@doubao-dev/cli 0.0.33 → 0.0.34

package/dist/prompt.js

@@ -1,4 +1,4 @@
-import"node:module";import{ignoreUpdateVersion as r,runtime_packageName as e}from"./2168.js";import{formatError as t,updateCli as n}from"./update.js";import{source as i}from"./9091.js";import{dist as o}from"./@inquirer/select.js";async function s(s){let a;process.stderr.write(`
+import"node:module";import{ignoreUpdateVersion as r,runtime_packageName as e}from"./2168.js";import{formatError as t,updateCli as n}from"./update.js";import{source as i}from"./3688.js";import{dist as o}from"./@inquirer/select.js";async function s(s){let a;process.stderr.write(`
 ${i.bold.yellow("✨ Update available!")} ${i.dim(s.current)} -> ${i.green(s.latest)}
 
 Release notes: ${i.underline("https://developer.doubao.com/docs/resources/changelog")}

package/dist/run.js

@@ -1,4 +1,4 @@
-import"node:module";import{__webpack_require__ as e}from"./rslib-runtime.js";import"./7945.js";import"./6831.js";import"./yaml.js";import{writeStructuredOutputDiagnostic as t,formatReferenceTimeContext as r,formatMs as i,isAppQuotaExhausted as a,parseJSONObject as s,classifyAttemptExecution as n,numberValue as o,LLMClient as l,shared_stringValue as c,getActiveBrowserStrategy as d,logger as u,LLMResponseError as p,requestStructuredOutput as m,parseMCPHTTPBody as f,setTerminalMuted as _,fetchMCPDescriptor as h,EvalDevRuntimeProcessError as g,genExplore as y,acquireEvalSandboxSession as v,limitedIntentCount as b,writeProjectEvalConfig as $,DEFAULT_EVAL_MODEL_ALIAS as k,genDirectTrajectory as x,isPostHookFailure as w,mergeEvalConfigs as S,timed as j,limitedEvalConcurrency as A,validateMCPHeaders as E,readManifestMcpUrl as M,limitedMaxRounds as C,StructuredOutputError as I,startEvalDevRuntime as P,registerStructuredOutputDiagnostics as N,buildAutoStartPayload as O,loadGlobalConfig as D,project_config_readEvalConfigFile as R,booleanValue as L,DEFAULT_EVAL_GATEWAY_ORIGIN as T,resolveProjectEvalConfig as F,LLMRequestTimeoutError as J,genTrajectory as q,launchHeadlessBrowser as z,isGatewayAuthFailure as K,validateProjectEvalConfig as U,limitPipelineRunConfig as H,integerConfigValue as B,projectEvalConfigReadPath as G,createReferenceTime as V}from"./3974.js";import{createKitDevOptions as Y}from"./kit-dev-options.js";import{readJsonIfExists as W,atomicWriteJson as X}from"./json-file.js";import{option_Option as Z,flushEvents as Q,sanitizeDetailField as ee,getAuthStatus as et,classifyError as er,setEventTrackingUid as ei,sanitizeProgressFailure as ea,reportEvalRun as es,reportCommand as en,command_Command as eo,sanitizeLogText as el,runCommand as ec,summarizeLogError as ed,reportEvalPhase as eu,classifyEventStatus as ep,isRecord as em}from"./2168.js";import{Text as ef,build_render as e_,Static as eh,components_Box as eg,render_to_string as ey}from"./ink.js";import{Chalk as ev}from"./9091.js";import{lookupRun as eb,registerRun as e$}from"./run-registry.js";import{resolveAppKey as ek}from"./workspace.js";import{requireRuntimeSkillEntries as ex}from"./project-layout.js";var ew={};e.r(ew),e.d(ew,{buildFreshConfig:()=>ot,configureEvalRunOptions:()=>n5,default:()=>n4,resolveEvalDevConfig:()=>oi,resolvePipelineRunConfig:()=>n7});let eS="dbx-eval-wasm-template-encoder";async function ej(e,t){let r=(await e())._encode(JSON.stringify(t)),i=r.buffer;try{if(0!==r.status)throw Error(`encode error: ${String(r.error_msg)}`);let e=Buffer.allocUnsafe(i.size());for(let t=0;t<e.length;t+=1)e[t]=i.get(t)??0;return{buffer:e,debugInfo:r.lepus_debug}}finally{eA(i),eA(r)}}function eA(e){"function"==typeof e.delete&&e.delete()}class eE{lynxTemplatePlugin;constructor(e){this.lynxTemplatePlugin=e}apply(t){t.hooks.thisCompilation.tap(eS,t=>{this.lynxTemplatePlugin.getLynxTemplatePluginHooks(t).encode.tapPromise({name:eS,stage:254},async({encodeOptions:t})=>ej((await import("./lepus.js").then(e.t.bind(e,"../../../node_modules/.pnpm/@byted-lynx+tasm@4.4.40_debug@4.4.3/node_modules/@byted-lynx/tasm/lepus.js",23))).default,t))})}}let eM=e("node:fs");async function eC(e){let t;process.env.DISABLE_APP_TEMPLATE_JS="1";let r=JSON.parse((0,eM.readFileSync)(e,"utf8")),i=!1,a=async()=>{i||(i=!0,t&&await t.stop(),eF({type:"stopped"}))},s=()=>{a().finally(()=>process.exit(0))};process.once("SIGINT",s),process.once("SIGTERM",s),process.on("message",e=>{var t;"object"==typeof(t=e)&&null!==t&&"type"in t&&"stop"===t.type&&s()});try{t=await eI(r),eF({type:"ready",webDebuggerUrl:t.webDebuggerUrl,sandboxSessionId:t.sandboxSessionId}),await new Promise(()=>void 0)}catch(t){let e=t instanceof Error?t:Error(String(t));eF({type:"failure",message:e.message,stack:e.stack}),await a().catch(()=>void 0),process.disconnect?.(),process.exitCode=1}finally{process.removeListener("SIGINT",s),process.removeListener("SIGTERM",s)}}async function eI(e){var t,r;let i,a,s,n=process.env.WATCHPACK_POLLING;n||(process.env.WATCHPACK_POLLING="1000");try{let r=await import("./sdk.js"),a=(t=r.DoubaoAppsSDK,class{delegate;constructor(e){let r=new t(e);this.delegate=r.launcher}dev(e){return this.delegate.dev({...e,rsbuildPlugins:[...e.rsbuildPlugins??[],{name:eS,setup(e){e.modifyBundlerChain((t,r)=>{if(r?.environment?.name==="web")return;let i=e.useExposed(Symbol.for("LynxTemplatePlugin"));t.plugin(eS).use(eE,[i.LynxTemplatePlugin])})}}]})}}),s=new r.DoubaoAppsSDK(e.miniappDir,a);eL("Launching isolated Kit dev runtime with WASM template encoding ..."),i=await s.dev(Y({appId:e.appId,mcpEndpoint:e.mcpEndpoint,simulatorManifestPath:e.simulatorManifestPath,skillsDirectory:e.skillsDirectory}))}finally{void 0===n?delete process.env.WATCHPACK_POLLING:process.env.WATCHPACK_POLLING=n}if(!i.webDebugger)throw await i.closeServer().catch(()=>void 0),Error("kit sdk.dev did not return webDebugger; simulator options may be misconfigured");let o=eR(i.webDebugger.url),l=(r=i.webDebugger).debuggerMcpEndpoint?eR(r.debuggerMcpEndpoint):eR(new URL("/__wsd_mcp",r.url).toString()),c=i.webDebugger.sandboxSessionId?.trim();if(!c)throw await eP(i),Error("Kit dev runtime did not return a Sandbox Session ID");eL(`Web debugger SPA: ${o}`),eL(`Debugger MCP:    ${l}`);try{({browser:a,page:s}=await eN(o,l))}catch(e){throw eT(`Headless SDK attach failed; rolling back dev runtime: ${e.message}`),await eP(i),e}let d=!1;return{webDebuggerUrl:l,sandboxSessionId:c,stop:async()=>{d||(d=!0,eL("stop: closing playwright page ..."),await eD(s.close(),5e3,"page.close").catch(e=>{eT(`page.close: ${e.message}`)}),eL("stop: closing playwright browser ..."),await eD(a.close(),5e3,"browser.close").catch(e=>{eT(`browser.close: ${e.message}`)}),await eP(i),eL("stop: done."))}}}async function eP(e){eL("stop: closing dev server ..."),await eD(e.closeServer(),1e4,"devHandle.closeServer").catch(e=>{eT(`closeServer failed: ${e.message}`)})}async function eN(e,t){let r,i=await z();try{if(r=await i.newPage(),await r.goto(e,{waitUntil:"domcontentloaded",timeout:3e4}),!await eO(t,3e4))throw Error(`SDK debugger did not attach at ${e} within 30s; widget bundle may not have loaded`);return await new Promise(e=>setTimeout(e,3e3)),eL("Headless SDK attached; widgets registered."),{browser:i,page:r}}catch(e){throw r&&await r.close().catch(()=>void 0),await i.close().catch(()=>void 0),e}}async function eO(e,t){let r=Date.now()+t,i=0;for(;Date.now()<r;){try{i+=1;let t=await fetch(e,{method:"POST",headers:{"Content-Type":"application/json",Accept:"application/json, text/event-stream"},body:JSON.stringify({jsonrpc:"2.0",id:i,method:"tools/call",params:{name:"list_render_targets",arguments:{type:"all"}}})});if(t.ok){let e=f(await t.text(),t.headers.get("content-type"),i),r=e.result?.structuredContent?.widgets;if(Array.isArray(r)&&r.length>0)return!0}}catch{}await new Promise(e=>setTimeout(e,800))}return!1}function eD(e,t,r){return new Promise((i,a)=>{let s=setTimeout(()=>a(Error(`${r} timed out after ${t}ms`)),t);e.then(e=>{clearTimeout(s),i(e)},e=>{clearTimeout(s),a(e)})})}function eR(e){try{let t=new URL(e);return"localhost"===t.hostname&&(t.hostname="127.0.0.1"),t.toString()}catch{return e}}function eL(e){eF({type:"log",level:"log",message:e})}function eT(e){eF({type:"log",level:"warn",message:e})}function eF(e){process.send?.(e)}let eJ={L1_platform:{shortLabel:"L1",promptName:"Sandbox 问题",legendLabel:"平台/环境",description:"工具可用性、时间上下文、状态隔离与注入或能力模拟失败。"},L2_tool:{shortLabel:"L2",promptName:"工具问题",legendLabel:"工具执行",description:"工具执行、结果、协议转换、参数传递或时间处理失败。"},L3_skill:{shortLabel:"L3",promptName:"Skill/模型能力问题",legendLabel:"Skill/模型能力",description:"任务理解、Skill 召回或遵循、结果核验、推理计算、模型幻觉编造或 Skill 定义失败。"},case_quality:{shortLabel:"CQ",promptName:"用例质量问题",legendLabel:"Case Quality",description:"用例剧本、前置条件或成功标准存在冲突。"},other_intent:{shortLabel:"O",promptName:"其他意图问题",legendLabel:"Other Intent",description:"任务路由错误或请求不属于当前 Skill 能力边界。"}},eq=Object.keys(eJ),ez={sandbox:new Set(["工具可用性","时间上下文","状态隔离与注入","能力模拟"]),tool:new Set(["执行稳定性","结果正确性","协议转换正确性","参数传递完整性","时间处理正确性"]),skill:new Set(["任务理解","Skill召回或选择","Skill遵循","结果核验","模型幻觉编造","推理计算","能力边界描述缺失或模糊","工具契约描述错误","规则冲突或不可执行","skill过拟合"]),case_quality:new Set(["剧本冲突"]),other_intent:new Set(["路由与任务交接","非当前 Skill 能力边界"])};function eK(e){return ez["platform"===e?"sandbox":e]}function eU(e){switch(e){case"L1_platform":case"L2_tool":case"L3_skill":case"case_quality":case"other_intent":return e;case"L3_model":case"L4_skill":return"L3_skill";default:return}}function eH(e){switch(e){case"sandbox":case"tool":case"case_quality":case"other_intent":return e;case"model":case"skill":return"skill";default:return}}function eB(e){return Array.isArray(e.skill_documents)?e.skill_documents:"string"==typeof e.skill_document?[e.skill_document]:[]}function eG(e){let t=eB(e);return t.length<=1?t[0]??"":t.map((e,t)=>`## Skill document ${t+1}
+import"node:module";import{__webpack_require__ as e}from"./rslib-runtime.js";import"./7945.js";import"./6831.js";import"./yaml.js";import{writeStructuredOutputDiagnostic as t,formatReferenceTimeContext as r,formatMs as i,isAppQuotaExhausted as a,parseJSONObject as s,classifyAttemptExecution as n,numberValue as o,LLMClient as l,shared_stringValue as c,getActiveBrowserStrategy as d,logger as u,LLMResponseError as p,requestStructuredOutput as m,parseMCPHTTPBody as f,setTerminalMuted as _,fetchMCPDescriptor as h,EvalDevRuntimeProcessError as g,genExplore as y,acquireEvalSandboxSession as v,limitedIntentCount as b,writeProjectEvalConfig as $,DEFAULT_EVAL_MODEL_ALIAS as k,genDirectTrajectory as x,isPostHookFailure as w,mergeEvalConfigs as S,timed as j,limitedEvalConcurrency as A,validateMCPHeaders as E,readManifestMcpUrl as M,limitedMaxRounds as C,StructuredOutputError as I,startEvalDevRuntime as P,registerStructuredOutputDiagnostics as N,buildAutoStartPayload as O,loadGlobalConfig as D,project_config_readEvalConfigFile as R,booleanValue as L,DEFAULT_EVAL_GATEWAY_ORIGIN as T,resolveProjectEvalConfig as F,LLMRequestTimeoutError as J,genTrajectory as q,launchHeadlessBrowser as z,isGatewayAuthFailure as K,validateProjectEvalConfig as U,limitPipelineRunConfig as H,integerConfigValue as B,projectEvalConfigReadPath as G,createReferenceTime as V}from"./3974.js";import{createKitDevOptions as Y}from"./kit-dev-options.js";import{readJsonIfExists as W,atomicWriteJson as X}from"./json-file.js";import{option_Option as Z,flushEvents as Q,sanitizeDetailField as ee,getAuthStatus as et,classifyError as er,setEventTrackingUid as ei,sanitizeProgressFailure as ea,reportEvalRun as es,reportCommand as en,command_Command as eo,sanitizeLogText as el,runCommand as ec,summarizeLogError as ed,reportEvalPhase as eu,classifyEventStatus as ep,isRecord as em}from"./2168.js";import{Text as ef,build_render as e_,Static as eh,components_Box as eg,render_to_string as ey}from"./ink.js";import{Chalk as ev}from"./3688.js";import{lookupRun as eb,registerRun as e$}from"./run-registry.js";import{resolveAppKey as ek}from"./workspace.js";import{requireRuntimeSkillEntries as ex}from"./project-layout.js";var ew={};e.r(ew),e.d(ew,{buildFreshConfig:()=>ot,configureEvalRunOptions:()=>n5,default:()=>n4,resolveEvalDevConfig:()=>oi,resolvePipelineRunConfig:()=>n7});let eS="dbx-eval-wasm-template-encoder";async function ej(e,t){let r=(await e())._encode(JSON.stringify(t)),i=r.buffer;try{if(0!==r.status)throw Error(`encode error: ${String(r.error_msg)}`);let e=Buffer.allocUnsafe(i.size());for(let t=0;t<e.length;t+=1)e[t]=i.get(t)??0;return{buffer:e,debugInfo:r.lepus_debug}}finally{eA(i),eA(r)}}function eA(e){"function"==typeof e.delete&&e.delete()}class eE{lynxTemplatePlugin;constructor(e){this.lynxTemplatePlugin=e}apply(t){t.hooks.thisCompilation.tap(eS,t=>{this.lynxTemplatePlugin.getLynxTemplatePluginHooks(t).encode.tapPromise({name:eS,stage:254},async({encodeOptions:t})=>ej((await import("./lepus.js").then(e.t.bind(e,"../../../node_modules/.pnpm/@byted-lynx+tasm@4.4.40_debug@4.4.3/node_modules/@byted-lynx/tasm/lepus.js",23))).default,t))})}}let eM=e("node:fs");async function eC(e){let t;process.env.DISABLE_APP_TEMPLATE_JS="1";let r=JSON.parse((0,eM.readFileSync)(e,"utf8")),i=!1,a=async()=>{i||(i=!0,t&&await t.stop(),eF({type:"stopped"}))},s=()=>{a().finally(()=>process.exit(0))};process.once("SIGINT",s),process.once("SIGTERM",s),process.on("message",e=>{var t;"object"==typeof(t=e)&&null!==t&&"type"in t&&"stop"===t.type&&s()});try{t=await eI(r),eF({type:"ready",webDebuggerUrl:t.webDebuggerUrl,sandboxSessionId:t.sandboxSessionId}),await new Promise(()=>void 0)}catch(t){let e=t instanceof Error?t:Error(String(t));eF({type:"failure",message:e.message,stack:e.stack}),await a().catch(()=>void 0),process.disconnect?.(),process.exitCode=1}finally{process.removeListener("SIGINT",s),process.removeListener("SIGTERM",s)}}async function eI(e){var t,r;let i,a,s,n=process.env.WATCHPACK_POLLING;n||(process.env.WATCHPACK_POLLING="1000");try{let r=await import("./sdk.js"),a=(t=r.DoubaoAppsSDK,class{delegate;constructor(e){let r=new t(e);this.delegate=r.launcher}dev(e){return this.delegate.dev({...e,rsbuildPlugins:[...e.rsbuildPlugins??[],{name:eS,setup(e){e.modifyBundlerChain((t,r)=>{if(r?.environment?.name==="web")return;let i=e.useExposed(Symbol.for("LynxTemplatePlugin"));t.plugin(eS).use(eE,[i.LynxTemplatePlugin])})}}]})}}),s=new r.DoubaoAppsSDK(e.miniappDir,a);eL("Launching isolated Kit dev runtime with WASM template encoding ..."),i=await s.dev(Y({appId:e.appId,mcpEndpoint:e.mcpEndpoint,simulatorManifestPath:e.simulatorManifestPath,skillsDirectory:e.skillsDirectory}))}finally{void 0===n?delete process.env.WATCHPACK_POLLING:process.env.WATCHPACK_POLLING=n}if(!i.webDebugger)throw await i.closeServer().catch(()=>void 0),Error("kit sdk.dev did not return webDebugger; simulator options may be misconfigured");let o=eR(i.webDebugger.url),l=(r=i.webDebugger).debuggerMcpEndpoint?eR(r.debuggerMcpEndpoint):eR(new URL("/__wsd_mcp",r.url).toString()),c=i.webDebugger.sandboxSessionId?.trim();if(!c)throw await eP(i),Error("Kit dev runtime did not return a Sandbox Session ID");eL(`Web debugger SPA: ${o}`),eL(`Debugger MCP:    ${l}`);try{({browser:a,page:s}=await eN(o,l))}catch(e){throw eT(`Headless SDK attach failed; rolling back dev runtime: ${e.message}`),await eP(i),e}let d=!1;return{webDebuggerUrl:l,sandboxSessionId:c,stop:async()=>{d||(d=!0,eL("stop: closing playwright page ..."),await eD(s.close(),5e3,"page.close").catch(e=>{eT(`page.close: ${e.message}`)}),eL("stop: closing playwright browser ..."),await eD(a.close(),5e3,"browser.close").catch(e=>{eT(`browser.close: ${e.message}`)}),await eP(i),eL("stop: done."))}}}async function eP(e){eL("stop: closing dev server ..."),await eD(e.closeServer(),1e4,"devHandle.closeServer").catch(e=>{eT(`closeServer failed: ${e.message}`)})}async function eN(e,t){let r,i=await z();try{if(r=await i.newPage(),await r.goto(e,{waitUntil:"domcontentloaded",timeout:3e4}),!await eO(t,3e4))throw Error(`SDK debugger did not attach at ${e} within 30s; widget bundle may not have loaded`);return await new Promise(e=>setTimeout(e,3e3)),eL("Headless SDK attached; widgets registered."),{browser:i,page:r}}catch(e){throw r&&await r.close().catch(()=>void 0),await i.close().catch(()=>void 0),e}}async function eO(e,t){let r=Date.now()+t,i=0;for(;Date.now()<r;){try{i+=1;let t=await fetch(e,{method:"POST",headers:{"Content-Type":"application/json",Accept:"application/json, text/event-stream"},body:JSON.stringify({jsonrpc:"2.0",id:i,method:"tools/call",params:{name:"list_render_targets",arguments:{type:"all"}}})});if(t.ok){let e=f(await t.text(),t.headers.get("content-type"),i),r=e.result?.structuredContent?.widgets;if(Array.isArray(r)&&r.length>0)return!0}}catch{}await new Promise(e=>setTimeout(e,800))}return!1}function eD(e,t,r){return new Promise((i,a)=>{let s=setTimeout(()=>a(Error(`${r} timed out after ${t}ms`)),t);e.then(e=>{clearTimeout(s),i(e)},e=>{clearTimeout(s),a(e)})})}function eR(e){try{let t=new URL(e);return"localhost"===t.hostname&&(t.hostname="127.0.0.1"),t.toString()}catch{return e}}function eL(e){eF({type:"log",level:"log",message:e})}function eT(e){eF({type:"log",level:"warn",message:e})}function eF(e){process.send?.(e)}let eJ={L1_platform:{shortLabel:"L1",promptName:"Sandbox 问题",legendLabel:"平台/环境",description:"工具可用性、时间上下文、状态隔离与注入或能力模拟失败。"},L2_tool:{shortLabel:"L2",promptName:"工具问题",legendLabel:"工具执行",description:"工具执行、结果、协议转换、参数传递或时间处理失败。"},L3_skill:{shortLabel:"L3",promptName:"Skill/模型能力问题",legendLabel:"Skill/模型能力",description:"任务理解、Skill 召回或遵循、结果核验、推理计算、模型幻觉编造或 Skill 定义失败。"},case_quality:{shortLabel:"CQ",promptName:"用例质量问题",legendLabel:"Case Quality",description:"用例剧本、前置条件或成功标准存在冲突。"},other_intent:{shortLabel:"O",promptName:"其他意图问题",legendLabel:"Other Intent",description:"任务路由错误或请求不属于当前 Skill 能力边界。"}},eq=Object.keys(eJ),ez={sandbox:new Set(["工具可用性","时间上下文","状态隔离与注入","能力模拟"]),tool:new Set(["执行稳定性","结果正确性","协议转换正确性","参数传递完整性","时间处理正确性"]),skill:new Set(["任务理解","Skill召回或选择","Skill遵循","结果核验","模型幻觉编造","推理计算","能力边界描述缺失或模糊","工具契约描述错误","规则冲突或不可执行","skill过拟合"]),case_quality:new Set(["剧本冲突"]),other_intent:new Set(["路由与任务交接","非当前 Skill 能力边界"])};function eK(e){return ez["platform"===e?"sandbox":e]}function eU(e){switch(e){case"L1_platform":case"L2_tool":case"L3_skill":case"case_quality":case"other_intent":return e;case"L3_model":case"L4_skill":return"L3_skill";default:return}}function eH(e){switch(e){case"sandbox":case"tool":case"case_quality":case"other_intent":return e;case"model":case"skill":return"skill";default:return}}function eB(e){return Array.isArray(e.skill_documents)?e.skill_documents:"string"==typeof e.skill_document?[e.skill_document]:[]}function eG(e){let t=eB(e);return t.length<=1?t[0]??"":t.map((e,t)=>`## Skill document ${t+1}
 ${e}`).join("\n\n")}let eV=["0.0","0.5","1.0"],eY=["0.0","1.0"],eW=new Set(["killer","strong","weak"]),eX=new Set(["tool_chain","constraint","terminal_state","text_quality","safety_boundary"]),eZ=new Set(["复合前置检索","根本性指代歧义","直接写操作","定时/预约"]);function eQ(e){if((0,eM.readFileSync)(e).byteLength>0x1400000)throw Error(`--eval-bundle exceeds 20 MiB: ${e}`);let t=W(e);if(!t)throw Error(`--eval-bundle is not valid JSON: ${e}`);if(1!==t.version)throw Error(`unsupported eval bundle version: ${String(t.version)}`);if(!Number.isFinite(new Date(t.reference_time).getTime()))throw Error("eval bundle reference_time must be an ISO timestamp");if(!Number.isInteger(t.pass_k)||t.pass_k<=0)throw Error("eval bundle pass_k must be a positive integer");if("string"!=typeof t.app_id||!t.app_id.trim())throw Error("eval bundle app_id must be a non-empty string");return function(e){for(let[t,r]of(delete e.global_principles,Object.entries(e.checklists||{})))e.checklists[t]=r.filter(e=>"principle"!==e.evaluator_type)}(t),function(e){if(!e1(e)||!e.name||!Array.isArray(e.tools))throw Error("eval bundle skill snapshot is invalid")}(t.skill),t.skill.skill_documents=eB(t.skill),delete t.skill.skill_document,function(e){if(!Array.isArray(e)||0===e.length)throw Error("eval bundle cases must be a non-empty array");let t=new Set;for(let r of e){if(!r.case_id||t.has(r.case_id))throw Error(`eval bundle has invalid or duplicate case_id: ${r.case_id}`);t.add(r.case_id)}}(t.cases),function(e,t){if(!e1(e))throw Error("eval bundle checklists must be an object");for(let r of t){let t=e[r];if(!Array.isArray(t)||0===t.length)throw Error(`eval bundle is missing checklists for case ${r}`);for(let e of t){if(!e1(e)||e.case_id!==r)throw Error(`eval bundle checklist case_id does not match key ${r}`);!function(e,t){if("string"!=typeof e.checklist_ref||!/^[A-Za-z0-9_-]{1,64}$/.test(e.checklist_ref))throw Error(`eval bundle checklist_ref is invalid for case ${t}`);if("rule"!==e.evaluator_type&&"rubric"!==e.evaluator_type&&"behavior"!==e.evaluator_type)throw Error(`eval bundle evaluator_type is invalid for case ${t}`);if(!Array.isArray(e.content)||0===e.content.length)throw Error(`eval bundle checklist content is empty for case ${t}`);if("rubric"===e.evaluator_type)return e0(e.content,`eval bundle rubric ${e.checklist_ref}`);for(let[r,i]of e.content.entries()){if(!e1(i))throw Error(`eval bundle ${e.evaluator_type} item ${r} must be an object`);if("string"!=typeof i.method||!i.method.trim()||!e1(i.params)||"string"!=typeof i.description||!i.description.trim())throw Error(`eval bundle ${e.evaluator_type} item ${r} is invalid for case ${t}`)}}(e,r)}}for(let r of Object.keys(e))if(!t.has(r))throw Error(`eval bundle has checklists for unknown case ${r}`)}(t.checklists,new Set(t.cases.map(e=>e.case_id))),function(e,t){if(!e1(e))throw Error("eval bundle trajectories must be an object");for(let r of t){let t=e[r.case_id];if(!Array.isArray(t)||0===t.length)throw Error(`eval bundle is missing trajectories for case ${r.case_id}`);for(let e=0;e<t.length;e++){let i=t[e];if(!e1(i)||!e1(i.metadata)||void 0!==i.metadata.case_id&&i.metadata.case_id!==r.case_id)throw Error(`eval bundle trajectory case_id does not match ${r.case_id}`);if(void 0!==i.metadata.attempt_k&&i.metadata.attempt_k!==e+1)throw Error(`eval bundle trajectory attempt_k for ${r.case_id} must be ${e+1}`)}}for(let r of Object.keys(e))if(!t.some(e=>e.case_id===r))throw Error(`eval bundle has trajectories for unknown case ${r}`)}(t.trajectories,t.cases),function(e){if(!e1(e.attempt_counts))throw Error("eval bundle attempt_counts must be an object");for(let t of e.cases){let r=e.trajectories[t.case_id].length,i=e.attempt_counts[t.case_id];if(!Number.isInteger(i)||i<=0||i!==r)throw Error(`eval bundle attempt_counts.${t.case_id} must equal ${r}`);if(i>e.pass_k)throw Error(`eval bundle attempt_counts.${t.case_id} exceeds pass_k ${e.pass_k}`)}}(t),t}function e0(e,t){let r=e.map((e,t)=>(function(e,t){if(!e1(e))throw Error(`injected rubrics[${t}] must be an object`);if(Buffer.byteLength(JSON.stringify(e))>32768)throw Error(`injected rubrics[${t}] exceeds 32768 bytes`);let r=e.step_index;if(!Number.isInteger(r)||0>=Number(r))throw Error(`injected rubrics[${t}].step_index must be positive`);if("string"!=typeof e.evidence||!e.evidence.trim())throw Error(`injected rubrics[${t}].evidence must be non-empty`);if(!e1(e.scoring_criteria))throw Error(`injected rubrics[${t}].scoring_criteria must be an object`);let i=Object.fromEntries(Object.entries(e.scoring_criteria).filter(([e,t])=>eV.includes(e)&&"string"==typeof t&&t.trim().length>0).map(([e,t])=>[e,t.trim()]));if(!eY.every(e=>i[e]))throw Error(`injected rubrics[${t}].scoring_criteria must contain non-empty 1.0 and 0.0 (0.5 is optional)`);let a=e.constraint_level;if(void 0!==a&&!eW.has(a))throw Error(`injected rubrics[${t}].constraint_level is invalid`);let s=function(e,t){if(void 0===e)return;if(!e1(e))throw Error(`injected rubrics[${t}].dimension_capacity must be an object`);let r="model"===e.l1?"skill":e.l1,i="回复忠实性"===e.atomic?"模型幻觉编造":"skill工具描述错误"===e.atomic?"工具契约描述错误":e.atomic;if("string"!=typeof i||!eK(r)?.has(i))throw Error(`injected rubrics[${t}].dimension_capacity is invalid`);return{l1:r,atomic:i}}(e.dimension_capacity,t),n="string"==typeof e.dimension&&e.dimension.trim()?e.dimension.trim():void 0;if(n&&!eX.has(n))throw Error(`injected rubrics[${t}].dimension is invalid`);let o=function(e,t){if(void 0===e)return[];if(!Array.isArray(e)||e.some(e=>"string"!=typeof e||!e.trim()))throw Error(`${t} must contain non-empty strings`);return e.map(e=>e.trim())}(e.dimension_biz_scene,`injected rubrics[${t}].dimension_biz_scene`);for(let e of o)if(!eZ.has(e))throw Error(`injected rubrics[${t}].dimension_biz_scene is invalid`);if(void 0!==e.weight&&("number"!=typeof e.weight||!Number.isFinite(e.weight)||e.weight<=0))throw Error(`injected rubrics[${t}].weight must be positive`);return{step_index:Number(r),evidence:e.evidence.trim(),scoring_criteria:i,...n?{dimension:n}:{},...a?{constraint_level:a}:{},...s?{dimension_capacity:s}:{},...o.length>0?{dimension_biz_scene:o}:{},..."number"==typeof e.weight&&Number.isFinite(e.weight)&&e.weight>0?{weight:e.weight}:{}}})(e,t)),i=new Set;for(let e of r){if(i.has(e.step_index))throw Error(`${t} contains duplicate step_index ${e.step_index}`);i.add(e.step_index)}return r}function e1(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function e2(e){try{e()}catch{}}let e3=/^后台剧本(?:描述)?\s*[:：]\s*/u,e6=/(?:[。；;]\s*|\r?\n+\s*)(?=(?:初始态|关键约束|系统需(?:要)?|用户可能|成功终态))/u,e8=/[。；;\s]+$/u;function e4(e){let t=e.trim();if(!t)return"";let r=t.replace(e3,"").trim();if(!r)return t;let i=e6.exec(r);return(i?r.slice(0,i.index):r).replace(e8,"").trim()||t}let e5=e("node:path");function e7(e){return e.length?`<ul>${e.map(e=>`<li>${tK(e)}</li>`).join("")}</ul>`:"—"}function e9(e){let t=function(e){let t=/^pass_rate ([\d.]+) < threshold ([\d.]+)/.exec(e);if(t)return{label:"通过率低于阈值",detail:`实际 ${tT(Number(t[1]))}，阈值 ${tT(Number(t[2]))}`};if("pass_rate is not a finite number"===e)return{label:"通过率无效",detail:"报告中的通过率不是有效数字，需要检查评分产物。"};let r=/^hard failure:\s*(.+)$/.exec(e);if(r){let e=r[1],t=/^user_experience veto:\s*(.+)$/.exec(e);return t?{label:"用户体验否决",detail:t[1]}:{label:"硬失败",detail:e}}let i=/^(\d+) L1_platform attribution item\(s\) indicate environment failures$/.exec(e);if(i)return{label:"平台/环境归因",detail:`${i[1]} 个失败项被归因为平台、环境或基础依赖问题。`};let a=/^skill_review score ([\d.]+) < ([\d.]+)/.exec(e);if(a)return{label:"Skill 审查低分",detail:`实际 ${a[1]}，阈值 ${a[2]}`};if("skill description 与实际 tool 行为不匹配"===e)return{label:"Skill 描述不匹配",detail:"Skill 文档描述与实际工具行为不一致。"};let s=/^max_attempts (\d+) > (\d+)$/.exec(e);return s?{label:"重试次数过多",detail:`当前 ${s[1]} 次，准出上限 ${s[2]} 次。`}:{label:"准出规则命中",detail:e}}(e);return`<li>
     <strong>${tK(t.label)}</strong>
     <span>${tK(t.detail)}</span>