工具调用并行化</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>减少重复</span></section></td><td style="font-size: 14px;padding-top: 5px;padding-right: 10px;padding-bottom: 5px;padding-left: 10px;min-width: 85px;border-top-style: solid;border-bottom-style: solid;border-left-style: solid;border-right-style: solid;border-top-width: 1px;border-bottom-width: 1px;border-left-width: 1px;border-right-width: 1px;border-top-color: rgba(204, 204, 204, 0.4);border-bottom-color: rgba(204, 204, 204, 0.4);border-left-color: rgba(204, 204, 204, 0.4);border-right-color: rgba(204, 204, 204, 0.4);border-top-left-radius: 0px;border-top-right-radius: 0px;border-bottom-right-radius: 0px;border-bottom-left-radius: 0px;"><section><span>无依赖调用合并为 1 轮，减少历史重复打包次数</span></section></td></tr></tbody></table></section><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>全流程反推：以实测的 Wave 消耗分布为基准，代入各分项已验证的降幅区间反推，一个中型需求跑完全流程，token 成本大致能降低 50%~65%。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span><br/></span></p><h3 style="padding-top: 0px;padding-bottom: 0px;padding-left: 0px;padding-right: 0px;display: block;text-align: left;margin-top: 0px;margin-right: 0px;margin-bottom: 10px;margin-left: 0px;"><span style="font-family: 'Microsoft YaHei';padding-left: 10px;border-left-width: 5px;border-left-style: solid;border-left-color: #0E88EB;font-size: 20px;color: #000000;line-height: 1.5em;letter-spacing: 0em;text-align: left;font-weight: bold;display: block;"><span>五、总结与展望</span></span></h3><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>回顾整个实践过程，我们沉淀了以下几点核心经验：</span></p><ol style="list-style-type: decimal;margin-top: 8px;margin-bottom: 8px;margin-left: 0px;margin-right: 0px;padding-top: 0px;padding-bottom: 0px;padding-left: 25px;padding-right: 0px;color: rgb(62, 71, 83);text-align: justify;" class="list-paddingleft-1"><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>省 token 不等于功能降级——只是调整"何时加载"“怎么表达”，没删任何功能，反而让工作流更清晰。</span></section></li><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>上游收集一次，通过文档传递——最贵的冗余是"每个 Agent 各自重新发现同一份信息"。</span></section></li><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>最省钱的调用是不调用——确定性操作用 CLI/数据预取解决，把 LLM 留给真正需要语义理解的地方。</span></section></li><li><section style="margin-top: 5px;margin-bottom: 5px;color: rgb(62, 71, 83);font-size: 16px;line-height: 1.8em;letter-spacing: 0em;text-align: left;font-weight: normal;"><span>能并行就不要串行——没有数据依赖的多次调用，合并到同一轮消息内发起，能省下的是历史被重复打包的那几轮，而不只是等待时间。</span></section></li></ol><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>落地优先级：先做规模预判和 Agent 拆分（架构前提）→ 度量 + SKILL.md 重排 + 全局接入 rtk（一个下午见效）→ 条件内容移出 SKILL.md、状态外化、排查无依赖调用改并行（逐步推进）→ 代码图谱、CLI 替代 MCP、工具裁剪、子 Agent 化、长期记忆索引化（中长期系统性梳理）。</span></p><p style="color: rgb(62, 71, 83);font-size: 16px;line-height: 1.75;text-align: justify;text-indent: 0em;padding-top: 8px;padding-bottom: 8px;padding-left: 0px;padding-right: 0px;margin-top: 0px;margin-right: 0px;margin-bottom: 0px;margin-left: 0px;"><span>目前这套"三原则、十方向"的优化方法已在 tech-leader 工作流全面落地，后续我们将补齐严格的端到端 A/B 复核，并把方法论沉淀为可复用的 checklist，应用到团队内其他 Multi-Agent 工作流的成本治理中。</span></p><section style="text-align: center;"></section></section> <div class="content-mask">
<a href="javascript:void (0);" class="mask-text login-trigger">
<i class="iconfont icon-suo"></i>
登录查看剩余 70% 内容
<i class="iconfont icon-arrow-right-o"></i>
</a>
</div>
</div>
</div>
想了解 AI 如何助力您的企业？
免费获取企业 AI 成熟度诊断报告，发现转型机会
