{"id":10021396,"date":"2026-07-06T21:36:38","date_gmt":"2026-07-06T13:36:38","guid":{"rendered":"https:\/\/itmax.cn\/?p=10021396"},"modified":"2026-07-06T21:36:38","modified_gmt":"2026-07-06T13:36:38","slug":"%e8%ae%a9-ai-agent-%e7%b3%bb%e7%bb%9f%e8%87%aa%e5%b7%b1%e5%8f%91%e7%8e%b0-bug%e3%80%81%e8%87%aa%e5%b7%b1%e6%8f%90%e4%bf%ae%e5%a4%8d-pr%ef%bc%9a%e8%87%aa%e6%88%91%e8%bf%9b%e5%8c%96%e7%9a%84-harness","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10021396","title":{"rendered":"\u8ba9 AI Agent \u7cfb\u7edf\u81ea\u5df1\u53d1\u73b0 bug\u3001\u81ea\u5df1\u63d0\u4fee\u590d PR\uff1a\u81ea\u6211\u8fdb\u5316\u7684 Harness"},"content":{"rendered":"<p>\u672c\u6587\u4ecb\u7ecd\u600e\u4e48\u8ba9 AI Agent \u7684\u5de5\u7a0b\u4ee3\u7801\uff08Harness\uff09\u5177\u5907\u81ea\u6211\u8fdb\u5316\u80fd\u529b\u2014\u2014\u81ea\u52a8\u8bb0\u5f55\u8fd0\u884c\u6570\u636e\u3001\u81ea\u52a8\u8bc6\u522b\u9519\u8bef\u6a21\u5f0f\u3001\u81ea\u52a8\u751f\u6210\u4fee\u590d PR\uff08Pull Request\uff0c\u5408\u5e76\u8bf7\u6c42\uff09\u3002\u5185\u5bb9\u8986\u76d6\u76d1\u63a7\u3001\u9519\u8bef\u6a21\u5f0f\u8bc6\u522b\u3001\u81ea\u52a8\u4fee\u590d\u3001\u884c\u4e3a\u5206\u6790\u548c\u751f\u4ea7\u843d\u5730\u65b9\u6848\uff0c\u6bcf\u4e00\u7ae0\u90fd\u4f1a\u914d\u5408 demo \u9879\u76ee <a href=\"https:\/\/link.segmentfault.com\/?enc=5gkJpUR1qkJopxTCllcX4w%3D%3D.yp96OEjHcJizkB6CUjabnBBMgfNqlJ1RpQaN2VoKKhszoU0MTWotvT6OTasZacLN\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">evo-agent-demo<\/a> \u7684\u4ee3\u7801\u548c\u8fd0\u884c\u7ed3\u679c\u6765\u8bb2\u89e3\u3002<\/p>\n<h2>\u4ece\u4e00\u4e2a bug \u8bf4\u8d77<\/h2>\n<p>\u5047\u8bbe\u4f60\u505a\u4e86\u4e00\u4e2a AI Agent \u4ea7\u54c1\uff0c\u5b83\u53ef\u4ee5\u641c\u7d22\u8d44\u6599\u3001\u67e5\u6570\u636e\u5e93\u3001\u6267\u884c\u4ee3\u7801\u3002\u4e0a\u7ebf\u524d\u4e5f\u5728\u5185\u90e8\u8fdb\u884c\u4e86\u53cd\u590d\u6d4b\u8bd5\uff0c\u5e76\u4e14\u6ca1\u53d1\u73b0\u4ec0\u4e48\u95ee\u9898\uff0c\u4e8e\u662f\u5c31\u6b63\u5f0f\u4e0a\u7ebf\u4e86\u3002<\/p>\n<p>\u4f46\u4ea7\u54c1\u4e0a\u7ebf\u4e00\u6bb5\u65f6\u95f4\u540e\uff0c\u5f00\u59cb\u9646\u7eed\u7684\u6536\u5230\u7528\u6237\u53cd\u9988\uff0c\u4ea7\u54c1\u4ea7\u751f\u4e86\u5404\u79cd\u5404\u6837\u7684\u9519\u8bef\u3002\u56e0\u4e3a\u5185\u90e8\u6d4b\u8bd5\u8986\u76d6\u4e0d\u5230\u6240\u6709\u7684\u4f7f\u7528\u573a\u666f\uff0c\u800c\u771f\u5b9e\u7528\u6237\u7684\u6570\u91cf\u8fdc\u8d85\u56e2\u961f\u6d4b\u8bd5\u4eba\u5458\uff0c\u5e76\u4e14\u4ed6\u4eec\u7684\u63d0\u95ee\u65b9\u5f0f\u3001\u4f7f\u7528\u4e60\u60ef\u3001\u8fb9\u754c\u64cd\u4f5c\u8fdc\u8fdc\u8d85\u51fa\u4e86\u6d4b\u8bd5\u7528\u4f8b\u7684\u8303\u56f4\uff0c\u5bfc\u81f4\u51fa\u73b0\u4e86\u5f88\u591a\u610f\u6599\u4e4b\u5916\u7684\u60c5\u51b5\u3002\u6bd4\u5982\u5ba2\u670d\u6536\u5230\u6295\u8bc9\u8bf4\u201c\u95ee\u4e86\u4e2a\u95ee\u9898\uff0cAI \u76f4\u63a5\u62a5\u9519\u4e86\u201d\uff0c\u6280\u672f\u8bba\u575b\u4e0a\u6709\u4eba\u53d1\u5e16\u8bf4\u591a\u8f6e\u5bf9\u8bdd\u540e\u7a0b\u5e8f\u5d29\u4e86\uff0c\u8fd0\u8425\u90a3\u8fb9\u4e5f\u53cd\u9988\u641c\u7d22\u529f\u80fd\u65f6\u597d\u65f6\u574f\u3002<\/p>\n<p>\u8fd9\u79cd\u9760\u7528\u6237\u53cd\u9988\u53d1\u73b0\u95ee\u9898\u7684\u65b9\u5f0f\u5f88\u88ab\u52a8\u2014\u2014\u7b49\u7528\u6237\u53cd\u9988\u65f6\uff0c\u5f80\u5f80\u5df2\u7ecf\u6709\u5f88\u591a\u7528\u6237\u53d7\u5230\u4e86\u5f71\u54cd\u3002\u5bf9\u5f00\u53d1\u6765\u8bf4\u66f4\u9ebb\u70e6\u7684\u662f\uff1a\u5149\u51ed\u4e00\u53e5\u53cd\u9988\u5f88\u96be\u5b9a\u4f4d\u5230 bug \u7684\u771f\u6b63\u539f\u56e0\uff0c\u56e0\u4e3a\u624b\u4e0a\u6ca1\u6709\u51fa\u95ee\u9898\u65f6\u7684\u6267\u884c\u4e0a\u4e0b\u6587\u3002<\/p>\n<p>\u5927\u591a\u6570 AI Agent \u4ea7\u54c1\u7684 bug \u57fa\u672c\u4e0a\u90fd\u51fa\u73b0\u5728\u5305\u88f9\u6a21\u578b\u7684\u90a3\u5c42\u5de5\u7a0b\u4ee3\u7801\u4e0a\uff0c\u8fd9\u5c42\u4ee3\u7801\u6709\u4e00\u4e2a\u4e13\u95e8\u7684\u540d\u5b57\uff1a<strong>Harness<\/strong>\u3002Harness\uff08\u610f\u601d\u662f\u9a6c\u5177\u3001\u9a6c\u978d\uff09 \u5c31\u662f\u9a71\u52a8\u6a21\u578b\u8fd0\u8f6c\u7684\u5de5\u7a0b\u7cfb\u7edf\u3002\u6253\u4e2a\u6bd4\u65b9\uff1a\u6a21\u578b\u662f CPU\uff0cHarness \u662f\u64cd\u4f5c\u7cfb\u7edf\u2014\u2014\u64cd\u4f5c\u7cfb\u7edf\u7ba1\u7406\u8fdb\u7a0b\u8c03\u5ea6\u548c\u5185\u5b58\u5206\u914d\uff0cHarness \u7ba1\u7406\u5de5\u5177\u8c03\u7528\u3001\u4e0a\u4e0b\u6587\u7a97\u53e3\u548c\u9519\u8bef\u6062\u590d\u3002<\/p>\n<h3>\u4ece\u88ab\u52a8\u5230\u4e3b\u52a8<\/h3>\n<p>\u65e2\u7136\u95ee\u9898\u51fa\u5728 Harness \u4e0a\uff0c\u90a3\u4e48\u6211\u4eec\u5f97\u77e5\u9053 Harness \u6bcf\u6b21\u6267\u884c\u7684\u65f6\u5019\u5230\u5e95\u53d1\u751f\u4e86\u4ec0\u4e48\u3002\u6211\u4eec\u53ef\u4ee5\u7ed9 Agent \u52a0\u4e00\u5c42 Tracing\uff0c\u8fd9\u76f8\u5f53\u4e8e\u7ed9 Agent \u88c5\u4e0a\u884c\u8f66\u8bb0\u5f55\u4eea\u3002Agent \u6bcf\u6b21\u6267\u884c\u7684\u65f6\u5019\uff0c\u8c03\u7528\u4e86\u4ec0\u4e48\u5de5\u5177\u3001\u4f20\u7684\u53c2\u6570\u662f\u4ec0\u4e48\u3001\u54cd\u5e94\u7ed3\u679c\u662f\u4ec0\u4e48\u7b49\u7b49\uff0c\u90fd\u4f1a\u8bb0\u5230\u6570\u636e\u5e93\u91cc\u3002\u8fd9\u6837\u5c31\u4e0d\u7528\u7b49\u7528\u6237\u6765\u53cd\u9988\uff0c\u6211\u4eec\u81ea\u5df1\u67e5\u6570\u636e\u5e93\u5c31\u80fd\u770b\u5230\u54ea\u4e9b\u8bf7\u6c42\u51fa\u4e86\u95ee\u9898\u3002<\/p>\n<p>\u4f46\u6bcf\u5929\u4ea7\u751f\u7684\u8bf7\u6c42\u8fd9\u4e48\u591a\uff0c\u53ef\u80fd\u6709\u51e0\u5343\u4e0a\u4e07\u6761\uff0c\u5176\u4e2d\u7684\u9519\u8bef\u4e5f\u4e0d\u5c11\uff0c\u4f9d\u8d56\u4eba\u5de5\u9010\u6761\u67e5\u770b\u4e0d\u592a\u73b0\u5b9e\u3002\u6240\u4ee5\u8fd8\u9700\u8981\u8ba9\u7cfb\u7edf\u81ea\u52a8\u6309\u7c7b\u578b\u628a\u9519\u8bef\u5206\u7ec4\uff08\u6bd4\u5982\u6240\u6709 429 \u9650\u6d41\u9519\u8bef\u5f52\u5230\u4e00\u7ec4\uff09\uff0c\u7136\u540e\u5224\u65ad\u6bcf\u4e00\u7c7b\u9519\u8bef\u662f\u7528\u6237\u7684\u95ee\u9898\/\u4f9b\u5e94\u5546\u7684\u95ee\u9898\u3001\u8fd8\u662f Harness \u81ea\u8eab\u7684 bug\u2014\u2014\u5982\u679c\u662f Harness \u7684 bug\uff0c\u9700\u8981\u8ba9 LLM \u6839\u636e\u9519\u8bef\u539f\u56e0\u53bb\u5206\u6790\u76f8\u5173\u7684\u6e90\u7801\u5e76\u4e14\u751f\u6210\u4fee\u590d\u4ee3\u7801\uff0c\u6700\u540e\u63d0\u4ea4 PR \u6765\u4fee\u590d\u8fd9\u4e2a\u95ee\u9898\u3002\u672c\u6587\u540e\u9762\u7684\u5185\u5bb9\u5c06\u6309\u8fd9\u4e2a\u987a\u5e8f\u6765\u5c55\u5f00\uff0c\u5e76\u914d\u4ee5 evo-agent-demo \u7684\u4ee3\u7801\u548c\u8fd0\u884c\u7ed3\u679c\u6765\u8bb2\u89e3\u3002<\/p>\n<h3>demo \u5feb\u901f\u4f53\u9a8c<\/h3>\n<p>\u6211\u4eec\u5148\u8fd0\u884c\u4e00\u6b21 evo-agent-demo \u6765\u770b\u770b\u6548\u679c\uff0c\u540e\u9762\u4f1a\u9010\u6b65\u62c6\u89e3\u6bcf\u4e2a\u73af\u8282\u7684\u5b9e\u73b0\uff1a<\/p>\n<pre><code class=\"bash\">git clone https:\/\/github.com\/woai3c\/evo-agent-demo\ncd evo-agent-demo &amp;&amp; pnpm install\ncp .env.example .env          # \u586b\u5165\u81f3\u5c11\u4e00\u4e2a LLM API Key\npnpm db:seed                  # \u586b\u5145\u6f14\u793a\u6570\u636e\npnpm simulate --mock          # \u6a21\u62df 100 \u6761\u7528\u6237\u8bf7\u6c42\uff08\u542b\u5404\u79cd\u9519\u8bef\uff0c\u4e0d\u8c03 API\uff09\npnpm dev                      # \u542f\u52a8\u524d\u540e\u7aef\u5f00\u53d1\u670d\u52a1\u5668<\/code><\/pre>\n<p>\u9879\u76ee\u542f\u52a8\u540e\uff0c\u6253\u5f00 <code>http:\/\/localhost:5173\/admin\/inspections<\/code> \u9875\u9762\uff0c\u70b9\u51fb\u201c\u5de1\u68c0\uff1a\u8bc6\u522b Pattern\u201d\u6309\u94ae\uff0c\u5373\u53ef\u89c2\u5bdf\u5230\u65e5\u5fd7\u9762\u677f\u5b9e\u65f6\u8f93\u51fa\u5de1\u68c0\u8fdb\u5ea6\u3002\u5de1\u68c0\u5b8c\u6210\u540e\u53ef\u4ee5\u5207\u6362\u5230\u6982\u89c8\u9875\u67e5\u770b\u5404\u9879\u6307\u6807\u7684\u53d8\u5316\u3002<\/p>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930611\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"><\/p>\n<h2>\u5feb\u901f\u4e86\u89e3 AI Agent Harness \u5de5\u7a0b\u5168\u8c8c<\/h2>\n<blockquote><p>\u5df2\u6709 AI Agent \u5f00\u53d1\u7ecf\u9a8c\u7684\u8bfb\u8005\u53ef\u4ee5\u8df3\u8fc7\u672c\u8282\uff0c\u76f4\u63a5\u4ece\u201cTracing\u201d\u5f00\u59cb\u3002<\/p><\/blockquote>\n<p>\u8fd9\u4e00\u8282\u5c06\u4ecb\u7ecd\u4e00\u4e2a Agent \u4ea7\u54c1\u7531\u54ea\u4e9b\u90e8\u4ef6\u7ec4\u6210\u3001\u5404\u81ea\u8d1f\u8d23\u4ec0\u4e48\uff0c\u8ba9\u8fd8\u6ca1\u63a5\u89e6\u8fc7 AI Agent \u5f00\u53d1\u7684\u8bfb\u8005\u5148\u6709\u4e2a\u6574\u4f53\u6982\u5ff5\u3002<\/p>\n<h3>\u4ec0\u4e48\u662f AI Agent<\/h3>\n<p>\u7b80\u5355\u6765\u8bf4\uff0cAI Agent \u5c31\u662f\u80fd\u8c03\u7528\u5de5\u5177\u6267\u884c\u4efb\u52a1\u7684 LLM \u5e94\u7528\u3002\u666e\u901a\u804a\u5929\u673a\u5668\u4eba\u53ea\u80fd\u56de\u590d\u6587\u672c\uff0c\u4f46 Agent \u8fd8\u53ef\u4ee5\u6267\u884c\u641c\u7d22\u3001\u67e5\u8be2\u6570\u636e\u5e93\u3001\u4fee\u6539\u4ee3\u7801\u7b49\u4efb\u52a1\u3002\u804a\u5929\u673a\u5668\u4eba\u7684\u6d41\u7a0b\u662f\u201c\u7528\u6237\u8bf4\u8bdd \u2192 \u6a21\u578b\u56de\u590d \u2192 \u7ed3\u675f\u201d\uff0cAgent \u5219\u662f\u201c\u7528\u6237\u8bf4\u8bdd \u2192 \u6a21\u578b\u51b3\u5b9a\u8c03\u4ec0\u4e48\u5de5\u5177 \u2192 \u62ff\u5230\u7ed3\u679c \u2192 \u7ee7\u7eed\u601d\u8003 \u2192 \u53ef\u80fd\u518d\u8c03\u5de5\u5177 \u2192 \u2026 \u2192 \u6700\u7ec8\u56de\u590d\u7528\u6237\u201d\u3002\u8fd9\u4e2a\u5faa\u73af\u8fc7\u7a0b\u53eb <strong>Agent Loop<\/strong>\u3002<\/p>\n<p>\u7528\u672c\u6587\u7684 demo\uff08\u4ee3\u53f7 Evo\uff09\u4e3e\u4e2a\u4f8b\u5b50\u3002\u7528\u6237\u95ee\u201c\u5e2e\u6211\u7814\u7a76\u4e00\u4e0b Hono \u6846\u67b6\u201d\uff0cAgent \u4f1a\u5148\u8c03\u7528 <code>webSearch<\/code> \u5de5\u5177\u641c\u7d22\u8d44\u6599\uff0c\u7136\u540e\u4ece\u7ed3\u679c\u91cc\u9009\u51e0\u4e2a URL\uff0c\u518d\u8c03\u7528 <code>webFetch<\/code> \u5de5\u5177\u6293\u53d6\u9875\u9762\u5185\u5bb9\uff0c\u6700\u540e\u518d\u7efc\u5408\u4fe1\u606f\u5199\u4e00\u4efd\u603b\u7ed3\u3002\u6574\u4e2a\u8fc7\u7a0b\u4e2d\u6a21\u578b\u81ea\u4e3b\u51b3\u5b9a\u5de5\u5177\u7684\u9009\u62e9\u548c\u8c03\u7528\u987a\u5e8f\uff0c\u4e0d\u9700\u8981\u4eba\u5de5\u5e72\u9884\u3002<\/p>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930612\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<p>\u7528\u65f6\u5e8f\u56fe\u62c6\u89e3\u4e00\u4e0b\u5185\u90e8\u6d41\u7a0b\uff1a<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">sequenceDiagram\n    autonumber\n    actor U as \u7528\u6237\n    participant AL as Agent Loop\n    participant LLM as LLM API\n    participant T as \u5de5\u5177\n\n    U-&gt;&gt;AL: &quot;\u5e2e\u6211\u7814\u7a76\u4e00\u4e0b Hono \u6846\u67b6&quot;\n    AL-&gt;&gt;LLM: \u53d1\u9001\u6d88\u606f + \u5de5\u5177\u5217\u8868\n    LLM--&gt;&gt;AL: tool_call: webSearch(&quot;Hono framework 2025&quot;)\n    AL-&gt;&gt;T: \u6267\u884c webSearch\n    T--&gt;&gt;AL: \u8fd4\u56de\u641c\u7d22\u7ed3\u679c\n    AL-&gt;&gt;LLM: \u5c06\u5de5\u5177\u7ed3\u679c\u56de\u4f20\n    LLM--&gt;&gt;AL: tool_call: webFetch(&quot;https:\/\/hono.dev&quot;)\n    AL-&gt;&gt;T: \u6267\u884c webFetch\n    T--&gt;&gt;AL: \u8fd4\u56de\u9875\u9762\u5185\u5bb9\n    AL-&gt;&gt;LLM: \u5c06\u5de5\u5177\u7ed3\u679c\u56de\u4f20\n    LLM--&gt;&gt;AL: \u751f\u6210\u6587\u672c\u56de\u590d\uff08\u6d41\u5f0f\u8f93\u51fa\uff09\n    AL-&gt;&gt;U: \u6d41\u5f0f\u5c55\u793a\u56de\u590d<\/code><\/pre>\n<p>Agent Loop \u5c31\u662f\u7528\u6237\u3001\u6a21\u578b\u548c\u5de5\u5177\u4e4b\u95f4\u7684\u8c03\u5ea6\u4e2d\u67a2\u3002\u7528\u6237\u7684\u6d88\u606f\u4ea4\u7ed9\u6a21\u578b\uff0c\u6a21\u578b\u51b3\u5b9a\u8c03\u7528\u4ec0\u4e48\u5de5\u5177\uff0cAgent Loop \u6267\u884c\u5de5\u5177\u5e76\u628a\u7ed3\u679c\u4ea4\u56de\u7ed9\u6a21\u578b\uff0c\u6a21\u578b\u518d\u51b3\u5b9a\u4e0b\u4e00\u6b65\u662f\u7ee7\u7eed\u8c03\u7528\u5de5\u5177\u8fd8\u662f\u76f4\u63a5\u56de\u590d\u7528\u6237\u3002\u8fd9\u4e2a\u5faa\u73af\u53ef\u80fd\u8f6c\u4e00\u6b21\uff0c\u4e5f\u53ef\u80fd\u8f6c\u5341\u6b21\uff0c\u53d6\u51b3\u4e8e\u4efb\u52a1\u7684\u590d\u6742\u5ea6\u3002<\/p>\n<h3>Agent \u7684\u6838\u5fc3\u7ec4\u6210<\/h3>\n<p>\u4e00\u4e2a Agent \u4ea7\u54c1\u5927\u6982\u7531\u8fd9\u51e0\u4e2a\u90e8\u5206\u7ec4\u6210\uff1a<\/p>\n<pre><code>\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n\u2502              Harness                \u2502\n\u2502                                     \u2502\n\u2502  \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510  \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510  \u2502\n\u2502  \u2502 Agent Loop\u2502  \u2502 \u5de5\u5177\u7cfb\u7edf       \u2502  \u2502\n\u2502  \u2502 (\u8c03\u5ea6\u5668)  \u2502  \u2502 (\u641c\u7d22\/\u6570\u636e\u5e93\/ \u2502  \u2502\n\u2502  \u2502           \u2502  \u2502  \u4ee3\u7801\u6267\u884c\/...) \u2502  \u2502\n\u2502  \u2514\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2518  \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518  \u2502\n\u2502        \u2502                \u2502          \u2502\n\u2502  \u250c\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510  \u2502\n\u2502  \u2502     \u4e0a\u4e0b\u6587\u7ba1\u7406 (\u8bb0\u5fc6)         \u2502  \u2502\n\u2502  \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518  \u2502\n\u2502                                     \u2502\n\u2502  \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510  \u2502\n\u2502  \u2502     Tracing (\u884c\u8f66\u8bb0\u5f55\u4eea)     \u2502  \u2502\n\u2502  \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518  \u2502\n\u2502                                     \u2502\n\u2502  \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510  \u2502\n\u2502  \u2502     \u8fdb\u5316\u5f15\u64ce (\u81ea\u6211\u5b8c\u5584)      \u2502  \u2502\n\u2502  \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518  \u2502\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n         \u2195\n    \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n    \u2502   LLM   \u2502  \u2190 \u53ef\u66ff\u6362\uff08DeepSeek \/ OpenAI \/ Anthropic \/ ...\uff09\n    \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/code><\/pre>\n<p><strong>Harness<\/strong> \u5c31\u662f\u5305\u88f9\u6a21\u578b\u7684\u6240\u6709\u5de5\u7a0b\u4ee3\u7801\u3002\u6a21\u578b\u63d0\u4f9b\u63a8\u7406\u80fd\u529b\uff0cHarness \u8d1f\u8d23\u628a\u8fd9\u79cd\u80fd\u529b\u53d8\u6210\u4e00\u4e2a\u80fd\u7528\u7684\u4ea7\u54c1\u3002<\/p>\n<p>Agent Loop \u662f\u8c03\u5ea6\u5668\uff0c\u9a71\u52a8\u201c\u95ee\u6a21\u578b \u2192 \u8c03\u7528\u5de5\u5177 \u2192 \u518d\u95ee\u6a21\u578b\u201d\u7684\u5faa\u73af\u3002demo \u91cc\u8fd9\u4e2a\u5faa\u73af\u7684\u5165\u53e3\u662f <code>agentLoop()<\/code> \u51fd\u6570\uff08<code>packages\/server\/src\/agent\/loop.ts<\/code>\uff09\uff0c\u901a\u8fc7 Vercel AI SDK \u5411\u6a21\u578b\u53d1\u9001\u6d41\u5f0f\u8bf7\u6c42\uff0c\u9010\u6b65\u6d88\u8d39\u6a21\u578b\u8fd4\u56de\u7684\u4e8b\u4ef6\u3002\u6a21\u578b\u8981\u6c42\u8c03\u7528\u5de5\u5177\u65f6\uff0cSDK \u81ea\u52a8\u6267\u884c\u5bf9\u5e94\u7684\u5de5\u5177\u5e76\u628a\u7ed3\u679c\u4f20\u56de\u3002\u540c\u65f6\u4e3a\u4e86\u9632\u6b62\u65e0\u9650\u5faa\u73af\uff0c\u5728\u672c demo \u91cc\u5355\u6b21\u5bf9\u8bdd\u6700\u591a\u6267\u884c 15 \u6b65\uff08\u751f\u4ea7\u73af\u5883\u91cc\u8fdc\u8fdc\u4e0d\u6b62 15 \u8f6e\uff09\u3002<\/p>\n<p>\u5de5\u5177\u7cfb\u7edf\u5b9a\u4e49\u4e86 Agent \u80fd\u505a\u4ec0\u4e48\u3002\u5728 Evo demo \u91cc\u6709 6 \u4e2a\u5de5\u5177\uff1a\u7f51\u7edc\u641c\u7d22\u3001\u6293\u53d6\u7f51\u9875\u3001\u8bfb\u53d6\u6587\u4ef6\u3001\u6267\u884c\u4ee3\u7801\u3001\u67e5\u8be2\u6570\u636e\u5e93\u3001\u53d1\u9001\u90ae\u4ef6\u3002\u6bcf\u4e2a\u5de5\u5177\u7531\u4e09\u90e8\u5206\u7ec4\u6210\uff1a\u4e00\u6bb5\u7ed9\u6a21\u578b\u770b\u7684\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\uff08\u6a21\u578b\u6839\u636e\u63cf\u8ff0\u51b3\u5b9a\u8981\u4e0d\u8981\u8c03\u8fd9\u4e2a\u5de5\u5177\uff09\uff0c\u4e00\u7ec4\u7528 Zod Schema \u5b9a\u4e49\u7684\u53c2\u6570\u683c\u5f0f\uff0c\u4ee5\u53ca\u4e00\u4e2a\u5b9e\u9645\u6267\u884c\u64cd\u4f5c\u7684\u51fd\u6570\u3002<\/p>\n<p>\u4e0a\u4e0b\u6587\u7ba1\u7406\u89e3\u51b3\u7684\u662f token \u9650\u5236\u95ee\u9898\u3002Agent \u6bcf\u8c03\u7528\u4e00\u6b21\u5de5\u5177\uff0c\u5de5\u5177\u7684\u8f93\u5165\u548c\u8f93\u51fa\u90fd\u8981\u653e\u8fdb\u5bf9\u8bdd\u5386\u53f2\u53d1\u7ed9\u6a21\u578b\u300210 \u8f6e\u5bf9\u8bdd\u52a0\u51e0\u6b21\u7f51\u9875\u641c\u7d22\uff0c\u4e0a\u4e0b\u6587\u7a97\u53e3\uff08Context Window\uff09\u53ef\u80fd\u5c31\u5feb\u6ee1\u4e86\u3002demo \u91cc\u7528\u4e24\u4e2a\u529e\u6cd5\u5e94\u5bf9\uff1a\u6d88\u606f\u538b\u7f29\uff08<code>packages\/server\/src\/context\/compression.ts<\/code>\uff09\u2014\u2014\u5bf9\u8bdd\u5386\u53f2\u7684 token \u6570\u8d85\u8fc7\u7a97\u53e3\u5bb9\u91cf\u7684 70% \u65f6\uff0c\u81ea\u52a8\u628a\u8001\u6d88\u606f\u538b\u7f29\u6210\u6458\u8981\uff1b\u5de5\u5177\u7ed3\u679c\u622a\u65ad\uff08<code>packages\/server\/src\/context\/truncation.ts<\/code>\uff09\u2014\u2014\u641c\u7d22\u8fd4\u56de\u7684\u7f51\u9875\u53ef\u80fd\u6709\u51e0\u5341 KB\uff0c\u53ea\u4fdd\u7559\u524d 30,000 \u5b57\u7b26\u52a0\u5c3e\u90e8\u7247\u6bb5\u3002<\/p>\n<p>Tracing \u8bb0\u5f55\u6bcf\u4e00\u6b65\u7684\u6267\u884c\u6570\u636e\u2014\u2014\u8c03\u7528\u4e86\u4ec0\u4e48\u5de5\u5177\u3001\u82b1\u4e86\u591a\u5c11 token\u3001\u8017\u65f6\u591a\u4e45\u3001\u6210\u529f\u6ca1\u6709\u3002\u8fd9\u4e9b\u6570\u636e\u662f\u540e\u9762\u6240\u6709\u5206\u6790\u7684\u57fa\u7840\uff0c\u4e0b\u4e00\u8282\u8be6\u7ec6\u8bb2\u3002<\/p>\n<p>\u8fdb\u5316\u5f15\u64ce\u4ece Tracing \u6570\u636e\u91cc\u81ea\u52a8\u53d1\u73b0\u9519\u8bef\u3001\u5206\u7c7b\u539f\u56e0\u3001\u751f\u6210\u4fee\u590d\u4ee3\u7801\u3002\u540e\u9762\u4f1a\u5206\u522b\u8bb2\u5b83\u7684\u4e09\u4e2a\u90e8\u5206\uff1a\u9519\u8bef Pattern\uff08\u6a21\u5f0f\uff09 \u8bc6\u522b\u3001\u81ea\u52a8\u4fee\u590d\u3001\u884c\u4e3a\u5206\u6790\u3002<\/p>\n<h3>demo \u9879\u76ee\u901f\u89c8<\/h3>\n<p>Evo \u662f\u4e00\u4e2a pnpm monorepo\uff08\u5355\u4ed3\u591a\u5305\uff09\uff0c\u5305\u542b\u4e09\u4e2a\u5305\uff1a<\/p>\n<ul>\n<li><code>packages\/shared<\/code>\uff08<code>@evo\/shared<\/code>\uff09\u2014\u2014TypeScript \u7c7b\u578b\u548c\u5e38\u91cf\uff0c\u88ab server \u548c web \u5171\u540c\u5f15\u7528<\/li>\n<li><code>packages\/server<\/code>\uff08<code>@evo\/server<\/code>\uff09\u2014\u2014Hono + Node.js \u540e\u7aef\uff0c\u5305\u542b Agent Loop\u3001\u5de5\u5177\u7cfb\u7edf\u3001Tracing\u3001\u8fdb\u5316\u5f15\u64ce\u3001REST API<\/li>\n<li><code>packages\/web<\/code>\uff08<code>@evo\/web<\/code>\uff09\u2014\u2014Vite + React 19 + Tailwind + shadcn\/ui \u524d\u7aef\uff0c\u5305\u542b\u804a\u5929\u754c\u9762\u548c\u7ba1\u7406\u540e\u53f0<\/li>\n<\/ul>\n<p>\u6570\u636e\u5e93\u7528 better-sqlite3\uff0cLLM \u63a5\u5165\u7528 Vercel AI SDK\uff0c\u652f\u6301 DeepSeek\u3001OpenAI\u3001Anthropic\u3001\u963f\u91cc\u767e\u70bc\u3001\u667a\u8c31\u3001Moonshot \u516d\u5bb6\u4f9b\u5e94\u5546\uff0c\u9ed8\u8ba4\u4f7f\u7528 DeepSeek\u3002<\/p>\n<p>\u5173\u952e\u4ee3\u7801\u4f4d\u7f6e\uff1a<code>packages\/server\/src\/agent\/loop.ts<\/code> \u662f Agent Loop \u7684\u5165\u53e3\uff0c<code>packages\/server\/src\/evolution\/<\/code> \u662f\u8fdb\u5316\u5f15\u64ce\u7684\u5168\u90e8\u5b9e\u73b0\u3002\u8bfb\u8005\u5728\u9605\u8bfb\u540e\u7eed\u5185\u5bb9\u65f6\uff0c\u53ef\u4ee5\u5bf9\u7167\u8fd9\u4e9b\u6587\u4ef6\u67e5\u770b\u5b8c\u6574\u4ee3\u7801\u3002<\/p>\n<h4>demo \u4e0e\u751f\u4ea7\u73af\u5883\u7684\u5dee\u5f02<\/h4>\n<p>demo \u505a\u4e86\u4e0d\u5c11\u7b80\u5316\uff0c\u548c\u751f\u4ea7\u73af\u5883\u7684\u4e3b\u8981\u5dee\u5f02\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u7ef4\u5ea6<\/th>\n<th>demo \u5b9e\u73b0<\/th>\n<th>\u751f\u4ea7\u73af\u5883\u66ff\u4ee3\u65b9\u6848<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u5de1\u68c0\u89e6\u53d1<\/td>\n<td>\u7ba1\u7406\u540e\u53f0\u624b\u52a8\u70b9\u51fb\u201c\u5f00\u59cb\u5de1\u68c0\u201d<\/td>\n<td>\u5b9a\u65f6\u4efb\u52a1\uff08\u6bcf 1-2 \u5c0f\u65f6\uff09\uff0c\u5206\u6790\u4e0a\u4e00\u5468\u671f\u7684\u65b0\u589e\u9519\u8bef<\/td>\n<\/tr>\n<tr>\n<td>\u81ea\u52a8\u4fee\u590d\u89e6\u53d1<\/td>\n<td>\u7ba1\u7406\u540e\u53f0\u624b\u52a8\u70b9\u51fb\u201c\u81ea\u52a8\u4fee\u590d\u201d<\/td>\n<td>\u5b9a\u65f6\u4efb\u52a1\uff08\u6bcf\u5929\u51cc\u6668\uff09\uff0c\u5bf9\u672a\u4fee\u590d\u7684 harness_bug \u751f\u6210 PR<\/td>\n<\/tr>\n<tr>\n<td>\u6570\u636e\u5e93<\/td>\n<td>SQLite \u5355\u6587\u4ef6\uff0c\u9002\u5408\u672c\u5730\u5f00\u53d1<\/td>\n<td>PostgreSQL \/ MySQL\uff0c\u652f\u6301\u5e76\u53d1\u548c\u591a\u5b9e\u4f8b<\/td>\n<\/tr>\n<tr>\n<td>\u6a21\u62df\u6570\u636e<\/td>\n<td><code>pnpm simulate --mock<\/code> \u6279\u91cf\u6ce8\u5165<\/td>\n<td>\u771f\u5b9e\u7528\u6237\u6d41\u91cf\u4ea7\u751f\u7684 trace \u6570\u636e<\/td>\n<\/tr>\n<tr>\n<td>\u9519\u8bef\u6ce8\u5165<\/td>\n<td><code>simulate-errors.ts<\/code> \u968f\u673a\u6ce8\u5165\u9519\u8bef<\/td>\n<td>\u751f\u4ea7\u73af\u5883\u4e2d\u81ea\u7136\u53d1\u751f\u7684\u9519\u8bef<\/td>\n<\/tr>\n<tr>\n<td>PR \u63d0\u4ea4<\/td>\n<td>\u63d0\u4ea4\u5230\u672c\u5730 Git \u4ed3\u5e93<\/td>\n<td>\u901a\u8fc7 GitHub\/GitLab API \u63d0\u4ea4\u5230\u4ee3\u7801\u4ed3\u5e93\uff0c\u8d70 CI\/CD\uff08\u6301\u7eed\u96c6\u6210\/\u6301\u7eed\u90e8\u7f72\uff09\u6d41\u7a0b<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>demo \u91cc\u6240\u6709\u64cd\u4f5c\u90fd\u662f\u624b\u52a8\u89e6\u53d1\u7684\u2014\u2014\u70b9\u51fb\u6309\u94ae\u6267\u884c\u4e00\u6b21\u5de1\u68c0\u3001\u70b9\u51fb\u6309\u94ae\u751f\u6210\u4fee\u590d PR\u3002\u8fd9\u662f\u4e3a\u4e86\u8ba9\u4f60\u80fd\u4e00\u6b65\u6b65\u770b\u5230\u6bcf\u4e2a\u73af\u8282\u7684\u8f93\u5165\u548c\u8f93\u51fa\u3002\u751f\u4ea7\u73af\u5883\u4e2d\u8fd9\u4e9b\u64cd\u4f5c\u5e94\u8be5\u7531\u5b9a\u65f6\u4efb\u52a1\u81ea\u52a8\u6267\u884c\uff0c\u8c03\u7528\u7684\u51fd\u6570\u548c demo \u91cc\u624b\u52a8\u89e6\u53d1\u7684\u662f\u540c\u4e00\u4e2a\u3002<\/p>\n<h2>Tracing\uff1a\u7ed9 Agent \u88c5\u4e0a\u201c\u884c\u8f66\u8bb0\u5f55\u4eea\u201d<\/h2>\n<p>\u884c\u8f66\u8bb0\u5f55\u4eea\u7684\u4f5c\u7528\u662f\u51fa\u4e86\u4e8b\u6545\u770b\u5f55\u50cf\uff0c\u6ca1\u51fa\u4e8b\u4e5f\u80fd\u590d\u76d8\u9a7e\u9a76\u4e60\u60ef\u2014\u2014Tracing \u5728 Agent \u7cfb\u7edf\u91cc\u7684\u4f5c\u7528\u5dee\u4e0d\u591a\u3002\u4e0d\u8fc7\u548c\u666e\u901a\u65e5\u5fd7\u4e0d\u540c\uff0cTracing \u8bb0\u5f55\u7684\u662f\u7ed3\u6784\u5316\u7684\u6570\u636e\uff0c\u800c\u4e14\u8fd9\u4e9b\u6570\u636e\u4f1a\u88ab\u540e\u9762\u7684\u81ea\u52a8\u5206\u6790\u7a0b\u5e8f\u6301\u7eed\u6d88\u8d39\uff0c\u4e0d\u53ea\u662f\u51fa\u95ee\u9898\u4e86\u624d\u7ffb\u770b\u3002<\/p>\n<h3>Trace \u6570\u636e\u7ed3\u6784<\/h3>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\u3002demo \u5185\u7f6e\u4e86\u4e00\u4e2a\u97f3\u4e50\u5e93\u4f5c\u4e3a\u793a\u4f8b\u6570\u636e\uff0c\u5047\u8bbe\u7528\u6237\u95ee\u201c\u54ea\u4e2a\u6b4c\u624b\u4e13\u8f91\u6700\u591a\u201d\uff0cTracing \u4f1a\u628a Agent \u7684\u6574\u4e2a\u6267\u884c\u8fc7\u7a0b\u8bb0\u5f55\u4e0b\u6765\uff1a<\/p>\n<pre><code>Operation: op_abc123\n  \u251c\u2500 Step 0: call_llm  \u2192 \u6a21\u578b\u51b3\u5b9a\u8c03 dbQuery     (\u8017\u65f6 800ms, tokens: 1200)\n  \u251c\u2500 Step 1: call_tool \u2192 dbQuery \u6267\u884c SQL        (\u8017\u65f6 50ms, \u6210\u529f)\n  \u251c\u2500 Step 2: call_llm  \u2192 \u6a21\u578b\u6839\u636e\u67e5\u8be2\u7ed3\u679c\u7ec4\u7ec7\u56de\u7b54 (\u8017\u65f6 600ms, tokens: 900)\n  \u2514\u2500 \u7ed3\u679c: success, \u603b\u8017\u65f6 1450ms<\/code><\/pre>\n<p>\u6574\u4e2a\u8fc7\u7a0b\u88ab\u62c6\u6210\u4e00\u4e2a Operation\uff08\u64cd\u4f5c\uff09\u548c\u82e5\u5e72 Step\uff08\u6b65\u9aa4\uff09\u3002Operation \u5bf9\u5e94\u4e00\u6b21\u5b8c\u6574\u7684\u7528\u6237\u8bf7\u6c42\uff0c\u8bb0\u5f55\u603b\u8017\u65f6\u3001token \u7528\u91cf\u3001\u6700\u7ec8\u72b6\u6001\u8fd9\u4e9b\u6c47\u603b\u4fe1\u606f\u3002\u6bcf\u4e2a Step \u5bf9\u5e94 Agent Loop \u91cc\u7684\u4e00\u6b21 LLM \u8c03\u7528\u6216\u4e00\u6b21\u5de5\u5177\u8c03\u7528\u3002<\/p>\n<p>\u6bcf\u4e2a Step \u8bb0\u5f55\u4ee5\u4e0b\u5b57\u6bb5\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u5b57\u6bb5<\/th>\n<th>\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>type<\/code><\/td>\n<td>\u6b65\u9aa4\u7c7b\u578b\uff1a<code>call_llm<\/code>\uff08\u8c03\u7528\u6a21\u578b\uff09\u6216 <code>call_tool<\/code>\uff08\u8c03\u7528\u5de5\u5177\uff09<\/td>\n<\/tr>\n<tr>\n<td><code>duration_ms<\/code><\/td>\n<td>\u8fd9\u4e00\u6b65\u7684\u8017\u65f6\uff08\u6beb\u79d2\uff09<\/td>\n<\/tr>\n<tr>\n<td><code>tokens<\/code><\/td>\n<td>\u6d88\u8017\u7684 token \u6570\uff08\u4ec5 <code>call_llm<\/code> \u6b65\u9aa4\uff09\uff0c\u5305\u542b input\u3001output\u3001cached \u4e09\u4e2a\u5206\u91cf<\/td>\n<\/tr>\n<tr>\n<td><code>tool_name<\/code><\/td>\n<td>\u8c03\u7528\u7684\u5de5\u5177\u540d\u79f0\uff08\u4ec5 <code>call_tool<\/code> \u6b65\u9aa4\uff09<\/td>\n<\/tr>\n<tr>\n<td><code>tool_success<\/code><\/td>\n<td>\u5de5\u5177\u8c03\u7528\u662f\u5426\u6210\u529f\uff08\u5e03\u5c14\u503c\uff09<\/td>\n<\/tr>\n<tr>\n<td><code>tool_input<\/code><\/td>\n<td>\u5de5\u5177\u8c03\u7528\u7684\u8f93\u5165\u53c2\u6570<\/td>\n<\/tr>\n<tr>\n<td><code>tool_output<\/code><\/td>\n<td>\u5de5\u5177\u8fd4\u56de\u7684\u5b8c\u6574\u7ed3\u679c<\/td>\n<\/tr>\n<tr>\n<td><code>error<\/code><\/td>\n<td>\u5982\u679c\u51fa\u9519\uff0c\u9519\u8bef\u7684\u8be6\u7ec6\u4fe1\u606f<\/td>\n<\/tr>\n<tr>\n<td><code>llm_response<\/code><\/td>\n<td>\u6a21\u578b\u7684\u5b8c\u6574\u6587\u672c\u56de\u590d\uff08\u4ec5 <code>call_llm<\/code> \u6b65\u9aa4\uff09<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><code>tool_input<\/code> \u548c <code>tool_output<\/code> \u4fdd\u7559\u4e86\u5de5\u5177\u8c03\u7528\u7684\u5b8c\u6574\u4e0a\u4e0b\u6587\uff0c\u7ba1\u7406\u9762\u677f\u91cc\u53ef\u4ee5\u67e5\u770b\u6bcf\u4e00\u6b65\u7684\u8f93\u5165\u8f93\u51fa\uff0c\u65b9\u4fbf\u6392\u67e5\u95ee\u9898\u3002<code>llm_response<\/code> \u4fdd\u7559\u4e86\u6a21\u578b\u56de\u590d\uff0c\u540e\u7eed\u91cd\u653e\u5f15\u64ce\u9700\u8981\u8fd9\u4e9b\u6570\u636e\u6765\u590d\u73b0\u6267\u884c\u8fc7\u7a0b\u3002<\/p>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930613\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<h3>Tracing \u7684\u63a5\u5165\u65b9\u5f0f<\/h3>\n<blockquote><p>\u6e90\u7801\uff1a<code>packages\/server\/src\/agent\/loop.ts<\/code><\/p><\/blockquote>\n<p>demo \u91cc Agent Loop \u6bcf\u6267\u884c\u5b8c\u4e00\u6b65\u5c31\u8bb0\u5f55\u4e00\u6761 trace \u6570\u636e\uff1a\u5de5\u5177\u8c03\u7528\u5b8c\u6210\u540e\uff0c\u6267\u884c\u4e00\u6b21 <code>tracer.onToolResult()<\/code> \u8bb0\u5f55\u8c03\u7528\u7ed3\u679c\uff1bLLM \u56de\u590d\u5b8c\u6210\u540e\uff0c\u6267\u884c <code>tracer.onStepFinish()<\/code> \u8bb0\u5f55 token \u7528\u91cf\u3002\u8fd9\u6837\u6267\u884c\u4ee3\u7801\u548c\u8bb0\u5f55\u4ee3\u7801\u59cb\u7ec8\u6210\u5bf9\u51fa\u73b0\uff0c\u4e0d\u4f1a\u51fa\u73b0\u67d0\u4e00\u6b65\u6267\u884c\u4e86\u4f46\u6f0f\u6389\u8bb0\u5f55\u7684\u60c5\u51b5\u3002<\/p>\n<p>trace \u8c03\u7528\u5728 Agent Loop \u4e2d\u7684\u4f4d\u7f6e\uff1a<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart TD\n    A[\u7528\u6237\u6d88\u606f\u5230\u8fbe] --&gt; B[\u521b\u5efa Tracer \u5b9e\u4f8b]\n    B --&gt; C[\u8c03\u7528 streamText \u5f00\u59cb\u5faa\u73af]\n    C --&gt; D{\u8bfb\u53d6 fullStream \u4e8b\u4ef6}\n    D --&gt;|step-start| E[&quot;tracer.onStepStart()&quot;]\n    E --&gt; D\n    D --&gt;|tool-call| F[&quot;tracer.onToolCallStart()&quot;]\n    F --&gt; D\n    D --&gt;|tool-result| G[&quot;tracer.onToolResult(toolName, input, success, ...)&quot;]\n    G --&gt; D\n    D --&gt;|step-finish| H[&quot;tracer.onStepFinish(tokenUsage, context, llmResponse)&quot;]\n    H --&gt; D\n    D --&gt;|error| I[&quot;tracer.onError(message)&quot;]\n    I --&gt; J[\u8fd4\u56de error \u4e8b\u4ef6]\n    D --&gt;|\u6d41\u7ed3\u675f| K[&quot;tracer.finish(status)&quot;]\n    K --&gt; L[\u8fd4\u56de done \u4e8b\u4ef6]<\/code><\/pre>\n<p>\u7b80\u5316\u540e\u7684\u4ee3\u7801\uff08\u5b8c\u6574\u5b9e\u73b0\u89c1\u6e90\u7801\uff09\uff1a<\/p>\n<pre><code class=\"typescript\">export async function* agentLoop(params: AgentLoopParams): AsyncGenerator&lt;StreamEvent&gt; {\n  const tracer = new Tracer({ userId, conversationId, provider, model: modelId })\n\n  const result = streamText({ model: llmModel, system: SYSTEM_PROMPT, messages, tools, maxSteps: 15 })\n\n  for await (const part of result.fullStream) {\n    switch (part.type) {\n      case 'step-start':\n        tracer.onStepStart()\n        break\n      case 'tool-call':\n        tracer.onToolCallStart()\n        yield { type: 'tool-call', toolName: part.toolName, input: part.args }\n        break\n      case 'tool-result':\n        tracer.onToolResult(part.toolName, toolInput, success, outputSize, resultObj, errorMsg)\n        yield { type: 'tool-result', toolName: part.toolName, success, outputSize }\n        break\n      case 'step-finish':\n        tracer.onStepFinish(tokenUsage, contextInfo, stepText)\n        break\n      case 'error':\n        tracer.onError(String(part.error))\n        yield { type: 'error', message: String(part.error) }\n        break\n    }\n  }\n\n  const operationId = tracer.finish(status)\n  yield { type: 'done', operationId }\n}<\/code><\/pre>\n<p>\u6bcf\u4e2a <code>case<\/code> \u5206\u652f\u91cc\u90fd\u6709\u4e00\u884c <code>tracer.onXxx()<\/code> \u8c03\u7528\u3002Tracer \u5728\u51fd\u6570\u5165\u53e3\u521b\u5efa\uff0c\u5728\u51fa\u53e3 <code>finish()<\/code>\u3002Agent Loop \u8dd1\u4e86\u591a\u5c11\u6b65\uff0ctrace \u5c31\u8bb0\u4e86\u591a\u5c11\u6b65\u3002<\/p>\n<p>demo \u91cc\u7684 Tracer \u662f\u81ea\u5df1\u5b9e\u73b0\u7684\uff0c\u56e0\u4e3a\u540e\u9762\u7684\u8fdb\u5316\u5f15\u64ce\u9700\u8981\u7528 SQL \u67e5\u8be2 trace \u6570\u636e\uff08\u5206\u6876\u3001\u5339\u914d\u3001\u56de\u626b\uff09\u3002\u5982\u679c\u4f60\u7684\u4ea7\u54c1\u5df2\u7ecf\u63a5\u5165\u4e86 Sentry\u3001PostHog \u7b49\u5e73\u53f0\uff0c\u53ef\u4ee5\u5728 Agent Loop \u7684\u76f8\u540c\u4f4d\u7f6e\u7528\u8fd9\u4e9b\u5e73\u53f0\u7684 API \u505a\u624b\u52a8\u57cb\u70b9\uff0c\u8bb0\u5f55\u7684\u6570\u636e\u5b57\u6bb5\u8ddf\u4e0a\u9762\u8868\u683c\u91cc\u7684\u4e00\u6837\u3002\u4f46\u57cb\u70b9\u6570\u636e\u6700\u7ec8\u8981\u540c\u6b65\u5230\u81ea\u5df1\u7684\u6570\u636e\u5e93\u91cc\uff0c\u6216\u8005\u901a\u8fc7\u5e73\u53f0 API \u5bfc\u51fa\uff0c\u56e0\u4e3a\u8fdb\u5316\u5f15\u64ce\u9700\u8981\u76f4\u63a5\u67e5\u8be2\u8fd9\u4e9b\u6570\u636e\u3002<\/p>\n<h3>\u4e24\u4e2a\u8bbe\u8ba1\u8981\u70b9\uff1a\u5373\u65f6\u5199\u5165\u4e0e token \u8bb0\u5f55<\/h3>\n<p>\u5373\u65f6\u5199\u5165\u3002Agent \u6bcf\u6267\u884c\u5b8c\u4e00\u6b65\uff0cTracer \u7acb\u523b\u628a\u8fd9\u4e00\u6b65\u7684\u6570\u636e\u5199\u5165 SQLite\uff0c\u4e0d\u5728\u5185\u5b58\u4e2d\u7f13\u5b58\u3002\u56e0\u4e3a\u5d29\u6e83\u65f6\u7684\u6570\u636e\u662f\u6700\u6709\u4ef7\u503c\u7684\uff0c\u5047\u8bbe Agent \u5728\u7b2c 5 \u6b65\u6267\u884c\u5d29\u6e83\uff0c\u4f46\u524d 4 \u6b65\u7684\u8bb0\u5f55\u5df2\u7ecf\u5728\u6570\u636e\u5e93\u91cc\u4e86\uff0c\u6240\u4ee5\u5206\u6790\u5d29\u6e83\u539f\u56e0\u65f6\u6709\u4e0a\u4e0b\u6587\u53ef\u67e5\u3002\u5982\u679c\u7f13\u5b58\u5728\u5185\u5b58\u4e2d\u6279\u91cf\u5199\u5165\uff0c\u7a0b\u5e8f\u5d29\u6e83\u6574\u6761 trace \u6570\u636e\u5c31\u4e22\u4e86\u3002demo \u91cc <code>Tracer<\/code> \u7684\u6bcf\u4e2a <code>onXxx()<\/code> \u65b9\u6cd5\u90fd\u4f1a\u7acb\u523b\u8c03\u7528 <code>traceStore.insertStep()<\/code> \u5199\u5165\u6570\u636e\u5e93\u3002SQLite \u5f00\u542f WAL\uff08Write-Ahead Logging\uff09\u6a21\u5f0f\u540e\uff0c\u6bcf\u6b65\u5199\u4e00\u6b21\u4e0d\u4f1a\u6709\u6027\u80fd\u95ee\u9898\u3002<\/p>\n<p>token \u8bb0\u5f55\u3002Tracer \u5728\u6bcf\u4e2a <code>step-finish<\/code> \u4e8b\u4ef6\u91cc\u8bb0\u5f55 <code>promptTokens<\/code>\u3001<code>completionTokens<\/code> \u548c\u7f13\u5b58\u547d\u4e2d\u6570\uff08\u4e0d\u540c\u4f9b\u5e94\u5546\u5b57\u6bb5\u540d\u4e0d\u4e00\u6837\u2014\u2014DeepSeek \u53eb <code>promptCacheHitTokens<\/code>\uff0cAnthropic \u53eb <code>cacheReadInputTokens<\/code>\uff0cOpenAI \u53eb <code>cachedPromptTokens<\/code>\uff09\u3002Operation \u7ed3\u675f\u65f6 <code>Tracer.finish()<\/code> \u6c47\u603b\u6240\u6709 step \u7684 token \u7528\u91cf\u3002\u540e\u9762\u884c\u4e3a\u5206\u6790\u4f1a\u7528\u5230\u8fd9\u4e9b\u6570\u636e\uff0c\u6bd4\u5982\u53d1\u73b0\u201c\u67d0\u7c7b\u64cd\u4f5c\u7684 token \u6d88\u8017\u662f\u5176\u4ed6\u64cd\u4f5c\u7684 10 \u500d\u201d\u3002<\/p>\n<h2>\u9519\u8bef Pattern \u81ea\u52a8\u8bc6\u522b<\/h2>\n<p>Trace \u80fd\u544a\u8bc9\u6211\u4eec\u201c\u51fa\u4e86\u4ec0\u4e48\u9519\u201d\uff0c\u4f46\u4e0d\u80fd\u544a\u8bc9\u6211\u4eec\u201c\u8be5\u600e\u4e48\u5904\u7406\u201d\u3002\u8fd9\u4e00\u8282\u8bb2\u7cfb\u7edf\u600e\u4e48\u81ea\u52a8\u628a\u9519\u8bef\u5206\u7c7b\uff0c\u4ee5\u53ca\u5206\u7c7b\u89c4\u5219\uff08Pattern\uff09\u662f\u600e\u4e48\u751f\u6210\u7684\u3002<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart LR\n    A[\u9519\u8bef\u5199\u5165] --&gt; B{\u5339\u914d\u5df2\u6709 Pattern\uff1f}\n    B --&gt;|\u547d\u4e2d| C[\u81ea\u52a8\u5f52\u7c7b]\n    B --&gt;|\u672a\u547d\u4e2d| D[\u6807\u8bb0\u4e3a\u672a\u5339\u914d]\n    D -.-&gt;|\u5de1\u68c0\u65f6| E[\u5206\u6876]\n    E --&gt; F[LLM \u5206\u6790]\n    F --&gt; G[\u751f\u6210\u65b0 Pattern]\n    G --&gt; H[\u56de\u626b\u5386\u53f2\u9519\u8bef]\n    H --&gt; B<\/code><\/pre>\n<h3>Trace \u91cc\u7684\u9519\u8bef\u8bb0\u5f55<\/h3>\n<p>Trace \u8bb0\u5f55\u7684\u6bcf\u6761\u9519\u8bef\u5305\u542b\u4f9b\u5e94\u5546\u3001\u9519\u8bef\u7c7b\u578b\u3001HTTP \u72b6\u6001\u7801\u3001\u5de5\u5177\u540d\u3001\u9519\u8bef\u6d88\u606f\u8fd9\u51e0\u4e2a\u5b57\u6bb5\u3002\u4e0b\u9762\u662f\u51e0\u79cd\u5178\u578b\u9519\u8bef\uff1a<\/p>\n<p>\u641c\u7d22 API \u9650\u6d41\uff1a<\/p>\n<pre><code>provider=deepseek, error_type=rate_limit, status_code=429, tool_name=null\nmessage: &quot;429 Too Many Requests&quot;<\/code><\/pre>\n<p>SQL \u67e5\u8be2\u4e86\u4e0d\u5b58\u5728\u7684\u8868\uff1a<\/p>\n<pre><code>provider=null, error_type=tool_error, status_code=null, tool_name=dbQuery\nmessage: &quot;no such table: orders&quot;<\/code><\/pre>\n<p>\u6a21\u578b\u8fd4\u56de\u4e86\u65e0\u6cd5\u89e3\u6790\u7684 JSON\uff1a<\/p>\n<pre><code>provider=openai, error_type=invalid_json, status_code=null, tool_name=null\nmessage: &quot;Unexpected token at position 1234&quot;<\/code><\/pre>\n<p>\u4e0a\u4e0b\u6587\u8d85\u51fa token \u9650\u5236\uff1a<\/p>\n<pre><code>provider=deepseek, error_type=context_overflow, status_code=400, tool_name=null\nmessage: &quot;This model's maximum context length is 65536 tokens&quot;<\/code><\/pre>\n<p>\u8fd9\u4e9b\u9519\u8bef\u6570\u636e\u90fd\u5e26\u7740\u4f9b\u5e94\u5546\u3001\u7c7b\u578b\u3001\u72b6\u6001\u7801\u8fd9\u4e9b\u5b57\u6bb5\uff0c\u4f46\u5b57\u6bb5\u672c\u8eab\u53ea\u63cf\u8ff0\u4e86\u201c\u53d1\u751f\u4e86\u4ec0\u4e48\u201d\uff0c\u6ca1\u6709\u56de\u7b54\u201c\u8be5\u600e\u4e48\u5904\u7406\u201d\u3002\u6240\u4ee5\u6211\u4eec\u9700\u8981\u4e00\u6761\u89c4\u5219\u6765\u5224\u65ad\uff1a\u8fd9\u79cd\u9519\u8bef\u5c5e\u4e8e\u54ea\u4e2a\u7c7b\u522b\uff1f\u8fd9\u5c31\u662f Pattern \u7684\u4f5c\u7528\u3002<\/p>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930614\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<h3>Pattern\uff1a\u9519\u8bef\u7684\u5206\u7c7b\u89c4\u5219<\/h3>\n<blockquote><p>\u6e90\u7801\uff1a<code>packages\/server\/src\/evolution\/pattern-matcher.ts<\/code><\/p><\/blockquote>\n<p>Pattern \u662f\u5b58\u5728\u6570\u636e\u5e93\u91cc\u7684\u4e00\u6761\u9519\u8bef\u5206\u7c7b\u89c4\u5219\u3002\u6bcf\u6761 Pattern \u7531\u4e09\u90e8\u5206\u7ec4\u6210\uff1a\u540d\u79f0\u3001\u5339\u914d\u6761\u4ef6\uff08\u54ea\u4e9b\u5b57\u6bb5\u8981\u6ee1\u8db3\u4ec0\u4e48\u503c\uff09\u3001\u5206\u7c7b\uff08\u8fd9\u79cd\u9519\u8bef\u5c5e\u4e8e\u54ea\u4e2a\u7c7b\u522b\uff09\u3002\u5f53\u4e00\u6761\u9519\u8bef\u7684\u5b57\u6bb5\u548c\u67d0\u4e2a Pattern \u7684\u5339\u914d\u6761\u4ef6\u543b\u5408\u65f6\uff0c\u7cfb\u7edf\u5c31\u77e5\u9053\u8fd9\u6761\u9519\u8bef\u5c5e\u4e8e\u54ea\u4e00\u7c7b\uff0c\u4e0d\u9700\u8981\u4eba\u5de5\u5224\u65ad\u3002\u6bd4\u5982\uff1a<\/p>\n<pre><code>Pattern \u793a\u4f8b\uff1a\n  \u540d\u79f0\uff1adeepseek-rate-limit-429\n  \u5206\u7c7b\uff1aprovider_error\uff08\u4f9b\u5e94\u5546\u7684\u95ee\u9898\uff0c\u4e0d\u662f Harness \u80fd\u4fee\u7684\uff09\n  \u5339\u914d\u89c4\u5219\uff1aprovider = deepseek AND statusCode = 429<\/code><\/pre>\n<p>\u4e0d\u540c\u7684\u9519\u8bef\u5206\u7c7b\u5bf9\u5e94\u56db\u79cd\u5904\u7406\u65b9\u5f0f\uff1a<code>user_error<\/code>\uff08\u7528\u6237\u7684\u95ee\u9898\uff0c\u6bd4\u5982 API Key \u8fc7\u671f\uff09\u4e0d\u7528\u6539\u4ee3\u7801\uff1b<code>provider_error<\/code>\uff08\u4f9b\u5e94\u5546\u7684\u95ee\u9898\uff0c\u6bd4\u5982\u9650\u6d41\uff09\u4e0d\u7528\u6539\u4ee3\u7801\uff0c\u4f46\u53ef\u80fd\u8981\u52a0\u91cd\u8bd5\uff1b<code>harness_bug<\/code>\uff08\u6211\u4eec\u81ea\u5df1\u7684 bug\uff09\u9700\u8981\u4fee\u4ee3\u7801\u2014\u2014\u540e\u9762\u201c\u81ea\u52a8\u4fee\u590d\u201d\u90a3\u8282\u5904\u7406\u7684\u5c31\u662f\u8fd9\u4e00\u7c7b\uff1b<code>ignore<\/code>\uff08\u7ba1\u7406\u5458\u8ba4\u4e3a\u4e0d\u9700\u8981\u5904\u7406\u7684\uff09\u76f4\u63a5\u8df3\u8fc7\uff0c\u81ea\u52a8\u4fee\u590d\u65f6\u4e5f\u4f1a\u5ffd\u7565\u3002<\/p>\n<p>\u6bcf\u6761\u9519\u8bef\u5199\u5165\u6570\u636e\u5e93\u65f6\uff0c\u7cfb\u7edf\u4f1a\u8c03 <code>matchError()<\/code> \u5c1d\u8bd5\u5339\u914d\u73b0\u6709 Pattern\u2014\u2014\u9010\u4e00\u5bf9\u6bd4 provider\u3001errorType\u3001statusCode\u3001toolName \u662f\u5426\u7b26\u5408\u6761\u4ef6\uff0c\u6700\u540e\u7528\u6b63\u5219\u68c0\u67e5\u9519\u8bef\u6d88\u606f\uff0c\u5168\u90e8\u5339\u914d\u4e0a\u624d\u7b97\u547d\u4e2d\u3002\u5339\u914d\u4e0d\u4e0a\u7684\u9519\u8bef\u7559\u5728\u90a3\u91cc\uff0c\u7b49\u4e0b\u4e00\u8f6e\u5de1\u68c0\u65f6\u88ab\u5206\u6876\u3001\u4ea4\u7ed9 LLM \u5206\u6790\u3001\u751f\u6210\u65b0 Pattern\u3002<\/p>\n<pre><code class=\"typescript\">export function matchError(error: ErrorToMatch): Pattern | null {\n  const patterns = loadPatterns()\n  for (const pattern of patterns) {\n    const rule = pattern.matchRule\n    if (rule.provider &amp;&amp; rule.provider !== '*' &amp;&amp; rule.provider !== error.provider) continue\n    if (rule.errorType &amp;&amp; rule.errorType !== error.errorType) continue\n    if (rule.statusCode != null &amp;&amp; rule.statusCode !== error.statusCode) continue\n    if (rule.toolName &amp;&amp; rule.toolName !== error.toolName) continue\n    if (rule.messageRegex) {\n      if (!new RegExp(rule.messageRegex, 'i').test(error.message)) continue\n    }\n    \/\/ \u547d\u4e2d\uff1a\u66f4\u65b0\u8ba1\u6570\n    return pattern\n  }\n  return null\n}<\/code><\/pre>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930615\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<h3>LLM \u5de1\u68c0\uff1a\u751f\u6210\u65b0 Pattern<\/h3>\n<blockquote><p>\u6e90\u7801\uff1a<code>packages\/server\/src\/evolution\/error-bucketer.ts<\/code>\u3001<code>packages\/server\/src\/evolution\/inspector.ts<\/code><\/p><\/blockquote>\n<p>\u4e0a\u4e00\u8282\u8bf4\u5230\uff0c\u5339\u914d\u4e0d\u4e0a\u7684\u9519\u8bef\u4f1a\u7559\u5728\u90a3\u91cc\u7b49\u5de1\u68c0\u5904\u7406\u3002\u7cfb\u7edf\u521a\u4e0a\u7ebf\u65f6 Pattern \u5e93\u662f\u7a7a\u7684\uff0c\u6240\u6709\u7684\u9519\u8bef\u90fd\u662f\u672a\u5339\u914d\u7684\u3002\u5de1\u68c0\u5c31\u662f\u7528 LLM \u6765\u5206\u6790\u8fd9\u4e9b\u9519\u8bef\u3001\u751f\u6210\u65b0 Pattern \u7684\u8fc7\u7a0b\u3002<\/p>\n<h4>\u5206\u6876<\/h4>\n<p>LLM \u4e0d\u9700\u8981\u9010\u6761\u770b\u9519\u8bef\u3002\u540c\u4e00\u79cd\u9519\u8bef\u51fa\u73b0 15 \u6b21\u548c\u51fa\u73b0 1 \u6b21\uff0c\u6839\u56e0\u90fd\u662f\u4e00\u6837\u7684\u3002\u6240\u4ee5\u5de1\u68c0\u5148\u628a\u672a\u5339\u914d\u7684\u9519\u8bef\u6309\u7c7b\u578b\u5f52\u7ec4\u2014\u2014<code>bucketErrors()<\/code> \u6309\u4f9b\u5e94\u5546\u3001\u9519\u8bef\u7c7b\u578b\u3001HTTP \u72b6\u6001\u7801\u3001\u5de5\u5177\u540d\u3001\u9519\u8bef\u6d88\u606f\u4e94\u4e2a\u7ef4\u5ea6\u505a GROUP BY \u5206\u7ec4\uff0c\u5e76\u8fd4\u56de\u6bcf\u4e2a\u6876\u7684\u9519\u8bef\u6761\u6570\uff1a<\/p>\n<pre><code class=\"typescript\">const rows = db.prepare(\n  `SELECT e.provider, e.error_type, e.status_code, e.tool_name, e.message, COUNT(*) as count\n   FROM errors e ${where}\n   GROUP BY e.provider, e.error_type, e.status_code, e.tool_name, e.message\n   ORDER BY count DESC`\n).all(...params)<\/code><\/pre>\n<p>\u5206\u6876\u7ed3\u679c\u793a\u4f8b\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6876<\/th>\n<th>\u6570\u91cf<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>deepseek \u00d7 rate_limit \u00d7 429 \u00d7 null<\/td>\n<td>15<\/td>\n<\/tr>\n<tr>\n<td>null \u00d7 tool_error \u00d7 null \u00d7 dbQuery<\/td>\n<td>8<\/td>\n<\/tr>\n<tr>\n<td>openai \u00d7 invalid_json \u00d7 null \u00d7 null<\/td>\n<td>3<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6bd4\u5982\u8868\u91cc\u7684 15 \u6761 deepseek \u9650\u6d41\u9519\u8bef\uff0c\u5206\u6876\u540e\u5c31\u662f\u4e00\u4e2a\u6876\uff0cLLM \u53ea\u9700\u8981\u5206\u6790\u4e00\u6b21\u3002<code>bucketErrors()<\/code> \u7684 <code>unmatched<\/code> \u53c2\u6570\u53ea\u8fd4\u56de\u8fd8\u6ca1\u88ab Pattern \u8986\u76d6\u7684\u9519\u8bef\u6876\uff0c\u5df2\u7ecf\u6709 Pattern \u7684\u4e0d\u9700\u8981\u518d\u5206\u6790\u3002<\/p>\n<h4>LLM \u5206\u6790<\/h4>\n<p>\u5206\u6876\u4e4b\u540e\uff0c\u7cfb\u7edf\u628a\u672a\u5339\u914d\u7684\u9519\u8bef\u6876\u7684\u6c47\u603b\u4fe1\u606f\u2014\u2014\u4f9b\u5e94\u5546\u3001\u9519\u8bef\u7c7b\u578b\u3001\u72b6\u6001\u7801\u3001\u5de5\u5177\u540d\u3001\u9519\u8bef\u6d88\u606f\u3001\u51fa\u73b0\u6b21\u6570\u2014\u2014\u53d1\u7ed9 LLM\uff0c\u8ba9\u5b83\u5224\u65ad\u6bcf\u4e2a\u6876\u7684\u6839\u56e0\u5e76\u751f\u6210\u5bf9\u5e94\u7684 Pattern\u3002<\/p>\n<p>LLM \u8fd4\u56de\u7684\u7ed3\u679c\u8981\u76f4\u63a5\u5199\u5165\u6570\u636e\u5e93\uff0c\u6240\u4ee5\u683c\u5f0f\u4e0d\u80fd\u968f\u610f\u3002\u8fd9\u91cc\u7528 Zod \u5b9a\u4e49\u4e00\u4e2a Schema\uff0c\u901a\u8fc7 Vercel AI SDK \u7684 <code>generateObject<\/code> \u4f20\u7ed9\u6a21\u578b\uff0c\u6a21\u578b\u5fc5\u987b\u6309\u8fd9\u4e2a\u7ed3\u6784\u8fd4\u56de JSON\uff0c\u683c\u5f0f\u4e0d\u5bf9\u65f6 SDK \u4f1a\u81ea\u52a8\u91cd\u8bd5\uff1a<\/p>\n<pre><code class=\"typescript\">const PatternSuggestionSchema = z.object({\n  patterns: z.array(z.object({\n    name: z.string().describe('Human-readable pattern name, e.g. deepseek-rate-limit-429'),\n    category: z.enum(['user_error', 'provider_error', 'harness_bug', 'ignore']),\n    errorType: z.string(),\n    matchRule: z.object({\n      statusCode: z.number().nullable().optional(),\n      provider: z.string().optional(),\n      toolName: z.string().nullable().optional(),\n      messageRegex: z.string().optional(),\n    }),\n    reasoning: z.string().describe('Why you classified it this way'),\n  })),\n  bugs: z.array(z.object({\n    title: z.string(),\n    description: z.string(),\n    rootCause: z.string(),\n    suggestedFix: z.string(),\n    severity: z.enum(['low', 'medium', 'high']),\n  })),\n  summary: z.string(),\n})\n\nconst result = await generateObject({\n  model,\n  schema: PatternSuggestionSchema,\n  prompt,  \/\/ \u5305\u542b\u672a\u5339\u914d\u9519\u8bef\u6876\u7684\u8be6\u7ec6\u4fe1\u606f\n  abortSignal: AbortSignal.timeout(120_000),\n})<\/code><\/pre>\n<p>\u4e0a\u9762\u4ee3\u7801\u4e2d\u7684 <code>patterns<\/code> \u5b57\u6bb5\u662f LLM \u4e3a\u6bcf\u4e2a\u9519\u8bef\u6876\u751f\u6210\u7684 Pattern\u2014\u2014\u540d\u79f0\u3001\u5206\u7c7b\u3001\u5339\u914d\u89c4\u5219\u548c\u5206\u7c7b\u7406\u7531\u3002\u800c <code>bugs<\/code> \u5b57\u6bb5\u662f <code>harness_bug<\/code> \u7684\u8be6\u7ec6\u5206\u6790\uff1a\u6839\u56e0\u3001\u4fee\u590d\u5efa\u8bae\u3001\u4e25\u91cd\u7a0b\u5ea6\u2014\u2014\u540e\u9762\u201c\u81ea\u52a8\u4fee\u590d\u201d\u90a3\u8282\u4f1a\u7528\u5230\u8fd9\u4e9b\u4fe1\u606f\u3002<code>summary<\/code> \u5b57\u6bb5\u662f\u672c\u8f6e\u5de1\u68c0\u7684\u603b\u7ed3\u3002<\/p>\n<p>LLM \u8fd4\u56de\u7ed3\u679c\u540e\uff0c<code>applyFixes()<\/code> \u628a\u65b0 Pattern \u5199\u5165\u6570\u636e\u5e93\uff0c\u5e76\u7acb\u523b\u56de\u626b\uff08Backfill\uff09\u6240\u6709\u5386\u53f2\u9519\u8bef\u2014\u2014\u627e\u5230\u6240\u6709 <code>pattern_id<\/code> \u4e3a\u7a7a\u7684\u8bb0\u5f55\uff0c\u7528\u65b0 Pattern \u9010\u4e00\u5bf9\u6bd4\uff0c\u80fd\u5339\u914d\u4e0a\u7684\u5c31\u6807\u8bb0\u3002\u4e4b\u524d\u6ca1\u88ab\u8bc6\u522b\u7684\u9519\u8bef\uff0c\u73b0\u5728\u6709\u4e86 Pattern \u4e4b\u540e\u5c31\u81ea\u52a8\u5f52\u7c7b\u4e86\u3002<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart TD\n    A[\u672a\u5339\u914d\u7684\u9519\u8bef\u6876] --&gt; B[\u6784\u9020 prompt]\n    B --&gt; C[&quot;generateObject() + Zod Schema&quot;]\n    C --&gt; D[\u7c7b\u578b\u5b89\u5168\u7684\u7ed3\u6784\u5316\u7ed3\u679c]\n    D --&gt; E[\u65b0 Pattern \u5199\u5165\u6570\u636e\u5e93]\n    E --&gt; F[\u56de\u626b\u5386\u53f2\u9519\u8bef]\n    F --&gt; G[\u4e4b\u524d\u672a\u5339\u914d\u7684\u9519\u8bef\u88ab\u81ea\u52a8\u5f52\u7c7b]<\/code><\/pre>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930616\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<p>\u8bfb\u8005\u53ef\u4ee5\u7528 demo \u81ea\u5df1\u9a8c\u8bc1\u3002\u6267\u884c <code>pnpm simulate --mock<\/code> \u6ce8\u5165\u6a21\u62df\u8bf7\u6c42\u540e\uff0c\u5728\u7ba1\u7406\u9762\u677f\u91cc\u624b\u52a8\u6267\u884c\u4e00\u6b21\u5de1\u68c0\u3002\u5de1\u68c0\u4f1a\u53d6\u51fa\u6240\u6709\u672a\u5339\u914d\u7684\u9519\u8bef\u6876\uff0cLLM \u4e3a\u6bcf\u4e2a\u6876\u751f\u6210 Pattern\uff0c\u7136\u540e\u7528\u65b0 Pattern \u56de\u626b\u5386\u53f2\u9519\u8bef\u3002\u5de1\u68c0\u6267\u884c\u5b8c\u6bd5\u540e\uff0c\u4e4b\u524d\u79ef\u7d2f\u7684\u672a\u5339\u914d\u9519\u8bef\u57fa\u672c\u90fd\u80fd\u88ab\u8986\u76d6\u3002<\/p>\n<h2>\u81ea\u52a8\u4fee\u590d\uff1a\u8ba9\u7cfb\u7edf\u81ea\u5df1\u63d0 PR<\/h2>\n<p>\u4e0a\u4e00\u8282\u8bc6\u522b\u51fa\u4e86 <code>harness_bug<\/code>\u3002\u8fd9\u4e00\u8282\u8ba9\u7cfb\u7edf\u81ea\u52a8\u751f\u6210\u4fee\u590d\u4ee3\u7801\u3001\u63d0\u4ea4 PR\u3002\u4ece\u5b9a\u4f4d\u6587\u4ef6\u5230\u521b\u5efa PR\uff0c\u4e2d\u95f4\u4e0d\u9700\u8981\u5f00\u53d1\u8005\u53c2\u4e0e\uff0c\u4f46\u5408\u5e76 PR \u8fd8\u662f\u8981\u5f00\u53d1\u8005\u5ba1\u6838\u3002<\/p>\n<h3>\u4fee\u590d\u6d41\u7a0b<\/h3>\n<blockquote><p>\u6e90\u7801\uff1a<code>packages\/server\/src\/evolution\/auto-pr.ts<\/code><\/p><\/blockquote>\n<p>\u81ea\u52a8\u4fee\u590d\u7528\u7684\u662f\u548c\u804a\u5929\u4e00\u6837\u7684 Agent Loop \u67b6\u6784\u2014\u2014LLM \u901a\u8fc7\u5de5\u5177\u81ea\u4e3b\u641c\u7d22\u6587\u4ef6\u3001\u8bfb\u53d6\u4ee3\u7801\u3001\u4fee\u6539\u4ee3\u7801\u3002<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart TD\n    A[harness_bug Pattern] --&gt; B[\u521b\u5efa git \u5206\u652f]\n    B --&gt; C[\u542f\u52a8\u4fee\u590d Agent]\n    C --&gt; D{Agent \u751f\u6210\u4e86\u6709\u6548\u4fee\u6539\uff1f}\n    D --&gt;|\u662f| E[git commit]\n    D --&gt;|\u5426| F[\u56de\u6eda\u5206\u652f]\n    E --&gt; G{\u6709\u8fdc\u7a0b\u4ed3\u5e93\uff1f}\n    G --&gt;|\u662f| H[push + \u521b\u5efa PR]\n    G --&gt;|\u5426| I[\u4ec5\u672c\u5730\u5206\u652f]\n    H --&gt; J[\u66f4\u65b0 fix_status = pr_created]\n    I --&gt; J<\/code><\/pre>\n<p>\u4e0a\u4e00\u8282\u7684\u5de1\u68c0\u4f1a\u628a Harness \u81ea\u8eab\u7684 bug \u6807\u8bb0\u4e3a <code>harness_bug<\/code> Pattern\uff0c\u5e76\u5c06\u72b6\u6001\u66f4\u65b0\u4e3a <code>fix_status = 'unfixed'<\/code>\u3002\u81ea\u52a8\u4fee\u590d\u4f1a\u9010\u4e2a\u5904\u7406\u8fd9\u4e9b\u672a\u4fee\u590d\u7684 Pattern\uff0c\u4e3a\u6bcf\u4e2a Pattern \u751f\u6210\u4fee\u590d\u4ee3\u7801\u5e76\u63d0\u4ea4 PR\u3002\u9664\u4e86 <code>harness_bug<\/code>\uff0c\u88ab\u6807\u4e3a <code>critical<\/code> \u7684\u4e0d\u5065\u5eb7\u884c\u4e3a\uff08\u4e0b\u4e00\u8282\u8bb2\uff09\u4e5f\u4f1a\u4e00\u5e76\u5904\u7406\u3002bug \u4fee\u590d\u7684\u5206\u652f\u7528 <code>fix\/<\/code> \u524d\u7f00\uff0c\u884c\u4e3a\u4f18\u5316\u7528 <code>improve\/<\/code>\u3002<\/p>\n<p>\u5bf9\u6bcf\u4e2a\u5f85\u4fee\u590d\u76ee\u6807\uff0c\u7cfb\u7edf\u5148\u521b\u5efa\u4e00\u4e2a git \u5206\u652f\uff0c\u7136\u540e\u542f\u52a8\u4e00\u4e2a\u4fee\u590d Agent\u3002\u8fd9\u4e2a Agent \u914d\u5907\u4e86\u6587\u4ef6\u641c\u7d22\uff08<code>glob<\/code>\u3001<code>grep<\/code>\uff09\u3001\u8bfb\u5199\uff08<code>readFile<\/code>\u3001<code>editFile<\/code>\u3001<code>writeFile<\/code>\uff09\u548c\u63d0\u4ea4\uff08<code>submitFix<\/code>\uff09\u7b49\u5de5\u5177\u3002Agent \u62ff\u5230\u9519\u8bef\u4fe1\u606f\u540e\uff0c\u81ea\u5df1\u641c\u7d22\u76f8\u5173\u6587\u4ef6\u3001\u8bfb\u53d6\u4ee3\u7801\u3001\u5b9a\u4f4d\u95ee\u9898\u3001\u5e94\u7528\u4fee\u6539\u2014\u2014\u548c\u804a\u5929 Agent \u4e00\u6837\uff0c\u90fd\u662f LLM \u901a\u8fc7\u5de5\u5177\u81ea\u4e3b\u5b8c\u6210\u7684\u3002<\/p>\n<p>Agent \u5b8c\u6210\u4fee\u6539\u540e\uff0c\u7cfb\u7edf\u4f1a\u63d0\u4ea4\u6240\u6709\u4fee\u6539\u8fc7\u7684\u6587\u4ef6\u3002\u5982\u679c\u5df2\u7ecf\u6709\u8fdc\u7a0b\u4ed3\u5e93\uff0c\u4f1a\u81ea\u52a8 push \u5e76\u521b\u5efa PR\uff1b\u5426\u5219\u53ea\u5728\u672c\u5730\u5efa\u5206\u652f\u3002\u6700\u540e\u66f4\u65b0 <code>fix_status<\/code> \u72b6\u6001\u4e3a <code>pr_created<\/code>\u3002<\/p>\n<p>\u4e0b\u9762\u662f\u4e00\u4e9b\u5173\u4e8e\u81ea\u52a8\u4fee\u590d\u8fd0\u884c\u8fc7\u7a0b\u7684\u622a\u56fe\u4ee5\u53ca\u81ea\u52a8\u4fee\u590d\u63d0\u4ea4\u7684 PR \u622a\u56fe\uff1a<br \/><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930617\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930618\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930619\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<h3>\u5b89\u5168\u8fb9\u754c\uff1a\u4e3a\u4ec0\u4e48\u4e0d\u8ba9\u5b83\u81ea\u5df1\u5408\u5e76\u4ee3\u7801<\/h3>\n<p>\u81ea\u52a8\u4fee\u590d\u53ea\u751f\u6210 PR\uff0c\u5408\u5e76\u4ecd\u7136\u9700\u8981\u5f00\u53d1\u8005\u5ba1\u6838\u3002LLM \u751f\u6210\u7684\u4ee3\u7801\u53ef\u80fd\u5f15\u5165\u65b0\u7684\u95ee\u9898\uff0c\u5728\u6ca1\u6709\u81ea\u52a8\u5316\u6d4b\u8bd5\u8986\u76d6\u4fee\u6539\u884c\u4e3a\u7684\u60c5\u51b5\u4e0b\uff0c\u8ba9\u7cfb\u7edf\u81ea\u5df1\u5408\u5e76\u4ee3\u7801\u98ce\u9669\u592a\u5927\u3002\u5982\u679c\u60f3\u8df3\u8fc7\u5f00\u53d1\u8005\u5ba1\u6838\uff0c\u81f3\u5c11\u9700\u8981\u4e24\u6837\u4e1c\u897f\uff1a\u4e00\u5957\u8986\u76d6\u4e86\u4fee\u6539\u884c\u4e3a\u7684\u81ea\u52a8\u5316\u6d4b\u8bd5\uff0c\u4ee5\u53ca\u4e00\u4e2a\u80fd\u590d\u73b0\u89e6\u53d1 bug \u7684\u64cd\u4f5c\u7684\u91cd\u653e\u5f15\u64ce\u3002\u540e\u9762\u201c\u843d\u5730\u6307\u5357\u201d\u90a3\u8282\u4f1a\u8bb2\u91cd\u653e\u5f15\u64ce\u7684\u8bbe\u8ba1\u3002<\/p>\n<h2>\u884c\u4e3a\u5206\u6790\uff1a\u6ca1\u62a5\u9519 \u2260 \u6ca1\u95ee\u9898<\/h2>\n<p>\u524d\u9762\u51e0\u8282\u89e3\u51b3\u7684\u90fd\u662f\u201c\u51fa\u4e86\u4ec0\u4e48\u9519\u201d\u3002\u8fd9\u4e00\u8282\u770b\u53e6\u4e00\u7c7b\u95ee\u9898\u2014\u2014\u201c\u6ca1\u62a5\u9519\uff0c\u4f46\u4e0d\u592a\u5bf9\u52b2\u201d\u3002<\/p>\n<h3>\u9519\u8bef\u6a21\u5f0f vs \u884c\u4e3a\u6a21\u5f0f<\/h3>\n<p>\u524d\u9762\u51e0\u8282\u5206\u6790\u7684\u662f\u5b9e\u9645\u62a5\u9519\u7684\u8bb0\u5f55\u2014\u2014\u9519\u8bef Pattern\u3002\u884c\u4e3a\u5206\u6790\u7684\u5bf9\u8c61\u4e0d\u540c\uff0c\u5b83\u770b\u6240\u6709\u64cd\u4f5c\uff0c\u5305\u62ec\u6210\u529f\u7684\u3002\u4e24\u8005\u7684\u533a\u522b\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th> <\/th>\n<th>\u9519\u8bef\u6a21\u5f0f<\/th>\n<th>\u884c\u4e3a\u6a21\u5f0f<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u5206\u6790\u5bf9\u8c61<\/td>\n<td>\u5b9e\u9645\u62a5\u9519\u7684\u8bb0\u5f55<\/td>\n<td>\u6240\u6709\u64cd\u4f5c\uff0c\u5305\u62ec\u6210\u529f\u7684<\/td>\n<\/tr>\n<tr>\n<td>\u53d1\u73b0\u4ec0\u4e48<\/td>\n<td>\u7cfb\u7edf\u5d29\u6e83\u6216\u5f02\u5e38<\/td>\n<td>\u6ca1\u5d29\u4f46\u6548\u7387\u4f4e\u3001\u8d28\u91cf\u5dee<\/td>\n<\/tr>\n<tr>\n<td>\u4e3e\u4f8b<\/td>\n<td>\u641c\u7d22 API 429 \u9650\u6d41\u3001SQL \u8bed\u6cd5\u9519\u8bef<\/td>\n<td>\u7b80\u5355\u95ee\u9898\u8c03\u7528\u4e86 8 \u6b21\u5de5\u5177\u3001\u641c\u7d22\u7c7b\u64cd\u4f5c token \u6d88\u8017\u8fc7\u9ad8<\/td>\n<\/tr>\n<tr>\n<td>\u5206\u7c7b\u65b9\u5f0f<\/td>\n<td>\u56db\u4e2a\u56fa\u5b9a\u7c7b\u522b\uff08user_error \/ provider_error \/ harness_bug \/ ignore\uff09<\/td>\n<td>LLM \u6309\u7528\u6237\u610f\u56fe\u52a8\u6001\u5206\u7ec4<\/td>\n<\/tr>\n<tr>\n<td>\u6700\u7ec8\u52a8\u4f5c<\/td>\n<td>harness_bug \u2192 \u81ea\u52a8\u751f\u6210\u4fee\u590d PR<\/td>\n<td>\u751f\u6210\u6539\u8fdb\u5efa\u8bae\uff0c\u4e25\u91cd\u7684\u4e5f\u81ea\u52a8\u751f\u6210\u4fee\u590d PR<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u4ec0\u4e48\u7b97\u201c\u4e0d\u5065\u5eb7\u201d<\/h3>\n<p>\u7528\u6237\u95ee\u4e86\u4e2a\u7b80\u5355\u95ee\u9898\uff08\u201c\u4ecb\u7ecd\u4e00\u4e0b\u4f60\u81ea\u5df1\u201d\uff09\uff0cAgent \u5374\u8c03\u7528\u4e86 3 \u6b21\u5de5\u5177\u3001\u82b1\u4e86 15 \u79d2\u624d\u56de\u7b54\u3002\u72b6\u6001\u662f\u201c\u6210\u529f\u201d\uff0c\u4f46\u6548\u7387\u660e\u663e\u4e0d\u5bf9\u3002\u53ef\u80fd\u662f system prompt \u5bf9\u5de5\u5177\u4f7f\u7528\u7684\u5f15\u5bfc\u4e0d\u591f\u7cbe\u786e\uff0c\u4e0d\u9700\u8981\u5de5\u5177\u7684\u65f6\u5019\u4e5f\u53bb\u641c\u7d22\u4e86\u3002<\/p>\n<p>\u6570\u636e\u5e93\u67e5\u8be2\u7c7b\u64cd\u4f5c\u7684\u6210\u529f\u7387\u53ea\u6709 65%\u3002Agent \u5de5\u5177\u5931\u8d25\u540e\u4f1a\u6362\u79cd\u65b9\u5f0f\u56de\u7b54\uff0c\u6240\u4ee5\u4e0d\u4f1a\u62a5\u9519\uff0c\u4f46\u4e09\u5206\u4e4b\u4e00\u7684\u67e5\u8be2\u7ed3\u679c\u662f\u4e0d\u6b63\u786e\u7684\u3002\u53ef\u80fd\u662f dbQuery \u5de5\u5177\u7684 Schema \u63cf\u8ff0\u91cc\u7f3a\u5c11\u8868\u7ed3\u6784\u4fe1\u606f\uff0c\u6a21\u578b\u5199\u51fa\u4e86\u9519\u8bef\u7684 SQL\u3002<\/p>\n<p>\u641c\u7d22\u7814\u7a76\u7c7b\u64cd\u4f5c\u7684\u5e73\u5747 token \u6d88\u8017\u662f\u5176\u4ed6\u64cd\u4f5c\u7684 10 \u500d\u3002\u8981\u4e48\u662f webFetch \u8fd4\u56de\u7684\u9875\u9762\u6ca1\u505a\u622a\u65ad\uff0c\u8981\u4e48\u662f Agent \u641c\u5b8c\u4e4b\u540e\u6293\u4e86\u592a\u591a\u4e0d\u76f8\u5173\u7684\u9875\u9762\u3002<\/p>\n<h3>\u5206\u6790\u6d41\u7a0b<\/h3>\n<blockquote><p>\u6e90\u7801\uff1a<code>packages\/server\/src\/evolution\/behavior-analyzer.ts<\/code><\/p><\/blockquote>\n<p>\u8981\u53d1\u73b0\u4e0a\u9762\u8fd9\u7c7b\u95ee\u9898\uff0c\u9700\u8981\u5148\u5bf9\u64cd\u4f5c\u6570\u636e\u505a\u805a\u7c7b\u5206\u6790\u3002\u7cfb\u7edf\u4ece\u6570\u636e\u5e93\u52a0\u8f7d\u6700\u8fd1 100 \u6761 operation\uff08\u4e00\u6761 operation \u5bf9\u5e94\u4e00\u6b21\u5b8c\u6574\u7684\u7528\u6237\u8bf7\u6c42\uff0c\u5305\u542b\u591a\u4e2a step\uff0c\u524d\u9762 Tracing \u7ae0\u8282\u4ecb\u7ecd\u8fc7\uff09\u7684\u6458\u8981\uff0c\u628a\u6bcf\u6761\u683c\u5f0f\u5316\u6210\u4e00\u884c\u6587\u672c\uff0c\u5168\u90e8\u62fc\u8fdb prompt \u53d1\u7ed9 LLM\uff1a<\/p>\n<pre><code>[0] \u201c\u5e2e\u6211\u7814\u7a76\u4e00\u4e0b Hono \u6846\u67b6\u201d \u2192 tools: webSearch\u2192webFetch\u2192webFetch \u2192 success (5 steps, 12.3s)\n[1] \u201c\u54ea\u4e2a\u6b4c\u624b\u4e13\u8f91\u6700\u591a\u201d \u2192 tools: dbQuery \u2192 success (3 steps, 2.1s)\n[2] \u201c\u603b\u7ed3\u4e00\u4e0b\u8fd9\u7bc7\u6587\u7ae0\u201d \u2192 tools: webFetch \u2192 success (3 steps, 4.5s)\n[3] \u201c\u67e5\u4e00\u4e0b\u9500\u91cf\u524d10\u7684\u4e13\u8f91\u201d \u2192 tools: dbQuery \u2192 error (4 steps, 3.2s)\n...<\/code><\/pre>\n<p>\u6bcf\u884c\u6587\u672c\u5305\u542b\u7528\u6237\u6d88\u606f\u3001\u5de5\u5177\u8c03\u7528\u94fe\u3001\u72b6\u6001\u3001\u6b65\u6570\u548c\u8017\u65f6\u3002LLM \u6839\u636e\u7528\u6237\u610f\u56fe\u548c\u5de5\u5177\u4f7f\u7528\u6a21\u5f0f\u5bf9\u8fd9\u4e9b operation \u505a\u5206\u7ec4\uff0c\u5e76\u8fd4\u56de\u6bcf\u7ec4\u7684\u540d\u79f0\u3001\u63cf\u8ff0\u3001\u4ee3\u8868\u6027\u5de5\u5177\u5e8f\u5217\uff0c\u4ee5\u53ca\u5c5e\u4e8e\u8be5\u7ec4\u7684\u7f16\u53f7\u5217\u8868\u3002\u6bd4\u5982\u4e0a\u9762\u7684 <code>[0]<\/code> \u548c <code>[2]<\/code> \u90fd\u6d89\u53ca\u7f51\u7edc\u641c\u7d22\u548c\u5185\u5bb9\u6293\u53d6\uff0c\u4f1a\u88ab\u5f52\u4e3a\u201c\u7f51\u7edc\u8c03\u7814 + \u603b\u7ed3\u201d\uff1b<code>[1]<\/code> \u548c <code>[3]<\/code> \u90fd\u53ea\u7528\u4e86 <code>dbQuery<\/code>\uff0c\u4f1a\u88ab\u5f52\u4e3a\u201c\u6570\u636e\u5e93\u67e5\u8be2\u201d\u3002\u5206\u7ec4\u6ca1\u6709\u9884\u5b9a\u4e49\u7684\u7c7b\u522b\uff0cLLM \u6839\u636e\u5b9e\u9645\u6570\u636e\u81ea\u884c\u51b3\u5b9a\u3002\u8f93\u51fa\u683c\u5f0f\u7528 <code>generateObject<\/code> + Zod Schema \u7ea6\u675f\u3002<\/p>\n<p>demo \u91cc\u53d6 100 \u6761\u662f\u56e0\u4e3a\u6240\u6709\u6458\u8981\u8981\u4e00\u6b21\u6027\u653e\u8fdb prompt\uff0c\u592a\u591a\u4f1a\u8d85\u51fa\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002\u751f\u4ea7\u73af\u5883\u5e94\u8be5\u6539\u6210\u6309\u65f6\u95f4\u7a97\u53e3\u589e\u91cf\u5206\u6790\u3002\u4e3a\u4e86\u907f\u514d\u91cd\u590d\u5206\u6790\uff0c\u7cfb\u7edf\u4f1a\u8bb0\u5f55\u4e0a\u4e00\u8f6e\u7684 operation \u96c6\u5408\u7b7e\u540d\uff0c\u6ca1\u6709\u65b0\u589e operation \u5c31\u8df3\u8fc7\u3002<\/p>\n<p>\u5206\u7ec4\u4e4b\u540e\u662f\u6253\u5206\u3002\u4e0d\u540c\u7c7b\u578b\u7684\u4efb\u52a1\u6027\u80fd\u5dee\u5f02\u5f88\u5927\u2014\u2014\u6d89\u53ca <code>webSearch<\/code>\/<code>webFetch<\/code> \u7684\u7f51\u7edc\u8c03\u7814\u5929\u7136\u6bd4\u672c\u5730\u7684 <code>dbQuery<\/code> \u6162\u5f97\u591a\uff0c\u7528\u540c\u4e00\u5957\u9608\u503c\u5bf9\u4e24\u8005\u6253\u5206\u4e0d\u5408\u7406\u3002\u6240\u4ee5\u7cfb\u7edf\u6839\u636e\u884c\u4e3a\u7ec4\u7684\u5de5\u5177\u5e8f\u5217\u81ea\u52a8\u9009\u62e9\u5bf9\u5e94\u7684\u9608\u503c\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u7ef4\u5ea6<\/th>\n<th>\u8f7b\u91cf\u4efb\u52a1\uff08\u672c\u5730\u5de5\u5177\uff09<\/th>\n<th>\u91cd\u5ea6\u4efb\u52a1\uff08\u542b\u7f51\u7edc\u5de5\u5177\uff09<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u5355\u6b21\u5e73\u5747\u8017\u65f6<\/td>\n<td>\u2264 15 \u79d2<\/td>\n<td>\u2264 60 \u79d2<\/td>\n<\/tr>\n<tr>\n<td>\u5355\u6b21\u5e73\u5747\u6b65\u6570<\/td>\n<td>\u2264 10<\/td>\n<td>\u2264 20<\/td>\n<\/tr>\n<tr>\n<td>\u5355\u6b21\u5e73\u5747 token<\/td>\n<td>\u2264 50k<\/td>\n<td>\u2264 150k<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6210\u529f\u7387\uff08\u2265 80%\uff09\u548c\u5de5\u5177\u9519\u8bef\u7387\uff08\u2264 20%\uff09\u4e0d\u533a\u5206\u4efb\u52a1\u7c7b\u578b\uff0c\u8fd9\u4e24\u4e2a\u6307\u6807\u8ddf\u590d\u6742\u5ea6\u65e0\u5173\u3002\u6bcf\u4e2a\u7ef4\u5ea6 0.2 \u5206\uff0c\u6ee1\u5206 1.0\u3002<\/p>\n<pre><code class=\"typescript\">const HEALTH_THRESHOLDS = {\n  minSuccessRate: 0.8,\n  maxToolErrorRate: 0.2,\n  light: { maxAvgDuration: 15_000, maxAvgSteps: 10, maxAvgTokens: 50_000 },\n  heavy: { maxAvgDuration: 60_000, maxAvgSteps: 20, maxAvgTokens: 150_000 },\n}\n\nconst WEB_TOOLS = ['webSearch', 'webFetch']\n\nfunction getThresholdTier(toolSequence: string) {\n  return WEB_TOOLS.some(t =&gt; toolSequence.includes(t))\n    ? HEALTH_THRESHOLDS.heavy : HEALTH_THRESHOLDS.light\n}<\/code><\/pre>\n<p>\u5224\u65ad\u903b\u8f91\u5f88\u7b80\u5355\uff1a\u770b\u884c\u4e3a\u7ec4\u7684\u4ee3\u8868\u6027\u5de5\u5177\u5e8f\u5217\u91cc\u6709\u6ca1\u6709\u7f51\u7edc\u5de5\u5177\uff0c\u6709\u7684\u8bdd\u7528 <code>heavy<\/code> \u6807\u51c6\uff0c\u6ca1\u6709\u7528 <code>light<\/code>\u3002\u5982\u679c\u4f60\u7684 Agent \u6709\u66f4\u591a\u7c7b\u578b\u7684\u5de5\u5177\uff0c\u52a0\u65b0\u7684\u5206\u7c7b\u5c31\u884c\u3002<\/p>\n<p>\u4e00\u4e2a\u884c\u4e3a\u7ec4\u7684\u5065\u5eb7\u5f97\u5206\u4f4e\u4e8e 0.8\uff08\u5373\u81f3\u5c11\u4e24\u4e2a\u7ef4\u5ea6\u672a\u8fbe\u6807\uff09\uff0c\u6216\u8005\u89e6\u53d1\u4e86 <code>low_success_rate<\/code> \u6216 <code>high_tool_error_rate<\/code> \u8fd9\u4e24\u4e2a\u5173\u952e\u6807\u8bb0\uff0c\u5c31\u88ab\u5224\u5b9a\u4e3a\u201d\u4e0d\u5065\u5eb7\u201d\u3002<\/p>\n<p>\u6700\u540e\u662f\u751f\u6210\u5efa\u8bae\u3002\u5bf9\u5f97\u5206\u4f4e\u7684\u7ec4\uff0c\u518d\u6b21\u8c03\u7528 LLM\uff0c\u8ba9\u5b83\u7ed9\u51fa\u5177\u4f53\u7684 Harness \u5c42\u6539\u8fdb\u5efa\u8bae\u3002LLM \u6536\u5230\u7684\u4fe1\u606f\u5305\u62ec\u884c\u4e3a\u540d\u79f0\u3001\u5177\u4f53\u7684\u5065\u5eb7\u6307\u6807\u6570\u503c\u3001\u5de5\u5177\u8c03\u7528\u5e8f\u5217\uff0c\u4ee5\u53ca\u54ea\u4e9b\u7ef4\u5ea6\u4e0d\u8fbe\u6807\u3002\u8981\u6c42\u5efa\u8bae\u9488\u5bf9 Harness \u4ee3\u7801\u672c\u8eab\uff08\u6bd4\u5982\u201c\u7ed9 webFetch \u52a0\u91cd\u8bd5\u201d\u3001\u201c\u4f18\u5316 dbQuery \u7684 Schema \u63cf\u8ff0\u201d\uff09\uff0c\u800c\u4e0d\u662f prompt \u5c42\u9762\u7684\u8c03\u6574\u3002<\/p>\n<p>LLM \u7ed9\u51fa\u7684\u6bcf\u6761\u5efa\u8bae\u4f1a\u6807\u6ce8\u4e25\u91cd\u7a0b\u5ea6\uff1a<code>critical<\/code> \u8868\u793a\u6709\u660e\u786e\u7684\u4ee3\u7801\u4fee\u6539\u65b9\u6848\uff0c\u53ef\u4ee5\u76f4\u63a5\u8fdb\u5165\u81ea\u52a8\u4fee\u590d\u6d41\u7a0b\uff1b<code>suggestion<\/code> \u8868\u793a\u5efa\u8bae\u6027\u6539\u8fdb\uff0c\u7531\u5f00\u53d1\u8005\u51b3\u5b9a\u662f\u5426\u5b9e\u65bd\u3002<code>critical<\/code> \u7ea7\u522b\u7684\u5efa\u8bae\u4f1a\u88ab\u6807\u8bb0\u4e3a <code>fix_status = 'unfixed'<\/code>\uff0c\u540e\u9762\u4f1a\u8fdb\u5165\u81ea\u52a8\u4fee\u590d\u7684\u5904\u7406\u961f\u5217\u3002<\/p>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047930620\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" title=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" loading=\"lazy\"><\/p>\n<h2>\u843d\u5730\u6307\u5357\uff1a\u4ece demo \u5230\u751f\u4ea7\uff0c\u4ee5\u53ca\u7528\u91cd\u653e\u5f15\u64ce\u9a8c\u8bc1\u4fee\u590d<\/h2>\n<p>\u524d\u9762\u51e0\u8282\u628a demo \u91cc\u7684\u6bcf\u4e2a\u6a21\u5757\u90fd\u8bb2\u4e86\u4e00\u904d\u3002\u8fd9\u4e00\u8282\u804a\u600e\u4e48\u628a\u8fd9\u4e9b\u4e1c\u897f\u8fc1\u79fb\u5230\u771f\u5b9e\u4ea7\u54c1\u91cc\uff0c\u4ee5\u53ca\u5f53\u524d\u6700\u5927\u7684\u77ed\u677f\u2014\u2014\u600e\u4e48\u81ea\u52a8\u9a8c\u8bc1\u4fee\u590d\u5230\u5e95\u6709\u6ca1\u6709\u7528\u3002<\/p>\n<h3>\u8fc1\u79fb\u5230\u4f60\u7684\u9879\u76ee<\/h3>\n<p>\u867d\u7136 demo \u7528\u7684\u662f TypeScript + Node.js\uff0c\u4f46\u5b9e\u73b0\u601d\u8def\u8ddf\u8bed\u8a00\u6ca1\u6709\u5173\u7cfb\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6b65\u9aa4<\/th>\n<th>\u505a\u4ec0\u4e48<\/th>\n<th>\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>1<\/td>\n<td>\u5728 Agent Loop \u91cc\u52a0 Tracing<\/td>\n<td>Agent \u6bcf\u6267\u884c\u5b8c\u4e00\u6b65\u5c31\u7acb\u523b\u5199\u5165\u6570\u636e\u5e93\u3002\u5173\u952e\u662f\u5373\u65f6\u5199\u5165\u2014\u2014\u5d29\u6e83\u4e86\u4e5f\u4e0d\u4f1a\u4e22\u5931\u5df2\u5b8c\u6210\u6b65\u9aa4\u7684\u8bb0\u5f55<\/td>\n<\/tr>\n<tr>\n<td>2<\/td>\n<td>\u9519\u8bef\u5206\u6876<\/td>\n<td>\u6309 provider \u00d7 \u9519\u8bef\u7c7b\u578b \u00d7 \u72b6\u6001\u7801 \u00d7 \u5de5\u5177\u540d \u00d7 \u9519\u8bef\u6d88\u606f\u505a GROUP BY \u5206\u7ec4\uff0c\u901a\u8fc7 SQL \u53ef\u4ee5\u770b\u5230\u51fa\u73b0\u6b21\u6570\u6700\u591a\u7684\u9519\u8bef\u7c7b\u578b<\/td>\n<\/tr>\n<tr>\n<td>3<\/td>\n<td>\u5b9a\u65f6 LLM \u5de1\u68c0<\/td>\n<td>\u5b9a\u65f6\u628a\u672a\u5339\u914d\u7684\u9519\u8bef\u6876\u53d1\u7ed9 LLM\uff0c\u901a\u8fc7 <code>generateObject<\/code> + Zod Schema\uff08\u6216\u5bf9\u5e94\u8bed\u8a00\u7684 JSON Schema \u5e93\uff09\u7ea6\u675f\u8f93\u51fa\u683c\u5f0f\uff0c\u81ea\u52a8\u751f\u6210\u65b0 Pattern<\/td>\n<\/tr>\n<tr>\n<td>4<\/td>\n<td>\u81ea\u52a8\u56de\u626b<\/td>\n<td>\u65b0 Pattern \u751f\u6210\u540e\u7acb\u523b\u56de\u626b\u5386\u53f2\u9519\u8bef\uff0c\u4e4b\u524d\u6ca1\u5339\u914d\u4e0a\u7684\u9519\u8bef\u73b0\u5728\u80fd\u88ab\u81ea\u52a8\u5f52\u7c7b<\/td>\n<\/tr>\n<tr>\n<td>5<\/td>\n<td>\u81ea\u52a8\u4fee\u590d<\/td>\n<td>\u5de1\u68c0\u8bc6\u522b\u51fa\u7684 <code>harness_bug<\/code> \u548c\u884c\u4e3a\u5206\u6790\u6807\u8bb0\u4e3a <code>critical<\/code> \u7684\u95ee\u9898\uff0c\u90fd\u4f1a\u8fdb\u5165\u81ea\u52a8\u4fee\u590d\u961f\u5217\u3002\u4fee\u590d\u65f6 Agent \u4f1a\u8bfb\u53d6\u76f8\u5173\u6e90\u7801\u3001\u751f\u6210\u4fee\u590d\u4ee3\u7801\u3001\u63d0\u4ea4 PR<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6700\u91cd\u8981\u7684\u662f\u7b2c 1 \u6b65\u3002\u6ca1\u6709 trace \u6570\u636e\uff0c\u540e\u9762\u4ec0\u4e48\u90fd\u5e72\u4e0d\u4e86\u3002<\/p>\n<h3>\u8c03\u5ea6\u7b56\u7565<\/h3>\n<p>demo \u91cc\u6240\u6709\u64cd\u4f5c\u90fd\u8981\u624b\u52a8\u70b9\u6309\u94ae\uff0c\u751f\u4ea7\u73af\u5883\u5e94\u8be5\u914d\u5b9a\u65f6\u4efb\u52a1\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u4efb\u52a1<\/th>\n<th>demo \u4e2d<\/th>\n<th>\u751f\u4ea7\u73af\u5883\u5efa\u8bae<\/th>\n<th>\u7406\u7531<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u9519\u8bef\u5de1\u68c0\uff08\u8bc6\u522b Pattern\uff09<\/td>\n<td>\u624b\u52a8\u70b9\u51fb<\/td>\n<td>\u6bcf 1-2 \u5c0f\u65f6\u81ea\u52a8\u6267\u884c<\/td>\n<td>\u9519\u8bef\u5bf9\u7528\u6237\u4f53\u9a8c\u5f71\u54cd\u5927\uff0c\u9700\u8981\u5feb\u901f\u53d1\u73b0\u3002\u6bcf\u6b21\u53ea\u5206\u6790\u4e0a\u4e00\u4e2a\u5468\u671f\u5185\u65b0\u4ea7\u751f\u7684\u672a\u5339\u914d\u9519\u8bef<\/td>\n<\/tr>\n<tr>\n<td>\u884c\u4e3a\u5206\u6790\uff08\u5065\u5eb7\u5ea6\u8bc4\u4f30\uff09<\/td>\n<td>\u624b\u52a8\u70b9\u51fb<\/td>\n<td>\u6bcf 24 \u5c0f\u65f6\u81ea\u52a8\u6267\u884c<\/td>\n<td>\u884c\u4e3a\u6a21\u5f0f\u662f\u6162\u53d8\u91cf\uff0c\u9700\u8981\u79ef\u7d2f\u8db3\u591f\u591a\u7684\u65b0\u64cd\u4f5c\u6570\u636e\u624d\u6709\u610f\u4e49<\/td>\n<\/tr>\n<tr>\n<td>\u81ea\u52a8\u4fee\u590d\uff08\u751f\u6210 PR\uff09<\/td>\n<td>\u624b\u52a8\u70b9\u51fb<\/td>\n<td>\u6bcf\u5929\u51cc\u6668\u56fa\u5b9a\u65f6\u95f4\u6267\u884c\u4e00\u6b21<\/td>\n<td>\u4ee3\u7801\u4fee\u6539\u9700\u8981\u5ba1\u614e\uff0c\u5b89\u6392\u5728\u4f4e\u5cf0\u671f\u6267\u884c<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6267\u884c\u987a\u5e8f\u6709\u4f9d\u8d56\uff1a\u5de1\u68c0\u5148\u6267\u884c \u2192 \u884c\u4e3a\u5206\u6790\u518d\u6267\u884c \u2192 \u6700\u540e\u81ea\u52a8\u4fee\u590d\u5904\u7406\u6240\u6709\u5f85\u4fee\u7684\u3002\u5b9e\u73b0\u65b9\u5f0f\u4e0d\u9650\uff0ccron job\u3001GitHub Actions\u3001Temporal \u90fd\u884c\uff0c\u6838\u5fc3\u5c31\u662f\u5b9a\u65f6\u8c03\u7528\u540c\u6837\u7684\u51fd\u6570\u3002<\/p>\n<h3>\u6210\u672c\u53c2\u8003<\/h3>\n<p>\u6309\u751f\u4ea7\u73af\u5883\u4f30\u7b97\uff1a\u6bcf\u5c0f\u65f6\u5de1\u68c0\u4e00\u6b21\uff0c\u6bcf\u5929\u884c\u4e3a\u5206\u6790\u548c\u81ea\u52a8\u4fee\u590d\u5404\u4e00\u6b21\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u4efb\u52a1<\/th>\n<th>\u5355\u8f6e token \u6d88\u8017<\/th>\n<th>\u6708\u5ea6\u9891\u6b21<\/th>\n<th>\u6708\u5ea6 token<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u5de1\u68c0<\/td>\n<td>~1 \u4e07\uff08\u9519\u8bef\u6876\u6458\u8981 + LLM \u5206\u6790\uff09<\/td>\n<td>24 \u6b21\/\u5929 \u00d7 30 \u5929 = 720 \u6b21<\/td>\n<td>~720 \u4e07<\/td>\n<\/tr>\n<tr>\n<td>\u884c\u4e3a\u5206\u6790<\/td>\n<td>~2 \u4e07\uff08100 \u6761 operation \u6458\u8981 + \u805a\u7c7b\uff09<\/td>\n<td>1 \u6b21\/\u5929 \u00d7 30 \u5929 = 30 \u6b21<\/td>\n<td>~60 \u4e07<\/td>\n<\/tr>\n<tr>\n<td>\u81ea\u52a8\u4fee\u590d<\/td>\n<td>20-100 \u4e07\/\u76ee\u6807\uff08\u5b8c\u6574 Agent Loop\uff0c\u542b\u591a\u8f6e\u6587\u4ef6\u8bfb\u53d6\u548c\u4ee3\u7801\u751f\u6210\uff09<\/td>\n<td>~10 \u4e2a\u76ee\u6807\/\u6708<\/td>\n<td>~200-1000 \u4e07<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u81ea\u52a8\u4fee\u590d\u6d88\u8017\u6700\u5927\uff0c\u56e0\u4e3a\u4fee\u590d Agent \u6bcf\u6b21 LLM \u8c03\u7528\u90fd\u5e26\u7740\u5b8c\u6574\u5bf9\u8bdd\u5386\u53f2\uff0c\u8bfb\u7684\u6587\u4ef6\u8d8a\u591a\u3001\u6539\u7684\u5730\u65b9\u8d8a\u591a\uff0c\u4e0a\u4e0b\u6587\u5c31\u8d8a\u5927\u3002\u7b80\u5355\u4fee\u590d\uff08\u6539\u4e00\u4e24\u4e2a\u6587\u4ef6\uff09\u7ea6 20 \u4e07 token\uff0c\u590d\u6742\u4fee\u590d\uff08\u8de8\u591a\u4e2a\u6587\u4ef6\u3001\u9700\u8981\u591a\u8f6e\u8fed\u4ee3\uff09\u53ef\u80fd\u8d85\u8fc7 100 \u4e07 token\u3002<\/p>\n<p>\u4ee5 DeepSeek V4-Pro\uff08\u8f93\u5165 \u00a53\/\u767e\u4e07 token\uff0c\u7f13\u5b58\u547d\u4e2d \u00a50.025\/\u767e\u4e07 token\uff0c\u8f93\u51fa \u00a56\/\u767e\u4e07 token\uff09\u4f30\u7b97\uff0c\u6708\u5ea6\u6210\u672c\u7ea6 \u00a530-80\uff0c\u53d6\u51b3\u4e8e\u4fee\u590d\u76ee\u6807\u7684\u6570\u91cf\u548c\u590d\u6742\u5ea6\u3002\u5de1\u68c0 prompt \u91cd\u590d\u5ea6\u9ad8\uff0c\u7f13\u5b58\u547d\u4e2d\u7387\u9ad8\u65f6\u6210\u672c\u4f1a\u66f4\u4f4e\u3002\u6362\u6210 GPT-4o \u6216 Claude \u7b49\u6a21\u578b\uff0c\u4ef7\u683c\u4f1a\u9ad8\u51fa\u4e00\u5230\u4e24\u4e2a\u6570\u91cf\u7ea7\u3002<\/p>\n<p>\u4ee5\u4e0a\u53ea\u662f\u57fa\u4e8e demo \u89c4\u6a21\u7684\u7c97\u7565\u4f30\u7b97\uff0c\u5b9e\u9645\u8d39\u7528\u53d6\u51b3\u4e8e\u7cfb\u7edf\u590d\u6742\u5ea6\u3001\u9519\u8bef\u6570\u91cf\u548c\u4fee\u590d\u9891\u7387\u3002<\/p>\n<h3>\u81ea\u8fdb\u5316\u7684\u4e09\u4e2a\u5c42\u6b21<\/h3>\n<p>\u524d\u9762\u8bb2\u7684\u5404\u4e2a\u73af\u8282\uff0c\u6309\u81ea\u52a8\u5316\u7a0b\u5ea6\u53ef\u4ee5\u5206\u6210\u4e09\u4e2a\u5c42\u6b21\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u5c42\u6b21<\/th>\n<th>\u4eba\u505a\u4ec0\u4e48<\/th>\n<th>\u7cfb\u7edf\u505a\u4ec0\u4e48<\/th>\n<th>\u672c\u6587\u8986\u76d6<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>L1 \u7eaf\u4eba\u5de5<\/td>\n<td>\u770b\u65e5\u5fd7\u3001\u5206\u7c7b\u9519\u8bef\u3001\u5199\u4fee\u590d\u4ee3\u7801<\/td>\n<td>\u65e0<\/td>\n<td>\u2014<\/td>\n<\/tr>\n<tr>\n<td>L2 \u8f85\u52a9<\/td>\n<td>\u786e\u8ba4\u5206\u7c7b\u7ed3\u679c + \u51b3\u5b9a\u662f\u5426\u4fee\u590d<\/td>\n<td>\u81ea\u52a8\u91c7\u96c6\u6570\u636e\u3001\u627e\u51fa\u53ef\u7591\u9519\u8bef<\/td>\n<td>\u2014<\/td>\n<\/tr>\n<tr>\n<td>L3 \u4e3b\u5bfc<\/td>\n<td>\u5ba1\u67e5 PR + \u9ad8\u98ce\u9669\u51b3\u7b56<\/td>\n<td>\u91c7\u96c6 \u2192 \u8bc6\u522b \u2192 \u5206\u7c7b \u2192 \u751f\u6210\u4fee\u6539 \u2192 \u63d0 PR<\/td>\n<td>\u672c\u6587\u4e3b\u8981\u6f14\u793a<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart LR\n    subgraph L1[L1 \u7eaf\u4eba\u5de5]\n        A1[\u4eba\u770b\u65e5\u5fd7] --&gt; A2[\u4eba\u5206\u7c7b] --&gt; A3[\u4eba\u4fee\u590d]\n    end\n    subgraph L2[L2 \u8f85\u52a9]\n        B1[\u7cfb\u7edf\u91c7\u96c6\u6570\u636e] --&gt; B2[\u7cfb\u7edf\u627e\u51fa\u53ef\u7591\u9519\u8bef] --&gt; B3[\u4eba\u786e\u8ba4 + \u4eba\u4fee\u590d]\n    end\n    subgraph L3[L3 \u4e3b\u5bfc]\n        C1[\u7cfb\u7edf\u91c7\u96c6] --&gt; C2[\u7cfb\u7edf\u8bc6\u522b] --&gt; C3[\u7cfb\u7edf\u5206\u7c7b] --&gt; C4[\u7cfb\u7edf\u751f\u6210\u4fee\u6539] --&gt; C5[\u7cfb\u7edf\u63d0 PR] --&gt; C6[\u4eba\u5ba1\u67e5 PR]\n    end\n    L1 -.-&gt;|\u52a0 Tracing| L2\n    L2 -.-&gt;|\u52a0 LLM \u5de1\u68c0 + \u81ea\u52a8\u4fee\u590d| L3<\/code><\/pre>\n<p>\u8868\u91cc\u7684\u4e09\u4e2a\u5c42\u6b21\u5bf9\u5e94\u4e0a\u9762\u201c\u8fc1\u79fb\u5230\u4f60\u7684\u9879\u76ee\u201d\u91cc\u7684\u6b65\u9aa4\uff1a\u52a0\u5b8c Tracing \u5c31\u5230 L2\uff0c\u518d\u63a5\u4e0a LLM \u5de1\u68c0\u548c\u81ea\u52a8\u4fee\u590d\u5c31\u5230 L3\u3002<\/p>\n<h3>\u91cd\u653e\u5f15\u64ce\uff1a\u81ea\u52a8\u9a8c\u8bc1\u4fee\u590d\u662f\u5426\u751f\u6548<\/h3>\n<p>\u76ee\u524d\u7cfb\u7edf\u6700\u5927\u7684\u77ed\u677f\u662f\uff1a\u81ea\u52a8\u4fee\u590d\u751f\u6210\u4e86 PR\uff0c\u4f46 PR \u5230\u5e95\u6709\u6ca1\u6709\u4fee\u597d bug\uff0c\u8fd8\u662f\u5f97\u9760\u5f00\u53d1\u6d4b\u8bd5\u548c\u5ba1\u6838\u3002\u91cd\u653e\u5f15\u64ce\u8981\u89e3\u51b3\u7684\u5c31\u662f\u8fd9\u4e2a\u95ee\u9898\uff0c\u8ba9\u7cfb\u7edf\u81ea\u5df1\u9a8c\u8bc1\u4fee\u590d\u662f\u5426\u6709\u6548\u3002<\/p>\n<h4>\u4e3a\u4ec0\u4e48\u786e\u5b9a\u6027\u91cd\u653e\u53ef\u884c<\/h4>\n<p>\u56de\u987e Tracing \u90a3\u8282\uff1asteps \u8868\u8bb0\u5f55\u4e86\u6bcf\u4e00\u6b65\u7684\u8f93\u5165\u8f93\u51fa\u2014\u2014<code>llm_response<\/code> \u5b58\u4e86\u6a21\u578b\u56de\u590d\uff0c<code>tool_output<\/code> \u5b58\u4e86\u5de5\u5177\u8fd4\u56de\u7684 JSON\u3002\u91cd\u653e\u65f6\u4e0d\u9700\u8981\u8c03\u7528\u771f\u5b9e\u7684 LLM\u3001\u4e0d\u6267\u884c\u771f\u5b9e\u7684\u5de5\u5177\uff0c\u628a\u5f55\u597d\u7684\u6570\u636e\u6309\u6b65\u9aa4\u987a\u5e8f\u4f20\u7ed9 Agent Loop \u5c31\u884c\uff0c\u6574\u4e2a\u6267\u884c\u8fc7\u7a0b\u5c31\u53d8\u6210\u4e86\u7eaf\u51fd\u6570\u3002<\/p>\n<p>\u8fd9\u5bf9 harness_bug \u7279\u522b\u6709\u6548\u3002\u8fd9\u7c7b bug \u51fa\u5728 Harness \u600e\u4e48\u5904\u7406\u8fd4\u56de\u503c\uff0c\u8ddf\u8fd4\u56de\u503c\u672c\u8eab\u65e0\u5173\u3002\u6bd4\u5982\u201ctoken \u63a5\u8fd1\u4e0a\u9650\u65f6\u6ca1\u89e6\u53d1\u538b\u7f29\uff0c\u4e0b\u4e00\u8f6e API \u8c03\u7528\u5c31\u5931\u8d25\u4e86\u201d\u2014\u2014\u4e0d\u7ba1\u6a21\u578b\u8bf4\u4e86\u4ec0\u4e48\uff0c\u53ea\u8981\u4e0a\u4e0b\u6587\u7ba1\u7406\u7684\u903b\u8f91\u6709\u95ee\u9898\uff0cbug \u5c31\u4f1a\u51fa\u73b0\u3002\u91cd\u653e\u540c\u6837\u7684 LLM \u56de\u590d\u548c\u5de5\u5177\u8fd4\u56de\u503c\uff0c\u5927\u6982\u7387\u80fd\u590d\u73b0\u3002<\/p>\n<h4>\u5b9e\u73b0\u65b9\u6848\uff1a\u590d\u7528\u540c\u4e00\u4e2a agentLoop<\/h4>\n<p>\u4e0d\u7528\u5199\u4e00\u5957\u65b0\u7684\u6267\u884c\u5f15\u64ce\u3002\u91cd\u653e\u8d70\u7684\u8fd8\u662f\u540c\u4e00\u4e2a <code>agentLoop()<\/code>\uff0c\u53ea\u662f\u628a LLM \u548c\u5de5\u5177\u66ff\u6362\u6210\u76f4\u63a5\u8fd4\u56de\u5f55\u597d\u6570\u636e\u7684 mock \u7248\u672c\u3002<\/p>\n<p>\u5148\u770b\u6b63\u5e38\u6267\u884c\u65f6 <code>agentLoop<\/code> \u600e\u4e48\u5de5\u4f5c\u7684\uff1a<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">sequenceDiagram\n    participant AL as agentLoop\n    participant LLM as \u771f\u5b9e LLM API\n    participant T as \u771f\u5b9e\u5de5\u5177\n\n    AL-&gt;&gt;LLM: \u53d1\u9001\u6d88\u606f\n    LLM--&gt;&gt;AL: &quot;\u8c03\u7528 dbQuery&quot;\n    AL-&gt;&gt;T: \u6267\u884c dbQuery(SQL)\n    T--&gt;&gt;AL: \u67e5\u8be2\u7ed3\u679c JSON\n    Note over AL: Harness \u5904\u7406\u8fd4\u56de\u503c&lt;br\/&gt;\uff08\u4e0a\u4e0b\u6587\u7ba1\u7406\u3001\u9519\u8bef\u6062\u590d\u7b49\uff09\n    AL-&gt;&gt;LLM: \u56de\u4f20\u5de5\u5177\u7ed3\u679c\n    LLM--&gt;&gt;AL: \u751f\u6210\u6700\u7ec8\u56de\u590d<\/code><\/pre>\n<p>\u91cd\u653e\u65f6\uff0c\u628a LLM \u548c\u5de5\u5177\u90fd\u6362\u6389\uff0c\u4f46 <code>agentLoop<\/code> \u672c\u8eab\u7684\u4ee3\u7801\u4e00\u884c\u4e0d\u6539\uff1a<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">sequenceDiagram\n    participant AL as agentLoop\uff08\u540c\u4e00\u4efd\u4ee3\u7801\uff09\n    participant SP as ScriptedProvider\n    participant TR as ToolReplayer\n\n    Note over SP: \u4ece steps \u8868\u8bfb\u53d6&lt;br\/&gt;\u5f55\u597d\u7684 llm_response\n    AL-&gt;&gt;SP: \u53d1\u9001\u6d88\u606f\n    SP--&gt;&gt;AL: \u8fd4\u56de\u5f55\u597d\u7684 &quot;\u8c03\u7528 dbQuery&quot;\n    Note over TR: \u4ece steps \u8868\u8bfb\u53d6&lt;br\/&gt;\u5f55\u597d\u7684 tool_output\n    AL-&gt;&gt;TR: \u6267\u884c dbQuery(SQL)\n    TR--&gt;&gt;AL: \u8fd4\u56de\u5f55\u597d\u7684\u67e5\u8be2\u7ed3\u679c JSON\n    Note over AL: Harness \u5904\u7406\u8fd4\u56de\u503c&lt;br\/&gt;\uff08\u88ab\u6d4b\u8bd5\u7684\u5c31\u662f\u8fd9\u90e8\u5206\uff09\n    AL-&gt;&gt;SP: \u56de\u4f20\u5de5\u5177\u7ed3\u679c\n    SP--&gt;&gt;AL: \u8fd4\u56de\u5f55\u597d\u7684\u6700\u7ec8\u56de\u590d<\/code><\/pre>\n<p>ScriptedProvider \u5b9e\u73b0\u4e86\u548c\u771f\u5b9e LLM \u4e00\u6837\u7684\u63a5\u53e3\uff08<code>LanguageModelV1<\/code>\uff09\uff0c\u4f46\u4e0d\u8fde\u4e0a\u6e38 API\uff0c\u76f4\u63a5\u4ece steps \u8868\u6309\u6b65\u9aa4\u987a\u5e8f\u8fd4\u56de\u5f55\u597d\u7684 <code>llm_response<\/code>\u3002ToolReplayer \u66ff\u6362\u5de5\u5177\u6ce8\u518c\u8868\uff0c\u6bcf\u6b21 <code>tool_call<\/code> \u76f4\u63a5\u8fd4\u56de\u5f55\u597d\u7684 <code>tool_output<\/code>\uff0c\u4e0d\u53d1\u90ae\u4ef6\u3001\u4e0d\u6267\u884c\u4ee3\u7801\u3001\u4e0d\u8bf7\u6c42\u5916\u90e8 API\u3002<\/p>\n<p>\u4ece <code>agentLoop<\/code> \u7684\u89c6\u89d2\u770b\uff0c\u91cd\u653e\u548c\u771f\u5b9e\u6267\u884c\u6ca1\u6709\u533a\u522b\u2014\u2014\u5b83\u4e0d\u77e5\u9053\u5bf9\u9762\u662f\u771f LLM \u8fd8\u662f\u5f55\u50cf\u56de\u653e\u3002\u88ab\u6d4b\u8bd5\u7684\u5c31\u662f <code>agentLoop<\/code> \u62ff\u5230\u8fd4\u56de\u503c\u4e4b\u540e\u7684\u5904\u7406\u903b\u8f91\uff1a\u4e0a\u4e0b\u6587\u7ba1\u7406\u6709\u6ca1\u6709\u6b63\u786e\u89e6\u53d1\u538b\u7f29\u3001\u5de5\u5177\u7ed3\u679c\u6709\u6ca1\u6709\u6b63\u786e\u622a\u65ad\u3001\u9519\u8bef\u6062\u590d\u6709\u6ca1\u6709\u6f0f\u6389\u8fb9\u754c\u60c5\u51b5\u3002<\/p>\n<h4>\u4e0e\u81ea\u52a8\u4fee\u590d\u7684\u96c6\u6210<\/h4>\n<p>\u5728 <code>auto-pr.ts<\/code> \u751f\u6210\u4ee3\u7801\u4fee\u6539\u540e\u3001push \u4e4b\u524d\uff0c\u52a0\u4e00\u9053\u9a8c\u8bc1\uff08Safety Gate\uff09\uff1a<\/p>\n<pre style=\"display:none;\"><code class=\"mermaid\">flowchart TD\n    A[LLM \u751f\u6210\u4ee3\u7801\u4fee\u6539] --&gt; B[\u5e94\u7528\u4fee\u6539\u5230\u5de5\u4f5c\u76ee\u5f55]\n    B --&gt; C[\u590d\u73b0\u9a8c\u8bc1]\n    C --&gt; D{&quot;\u672a\u6253\u8865\u4e01\u7684\u4ee3\u7801\/n\u91cd\u653e\u89e6\u53d1 bug \u7684 trace&quot;}\n    D --&gt;|FAIL \u2713| E{&quot;\u5df2\u6253\u8865\u4e01\u7684\u4ee3\u7801\/n\u91cd\u653e\u540c\u4e00\u6761 trace&quot;}\n    D --&gt;|PASS \u2717| F[trace \u65e0\u6cd5\u590d\u73b0 bug\/n\u8df3\u8fc7\u9a8c\u8bc1]\n    E --&gt;|PASS \u2713| G[\u56de\u5f52\u9a8c\u8bc1]\n    E --&gt;|FAIL \u2717| H[\u4fee\u590d\u65e0\u6548\/n\u56de\u6eda\uff0c\u6807\u8bb0 verify_failed]\n    G --&gt; I{&quot;\u91cd\u653e\u6240\u6709\u56de\u5f52\u7528\u4f8b&quot;}\n    I --&gt;|\u5168\u90e8 PASS| J[\u9a8c\u8bc1\u901a\u8fc7\/ncommit + push + PR]\n    I --&gt;|\u6709\u5931\u8d25| H<\/code><\/pre>\n<p>\u590d\u73b0\u9a8c\u8bc1\u5206\u4e24\u6b65\uff1a\u5148\u7528\u6ca1\u6253\u8865\u4e01\u7684\u4ee3\u7801\u91cd\u653e\u89e6\u53d1 bug \u7684 operation\u2014\u2014\u5fc5\u987b\u5931\u8d25\uff08\u786e\u8ba4 trace \u80fd\u590d\u73b0 bug\uff09\uff1b\u518d\u7528\u6253\u4e86\u8865\u4e01\u7684\u4ee3\u7801\u91cd\u653e\u540c\u4e00\u6761\u2014\u2014\u5fc5\u987b\u6210\u529f\uff08\u786e\u8ba4\u4fee\u590d\u6709\u6548\uff09\u3002\u5982\u679c\u7b2c\u4e00\u6b65\u5c31\u6210\u529f\u4e86\uff0c\u8bf4\u660e\u8fd9\u4e2a bug \u53ef\u80fd\u8ddf Harness \u4e4b\u5916\u7684\u6761\u4ef6\u6709\u5173\uff0c\u91cd\u653e\u9a8c\u8bc1\u4e0d\u4e86\uff0c\u8df3\u8fc7\u3002\u56de\u5f52\u9a8c\u8bc1\u662f\u62ff\u6253\u8fc7\u8865\u4e01\u7684\u4ee3\u7801\u53bb\u8dd1\u4e4b\u524d\u4fee\u8fc7\u7684 bug \u5bf9\u5e94\u7684 trace\uff0c\u5168\u90e8\u901a\u8fc7\u624d\u5141\u8bb8\u63d0\u4ea4\u3002<\/p>\n<h4>\u56de\u5f52\u8bed\u6599\u5e93\u7684\u81ea\u52a8\u79ef\u7d2f<\/h4>\n<p>\u6bcf\u5f53\u4e00\u4e2a harness_bug \u7684\u4fee\u590d PR \u5408\u5e76\u540e\uff08<code>fix_status<\/code> \u2192 <code>merged<\/code>\uff09\uff0c\u89e6\u53d1\u8fd9\u4e2a bug \u7684\u90a3\u6761 trace \u4f1a\u81ea\u52a8\u5b58\u8fdb\u4e00\u4e2a\u56de\u5f52\u7528\u4f8b\u96c6\u5408\u3002\u4e4b\u540e\u6bcf\u6b21\u751f\u6210\u65b0\u7684\u4fee\u590d PR\uff0c\u9664\u4e86\u9a8c\u8bc1\u65b0 PR \u672c\u8eab\u80fd\u4fee\u597d\u76ee\u6807 bug\uff0c\u8fd8\u8981\u628a\u56de\u5f52\u7528\u4f8b\u96c6\u5408\u91cc\u5df2\u6709\u7684 trace \u5168\u90e8\u91cd\u653e\u4e00\u904d\uff0c\u786e\u4fdd\u65b0\u6539\u52a8\u6ca1\u6709\u628a\u4ee5\u524d\u4fee\u597d\u7684 bug \u91cd\u65b0\u5f15\u5165\u3002\u4fee\u597d\u7684 bug \u8d8a\u591a\uff0c\u56de\u5f52\u7528\u4f8b\u96c6\u5408\u8d8a\u5927\uff0c\u56de\u5f52\u4fdd\u62a4\u7684\u8986\u76d6\u9762\u4e5f\u8d8a\u5e7f\u3002<\/p>\n<h4>\u8fb9\u754c<\/h4>\n<p>\u91cd\u653e\u5f15\u64ce\u9a8c\u8bc1\u7684\u662f Harness \u7684\u5904\u7406\u903b\u8f91\uff0c\u4e0d\u662f LLM \u7684\u56de\u7b54\u8d28\u91cf\u3002\u56e0\u4e3a\u91cd\u653e\u65f6 LLM \u7684\u56de\u590d\u662f\u5f55\u597d\u7684\u56fa\u5b9a\u6570\u636e\uff0c\u6240\u4ee5\u6a21\u578b\u56de\u7b54\u5f97\u597d\u4e0d\u597d\u4e0d\u5728\u9a8c\u8bc1\u8303\u56f4\u5185\u3002\u91cd\u653e\u5f15\u64ce\u53ea\u5173\u5fc3\u4e00\u4ef6\u4e8b\uff1aHarness \u62ff\u5230\u8fd9\u4e9b\u8fd4\u56de\u503c\u4e4b\u540e\uff0c\u4e0a\u4e0b\u6587\u7ba1\u7406\u3001\u5de5\u5177\u7ed3\u679c\u5904\u7406\u3001\u9519\u8bef\u6062\u590d\u8fd9\u4e9b\u4ee3\u7801\u903b\u8f91\u6709\u6ca1\u6709\u6b63\u786e\u6267\u884c\u3002<\/p>\n<p>\u7531\u4e8e\u590d\u6742\u5ea6\u7684\u95ee\u9898\uff0c\u6211\u6ca1\u6709\u5728 demo \u91cc\u5b9e\u73b0\u91cd\u653e\u5f15\u64ce\uff0c\u4e0d\u8fc7\u8fd9\u4e2a\u786e\u5b9e\u662f\u53ef\u884c\u7684\uff0c\u56e0\u4e3a\u6211\u4e4b\u524d\u5728\u522b\u7684\u4ea7\u54c1\u5df2\u7ecf\u5b9e\u73b0\u8fc7\u4e86\u3002<\/p>\n<h2>\u505a\u4e86\u8fd9\u4e9b\u4e4b\u540e\u80fd\u5f97\u5230\u4ec0\u4e48<\/h2>\n<p>\u8bfb\u8005\u53ef\u4ee5\u5728 demo \u91cc\u76f4\u63a5\u9a8c\u8bc1\u8fd9\u5957\u6d41\u7a0b\u7684\u6548\u679c\u3002\u5148\u6267\u884c <code>pnpm simulate --mock<\/code> \u6ce8\u5165\u6a21\u62df\u6570\u636e\uff08\u4e5f\u53ef\u4ee5\u4e0d\u52a0 <code>--mock<\/code> \u53c2\u6570\uff0c\u7528\u771f\u5b9e API \u8c03\u7528\uff09\uff0c\u7136\u540e\u5728\u7ba1\u7406\u9762\u677f\u8dd1\u4e00\u6b21\u5de1\u68c0\u3002\u56e0\u4e3a\u8fd9\u65f6\u5019 Pattern \u5e93\u8fd8\u662f\u7a7a\u7684\uff0c\u6240\u6709\u9519\u8bef\u90fd\u5904\u4e8e&#8221;\u672a\u5339\u914d&#8221;\u72b6\u6001\u3002\u5de1\u68c0\u6267\u884c\u540e\uff0c\u7cfb\u7edf\u4f1a\u628a\u76f8\u4f3c\u7684\u9519\u8bef\u5f52\u5230\u4e00\u8d77\uff0c\u4e3a\u6bcf\u7ec4\u9519\u8bef\u751f\u6210\u5bf9\u5e94\u7684 Pattern\uff0c\u518d\u62ff\u65b0 Pattern \u53bb\u56de\u626b\u4e4b\u524d\u7684\u5386\u53f2\u9519\u8bef\uff0c\u628a\u80fd\u5339\u914d\u4e0a\u7684\u90fd\u6807\u8bb0\u4e0a\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u539f\u5148\u5168\u662f&#8221;\u672a\u5339\u914d&#8221;\u7684\u9519\u8bef\u5217\u8868\u91cc\u5927\u90e8\u5206\u9519\u8bef\u90fd\u5df2\u7ecf\u88ab\u5206\u7c7b\uff0c\u672a\u5339\u914d\u9519\u8bef\u7387\u5927\u5e45\u4e0b\u964d\u3002<\/p>\n<p>\u4f46\u5de1\u68c0\u53ea\u662f\u8bc6\u522b\u548c\u5206\u7c7b\u9519\u8bef\uff0c\u771f\u6b63\u8ba9\u9519\u8bef\u51cf\u5c11\u7684\u662f\u540e\u7eed\u7684\u81ea\u52a8\u4fee\u590d\u3002\u5de1\u68c0\u628a <code>harness_bug<\/code> \u7c7b\u578b\u7684 Pattern \u6807\u8bb0\u51fa\u6765\u4e4b\u540e\uff0c\u81ea\u52a8\u4fee\u590d\u4f1a\u4e3a\u8fd9\u4e9b bug \u751f\u6210\u4fee\u590d PR\u3002\u5f00\u53d1\u8005\u5ba1\u6838\u5e76\u5408\u5e76 PR \u540e\uff0c\u5bf9\u5e94\u7684 bug \u5728\u540e\u7eed\u7684\u8bf7\u6c42\u4e2d\u5c31\u4e0d\u4f1a\u518d\u51fa\u73b0\u4e86\u3002\u968f\u7740\u5de1\u68c0\u4e0d\u65ad\u53d1\u73b0\u65b0\u7684 bug\u3001\u81ea\u52a8\u4fee\u590d\u4e0d\u65ad\u751f\u6210 PR\u3001\u5f00\u53d1\u8005\u4e0d\u65ad\u5408\u5e76\u4fee\u590d\uff0c\u7cfb\u7edf\u4ea7\u751f\u7684\u9519\u8bef\u4f1a\u8d8a\u6765\u8d8a\u5c11\uff0c\u9700\u8981\u4eba\u5de5\u5904\u7406\u7684\u95ee\u9898\u4e5f\u8d8a\u6765\u8d8a\u5c11\u3002<\/p>\n<p>\u8fd9\u5957\u601d\u8def\u6765\u81ea LobeHub \u7684\u751f\u4ea7\u5b9e\u8df5\uff08<a href=\"https:\/\/link.segmentfault.com\/?enc=7ZP3JiJoO1AVdv%2BOo17aQg%3D%3D.M8sA89bqyA3T1J44c5DZ%2FfEuRcAJdNR%2Bx5Csf%2BxIT1a144%2B6%2Br0o%2FfpJ0kob799r8NhV16%2Be9zNQXgGL%2Bom3fw%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u300a\u9700\u8981\u81ea\u8fdb\u5316\u7684\u4e0d\u662f Agent\uff0c\u800c\u662f Harness\u300b<\/a>\uff09\uff0c\u4ed6\u4eec\u516c\u5e03\u7684\u6570\u636e\uff1a<\/p>\n<ul>\n<li>\u7ecf\u8fc7 9 \u8f6e\u5de1\u68c0\uff0cPattern \u4ece 31 \u6761\u589e\u957f\u5230 104 \u6761\u540e\u8d8b\u4e8e\u9971\u548c\uff0c\u6bcf\u8f6e\u65b0\u589e\u7684 Pattern \u6570\u91cf\u4ece 31 \u6301\u7eed\u4e0b\u964d\u5230 0<\/li>\n<li>\u5de1\u68c0\u8fc7\u7a0b\u4e2d\u53d1\u73b0\u4e86 20 \u591a\u4e2a Harness \u81ea\u8eab\u7684\u7f3a\u9677\uff0c\u5305\u62ec Schema \u4e0d\u517c\u5bb9\u3001\u8d1f\u6570 max_tokens\u3001reasoning_content \u4e22\u5931\u3001Context Window \u8fc7\u8f7d\u7b49<\/li>\n<li>Agent \u6210\u529f\u7387\u4ece\u65e9\u671f\u7684\u7ea6 75% \u63d0\u5347\u5230 95% \u4ee5\u4e0a<\/li>\n<\/ul>\n<p>\u8fd9\u5957\u673a\u5236\u8fd0\u884c\u8d77\u6765\u4e4b\u540e\uff0c\u6bcf\u6b21\u8bf7\u6c42\u4ea7\u751f\u7684\u9519\u8bef\u90fd\u4f1a\u88ab trace \u8bb0\u5f55\u4e0b\u6765\uff0c\u4e0b\u4e00\u8f6e\u5b9a\u65f6\u5de1\u68c0\u81ea\u52a8\u628a\u5b83\u5f52\u7c7b\u5230\u5df2\u6709\u7684 Pattern \u6216\u8005\u751f\u6210\u65b0\u7684 Pattern\u3002\u540c\u4e00\u7c7b\u9519\u8bef\u53ea\u9700\u8981\u5f00\u53d1\u8005\u5ba1\u6838\u4e00\u6b21\uff08\u786e\u8ba4 Pattern \u5206\u7c7b\u662f\u5426\u51c6\u786e\uff09\uff0c\u4e4b\u540e\u518d\u51fa\u73b0\u540c\u7c7b\u9519\u8bef\u65f6\u7cfb\u7edf\u76f4\u63a5\u5339\u914d\uff0c\u4e0d\u9700\u8981\u5f00\u53d1\u8005\u518d\u4ecb\u5165\u3002Pattern \u5e93\u968f\u7740\u5de1\u68c0\u8f6e\u6b21\u4e0d\u65ad\u79ef\u7d2f\uff0c\u672a\u5339\u914d\u7684\u9519\u8bef\u4f1a\u8d8a\u6765\u8d8a\u5c11\u3002<\/p>\n<h3>\u6211\u7684\u5176\u4ed6 AI Agent \u6587\u7ae0<\/h3>\n<ul>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=OEal8xfd4T7ikCZXrWoMTg%3D%3D.xL9nYih46pcgJt4JpB0fngRVGITMkOiBfj%2BojWrLesXf71TmYTh603WdMj%2B%2FFcQ0Zy1bCCjizHCggpdVzz8mK5JDBN1vyCwaQ%2FhTLR00C3fOvuvKEbLzQ9olhI14NIbn\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">AI Agent \u5de5\u7a0b\u5e08\u5165\u95e8\u6307\u5357<\/a><\/li>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=Vzz8Pi9fNn5PluFse0I2AQ%3D%3D.vH73TLCw61ikSSw7k54od7SGUESq31tlSU4KLt5nO8i7tot%2Fwod7RLRCi8rMClqPvjs7EQTNBlC4A%2BeA1ZoGTMIRB0s1hlMIukoH4CBl8hI5tkmHSdT%2F1u0ti7M5Szdl\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u5982\u4f55\u4ece\u96f6\u5f00\u59cb\u5b9e\u73b0\u4e00\u4e2a AI Agent CLI<\/a><\/li>\n<li>\n<p><a href=\"https:\/\/link.segmentfault.com\/?enc=aKhKjif1sv%2B7hIycsSg%2BYg%3D%3D.Nseaw9905FTw90S6bNWp2OckxI5dn76Q%2BhvDrPMNwWwGMmmqfsgE1xK0t6SbzPPtZjk2Cunlwv0YGZ4U4bHTwqK%2Bidytotx%2FVKCyyT6jIZo%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u6398\u91d1\u5c0f\u518c&#8211;\u4ece\u96f6\u6253\u9020\u4e00\u4e2a AI Agent CLI<\/a><\/p>\n<h2>\u53c2\u8003\u8d44\u6599<\/h2>\n<\/li>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=4yvNewK2cm%2BeWM6qd8XvMA%3D%3D.59PDstu%2FGsO1wLIBN3iiFLZOHjIkll%2BaK9VYgvnyzwdrXhuBIPxwxSQj89C%2BqD1h1cW0zNPMtWiZAv20ZdFKww%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u9700\u8981\u81ea\u8fdb\u5316\u7684\u4e0d\u662f Agent\uff0c\u800c\u662f Harness<\/a><\/li>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=CdgkteBHOZKIKyzJB4RKkA%3D%3D.W5h2h6In3JX%2BGupiX%2BvrpV9hdifWspxf%2FbqAwS1pbJ3IMhu%2FGAHVQHn%2FjTjotsbF\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">evo-agent-demo<\/a><\/li>\n<li><a href=\"https:\/\/link.segmentfault.com\/?enc=wb61IO2KeSIlcAFvj0jcWQ%3D%3D.q%2BcSdKpoqEBwWBPStS1%2BXbp2UkEhB%2BrTcq479yY72SA%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Vercel AI SDK<\/a><\/li>\n<\/ul>\n<p><a href=\"https:\/\/segmentfault.com\/a\/1190000047930609\">https:\/\/segmentfault.com\/a\/1190000047930609<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u4ecb\u7ecd\u600e\u4e48\u8ba9 AI Agent \u7684\u5de5\u7a0b\u4ee3\u7801\uff08Harness\uff09\u5177\u5907\u81ea\u6211\u8fdb\u5316\u80fd\u529b\u2014\u2014\u81ea\u52a8\u8bb0\u5f55\u8fd0\u884c\u6570\u636e\u3001\u81ea\u52a8\u8bc6\u522b\u9519 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10021396","post","type-post","status-publish","format-standard","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021396","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10021396"}],"version-history":[{"count":0,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021396\/revisions"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10021396"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10021396"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10021396"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}