@witty-ai/skill-insight 0.5.0-beta → 0.6.0-beta

package/.next/standalone/.next/server/chunks/src_98433cb8._.js

@@ -0,0 +1,175 @@
+module.exports=[50374,e=>{"use strict";function t(e){if(!e||!Array.isArray(e)||0===e.length)return[];if(e.some(e=>e&&(e.requestMessages||e.responseMessage)))return e;let t=[],n=[],s=e=>{if(0===e.length)return;let n=-1;for(let t=e.length-1;t>=0;t--)if("assistant"===e[t].role){n=t;break}-1!==n?t.push({requestMessages:e.slice(0,n),responseMessage:e[n]}):t.push({requestMessages:e,responseMessage:null})};for(let t of e)t&&("user"===(t.role||"unknown")&&n.length>0&&(s(n),n=[]),n.push(t));return s(n),t}function n(e){let t=new Set,n=[],s=/^[a-zA-Z0-9_\-\.]+$/,i=e=>{if(e)for(let i of e.tool_calls||e.toolCalls||[]){let e=(i?.function?.name??i?.name??"").toLowerCase(),l=i?.function?.arguments??i?.arguments??"";try{let i="string"==typeof l?JSON.parse(l):l;if("skill"===e||"load_skill"===e){let e=i?.name??i?.skill_name??i?.skillName??i?.skill;if(null!=e&&String(e).trim()){let l=String(e).trim().replace(/^['"]+|['"]+$/g,"");if(s.test(l)&&!t.has(l)){t.add(l);let e=i?.version!=null?Number(i.version):null;n.push({name:l,version:null===e||isNaN(e)?null:e})}}continue}if("task"===e){let e=i?.load_skills??i?.loadSkills??i?.skills??[];if(Array.isArray(e))for(let i of e){let e="string"==typeof i?i:i?.name??i?.skill??i?.skill_name??i?.skillName;if(null==e||!String(e).trim())continue;let l=String(e).trim().replace(/^['"]+|['"]+$/g,"");if(!s.test(l)||t.has(l))continue;t.add(l);let r="object"==typeof i?i?.version:null,o=null!=r?Number(r):null;n.push({name:l,version:null===o||isNaN(o)?null:o})}}}catch{}}};for(let t of e)for(let e of(i(t.responseMessage),t.requestMessages||[]))"assistant"===e.role&&i(e);return n}function s(e){let t=new Set,n=[],s=e=>{if(e&&Array.isArray(e))for(let s of e){if(s?.type!=="tool_use")continue;let e=(s?.name||"").toLowerCase();if("skill"!==e&&"load_skill"!==e)continue;let i=s.input,l=i?.skill??i?.skill_name??i?.skillName??i?.name;if(null==l||!String(l).trim())continue;let r=String(l).trim().replace(/^['"]+|['"]+$/g,"");if(/^[a-zA-Z0-9_\-\.]+$/.test(r)&&!t.has(r)){t.add(r);let e=i?.version!=null?Number(i.version):null;n.push({name:r,version:null===e||isNaN(e)?null:e})}}};for(let t of e)if(t.responseMessage?.content&&s(t.responseMessage.content),t.requestMessages)for(let e of t.requestMessages)"assistant"===e.role&&e.content&&s(e.content);return n}function i(e){let t=new Set,n=[],s=e=>{if(e&&Array.isArray(e))for(let s of e){if(s?.type!=="toolCall")continue;let e=(s?.name||"").toLowerCase();if("skill"!==e&&"load_skill"!==e)continue;let i=s?.arguments,l=i?.skill??i?.skill_name??i?.skillName??i?.name;if(null==l||!String(l).trim())continue;let r=String(l).trim().replace(/^['"]+|['"]+$/g,"");if(/^[a-zA-Z0-9_\-\.]+$/.test(r)&&!t.has(r)){t.add(r);let e=i?.version!=null?Number(i.version):null;n.push({name:r,version:null===e||isNaN(e)?null:e})}}};for(let t of e)if(t.responseMessage?.content&&s(t.responseMessage.content),t.requestMessages)for(let e of t.requestMessages)"assistant"===e.role&&e.content&&s(e.content);return n}e.s(["extractSkillsWithVersionsFromClaudeSession",()=>s,"extractSkillsWithVersionsFromOpenClawSession",()=>i,"extractSkillsWithVersionsFromOpencodeSession",()=>n,"normalizeInteractions",()=>t])},62414,e=>{"use strict";function t(e,t,n,s,i){let l=i?`Reference Skill Definition (Using this skill as the context for evaluation):
+${i}
+
+`:"";return`
+You are an objective and strict judge. Your task is to evaluate a "User Answer" against a set of weighted criteria for a given "User Query".
+
+${l}User Query: ${e}
+User Answer: ${t}
+
+Evaluation Criteria (Score strictly based on these weighted items):
+1. Root Causes (Must identify these issues):
+${n.map(e=>`   - [ID: ${e.id}] [Weight: ${e.weight}] ${e.content}`).join("\n")||"   (None)"}
+
+2. Key Actions (Must perform these actions):
+${s.map(e=>`   - [ID: ${e.id}] [Weight: ${e.weight}] ${e.content}`).join("\n")||"   (None)"}
+
+Evaluation Steps:
+1. For each item listed above (marked with [ID: ...]), determine the degree of match (0.0 to 1.0).
+   - 0.0 = Not mentioned or completely wrong.
+   - 0.5 = Partially mentioned or vague.
+   - 1.0 = Clearly and correctly addressed.
+   **CRITICAL**: For Key Actions involving specific operations (e.g., "backup", "modify", "restart"), you must find EXPLICT EVIDENCE in the User Answer that these actions were performed (checking/reading is NOT the same as backing up).
+2. Provide a brief explanation (in Chinese) for your evaluation of each item.
+3. If a Reference Skill Definition is provided, consider whether the answer aligns with the skill's capabilities and instructions, but primarily score based on the specific Root Causes and Key Actions listed above.
+
+Respond ONLY with a JSON object in the following format:
+{
+  "evaluations": [
+    { "id": "RC-0", "match_score": 0.5, "explanation": "此处用中文简要解释评分理由..." },
+    { "id": "KA-0", "match_score": 1.0, "explanation": "..." }
+    ...
+  ]
+}
+`}e.s(["generateJudgePrompt",()=>t])},33526,e=>{"use strict";let t=`
+你是一个专业的 Skill 优化分析师。你的任务是：
+
+1. 分析 Agent 在某个评分项上扣分的原因
+2. 判断该扣分是否可以通过优化 Skill 定义来解决
+3. 检查 Agent 的实际执行过程是否符合 Skill 中定义的步骤
+
+## 判断标准
+
+**是 Skill 问题的情况**：
+- Skill 中完全没有提到该知识点或操作步骤
+- Skill 中的描述不够清晰或不够详细，导致 Agent 无法正确理解
+- Skill 中缺少关键的参数说明、工具使用方法或注意事项
+- Skill 中的信息过时或与实际情况不符
+- Agent 的执行步骤与 Skill 定义不一致，但 Skill 的指导本身存在问题（如步骤顺序不合理、缺少前置条件说明等）
+
+**不是 Skill 问题的情况**：
+- Skill 中已经清楚地包含了相关信息，但 Agent 没有正确使用（模型能力问题）
+- 该评分项要求的是通用知识，不在 Skill 的职责范围内
+- 该评分项涉及的是运行时环境问题（网络、权限等）
+- Agent 的执行过程偏离了 Skill 的指导，但 Skill 的指导本身是正确和清晰的
+
+## 分析步骤
+
+1. **阅读 Skill 定义**：理解 Skill 规定的操作步骤和知识要点
+2. **查看交互历史**：观察 Agent 实际执行了哪些操作
+3. **对比分析**：
+   - Agent 的执行是否遵循了 Skill 的指导？
+   - 如果没有遵循，是因为 Skill 不清晰，还是 Agent 自身的问题？
+   - 评分项要求的内容在 Skill 中是否有足够的覆盖？
+4. **得出结论**：判断是否是 Skill 问题，并给出改进建议
+
+## 输出格式
+
+请以 JSON 格式返回：
+{
+    "id": "评分项ID",
+    "is_skill_issue": true或false,
+    "reasoning": "使用中文详细解释判断依据，需要：1) 引用 Skill 定义中的相关内容 2) 说明 Agent 执行过程中的表现 3) 解释为什么是/不是 Skill 问题",
+    "improvement_suggestion": "如果是 Skill 问题，给出具体的改进建议，说明应该在 Skill 中添加或修改什么内容；否则留空或不填"
+}
+`;function n(e,n,s,i,l){return`${t}
+
+---
+
+## 分析材料
+
+### [当前 Skill 定义]
+\`\`\`
+${e}
+\`\`\`
+
+### [用户问题]
+${s}
+
+### [Agent 执行过程 - 完整交互历史]
+\`\`\`
+${l}
+\`\`\`
+
+### [Agent 最终回答]
+${i}
+
+### [待分析的评分项]
+- **ID**: ${n.id}
+- **类型**: ${"root_cause"===n.type?"Root Cause (根因分析)":"Key Action (关键操作)"}
+- **评分标准**: ${n.content}
+- **得分**: ${(100*n.match_score).toFixed(0)}% (满分 100%)
+- **扣分原因**: ${n.explanation}
+- **权重**: ${n.weight}
+
+---
+
+请根据以上材料，判断此评分项的扣分是否可以通过优化 Skill 定义来改善，并输出 JSON 结果：`}e.s(["SKILL_ISSUE_ANALYSIS_PROMPT",0,t,"generateSkillIssuePrompt",()=>n])},44087,e=>{"use strict";let t=`
+你是一位专家级的日志分析师。你的任务是分析用户与 AI 助手之间的对话历史（包括工具输出），并提取任何“中间故障”或“异常过程”。
+
+“中间故障”或“异常过程”定义如下：
+1. **工具执行错误**：代理尝试运行工具（例如 bash 命令、python 脚本）但失败了（非零退出代码、堆栈跟踪、错误消息）。
+2. **逻辑/推理修正**：代理意识到自己犯了错误并明确纠正自己（例如，“我犯了一个错误...”，“之前的方法失败了...”）。
+3. **超时/卡住**：代理提到等待太久或进程卡住。
+4. **无效参数**：代理尝试使用带有无效参数的工具并被系统拒绝。
+
+你将获得包含完整对话历史的最后一次交互内容。
+
+逐步分析历史记录。对于发现的每个故障，提取：
+- failure_type: (Tool Error / Reasoning Error / Timeout / Invalid Usage)
+- description: 用中文简要总结出了什么问题。
+- context: 导致失败的具体命令或推理内容。
+- recovery: 代理如何尝试恢复（如果有）。
+
+仅以以下 JSON 格式响应：
+{
+  "failures": [
+    {
+      "failure_type": "Tool Error",
+      "description": "无法安装包 'xyz'",
+      "context": "pip install xyz",
+      "recovery": "代理尝试改用 apt-get。"
+    },
+    ...
+  ]
+}
+
+如果未发现故障，请返回：
+{
+  "failures": []
+}
+`;function n(e){return`
+${t}
+
+Conversation History:
+${e}
+`}e.s(["FAILURE_EXTRACTION_PROMPT",0,t,"generateFailureAnalysisPrompt",()=>n])},18868,e=>{"use strict";let t=`
+你是一位专家级的日志分析师。你的任务是分析用户与 AI 助手之间的完整对话历史，并精准提取本次会话的**最终执行结果**。
+
+**提取规则：**
+1. **必须完整保留**：如果最终结果分散在多个连续的回复中（例如先输出长篇详细报告，紧接着输出总结摘要），必须将它们**全部提取并按顺序拼接**，形成一个完整的报告。
+2. **剔除过程噪声**：只提取最终交付给用户的有效信息（如分析报告、代码方案、问题解答）。必须剔除之前的思考过程、工具调用命令（如 bash/python）、工具执行日志、自我纠正的尝试等中间过程。
+3. **保持原样**：提取的内容必须**完全保持原样（Verbatim）**，不要进行总结、改写或缩减。包括 Markdown 格式（标题、列表、代码块）都要保留。
+4. **识别头部**：通常最终结果会以明确的标题开始（如 "# 分析报告"、"## 结论"），或者直接给出用户请求的答案。
+
+**输入数据：**
+完整的对话历史记录。
+
+**输出格式：**
+请直接返回提取后的最终结果内容。不要包含 JSON 格式，不要包含 "Here is the result:" 等废话，由于内容可能包含 Markdown，请确保输出是纯文本格式的 Markdown 内容。
+`;function n(e){return`
+${t}
+
+**对话历史：**
+${e}
+
+**请提取最终结果：**
+`}e.s(["EXTRACTION_PROMPT",0,t,"generateExtractionPrompt",()=>n])},85874,e=>e.a(async(t,n)=>{try{var s=e.i(22734);e.i(89228);var i=e.i(88045),l=e.i(14747),r=e.i(74008),o=e.i(22965),a=e.i(50374),c=t([o]);[o]=c.then?(await c)():c;let k=l.default.join(process.cwd(),"data","model_debug.jsonl");function u(e,t,n){try{let i=l.default.dirname(k);s.default.existsSync(i)||s.default.mkdirSync(i,{recursive:!0});let r={timestamp:new Date().toISOString(),stage:e,input:t,output:n};s.default.appendFileSync(k,JSON.stringify(r)+"\n")}catch(e){console.error("Log error",e)}}async function f(e){let t=await (0,o.getActiveConfig)(e);if(!t)return{client:null,model:null};let n=t.apiKey||"no-api-key-required",s=t.baseUrl||"https://api.deepseek.com",{customFetch:l}=(0,r.getProxyConfig)();return{client:new i.OpenAI({apiKey:n,baseURL:s,fetch:l}),model:t.model||"deepseek-chat"}}async function g(t,n,s,i){let{client:l,model:r}=await f(i);if(!l||!l.apiKey)return console.warn("LLM Evaluation disabled or missing config. Skipping."),{is_correct:!1,score:0,reason:"请在首页左上角的设置中配置 LLM"};try{let i=n.root_causes||[],o=n.key_actions||[],a=i.map((e,t)=>({id:`RC-${t}`,...e})),c=o.map((e,t)=>({id:`KA-${t}`,...e})),{generateJudgePrompt:f}=e.r(62414),g=f(t,s,a,c,n.skill_definition),m=await l.chat.completions.create({messages:[{role:"user",content:g}],model:r});console.log(`[Judge API Debug] Model: ${r}. Received response choices:`,m?.choices?.length);let d=m.choices?.[0]?.message?.content;if(u("result_evaluation",{prompt:g},{raw_output:d}),!d)throw console.error("\n[Judge API Error 🚨] LLM content is empty or undefined!"),console.error(">>> Full LLM Response:"),console.error(JSON.stringify(m,null,2)),console.error("<<<\n"),Error("No content from evaluation model");let h=d.trim(),p=h.match(/```(?:json)?\\s*([\\s\\S]*?)\\s*```/i);if(p)h=p[1];else{let e=h.indexOf("{"),t=h.lastIndexOf("}");-1!==e&&-1!==t&&t>=e&&(h=h.substring(e,t+1))}let y=JSON.parse(h).evaluations||[],k=0,S=0,A=[];a.forEach(e=>{let t=y.find(t=>t.id===e.id),n=t?Math.max(0,Math.min(1,Number(t.match_score))):0,s=t?.explanation||"未找到评分结果";k+=n*e.weight,S+=e.weight,A.push(`1. **Root Cause** [${e.content.replace(/\n/g," ")}]: ${(100*n).toFixed(0)}% match. ${s} (Weight: ${e.weight})`)}),c.forEach(e=>{let t=y.find(t=>t.id===e.id),n=t?Math.max(0,Math.min(1,Number(t.match_score))):0,s=t?.explanation||"未找到评分结果";k+=n*e.weight,S+=e.weight,A.push(`2. **Key Action** [${e.content.replace(/\n/g," ")}]: ${(100*n).toFixed(0)}% match. ${s} (Weight: ${e.weight})`)});let $=0;$=S>0?k/S:0;let _=A.map(e=>{let t=e.match(/(\d+)% match/),n=e.match(/\(Weight: ([\d\.]+)\)/);if(t&&n){let e=(parseInt(t[1])/100).toFixed(1),s=parseFloat(n[1]).toFixed(1);return`${e}*${s}`}return null}).filter(e=>e).join(" + "),w=A.map(e=>{let t=e.match(/\(Weight: ([\d\.]+)\)/);return t?parseFloat(t[1]).toFixed(1):null}).filter(e=>e).join(" + "),N=`**Calculation**: (${_}) / (${w}) = ${k.toFixed(2)} / ${S.toFixed(2)} = ${$.toFixed(2)}`;A.push("",N);let x=A.join("\n");return{is_correct:$>=.8,score:$,reason:x}}catch(e){return console.error("LLM Judgment Error:",e),{is_correct:!1,score:0,reason:"Judgment API failed"}}}async function m(t,n,s,i,l,r){console.log("[ItemAttribution] Starting analysis...");let{client:o,model:a}=await f(r);if(!o||!t||!n)return console.warn(`[ItemAttribution] ✗ Early return: client=${o?"present":"absent"}, skillDef=${t?"present":"absent"}, judgmentReason=${n?"present":"absent"}`),[];let c=function(e){let t=[];if(!e)return t;let n=e.split("\n"),s={rc:0,ka:0};for(let e of n){let n=e.match(/\*\*Root Cause\*\*\s*\[(.*?)\]\s*.*?:\s*(\d+)%\s*match\.\s*(.+?)\s*\(Weight:\s*([\d.]+)\)/);if(n){t.push({id:`RC-${s.rc++}`,type:"root_cause",content:n[1].replace(/\.{3}$/,""),match_score:parseInt(n[2])/100,explanation:n[3].trim(),weight:parseFloat(n[4])});continue}let i=e.match(/\*\*Key Action\*\*\s*\[(.*?)\]\s*.*?:\s*(\d+)%\s*match\.\s*(.+?)\s*\(Weight:\s*([\d.]+)\)/);i&&t.push({id:`KA-${s.ka++}`,type:"key_action",content:i[1].replace(/\.{3}$/,""),match_score:parseInt(i[2])/100,explanation:i[3].trim(),weight:parseFloat(i[4])})}return t}(n);console.log(`[ItemAttribution] Parsed ${c.length} evaluation items from judgment reason`);let g=c.filter(e=>e.match_score<1);if(console.log(`[ItemAttribution] ${g.length} items need analysis (score < 100%)`),0===g.length)return console.log(`[ItemAttribution] ✗ No imperfect items found, returning empty array`),[];let{generateSkillIssuePrompt:m}=e.r(33526),d=[];for(let e of g)try{let n=m(t,e,s,i,l),r=await o.chat.completions.create({messages:[{role:"user",content:n}],model:a}),c=r.choices?.[0]?.message?.content;if(!c){console.error(`
+[SkillAnalysis API Error 🚨] LLM content is empty for item: ${e.id}`),console.error(">>> Full LLM Response:"),console.error(JSON.stringify(r,null,2)),console.error("<<<\n");continue}u("skill_issue_analysis",{item_id:e.id,prompt:n},{raw_output:c});let f=c.trim(),g=f.match(/```(?:json)?\\s*([\\s\\S]*?)\\s*```/i);if(g)f=g[1];else{let e=f.indexOf("{"),t=f.lastIndexOf("}");-1!==e&&-1!==t&&t>=e&&(f=f.substring(e,t+1))}let h=JSON.parse(f);!0===h.is_skill_issue?(d.push({...e,is_skill_issue:!0,reasoning:h.reasoning??"",improvement_suggestion:h.improvement_suggestion}),console.log(`[SkillAnalysis] ${e.id}: IS Skill Issue - ${h.reasoning?.substring(0,50)}...`)):console.log(`[SkillAnalysis] ${e.id}: NOT Skill Issue - ${h.reasoning?.substring(0,50)}...`)}catch(t){console.error(`[SkillAnalysis] Error analyzing ${e.id}:`,t.message),u("skill_issue_analysis_error",{item_id:e.id},{error:t.message})}return d}async function d(t,n,s,i,l,r,o,a){let c=p(t);if(!c||0===c.length)return{failures:[]};try{let{client:t,model:g}=await f(a);if(!t||!t.apiKey)return console.warn("LLM Analysis disabled. Skipping."),{failures:[],skill_issues:[]};let d="",h=c[c.length-1];(h.requestMessages||[]).forEach(e=>{let t="";"string"==typeof e.content?t=e.content:Array.isArray(e.content)&&(t=JSON.stringify(e.content)),d+=`[${(e.role||"UNKNOWN").toUpperCase()}]: ${t}
+`});let p=h.responseMessage;if(p){let e="";"string"==typeof p.content?e=p.content:Array.isArray(p.content)&&(e=JSON.stringify(p.content)),d+=`[ASSISTANT]: ${e}
+`}else h.debug_raw_stream&&(d+=`[SYSTEM/TOOL OUTPUTS]: (Check raw logs for full details)
+`);let{generateFailureAnalysisPrompt:y}=e.r(44087),k=y(d),S=await t.chat.completions.create({messages:[{role:"user",content:k}],model:g}),A=S.choices?.[0]?.message?.content;if(!A)return console.error("\n[Failure Analysis API Error 🚨] LLM content is empty!"),console.error(">>> Full LLM Response:"),console.error(JSON.stringify(S,null,2)),console.error("<<<\n"),{failures:[]};u("failure_analysis",{prompt:k,history_length:d.length},{raw_output:A});let $=A.trim(),_=$.match(/```(?:json)?\\s*([\\s\\S]*?)\\s*```/i);if(_)$=_[1];else{let e=$.indexOf("{"),t=$.lastIndexOf("}");-1!==e&&-1!==t&&t>=e&&($=$.substring(e,t+1))}let w=JSON.parse($).failures||[],N=[];if(console.log(`[SkillAnalysis] Checking: skillName=${n||"none"}, skillDef=${s?"present":"absent"}, failuresCount=${w.length}, answerScore=${i}`),n&&s){console.log(`[SkillAnalysis] ✓ Skill name and definition found`);let e=o||"(见交互历史)",t=r||"(未知)";console.log(`[SkillAnalysis] Conditions check: answerScore=${i}, judgmentReason=${l?"present":"absent"}, history=${d?"present":"absent"}`),void 0!==i&&i<1&&l&&d?(console.log(`[SkillAnalysis] ✓ All conditions met. Score is imperfect (${i}). Analyzing which items are Skill issues...`),console.log(`[SkillAnalysis] Using: query=${t.substring(0,50)}..., answer=${e.substring(0,50)}..., history_len=${d.length}`),N=await m(s,l,t,e,d,a),console.log(`[SkillAnalysis] Analysis complete: ${N.length} items identified as Skill issues`)):void 0!==i&&i>=1?console.log(`[SkillAnalysis] ✗ Skipped: Perfect score (${i}). No Skill analysis needed.`):console.log(`[SkillAnalysis] ✗ Skipped: Conditions not met - answerScore=${i}, judgmentReason=${l?"present":"absent"}, history=${d?"present":"absent"}`)}else console.warn(`[SkillAnalysis] ✗ Skipped: Missing skillName (${n||"none"}) or skillDef (${s?"present":"absent"})`);return{failures:w,skill_issues:N.length>0?N:void 0}}catch(e){return console.error("Failure Analysis Error:",e),{failures:[]}}}function h(e){if(!e||"string"!=typeof e)return!1;try{let t=JSON.parse(e);if(t&&"object"==typeof t)return!0}catch(e){}return!1}function p(e){return(0,a.normalizeInteractions)(e)}async function y(t,n){let s=p(t);if(!s||0===s.length)return{query:"",skill:"",final_result:""};let i="",l="",r="";for(let e of s){let t=(e.requestMessages||[]).find(e=>"user"===e.role);if(t){let e="";if("string"==typeof t.content)e=t.content;else if(Array.isArray(t.content)){let n=t.content.filter(e=>"text"===e.type&&!e.text.trim().startsWith("<system-reminder>")&&!e.text.includes("[SUGGESTION MODE:"));n.length>0&&(e=n.map(e=>e.text).join("\n").trim())}if(e){let t=(e=e.replace(/<EXTRA_INFO>[\s\S]*?<\/EXTRA_INFO>/g,"")).trim();if("count"===t||"ping"===t||t.startsWith("Please write a 5-10 word title")||"hi"===t||"hello"===t&&s.length>2)continue;if(i||(i=t),i)break}}}for(let e=s.length-1;e>=0;e--){let t=s[e];if((t.requestMessages||[]).some(e=>"string"==typeof e.content?e.content.includes("[SUGGESTION MODE:"):!!Array.isArray(e.content)&&e.content.some(e=>e.text&&e.text.includes("[SUGGESTION MODE:"))))continue;let n=t.responseMessage,i="";if(n&&n.content&&"string"==typeof n.content)i=n.content;else if(n&&Array.isArray(n.content)){let e=n.content.find(e=>"text"===e.type);e&&(i=e.text)}if((!i||""===i.trim())&&t.debug_raw_stream)try{let e="";for(let n of t.debug_raw_stream)n.choices&&n.choices[0]&&n.choices[0].delta&&n.choices[0].delta.content&&(e+=n.choices[0].delta.content),"content_block_delta"===n.type&&n.delta&&"text_delta"===n.delta.type&&(e+=n.delta.text);e&&(i=e)}catch(e){}let r=[];if(t.requestMessages&&t.requestMessages.length>0)for(let e=t.requestMessages.length-1;e>=0;e--){let n=t.requestMessages[e];if("assistant"===n.role){let e="";if("string"==typeof n.content)e=n.content;else if(Array.isArray(n.content)){let t=n.content.find(e=>"text"===e.type);t&&(e=t.text)}if(e&&e.trim()){let t=e.trim();if(!h(t)&&(r.unshift(t),t.startsWith("#")||t.startsWith("##")||t.startsWith("###")))break}else break}else break}if(i&&i.trim()){let e=i.trim();!h(e)&&(r.push(e),e.startsWith("#")||e.startsWith("##")||e.startsWith("###"))}if(r.length>0){l=r.join("\n\n");break}}let o=new Set,a="";for(let e of s){(e.requestMessages||[]).forEach(e=>{"string"==typeof e.content?a+=e.content+"\n":Array.isArray(e.content)&&e.content.forEach(e=>{"text"===e.type&&(a+=e.text+"\n")})});let t=e.responseMessage;t&&("string"==typeof t.content?a+=t.content+"\n":Array.isArray(t.content)&&t.content.forEach(e=>{"text"===e.type&&(a+=e.text+"\n")}),t.tool_calls&&Array.isArray(t.tool_calls)&&t.tool_calls.forEach(e=>{let t=e.function?.name??e.name;if(t)if("skill"===t||"load_skill"===t)try{let t=e.function?.arguments??e.arguments,n="string"==typeof t?JSON.parse(t):t,s=n?.name??n?.skill_name??n?.skillName??n?.skill;s&&o.add(s)}catch(e){}else o.add(t)}),t.function_call&&t.function_call.name&&o.add(t.function_call.name))}let c=a.match(/(?:Loading skill|Load skill)[:\s]+([a-zA-Z0-9_\-\.]+)/i);c&&c[1]&&o.add(c[1].trim());let g=a.match(/Skill\s+([a-zA-Z0-9_\-\.]+)\s+loaded/i);g&&g[1]&&o.add(g[1].trim()),o.size>0&&(r=Array.from(o).join(", ")),console.log(`[Rule-Based Analysis] Query: ${i.substring(0,20)}..., Skill: ${r}, Result Length: ${l.length}`),u("extraction",{messages_summary:`Total ${s.length} interactions`,notes:"Scanned full history"},{query:i,skill:r,final_result:l});let m="";try{let{client:t,model:i}=await f(n);if(t&&t.apiKey){let n="";s.forEach(e=>{(e.requestMessages||[]).forEach(e=>{let t="";if("string"==typeof e.content)t=e.content;else if(Array.isArray(e.content)){let n=e.content.find(e=>"text"===e.type);n&&(t=n.text)}n+=`[${(e.role||"UNKNOWN").toUpperCase()}]: ${t}
+`});let t=e.responseMessage;if(t){let e="";if("string"==typeof t.content)e=t.content;else if(Array.isArray(t.content)){let n=t.content.find(e=>"text"===e.type);n&&(e=n.text)}n+=`[ASSISTANT]: ${e}
+`}});let{generateExtractionPrompt:l}=e.r(18868),r=l(n),o=(await t.chat.completions.create({messages:[{role:"user",content:r}],model:i,temperature:.1})).choices[0].message.content;o&&(m=o.trim(),u("result_extraction_llm",{history_length:n.length},{extracted:m}))}}catch(e){console.error("LLM Extraction Failed",e)}return m&&m.length>20?(console.log(`[Judge] LLM extraction preferred (Length: ${m.length} vs Rule: ${l.length})`),l=m):l&&console.log(`[Judge] Rule-Based extraction used (Length: ${l.length})`),{query:i,skill:r,final_result:l}}e.s(["analyzeFailures",()=>d,"analyzeSession",()=>y,"judgeAnswer",()=>g,"normalizeInteractions",()=>p]),n()}catch(e){n(e)}},!1)];
+
+//# sourceMappingURL=src_98433cb8._.js.map

package/.next/standalone/.next/server/chunks/src_lib_12408140._.js

@@ -1,3 +1,3 @@
-module.exports=[11811,e=>{"use strict";var t=e.i(22734),l=e.i(14747);let i={"claude-opus-4-6":{inputTokenPrice:5,outputTokenPrice:25,cacheReadInputTokenPrice:.5,cacheCreationInputTokenPrice:6.25},"claude-sonnet-4-6":{inputTokenPrice:3,outputTokenPrice:15,cacheReadInputTokenPrice:.3,cacheCreationInputTokenPrice:3.75},"deepseek-chat":{inputTokenPrice:.28,outputTokenPrice:.42,cacheReadInputTokenPrice:.028},"deepseek-reasoner":{inputTokenPrice:.28,outputTokenPrice:.42,cacheReadInputTokenPrice:.028},"minimax-m2.5-free":{inputTokenPrice:0,outputTokenPrice:0}},n={"claude-opus-4-6":1e6,"claude-sonnet-4-6":1e6,"deepseek-chat":128e3,"deepseek-reasoner":128e3,"minimax-m2.5-free":196608},s=l.default.join(process.cwd(),"custom-models.json"),o={},r={},a=-1;function u(){try{let e=t.default.statSync(s).mtimeMs;if(e===a)return{pricing:o,contextWindows:r};let l=JSON.parse(t.default.readFileSync(s,"utf-8")),i={},n={};for(let[e,t]of Object.entries(l))e.startsWith("_")||("number"==typeof t.inputTokenPrice&&"number"==typeof t.outputTokenPrice&&(i[e]=t),"number"==typeof t.contextWindow&&(n[e]=t.contextWindow));o=i,r=n,a=e}catch(e){e instanceof SyntaxError&&console.warn("[model-config] Failed to parse custom-models.json:",e.message),o={},r={},a=-1}return{pricing:o,contextWindows:r}}function c(e,t){if(t[e])return t[e];for(let[l,i]of Object.entries(t).sort((e,t)=>t[0].length-e[0].length))if(e.startsWith(l))return i;return null}function d(e){let{pricing:t}=u(),l=c(e,t);if(l)return{pricing:l,source:"custom"};let n=c(e,i);return n?{pricing:n,source:"default"}:null}function k(e){let{contextWindows:t}=u(),l=c(e,t);if(null!=l)return{contextWindow:l,source:"custom"};let i=c(e,n);return null!=i?{contextWindow:i,source:"default"}:null}function _(e,t,l,i,n){let s=l.cacheReadInputTokenPrice??.1*l.inputTokenPrice,o=l.cacheCreationInputTokenPrice??1.25*l.inputTokenPrice;return(e*l.inputTokenPrice+(i??0)*s+(n??0)*o+t*l.outputTokenPrice)/1e6}e.s(["DEFAULT_CACHE_CREATION_RATIO",0,1.25,"DEFAULT_CACHE_READ_RATIO",0,.1,"calculateCost",()=>_,"getModelContextWindow",()=>k,"getModelPricing",()=>d])},85765,e=>e.a(async(t,l)=>{try{var i=e.i(22734),n=e.i(14747),s=e.i(31390),o=e.i(98043),r=e.i(11811),a=t([s,o]);function u(e){let t=e.trim(),l=[['"','"'],["'","'"],["“","”"],["‘","’"],["`","`"],["《","》"],["（","）"],["(",")"],["【","】"],["[","]"],["{","}"],["<",">"]];for(let e=0;e<6;e++){let e=t;for(let[e,i]of(t=t.trim(),l))t.startsWith(e)&&t.endsWith(i)&&t.length>=e.length+i.length+1&&(t=t.slice(e.length,-i.length));if(t===e)break}return t=(t=(t=t.replace(/[\s"'“”‘’`。.]/g,"")).replace(/^[\s.,，。!?！？;；:：、·…]+|[\s.,，。!?！？;；:：、·…]+$/g,"")).replace(/\s+/g," ").trim()}function c(e,t){if(!t)return;let l=u(t);if(!l)return;let i=e.filter(e=>e.query&&e.query.trim()).filter(e=>{let t=u(e.query);return!!t&&l.endsWith(t)});if(0!==i.length)return i.reduce((e,t)=>{let l=u(e.query).length;return u(t.query).length>l?t:e})}[s,o]=a.then?(await a)():a;let p=n.default.join(process.cwd(),"data"),f=n.default.join(p,"evaluation_result.json");async function d(e,t){let l={};if(e&&(l.OR=[{user:e},{user:null}]),!t?.query&&t?.taskId){let e=await o.db.findExecutionById(t.taskId);e&&e.query?(l.query=e.query,t.framework&&(l.framework=t.framework)):l.id=t.taskId}else t?.query&&(l.query=t.query,t.framework&&(l.framework=t.framework));t?.skill!==void 0&&(l.skill=t.skill),t?.skillVersion!==void 0&&(l.skillVersion=t.skillVersion);let i=await o.db.findExecutions(l,{timestamp:"desc"}),n=new Map;for(let e of i){let t=e.taskId||null;t&&(n.has(t)||n.set(t,[]),n.get(t).push(e))}let s=new Set;for(let[e,t]of n.entries()){if(1===t.length){s.add(t[0].id);continue}let l=t.find(t=>t.id===e);if(l){s.add(l.id);continue}let i=t.slice().sort((e,t)=>{let l=new Date(e.timestamp).getTime(),i=new Date(t.timestamp).getTime();if(i!==l)return i-l;let n=String(e.finalResult||"").length;return String(t.finalResult||"").length-n});s.add(i[0].id)}let a=i.filter(e=>!e.taskId||s.has(e.id));for(let[e,t]of n.entries())if(!(t.length<=1))for(let e of t)s.has(e.id)||o.db.deleteExecution(e.id).catch(()=>{});return a.map(e=>{let t=e.model??null,l=t?(0,r.getModelPricing)(t):null,i=l?.pricing??null,n=t&&null!=e.maxSingleCallTokens?(0,r.getModelContextWindow)(t):null;return{...e,upload_id:e.id,task_id:e.taskId||void 0,query:e.query||void 0,framework:e.framework||void 0,tokens:e.tokens||void 0,cost:i&&null!=e.inputTokens&&null!=e.outputTokens?(0,r.calculateCost)(e.inputTokens,e.outputTokens,i,e.cacheReadInputTokens??void 0,e.cacheCreationInputTokens??void 0):void 0,latency:e.latency||void 0,timestamp:e.timestamp?.toISOString?.()||e.timestamp,final_result:e.finalResult||void 0,skill:e.skill||void 0,skills:e.skills?JSON.parse(e.skills):void 0,invokedSkills:e.invokedSkills?JSON.parse(e.invokedSkills):void 0,is_skill_correct:e.isSkillCorrect||!1,is_answer_correct:e.isAnswerCorrect||!1,answer_score:void 0!==e.answerScore?e.answerScore:void 0,skill_score:void 0!==e.skillScore?e.skillScore:void 0,judgment_reason:e.judgmentReason||void 0,failures:e.failures?JSON.parse(e.failures):void 0,label:e.label??null,user:e.user??null,skill_issues:e.skillIssues?JSON.parse(e.skillIssues):[],skill_version:e.skillVersion??null,model:t,tool_call_count:e.toolCallCount??void 0,llm_call_count:e.llmCallCount??void 0,input_tokens:e.inputTokens??void 0,output_tokens:e.outputTokens??void 0,tool_call_error_count:e.toolCallErrorCount??void 0,cache_read_input_tokens:e.cacheReadInputTokens??void 0,cache_creation_input_tokens:e.cacheCreationInputTokens??void 0,max_single_call_tokens:e.maxSingleCallTokens??void 0,expected_skill_version:e.expectedSkillVersion??null,skill_recall_rate:e.skillRecallRate??null,context_window_pct:null!=e.maxSingleCallTokens&&n?Math.round(e.maxSingleCallTokens/n.contextWindow*1e3)/10:void 0,context_window_limit:n?.contextWindow,context_window_source:n?.source,cost_pricing:i?{inputTokenPrice:i.inputTokenPrice,outputTokenPrice:i.outputTokenPrice,cacheReadInputTokenPrice:i.cacheReadInputTokenPrice??i.inputTokenPrice*r.DEFAULT_CACHE_READ_RATIO,cacheCreationInputTokenPrice:i.cacheCreationInputTokenPrice??i.inputTokenPrice*r.DEFAULT_CACHE_CREATION_RATIO,source:l?.source??"default"}:null}})}async function k(e){let t={};return e&&(t.OR=[{user:e},{user:null}]),(await o.db.findConfigs(t)).map(e=>{let t=(t,l)=>{if(t)try{return JSON.parse(t)}catch(t){console.error(`[readConfig] Failed to parse ${l} for config ${e.id}:`,t);return}};return{id:e.id,query:e.query,skill:e.skill,skillVersion:e.skillVersion,expectedSkills:t(e.expectedSkills,"expectedSkills"),standard_answer:e.standardAnswer,root_causes:t(e.rootCauses,"rootCauses"),key_actions:t(e.keyActions,"keyActions"),parse_status:e.parseStatus||"completed"}})}async function _(e){let t=e.upload_id||e.task_id||crypto.randomUUID();if(e.task_id)try{let l={taskId:e.task_id};e.framework&&(l.framework=e.framework);let i=await o.db.findExecutions(l,{timestamp:"desc"});if(i&&i.length>0&&i[0]?.id){let l=i.find(t=>t.id===e.task_id),n=l&&l.id?l.id:i[0].id;n!==t&&(t=n)}}catch{}let l=null,n=await o.db.findExecutionById(t);n&&(l={...n,upload_id:n.id,task_id:n.taskId||void 0,query:n.query||void 0,framework:n.framework||void 0,tokens:n.tokens||void 0,cost:n.cost||void 0,latency:n.latency||void 0,timestamp:n.timestamp?.toISOString?.()||n.timestamp,final_result:n.finalResult||void 0,skill:n.skill||void 0,skills:n.skills?JSON.parse(n.skills):void 0,invokedSkills:n.invokedSkills?(()=>{try{return JSON.parse(n.invokedSkills)}catch{return}})():void 0,is_skill_correct:n.isSkillCorrect||!1,is_answer_correct:n.isAnswerCorrect||!1,answer_score:n.answerScore||void 0,skill_score:n.skillScore||void 0,judgment_reason:n.judgmentReason||void 0,failures:n.failures?JSON.parse(n.failures):void 0,skill_issues:n.skillIssues?JSON.parse(n.skillIssues):void 0,label:n.label||void 0,user:n.user||void 0,skill_version:n.skillVersion??void 0,expected_skill_version:n.expectedSkillVersion??null,skill_recall_rate:n.skillRecallRate??null,model:n.model||void 0,tool_call_count:n.toolCallCount??void 0,llm_call_count:n.llmCallCount??void 0,input_tokens:n.inputTokens??void 0,output_tokens:n.outputTokens??void 0,tool_call_error_count:n.toolCallErrorCount??void 0,cache_read_input_tokens:n.cacheReadInputTokens??void 0,cache_creation_input_tokens:n.cacheCreationInputTokens??void 0,max_single_call_tokens:n.maxSingleCallTokens??void 0});let r=l?{...l}:{},a=!!l;a||r.timestamp||e.timestamp?e.timestamp&&(r.timestamp=e.timestamp):r.timestamp=new Date().toISOString();let u=!!e.force_query_update,d="string"==typeof l?.query?l.query.trim():"",_="string"==typeof e.query?e.query.trim():"";if(r={...r,...e},d&&!u?r.query=d:!d&&_?r.query=_:"string"!=typeof r.query||r.query.trim()?"string"==typeof r.query&&(r.query=r.query.trim()):r.query=void 0,!r.upload_id&&r.task_id&&(r.upload_id=r.task_id),!r.task_id&&r.upload_id&&(r.task_id=r.upload_id),r.upload_id=t,(!r.label||!r.model||!r.user)&&r.task_id){let e=await o.db.findSessionByTaskId(r.task_id);e&&(!r.label&&e.label&&(r.label=e.label),!r.model&&e.model&&(r.model=e.model),!r.user&&e.user&&(r.user=e.user))}if(!r.user)try{let e=o.db.getClient();if("query"in e){let t=await e.query('SELECT username FROM "User" LIMIT 1');t.rows[0]&&(r.user=t.rows[0].username,console.log(`[Data-Service] Fallback resolved user for task ${r.task_id} to: ${r.user}`))}}catch(e){console.warn("[Data-Service] Fallback user lookup failed:",e)}let p=e.Token||e.token||e.tokens;void 0!==p&&(r.tokens=Number(p)),void 0!==e.tool_call_count&&(r.tool_call_count=Number(e.tool_call_count)),void 0!==e.llm_call_count&&(r.llm_call_count=Number(e.llm_call_count)),void 0!==e.input_tokens&&(r.input_tokens=Number(e.input_tokens)),void 0!==e.output_tokens&&(r.output_tokens=Number(e.output_tokens)),void 0!==e.tool_call_error_count&&(r.tool_call_error_count=Number(e.tool_call_error_count)),void 0!==e.cache_read_input_tokens&&(r.cache_read_input_tokens=Number(e.cache_read_input_tokens)),void 0!==e.cache_creation_input_tokens&&(r.cache_creation_input_tokens=Number(e.cache_creation_input_tokens)),void 0!==e.max_single_call_tokens&&(r.max_single_call_tokens=Number(e.max_single_call_tokens));let m="未找到匹配的评测配置",v=!1,y=r.is_answer_correct||!1,g=r.judgment_reason||m,w=await k(r.user);if(r.query&&w.length>0){let t=c(w,r.query);if(t){let i=Array.isArray(r.invokedSkills)?r.invokedSkills:[],n=(Array.isArray(r.skills)?r.skills:[]).map(e=>({name:e,version:null})),u=t.expectedSkills||[];if(u.length>0){let e=i.length>0?i:n;if(e.length>0){let t=0,l=u.filter(e=>e.skill?.trim()),i=l.map(e=>e.skill.trim()),n=new Map;if(i.length>0)try{for(let e of(await o.db.findSkills({name:{in:i},user:r.user||null})))n.set(e.name,e)}catch(e){console.error("[Judgment] Error fetching skills for version check:",e)}for(let i of l){let l=i.skill.trim(),s=i.version??null,o=e.find(e=>e.name===l);if(o){let e=!1;if(null===s)e=!0;else if(null!==o.version)e=o.version===s;else{let t=n.get(l);e=!!t&&(t.activeVersion||0)===s}e&&(t++,v||(v=!0))}}l.length>0&&(r.skill_recall_rate=t/l.length)}}if(r.is_skill_correct=v,void 0!==r.final_result){let i=!0;if(a&&!e.force_judgment&&l&&l.query===r.query&&l.final_result===r.final_result&&(i=!1),i&&!r.skip_evaluation){let e,l=(r.skill||t.skill||"").trim();if(l)try{let t=await o.db.findSkill(l,r.user||null);if(t){let l=t.activeVersion||0,i=t.versions?.find(e=>e.version===l);if(i&&i.content)e=i.content,r.skill_version=i.version;else if(t.versions&&t.versions.length>0){let l=t.versions[0];l&&l.content&&(e=l.content,r.skill_version=l.version)}}}catch(e){console.error("[Judgment] Error fetching skill definition:",e)}let i=await (0,s.judgeAnswer)(r.query||"",{standard_answer_example:t.standard_answer,root_causes:t.root_causes,key_actions:t.key_actions,skill_definition:e},r.final_result,r.user);y=i.is_correct,r.answer_score=i.score,g=i.reason||"Judged by Evaluation Model"}}}else a&&!e.force_judgment||r.answer_score||(y=!1,g=m,r.answer_score=null)}else r.query&&(!a||e.force_judgment)&&!r.answer_score&&(y=!1,g=m,r.answer_score=null);e.skip_evaluation&&(r.answer_score=null,g="结果评估中..."),r.is_skill_correct=v,r.is_answer_correct=y,r.judgment_reason=g;let h=Array.isArray(r.skills)&&r.skills.length>0?r.skills[0]:void 0;if(h){let e=function(){if(!i.default.existsSync(f))return{};try{return JSON.parse(i.default.readFileSync(f,"utf-8"))}catch{return{}}}()[h];e&&(r.skill_score=parseFloat(e))}if(r.skill&&void 0!==r.skill_version&&null!==r.skill_version?r.label=`${r.skill}-v${r.skill_version}`:r.skill?r.label=`${r.skill}-v1`:r.label="without-skill",await o.db.upsertExecution({where:{id:t},create:{id:t,taskId:r.task_id,query:r.query,framework:r.framework,tokens:r.tokens,cost:r.cost,latency:r.latency,timestamp:r.timestamp?new Date(r.timestamp):new Date,finalResult:r.final_result,skill:r.skill,skills:r.skills?JSON.stringify(r.skills):null,invokedSkills:r.invokedSkills?JSON.stringify(r.invokedSkills):null,isSkillCorrect:r.is_skill_correct,isAnswerCorrect:r.is_answer_correct,answerScore:r.answer_score,skillScore:r.skill_score,judgmentReason:r.judgment_reason,failures:r.failures?JSON.stringify(r.failures):null,skillIssues:r.skill_issues?JSON.stringify(r.skill_issues):null,label:r.label,user:r.user,skillVersion:r.skill_version,model:r.model,toolCallCount:r.tool_call_count,llmCallCount:r.llm_call_count,inputTokens:r.input_tokens,outputTokens:r.output_tokens,toolCallErrorCount:r.tool_call_error_count,skillRecallRate:r.skill_recall_rate,cacheReadInputTokens:r.cache_read_input_tokens,cacheCreationInputTokens:r.cache_creation_input_tokens,maxSingleCallTokens:r.max_single_call_tokens},update:{taskId:r.task_id,query:r.query,framework:r.framework,tokens:r.tokens,cost:r.cost,latency:r.latency,timestamp:r.timestamp?new Date(r.timestamp):new Date,finalResult:r.final_result,skill:r.skill,skills:r.skills?JSON.stringify(r.skills):null,invokedSkills:r.invokedSkills?JSON.stringify(r.invokedSkills):null,isSkillCorrect:r.is_skill_correct,isAnswerCorrect:r.is_answer_correct,answerScore:r.answer_score,skillScore:r.skill_score,judgmentReason:r.judgment_reason,failures:r.failures?JSON.stringify(r.failures):null,skillIssues:r.skill_issues?JSON.stringify(r.skill_issues):null,label:r.label,user:r.user,skillVersion:r.skill_version,model:r.model,toolCallCount:r.tool_call_count,llmCallCount:r.llm_call_count,inputTokens:r.input_tokens,outputTokens:r.output_tokens,toolCallErrorCount:r.tool_call_error_count,skillRecallRate:r.skill_recall_rate,cacheReadInputTokens:r.cache_read_input_tokens,cacheCreationInputTokens:r.cache_creation_input_tokens,maxSingleCallTokens:r.max_single_call_tokens}}),e.upload_id&&e.task_id&&e.upload_id!==t)try{let t=await o.db.findExecutionById(e.upload_id);t&&t.taskId===e.task_id&&await o.db.deleteExecution(e.upload_id)}catch{}if(r.task_id&&r.interactions){let e="string"==typeof r.interactions?(()=>{try{return JSON.parse(r.interactions)}catch{return[]}})():r.interactions,t=e;try{let l=await o.db.findSessionByTaskId(r.task_id),i=l?.interactions?(()=>{try{return JSON.parse(l.interactions)}catch{return[]}})():[];Array.isArray(i)&&i.length>0&&(t=!Array.isArray(e)||e.length<i.length?i:e.map((e,t)=>{let l=i[t],n=e?.content===""||e?.content==null,s="string"==typeof l?.content&&l.content.length>0;return n&&s&&l?.role===e?.role?{...e,content:l.content}:e}))}catch{}await o.db.upsertSession(r.task_id,{taskId:r.task_id,query:r.query,label:r.label,user:r.user,model:r.model,interactions:JSON.stringify(t)},{query:r.query,label:r.label,user:r.user,model:r.model,interactions:JSON.stringify(t)})}return{success:!0,record:r}}e.s(["findBestMatchConfig",()=>c,"readConfig",()=>k,"readRecords",()=>d,"saveExecutionRecord",()=>_]),l()}catch(e){l(e)}},!1)];
+module.exports=[11811,e=>{"use strict";var t=e.i(22734),l=e.i(14747);let i={"claude-opus-4-6":{inputTokenPrice:5,outputTokenPrice:25,cacheReadInputTokenPrice:.5,cacheCreationInputTokenPrice:6.25},"claude-sonnet-4-6":{inputTokenPrice:3,outputTokenPrice:15,cacheReadInputTokenPrice:.3,cacheCreationInputTokenPrice:3.75},"deepseek-chat":{inputTokenPrice:.28,outputTokenPrice:.42,cacheReadInputTokenPrice:.028},"deepseek-reasoner":{inputTokenPrice:.28,outputTokenPrice:.42,cacheReadInputTokenPrice:.028},"minimax-m2.5-free":{inputTokenPrice:0,outputTokenPrice:0}},n={"claude-opus-4-6":1e6,"claude-sonnet-4-6":1e6,"deepseek-chat":128e3,"deepseek-reasoner":128e3,"minimax-m2.5-free":196608},o=l.default.join(process.cwd(),"custom-models.json"),s={},r={},a=-1;function u(){try{let e=t.default.statSync(o).mtimeMs;if(e===a)return{pricing:s,contextWindows:r};let l=JSON.parse(t.default.readFileSync(o,"utf-8")),i={},n={};for(let[e,t]of Object.entries(l))e.startsWith("_")||("number"==typeof t.inputTokenPrice&&"number"==typeof t.outputTokenPrice&&(i[e]=t),"number"==typeof t.contextWindow&&(n[e]=t.contextWindow));s=i,r=n,a=e}catch(e){e instanceof SyntaxError&&console.warn("[model-config] Failed to parse custom-models.json:",e.message),s={},r={},a=-1}return{pricing:s,contextWindows:r}}function c(e,t){if(t[e])return t[e];for(let[l,i]of Object.entries(t).sort((e,t)=>t[0].length-e[0].length))if(e.startsWith(l))return i;return null}function d(e){let{pricing:t}=u(),l=c(e,t);if(l)return{pricing:l,source:"custom"};let n=c(e,i);return n?{pricing:n,source:"default"}:null}function k(e){let{contextWindows:t}=u(),l=c(e,t);if(null!=l)return{contextWindow:l,source:"custom"};let i=c(e,n);return null!=i?{contextWindow:i,source:"default"}:null}function _(e,t,l,i,n){let o=l.cacheReadInputTokenPrice??.1*l.inputTokenPrice,s=l.cacheCreationInputTokenPrice??1.25*l.inputTokenPrice;return(e*l.inputTokenPrice+(i??0)*o+(n??0)*s+t*l.outputTokenPrice)/1e6}e.s(["DEFAULT_CACHE_CREATION_RATIO",0,1.25,"DEFAULT_CACHE_READ_RATIO",0,.1,"calculateCost",()=>_,"getModelContextWindow",()=>k,"getModelPricing",()=>d])},85765,e=>e.a(async(t,l)=>{try{var i=e.i(22734),n=e.i(14747),o=e.i(85874),s=e.i(98043),r=e.i(11811),a=t([o,s]);function u(e){let t=e.trim(),l=[['"','"'],["'","'"],["“","”"],["‘","’"],["`","`"],["《","》"],["（","）"],["(",")"],["【","】"],["[","]"],["{","}"],["<",">"]];for(let e=0;e<6;e++){let e=t;for(let[e,i]of(t=t.trim(),l))t.startsWith(e)&&t.endsWith(i)&&t.length>=e.length+i.length+1&&(t=t.slice(e.length,-i.length));if(t===e)break}return t=(t=(t=t.replace(/[\s"'“”‘’`。.]/g,"")).replace(/^[\s.,，。!?！？;；:：、·…]+|[\s.,，。!?！？;；:：、·…]+$/g,"")).replace(/\s+/g," ").trim()}function c(e,t){if(!t)return;let l=u(t);if(!l)return;let i=e.filter(e=>e.query&&e.query.trim()).filter(e=>{let t=u(e.query);return!!t&&l.endsWith(t)});if(0!==i.length)return i.reduce((e,t)=>{let l=u(e.query).length;return u(t.query).length>l?t:e})}[o,s]=a.then?(await a)():a;let p=n.default.join(process.cwd(),"data"),f=n.default.join(p,"evaluation_result.json");async function d(e,t){let l={};if(e&&(l.OR=[{user:e},{user:null}]),!t?.query&&t?.taskId){let e=await s.db.findExecutionById(t.taskId);e&&e.query?(l.query=e.query,t.framework&&(l.framework=t.framework)):l.id=t.taskId}else t?.query&&(l.query=t.query,t.framework&&(l.framework=t.framework));t?.skill!==void 0&&(l.skill=t.skill),t?.skillVersion!==void 0&&(l.skillVersion=t.skillVersion);let i=await s.db.findExecutions(l,{timestamp:"desc"}),n=new Map;for(let e of i){let t=e.taskId||null;t&&(n.has(t)||n.set(t,[]),n.get(t).push(e))}let o=new Set;for(let[e,t]of n.entries()){if(1===t.length){o.add(t[0].id);continue}let l=t.find(t=>t.id===e);if(l){o.add(l.id);continue}let i=t.slice().sort((e,t)=>{let l=new Date(e.timestamp).getTime(),i=new Date(t.timestamp).getTime();if(i!==l)return i-l;let n=String(e.finalResult||"").length;return String(t.finalResult||"").length-n});o.add(i[0].id)}let a=i.filter(e=>!e.taskId||o.has(e.id));for(let[e,t]of n.entries())if(!(t.length<=1))for(let e of t)o.has(e.id)||s.db.deleteExecution(e.id).catch(()=>{});return a.map(e=>{let t=e.model??null,l=t?(0,r.getModelPricing)(t):null,i=l?.pricing??null,n=t&&null!=e.maxSingleCallTokens?(0,r.getModelContextWindow)(t):null;return{...e,upload_id:e.id,task_id:e.taskId||void 0,query:e.query||void 0,framework:e.framework||void 0,tokens:e.tokens||void 0,cost:i&&null!=e.inputTokens&&null!=e.outputTokens?(0,r.calculateCost)(e.inputTokens,e.outputTokens,i,e.cacheReadInputTokens??void 0,e.cacheCreationInputTokens??void 0):void 0,latency:e.latency||void 0,timestamp:e.timestamp?.toISOString?.()||e.timestamp,final_result:e.finalResult||void 0,skill:e.skill||void 0,skills:e.skills?JSON.parse(e.skills):void 0,invokedSkills:e.invokedSkills?JSON.parse(e.invokedSkills):void 0,is_skill_correct:e.isSkillCorrect||!1,is_answer_correct:e.isAnswerCorrect||!1,answer_score:void 0!==e.answerScore?e.answerScore:void 0,skill_score:void 0!==e.skillScore?e.skillScore:void 0,judgment_reason:e.judgmentReason||void 0,failures:e.failures?JSON.parse(e.failures):void 0,label:e.label??null,user:e.user??null,skill_issues:e.skillIssues?JSON.parse(e.skillIssues):[],skill_version:e.skillVersion??null,model:t,tool_call_count:e.toolCallCount??void 0,llm_call_count:e.llmCallCount??void 0,input_tokens:e.inputTokens??void 0,output_tokens:e.outputTokens??void 0,tool_call_error_count:e.toolCallErrorCount??void 0,cache_read_input_tokens:e.cacheReadInputTokens??void 0,cache_creation_input_tokens:e.cacheCreationInputTokens??void 0,max_single_call_tokens:e.maxSingleCallTokens??void 0,reasoning_tokens:e.reasoningTokens??void 0,expected_skill_version:e.expectedSkillVersion??null,skill_recall_rate:e.skillRecallRate??null,context_window_pct:null!=e.maxSingleCallTokens&&n?Math.round(e.maxSingleCallTokens/n.contextWindow*1e3)/10:void 0,context_window_limit:n?.contextWindow,context_window_source:n?.source,cost_pricing:i?{inputTokenPrice:i.inputTokenPrice,outputTokenPrice:i.outputTokenPrice,cacheReadInputTokenPrice:i.cacheReadInputTokenPrice??i.inputTokenPrice*r.DEFAULT_CACHE_READ_RATIO,cacheCreationInputTokenPrice:i.cacheCreationInputTokenPrice??i.inputTokenPrice*r.DEFAULT_CACHE_CREATION_RATIO,source:l?.source??"default"}:null}})}async function k(e){let t={};return e&&(t.OR=[{user:e},{user:null}]),(await s.db.findConfigs(t)).map(e=>{let t=(t,l)=>{if(t)try{return JSON.parse(t)}catch(t){console.error(`[readConfig] Failed to parse ${l} for config ${e.id}:`,t);return}};return{id:e.id,query:e.query,skill:e.skill,skillVersion:e.skillVersion,expectedSkills:t(e.expectedSkills,"expectedSkills"),standard_answer:e.standardAnswer,root_causes:t(e.rootCauses,"rootCauses"),key_actions:t(e.keyActions,"keyActions"),parse_status:e.parseStatus||"completed"}})}async function _(e){let t=e.upload_id||e.task_id||crypto.randomUUID();if(e.task_id)try{let l={taskId:e.task_id};e.framework&&(l.framework=e.framework);let i=await s.db.findExecutions(l,{timestamp:"desc"});if(i&&i.length>0&&i[0]?.id){let l=i.find(t=>t.id===e.task_id),n=l&&l.id?l.id:i[0].id;n!==t&&(t=n)}}catch{}let l=null,n=await s.db.findExecutionById(t);n&&(l={...n,upload_id:n.id,task_id:n.taskId||void 0,query:n.query||void 0,framework:n.framework||void 0,tokens:n.tokens||void 0,cost:n.cost||void 0,latency:n.latency||void 0,timestamp:n.timestamp?.toISOString?.()||n.timestamp,final_result:n.finalResult||void 0,skill:n.skill||void 0,skills:n.skills?JSON.parse(n.skills):void 0,invokedSkills:n.invokedSkills?(()=>{try{return JSON.parse(n.invokedSkills)}catch{return}})():void 0,is_skill_correct:n.isSkillCorrect||!1,is_answer_correct:n.isAnswerCorrect||!1,answer_score:n.answerScore||void 0,skill_score:n.skillScore||void 0,judgment_reason:n.judgmentReason||void 0,failures:n.failures?JSON.parse(n.failures):void 0,skill_issues:n.skillIssues?JSON.parse(n.skillIssues):void 0,label:n.label||void 0,user:n.user||void 0,skill_version:n.skillVersion??void 0,expected_skill_version:n.expectedSkillVersion??null,skill_recall_rate:n.skillRecallRate??null,model:n.model||void 0,tool_call_count:n.toolCallCount??void 0,llm_call_count:n.llmCallCount??void 0,input_tokens:n.inputTokens??void 0,output_tokens:n.outputTokens??void 0,tool_call_error_count:n.toolCallErrorCount??void 0,cache_read_input_tokens:n.cacheReadInputTokens??void 0,cache_creation_input_tokens:n.cacheCreationInputTokens??void 0,max_single_call_tokens:n.maxSingleCallTokens??void 0,reasoning_tokens:n.reasoningTokens??void 0});let r=l?{...l}:{},a=!!l;a||r.timestamp||e.timestamp?e.timestamp&&(r.timestamp=e.timestamp):r.timestamp=new Date().toISOString();let u=!!e.force_query_update,d="string"==typeof l?.query?l.query.trim():"",_="string"==typeof e.query?e.query.trim():"";if(r={...r,...e},d&&!u?r.query=d:!d&&_?r.query=_:"string"!=typeof r.query||r.query.trim()?"string"==typeof r.query&&(r.query=r.query.trim()):r.query=void 0,!r.upload_id&&r.task_id&&(r.upload_id=r.task_id),!r.task_id&&r.upload_id&&(r.task_id=r.upload_id),r.upload_id=t,(!r.label||!r.model||!r.user)&&r.task_id){let e=await s.db.findSessionByTaskId(r.task_id);e&&(!r.label&&e.label&&(r.label=e.label),!r.model&&e.model&&(r.model=e.model),!r.user&&e.user&&(r.user=e.user))}if(!r.user)try{let e=s.db.getClient();if("query"in e){let t=await e.query('SELECT username FROM "User" LIMIT 1');t.rows[0]&&(r.user=t.rows[0].username,console.log(`[Data-Service] Fallback resolved user for task ${r.task_id} to: ${r.user}`))}}catch(e){console.warn("[Data-Service] Fallback user lookup failed:",e)}let p=e.Token||e.token||e.tokens;void 0!==p&&(r.tokens=Number(p)),void 0!==e.tool_call_count&&(r.tool_call_count=Number(e.tool_call_count)),void 0!==e.llm_call_count&&(r.llm_call_count=Number(e.llm_call_count)),void 0!==e.input_tokens&&(r.input_tokens=Number(e.input_tokens)),void 0!==e.output_tokens&&(r.output_tokens=Number(e.output_tokens)),void 0!==e.tool_call_error_count&&(r.tool_call_error_count=Number(e.tool_call_error_count)),void 0!==e.cache_read_input_tokens&&(r.cache_read_input_tokens=Number(e.cache_read_input_tokens)),void 0!==e.cache_creation_input_tokens&&(r.cache_creation_input_tokens=Number(e.cache_creation_input_tokens)),void 0!==e.max_single_call_tokens&&(r.max_single_call_tokens=Number(e.max_single_call_tokens)),void 0!==e.reasoning_tokens&&(r.reasoning_tokens=Number(e.reasoning_tokens));let m="未找到匹配的评测配置",v=!1,g=r.is_answer_correct||!1,y=r.judgment_reason||m,w=await k(r.user);if(r.query&&w.length>0){let t=c(w,r.query);if(t){let i=Array.isArray(r.invokedSkills)?r.invokedSkills:[],n=(Array.isArray(r.skills)?r.skills:[]).map(e=>({name:e,version:null})),u=t.expectedSkills||[];if(u.length>0){let e=i.length>0?i:n;if(e.length>0){let t=0,l=u.filter(e=>e.skill?.trim()),i=l.map(e=>e.skill.trim()),n=new Map;if(i.length>0)try{for(let e of(await s.db.findSkills({name:{in:i},user:r.user||null})))n.set(e.name,e)}catch(e){console.error("[Judgment] Error fetching skills for version check:",e)}for(let i of l){let l=i.skill.trim(),o=i.version??null,s=e.find(e=>e.name===l);if(s){let e=!1;if(null===o)e=!0;else if(null!==s.version)e=s.version===o;else{let t=n.get(l);e=!!t&&(t.activeVersion||0)===o}e&&(t++,v||(v=!0))}}l.length>0&&(r.skill_recall_rate=t/l.length)}}if(r.is_skill_correct=v,void 0!==r.final_result){let i=!0;if(a&&!e.force_judgment&&l&&l.query===r.query&&l.final_result===r.final_result&&(i=!1),i&&!r.skip_evaluation){let e,l=(r.skill||t.skill||"").trim();if(l)try{let t=await s.db.findSkill(l,r.user||null);if(t){let l=t.activeVersion||0,i=t.versions?.find(e=>e.version===l);if(i&&i.content)e=i.content,r.skill_version=i.version;else if(t.versions&&t.versions.length>0){let l=t.versions[0];l&&l.content&&(e=l.content,r.skill_version=l.version)}}}catch(e){console.error("[Judgment] Error fetching skill definition:",e)}let i=await (0,o.judgeAnswer)(r.query||"",{standard_answer_example:t.standard_answer,root_causes:t.root_causes,key_actions:t.key_actions,skill_definition:e},r.final_result,r.user);g=i.is_correct,r.answer_score=i.score,y=i.reason||"Judged by Evaluation Model"}}}else a&&!e.force_judgment||r.answer_score||(g=!1,y=m,r.answer_score=null)}else r.query&&(!a||e.force_judgment)&&!r.answer_score&&(g=!1,y=m,r.answer_score=null);e.skip_evaluation&&(r.answer_score=null,y="结果评估中..."),r.is_skill_correct=v,r.is_answer_correct=g,r.judgment_reason=y;let h=Array.isArray(r.skills)&&r.skills.length>0?r.skills[0]:void 0;if(h){let e=function(){if(!i.default.existsSync(f))return{};try{return JSON.parse(i.default.readFileSync(f,"utf-8"))}catch{return{}}}()[h];e&&(r.skill_score=parseFloat(e))}if(r.skill&&void 0!==r.skill_version&&null!==r.skill_version?r.label=`${r.skill}-v${r.skill_version}`:r.skill?r.label=`${r.skill}-v1`:r.label="without-skill",await s.db.upsertExecution({where:{id:t},create:{id:t,taskId:r.task_id,query:r.query,framework:r.framework,tokens:r.tokens,cost:r.cost,latency:r.latency,timestamp:r.timestamp?new Date(r.timestamp):new Date,finalResult:r.final_result,skill:r.skill,skills:r.skills?JSON.stringify(r.skills):null,invokedSkills:r.invokedSkills?JSON.stringify(r.invokedSkills):null,isSkillCorrect:r.is_skill_correct,isAnswerCorrect:r.is_answer_correct,answerScore:r.answer_score,skillScore:r.skill_score,judgmentReason:r.judgment_reason,failures:r.failures?JSON.stringify(r.failures):null,skillIssues:r.skill_issues?JSON.stringify(r.skill_issues):null,label:r.label,user:r.user,skillVersion:r.skill_version,model:r.model,toolCallCount:r.tool_call_count,llmCallCount:r.llm_call_count,inputTokens:r.input_tokens,outputTokens:r.output_tokens,toolCallErrorCount:r.tool_call_error_count,skillRecallRate:r.skill_recall_rate,cacheReadInputTokens:r.cache_read_input_tokens,cacheCreationInputTokens:r.cache_creation_input_tokens,maxSingleCallTokens:r.max_single_call_tokens,reasoningTokens:r.reasoning_tokens},update:{taskId:r.task_id,query:r.query,framework:r.framework,tokens:r.tokens,cost:r.cost,latency:r.latency,timestamp:r.timestamp?new Date(r.timestamp):new Date,finalResult:r.final_result,skill:r.skill,skills:r.skills?JSON.stringify(r.skills):null,invokedSkills:r.invokedSkills?JSON.stringify(r.invokedSkills):null,isSkillCorrect:r.is_skill_correct,isAnswerCorrect:r.is_answer_correct,answerScore:r.answer_score,skillScore:r.skill_score,judgmentReason:r.judgment_reason,failures:r.failures?JSON.stringify(r.failures):null,skillIssues:r.skill_issues?JSON.stringify(r.skill_issues):null,label:r.label,user:r.user,skillVersion:r.skill_version,model:r.model,toolCallCount:r.tool_call_count,llmCallCount:r.llm_call_count,inputTokens:r.input_tokens,outputTokens:r.output_tokens,toolCallErrorCount:r.tool_call_error_count,skillRecallRate:r.skill_recall_rate,cacheReadInputTokens:r.cache_read_input_tokens,cacheCreationInputTokens:r.cache_creation_input_tokens,maxSingleCallTokens:r.max_single_call_tokens,reasoningTokens:r.reasoning_tokens}}),e.upload_id&&e.task_id&&e.upload_id!==t)try{let t=await s.db.findExecutionById(e.upload_id);t&&t.taskId===e.task_id&&await s.db.deleteExecution(e.upload_id)}catch{}if(r.task_id&&r.interactions){let e="string"==typeof r.interactions?(()=>{try{return JSON.parse(r.interactions)}catch{return[]}})():r.interactions,t=e;try{let l=await s.db.findSessionByTaskId(r.task_id),i=l?.interactions?(()=>{try{return JSON.parse(l.interactions)}catch{return[]}})():[];Array.isArray(i)&&i.length>0&&(t=!Array.isArray(e)||e.length<i.length?i:e.map((e,t)=>{let l=i[t],n=e?.content===""||e?.content==null,o="string"==typeof l?.content&&l.content.length>0;return n&&o&&l?.role===e?.role?{...e,content:l.content}:e}))}catch{}await s.db.upsertSession(r.task_id,{taskId:r.task_id,query:r.query,label:r.label,user:r.user,model:r.model,interactions:JSON.stringify(t)},{query:r.query,label:r.label,user:r.user,model:r.model,interactions:JSON.stringify(t)})}return{success:!0,record:r}}e.s(["findBestMatchConfig",()=>c,"readConfig",()=>k,"readRecords",()=>d,"saveExecutionRecord",()=>_]),l()}catch(e){l(e)}},!1)];
 
 //# sourceMappingURL=src_lib_12408140._.js.map