@zibby/core 0.3.11 → 0.3.13
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/index.js +16 -16
- package/dist/package.json +1 -1
- package/dist/register-built-in-strategies.js +1 -1
- package/dist/strategies/claude-strategy.js +2 -2
- package/dist/strategies/index.js +10 -10
- package/dist/templates/browser-test-automation/graph.mjs +14 -2
- package/dist/templates/browser-test-automation/nodes/cache-replay.mjs +213 -0
- package/dist/templates/browser-test-automation/nodes/index.mjs +1 -0
- package/package.json +1 -1
- package/templates/browser-test-automation/graph.mjs +14 -2
- package/templates/browser-test-automation/nodes/cache-replay.mjs +213 -0
- package/templates/browser-test-automation/nodes/index.mjs +1 -0
|
@@ -1,3 +1,3 @@
|
|
|
1
|
-
import{AgentStrategy as te,getSkill as oe}from"@zibby/agent-workflow";import{query as ne}from"@anthropic-ai/claude-agent-sdk";import{zodToJsonSchema as se}from"zod-to-json-schema";import{writeFileSync as
|
|
1
|
+
import{AgentStrategy as te,getSkill as oe}from"@zibby/agent-workflow";import{query as ne}from"@anthropic-ai/claude-agent-sdk";import{zodToJsonSchema as se}from"zod-to-json-schema";import{writeFileSync as x,mkdirSync as P,existsSync as q,readFileSync as X}from"fs";import{homedir as z}from"os";import{join as A}from"path";import C from"chalk";var f={debug:0,info:1,warn:2,error:3,silent:4},v=class{constructor(){this._level=this._getLogLevel()}_getLogLevel(){if(process.env.ZIBBY_DEBUG==="true")return f.debug;if(process.env.ZIBBY_VERBOSE==="true")return f.info;let e=process.env.LOG_LEVEL?.toLowerCase();return e&&e in f?f[e]:f.info}_shouldLog(e){return f[e]>=this._level}_formatMessage(e,n,r={}){let s=new Date().toISOString(),a=`${this._getPrefix(e)} ${n}`;return Object.keys(r).length>0&&(a+=C.dim(` ${JSON.stringify(r)}`)),a}_getPrefix(e){return{debug:C.gray("[DEBUG]"),info:C.cyan("[INFO]"),warn:C.yellow("[WARN]"),error:C.red("\u274C [ERROR]")}[e]||""}debug(e,n){this._shouldLog("debug")&&console.log(this._formatMessage("debug",e,n))}info(e,n){this._shouldLog("info")&&console.log(this._formatMessage("info",e,n))}warn(e,n){this._shouldLog("warn")&&console.warn(this._formatMessage("warn",e,n))}error(e,n){this._shouldLog("error")&&console.error(this._formatMessage("error",e,n))}setLevel(e){e in f&&(this._level=f[e])}getLevel(){return Object.keys(f).find(e=>f[e]===this._level)}},i=new v;import{timeline as b,Timeline as pe,WORKFLOW_GRAPH_LOG_MARKER_PREFIX as ge}from"@zibby/agent-workflow";var I={ASSISTANT:"gpt-5.4-nano-2026-03-17",CLAUDE:"claude-sonnet-4-6",CURSOR:"auto",CODEX:"o4-mini",GEMINI:"gemini-2.5-pro",OPENAI_POSTPROCESSING:"gpt-4o-mini"};var $={auto:"claude-sonnet-4-6","sonnet-4.6":"claude-sonnet-4-6","sonnet-4-6":"claude-sonnet-4-6","opus-4.6":"claude-opus-4-6","opus-4-6":"claude-opus-4-6","sonnet-4.5":"claude-sonnet-4-5-20250929","sonnet-4-5":"claude-sonnet-4-5-20250929","opus-4.5":"claude-opus-4-20250514","opus-4-5":"claude-opus-4-20250514","claude-sonnet-4-6":"claude-sonnet-4-6","claude-opus-4-6":"claude-opus-4-6","claude-sonnet-4-5-20250929":"claude-sonnet-4-5-20250929","claude-opus-4-20250514":"claude-opus-4-20250514"};var fe={CURSOR_AGENT_DEFAULT:1200*1e3,OPENAI_REQUEST:18e4};function D(l){let e=`${l}_POOL`,n=process.env[e];if(!n||typeof n!="string")return{picked:null,count:0,dispose:()=>{}};let r=n.split(/[,\n]+/).map(a=>a.trim()).filter(Boolean);if(r.length===0)return{picked:null,count:0,dispose:()=>{}};let s=r[Math.floor(Math.random()*r.length)],o=process.env[l];return process.env[l]=s,i.debug(`[auth-pool] ${l}: picked 1 of ${r.length} from pool (***${s.slice(-4)})`),{picked:s,count:r.length,dispose(){o===void 0?delete process.env[l]:process.env[l]=o}}}var Z=class extends te{constructor(){super("claude","Claude (Anthropic API)",50)}canHandle(e){let n=!!process.env.ANTHROPIC_API_KEY||!!process.env.ANTHROPIC_API_KEY_POOL,r=!!process.env.CLAUDE_CODE_OAUTH_TOKEN||!!process.env.CLAUDE_CODE_OAUTH_TOKEN_POOL,s=!!process.env.ANTHROPIC_AUTH_TOKEN||!!process.env.ANTHROPIC_AUTH_TOKEN_POOL,o=n||r||s;return o||i.debug("ClaudeAgentStrategy: no credentials. Set one of ANTHROPIC_API_KEY / CLAUDE_CODE_OAUTH_TOKEN / ANTHROPIC_AUTH_TOKEN (or a *_POOL variant for multi-token rotation)."),o}async invoke(e,n={}){let{model:r,workspace:s=process.cwd(),schema:o=null,images:a=[],skills:d=null,sessionPath:m=null,nodeName:N=null,timeout:R,config:M={},signal:E=null}=n,g=r;(!g||g==="auto")&&(i.debug(`Model is '${g||"undefined"}', using default: ${I.CLAUDE}`),g=I.CLAUDE);let S=$[g]||g;$[g]&&g!==S&&i.debug(`Mapped model: ${g} \u2192 ${S}`);let Q=[D("ANTHROPIC_API_KEY"),D("CLAUDE_CODE_OAUTH_TOKEN"),D("ANTHROPIC_AUTH_TOKEN")];try{i.debug(`Invoking Claude Agent SDK with model: ${S}, skills: ${JSON.stringify(d)}`);let O=process.env.CLAUDE_CODE_OAUTH_TOKEN,U=process.env.ANTHROPIC_API_KEY,K=process.env.ANTHROPIC_AUTH_TOKEN;process.env.CLAUDE_CODE_CREDENTIALS_JSON?ce(process.env.CLAUDE_CODE_CREDENTIALS_JSON):O&&ie(O,process.env.__ZIBBY_CLAUDE_PLAN);let k=process.env.__ZIBBY_CLAUDE_PLAN,J=k?k.charAt(0).toUpperCase()+k.slice(1):null,T;O?T=J?`Claude ${J} plan \xB7 ***${O.slice(-4)}`:`Subscription \xB7 ***${O.slice(-4)}`:U?T=`API key \xB7 ***${U.slice(-4)}`:K?T=`Bearer \xB7 ***${K.slice(-4)}`:T="NONE \u2014 request will fail";let _=(await import("chalk")).default;console.log(""),console.log(` ${_.cyan("\u25C6")} ${_.bold(S)}`),console.log(` ${_.dim("Auth: ")}${T}`),console.log(` ${_.dim("Workdir: ")}${s}`),console.log(""),console.log(_.bold("Prompt sent to LLM:")),console.log(_.dim("\u2500".repeat(60))),console.log(_.dim(e)),console.log(_.dim("\u2500".repeat(60)));let{allowedTools:F,mcpServers:H}=this._resolveSkills(d,{sessionPath:m,workspace:s,nodeName:N});try{let p={cwd:s,allowedTools:F,settings:{permissions:{allow:["Bash(*)","Read(*)","Write(*)","Edit(*)","MultiEdit(*)","NotebookEdit(*)","Grep(*)","Glob(*)","WebFetch(*)","WebSearch(*)","Task(*)","TodoWrite(*)","Skill(*)"]}},model:S,...Object.keys(H).length>0&&{mcpServers:H}};if(o){let h=typeof o.parse=="function"?se(o,{target:"openApi3"}):o;p.outputFormat={type:"json_schema",schema:h},i.debug("Structured output enforced via SDK outputFormat")}i.debug(`Agent SDK options: ${JSON.stringify({cwd:p.cwd,toolCount:F.length,permissionMode:p.permissionMode,model:p.model,hasOutputFormat:!!p.outputFormat})}`);let L="",B=0,G=[];i.debug("Starting Claude Agent SDK query stream");let j;try{j=ne({prompt:e,options:p})}catch(t){throw i.error(`Failed to initialize Claude Agent SDK: ${t.message}`),t}let W=null,w=0,V=3;if(E?.aborted){let t=new Error("Aborted via signal");throw t.name="AbortError",t}try{for await(let t of j){if(E?.aborted){let u=new Error("Aborted via signal");throw u.name="AbortError",u}if(G.push(t),t.type==="error"||t.error){let u=t.error?.message||t.error||t.message||"Unknown API error";try{i.error("SDK error event raw payload",{fullMessage:JSON.stringify(t,null,2)})}catch{}throw new Error(typeof u=="string"?u:JSON.stringify(u))}let h=JSON.stringify(t.message?.content||t.text||"").slice(0,200);if(h===W){if(w++,w>=V){let u=(t.message?.content?.[0]?.text||t.text||"unknown").slice(0,100);throw new Error(`API stuck in loop (${w}x repeated): ${u}`)}}else W=h,w=1;if(t.type==="assistant"||t.constructor?.name==="AssistantMessage"){let u=t.message?.content||t.content||[];for(let c of u)if(c.type==="thinking"&&c.thinking)console.log(`${c.thinking.substring(0,200)}${c.thinking.length>200?"...":""}`);else if(c.type==="text"&&c.text)L+=c.text,c.text.length<500?console.log(`${c.text}`):console.log(`${c.text.substring(0,200)}... (${c.text.length} chars)`);else if(c.type==="tool_use"){B++,c.name.includes("memory")?b.stepMemory(`Tool: ${c.name}`):b.stepTool(`Tool: ${c.name}`);let ee=JSON.stringify(c.input).substring(0,100);console.log(` Input: ${ee}${JSON.stringify(c.input).length>100?"...":""}`)}}else if(!(t.type==="user"&&t.tool_use_result)){if(t.type==="result"||t.constructor?.name==="ResultMessage"){let u=t.result||t.text||t.content||L;try{re({message:t,sessionPath:m,nodeName:N,model:g,toolCallCount:B})}catch(c){i.debug(`usage.json write failed: ${c.message}`)}if(o){if(t.structured_output){i.debug("Using SDK native structured_output");let Y=typeof o.parse=="function"?o.parse(t.structured_output):t.structured_output;return{raw:u,structured:Y}}if(u){let c=this._extractJson(u,o);if(c)return{raw:u,structured:c}}i.warn(`Could not extract structured output \u2014 returning raw text (${(u||"").length} chars)`)}return u||""}}}if(i.warn(`Agent SDK ended without result. Collected ${G.length} messages`),L.length>0)return i.debug("Returning accumulated text from messages"),L;throw new Error("Claude Agent SDK query ended without result")}catch(t){let h={message:t?.message,name:t?.name,code:t?.code,cause:t?.cause?.message||t?.cause,stack:t?.stack?.split(`
|
|
2
2
|
`).slice(0,5).join(`
|
|
3
|
-
`),stringified:(()=>{try{return JSON.stringify(t,Object.getOwnPropertyNames(Object(t)))}catch{return String(t)}})()};throw i.error(`Error during query stream: ${t?.message||t}`,h),t}}catch(p){throw i.error("Claude Agent SDK call failed",{message:p?.message,name:p?.name,code:p?.code,cause:p?.cause?.message||p?.cause}),p}}finally{for(let O of Q)O.dispose()}}_resolveSkills(e,n){let r=["Read","Write","Edit","Bash","Grep","Glob"];if(e===null)return i.debug("No skills \u2014 pure LLM mode"),{allowedTools:[],mcpServers:{}};if(!Array.isArray(e)||e.length===0)return i.debug("Default IDE skills for code generation"),{allowedTools:[...r],mcpServers:{}};let s=[...r],o={};for(let a of e){let d=oe(a);if(!d){i.warn(`Unknown skill "${a}" \u2014 skipping`);continue}if(d.allowedTools&&s.push(...d.allowedTools),typeof d.resolve=="function"){let m=d.resolve(n);m&&(o[d.serverName]=m,i.debug(`MCP: ${d.serverName} \u2192 ${m.command} ${m.args[0]}`))}}return{allowedTools:s,mcpServers:o}}_extractJson(e,n){let r=[()=>{if(e.includes("===JSON_START===")){let s=e.indexOf("===JSON_START===")+16,o=e.indexOf("===JSON_END===");return e.substring(s,o).trim()}},()=>e.match(/```json\s*\n([\s\S]*?)\n```/)?.[1]?.trim(),()=>{if(!e.startsWith("{"))return e.match(/```\s*\n([\s\S]*?)\n```/)?.[1]?.trim()},()=>e.trim(),()=>{let s=e.indexOf("{"),o=e.lastIndexOf("}");if(s!==-1&&o>s)return e.substring(s,o+1)}];for(let s of r)try{let o=s();if(!o)continue;let a=JSON.parse(o);if(typeof a!="object"||a===null)continue;return typeof n.parse=="function"?n.parse(a):a}catch{}return null}};function re({message:l,sessionPath:e,nodeName:n,model:r,toolCallCount:s}){if(!e||!n)return;let o=l?.usage||l?.message?.usage||{},a=
|
|
3
|
+
`),stringified:(()=>{try{return JSON.stringify(t,Object.getOwnPropertyNames(Object(t)))}catch{return String(t)}})()};throw i.error(`Error during query stream: ${t?.message||t}`,h),t}}catch(p){throw i.error("Claude Agent SDK call failed",{message:p?.message,name:p?.name,code:p?.code,cause:p?.cause?.message||p?.cause}),p}}finally{for(let O of Q)O.dispose()}}_resolveSkills(e,n){let r=["Read","Write","Edit","Bash","Grep","Glob"];if(e===null)return i.debug("No skills \u2014 pure LLM mode"),{allowedTools:[],mcpServers:{}};if(!Array.isArray(e)||e.length===0)return i.debug("Default IDE skills for code generation"),{allowedTools:[...r],mcpServers:{}};let s=[...r],o={};for(let a of e){let d=oe(a);if(!d){i.warn(`Unknown skill "${a}" \u2014 skipping`);continue}if(d.allowedTools&&s.push(...d.allowedTools),typeof d.resolve=="function"){let m=d.resolve(n);m&&(o[d.serverName]=m,i.debug(`MCP: ${d.serverName} \u2192 ${m.command} ${m.args[0]}`))}}return{allowedTools:s,mcpServers:o}}_extractJson(e,n){let r=[()=>{if(e.includes("===JSON_START===")){let s=e.indexOf("===JSON_START===")+16,o=e.indexOf("===JSON_END===");return e.substring(s,o).trim()}},()=>e.match(/```json\s*\n([\s\S]*?)\n```/)?.[1]?.trim(),()=>{if(!e.startsWith("{"))return e.match(/```\s*\n([\s\S]*?)\n```/)?.[1]?.trim()},()=>e.trim(),()=>{let s=e.indexOf("{"),o=e.lastIndexOf("}");if(s!==-1&&o>s)return e.substring(s,o+1)}];for(let s of r)try{let o=s();if(!o)continue;let a=JSON.parse(o);if(typeof a!="object"||a===null)continue;return typeof n.parse=="function"?n.parse(a):a}catch{}return null}};function re({message:l,sessionPath:e,nodeName:n,model:r,toolCallCount:s}){if(!e||!n)return;let o=l?.usage||l?.message?.usage||{},a=y(o.input_tokens,o.inputTokens),d=y(o.output_tokens,o.outputTokens),m=y(o.cache_read_input_tokens,o.cacheReadInputTokens),N=y(o.cache_creation_input_tokens,o.cacheCreationInputTokens),R=y(l?.duration_ms,l?.durationMs),M={model:r||null,input_tokens:a,output_tokens:d,cache_read_tokens:m,cache_creation_tokens:N,tool_call_count:typeof s=="number"?s:null,duration_ms:R,captured_at:new Date().toISOString()},E=A(e,n);try{P(E,{recursive:!0})}catch{}x(A(E,"usage.json"),JSON.stringify(M,null,2),"utf-8")}function y(...l){for(let e of l){if(typeof e=="number"&&Number.isFinite(e))return e;if(typeof e=="string"&&e.trim()!==""&&!isNaN(Number(e)))return Number(e)}return null}function ie(l,e){let n=A(z(),".claude"),r=A(n,".credentials.json"),s={claudeAiOauth:{accessToken:l,...e?{subscriptionType:String(e).toLowerCase()}:{}}};try{if(q(r)&&JSON.parse(X(r,"utf-8"))?.claudeAiOauth?.accessToken===l)return;P(n,{recursive:!0}),x(r,JSON.stringify(s,null,2),"utf-8"),i.debug(`Wrote Claude Code OAuth credentials to ${r}`)}catch(o){i.warn(`Could not write ${r}: ${o.message} \u2014 Claude SDK may fail to authenticate.`)}}function ce(l){let e=A(z(),".claude"),n=A(e,".credentials.json"),r;try{r=JSON.parse(l)}catch(s){i.warn(`CLAUDE_CODE_CREDENTIALS_JSON not valid JSON: ${s.message}`);return}if(!r?.claudeAiOauth?.accessToken){i.warn("CLAUDE_CODE_CREDENTIALS_JSON missing claudeAiOauth.accessToken \u2014 skipping write.");return}try{if(q(n)){let s=JSON.parse(X(n,"utf-8"));if(s?.claudeAiOauth?.accessToken===r.claudeAiOauth.accessToken&&s?.claudeAiOauth?.expiresAt===r.claudeAiOauth.expiresAt)return}P(e,{recursive:!0}),x(n,JSON.stringify(r,null,2),"utf-8"),i.debug(`Wrote full Claude Code credentials blob to ${n}`)}catch(s){i.warn(`Could not write ${n}: ${s.message} \u2014 Claude SDK may fail to authenticate.`)}}export{Z as ClaudeAgentStrategy};
|
package/dist/strategies/index.js
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import{AgentStrategy as Ct,DEFAULT_OUTPUT_BASE as kt,SESSION_INFO_FILE as Pt,STOP_REQUEST_FILE as Nt,getAllSkills as Rt,getSkill as Xe}from"@zibby/agent-workflow";import{spawn as Mt,execSync as
|
|
1
|
+
import{AgentStrategy as Ct,DEFAULT_OUTPUT_BASE as kt,SESSION_INFO_FILE as Pt,STOP_REQUEST_FILE as Nt,getAllSkills as Rt,getSkill as Xe}from"@zibby/agent-workflow";import{spawn as Mt,execSync as W}from"node:child_process";import{writeFileSync as We,readFileSync as Ze,mkdirSync as qe,existsSync as le,accessSync as Ve,constants as Qe,unlinkSync as Lt}from"node:fs";import{join as j,resolve as Dt}from"node:path";import{homedir as ae}from"node:os";import ie from"chalk";var F={debug:0,info:1,warn:2,error:3,silent:4},ve=class{constructor(){this._level=this._getLogLevel()}_getLogLevel(){if(process.env.ZIBBY_DEBUG==="true")return F.debug;if(process.env.ZIBBY_VERBOSE==="true")return F.info;let e=process.env.LOG_LEVEL?.toLowerCase();return e&&e in F?F[e]:F.info}_shouldLog(e){return F[e]>=this._level}_formatMessage(e,t,r={}){let o=new Date().toISOString(),i=`${this._getPrefix(e)} ${t}`;return Object.keys(r).length>0&&(i+=ie.dim(` ${JSON.stringify(r)}`)),i}_getPrefix(e){return{debug:ie.gray("[DEBUG]"),info:ie.cyan("[INFO]"),warn:ie.yellow("[WARN]"),error:ie.red("\u274C [ERROR]")}[e]||""}debug(e,t){this._shouldLog("debug")&&console.log(this._formatMessage("debug",e,t))}info(e,t){this._shouldLog("info")&&console.log(this._formatMessage("info",e,t))}warn(e,t){this._shouldLog("warn")&&console.warn(this._formatMessage("warn",e,t))}error(e,t){this._shouldLog("error")&&console.error(this._formatMessage("error",e,t))}setLevel(e){e in F&&(this._level=F[e])}getLevel(){return Object.keys(F).find(e=>F[e]===this._level)}},a=new ve;var B={ASSISTANT:"gpt-5.4-nano-2026-03-17",CLAUDE:"claude-sonnet-4-6",CURSOR:"auto",CODEX:"o4-mini",GEMINI:"gemini-2.5-pro",OPENAI_POSTPROCESSING:"gpt-4o-mini"};var Ce={auto:"claude-sonnet-4-6","sonnet-4.6":"claude-sonnet-4-6","sonnet-4-6":"claude-sonnet-4-6","opus-4.6":"claude-opus-4-6","opus-4-6":"claude-opus-4-6","sonnet-4.5":"claude-sonnet-4-5-20250929","sonnet-4-5":"claude-sonnet-4-5-20250929","opus-4.5":"claude-opus-4-20250514","opus-4-5":"claude-opus-4-20250514","claude-sonnet-4-6":"claude-sonnet-4-6","claude-opus-4-6":"claude-opus-4-6","claude-sonnet-4-5-20250929":"claude-sonnet-4-5-20250929","claude-opus-4-20250514":"claude-opus-4-20250514"},ke={auto:"o4-mini","o4-mini":"o4-mini",o3:"o3","o3-mini":"o3-mini","codex-mini":"codex-mini-latest","gpt-4o":"gpt-4o","gpt-4o-mini":"gpt-4o-mini","gpt-5.2-codex":"gpt-5.2-codex","gpt-5.2":"gpt-5.2","gpt-5.3":"gpt-5.3","gpt-5.4":"gpt-5.4"},Ge={auto:"gemini-2.5-pro","gemini-2.5-pro":"gemini-2.5-pro","gemini-2.5-flash":"gemini-2.5-flash"},de={CURSOR_AGENT_DEFAULT:1200*1e3,OPENAI_REQUEST:18e4};var q=class s{constructor(){this.buffer="",this.extractedResult=null,this.rawText="",this.zodSchema=null,this.lastOutputLength=0,this.onToolCall=null,this._lastToolEmit=null}processChunk(e){if(!e)return null;this.buffer+=e;let t=this.buffer.split(`
|
|
2
2
|
`);this.buffer=t.pop()||"";let r="";for(let o of t)if(o.trim())try{let n=JSON.parse(o);this._emitToolCalls(n);let i=this.extractText(n);if(i){if(this.rawText&&i.startsWith(this.rawText)){let l=i.substring(this.rawText.length);this.rawText=i,r+=l}else(!this.rawText.includes(i)||i.length<20)&&(this.rawText+=i,r+=i);this.tryExtractResult(this.rawText)}else this.isValidResult(n)&&(this.rawText+=`${o}
|
|
3
3
|
`,r+=`${o}
|
|
4
4
|
`,this.extractedResult=n)}catch{if(o.includes('"text"')||o.includes('"content"')){let i=o.match(/"text"\s*:\s*"([^"]*)/),l=o.match(/"content"\s*:\s*"([^"]*)/),c=i?i[1]:l?l[1]:null;c&&!this.rawText.includes(c)&&(r+=c,this.rawText+=c)}}return r||null}flush(){if(!this.buffer.trim())return null;let e="";try{let t=JSON.parse(this.buffer);this._emitToolCalls(t);let r=this.extractText(t);r&&(this.rawText+=r,e+=r,this.tryExtractResult(r))}catch{this.rawText+=this.buffer,e+=this.buffer,this.tryExtractResult(this.buffer)}return this.buffer="",e||null}_emitToolCalls(e){if(!this.onToolCall)return;let t=(i,l)=>{if(!i)return;let c=`${i}:${JSON.stringify(l??{})}`;this._lastToolEmit!==c&&(this._lastToolEmit=c,this.onToolCall(i,l??void 0))},r=i=>{if(i!=null){if(typeof i=="object"&&!Array.isArray(i))return i;if(typeof i=="string")try{return JSON.parse(i)}catch{return}}};if(e.type==="tool_use"||e.type==="tool_call"){if(e.name){t(e.name,r(e.input??e.arguments));return}let i=e.tool_call;if(i&&typeof i=="object"&&!Array.isArray(i)){let l=Object.keys(i);if(l.length===1){let c=l[0],p=i[c],f=p&&typeof p=="object"?p.args??p.input??p:void 0;t(c,r(f))}return}return}if(Array.isArray(e.tool_calls)){for(let i of e.tool_calls)t(i.name,r(i.input??i.arguments));return}let o=e.message??e;if(Array.isArray(o?.tool_calls)){for(let i of o.tool_calls)t(i.name,r(i.input??i.arguments));return}let n=o?.content??e.content;if(Array.isArray(n))for(let i of n)(i.type==="tool_use"||i.type==="tool_call")&&i.name&&t(i.name,r(i.input??i.arguments))}extractText(e){if(e.type==="assistant"&&e.message?.content){let t=e.message.content;if(Array.isArray(t))return t.filter(r=>r.type==="text"&&r.text).map(r=>r.text).join("")}return e.type==="thinking"&&e.text||e.text?e.text:e.content&&typeof e.content=="string"?e.content:e.delta?e.delta:null}tryExtractResult(e){if(!e||typeof e!="string")return;let t=[],r=/```json\s*\n?([\s\S]*?)\n?```/g,o;for(;(o=r.exec(e))!==null;){let _=o[1].trim();try{JSON.parse(_),t.push({text:_,source:"markdown"})}catch{}}let n=0,i=0;for(;n<e.length&&(n=e.indexOf("{",n),n!==-1);){let _=0,m=n;for(let g=n;g<e.length;g++)if(e[g]==="{")_++;else if(e[g]==="}"&&(_--,_===0)){m=g,t.push({text:e.substring(n,m+1),source:"brace"}),i++;break}n=m+1}let l=this.extractedResult,c=l?JSON.stringify(l).length:0,p=0,f=-1;for(let _=0;_<t.length;_++){let m=t[_];try{let g=m.text.replace(/,(\s*[}\]])/g,"$1"),S=JSON.parse(g);this.isValidResult(S)&&(p++,c=JSON.stringify(S).length,l=S,f=_)}catch{}}l&&(this.extractedResult=l)}isValidResult(e){if(!e||typeof e!="object"||Array.isArray(e)||e.session_id||e.timestamp_ms||e.type||e.call_id||e.tool_call||e.result&&typeof e.result=="object"&&(e.result.success&&typeof e.result.success=="object"||e.result.error&&typeof e.result.error=="object")||e.args&&typeof e.args=="object")return!1;if(this.zodSchema)try{return this.zodSchema.parse(e),!0}catch{return!1}return!0}getResult(){return this.extractedResult}getRawText(){return this.rawText}static extractResult(e,t=null){let r=new s;r.zodSchema=t,r.processChunk(e),r.flush();let o=r.getResult();return!o&&process.env.LOG_LEVEL==="debug"&&console.error("[StreamingParser] No result extracted from",e?.length||0,"chars"),o}};import{zodToJsonSchema as dt}from"zod-to-json-schema";var me=class{static generateFileOutputInstructions(e,t){let r;typeof e?.parse=="function"?r=dt(e,{target:"openApi3"}):r=e;let o=this._buildExample(r);return`
|
|
@@ -39,7 +39,7 @@ ${l}
|
|
|
39
39
|
|
|
40
40
|
Extract all relevant information and format it according to the schema. If any required fields are missing, do your best to infer them from the content.`,p={model:B.OPENAI_POSTPROCESSING,messages:[{role:"user",content:c}],response_format:{type:"json_schema",json_schema:{name:"extract",schema:n,strict:!0}}};a.info(`\u{1F4E4} Sending to OpenAI proxy: model=${B.OPENAI_POSTPROCESSING}, schema keys=${Object.keys(n.properties||{}).join(", ")}`),a.debug(` Schema size: ${JSON.stringify(n).length} chars`),a.debug(` Prompt size: ${c.length} chars`);try{let f={"Content-Type":"application/json"};process.env.OPENAI_PROXY_TOKEN?(f["x-proxy-token"]=t,f["x-execution-id"]=process.env.EXECUTION_ID||""):(f.Authorization=`Bearer ${t}`,f["x-api-key"]=process.env.ZIBBY_API_KEY||"",f["x-execution-id"]=process.env.EXECUTION_ID||"");let m=(await mt.post(r,p,{headers:f,timeout:de.OPENAI_REQUEST})).data?.choices?.[0]?.message?.content;if(!m)throw new Error("OpenAI proxy returned empty response");let g=JSON.parse(m);return a.info("\u2705 Successfully formatted with OpenAI proxy"),{structured:g,raw:s}}catch(f){if(f.response){let _=f.response.status,m=f.response.data;throw a.error(`\u274C OpenAI proxy request failed: ${_}`),a.error(` Status: ${_}`),a.error(` Response: ${JSON.stringify(m,null,2)}`),_===401||_===403?new Error(`Authentication failed for OpenAI proxy.
|
|
41
41
|
Run \`zibby login\` or set ZIBBY_USER_TOKEN environment variable.
|
|
42
|
-
Response: ${JSON.stringify(m)}`,{cause:f}):new Error(`Failed to format Cursor output: ${m?.error?.message||"Unknown error"}`,{cause:f})}throw a.error(`\u274C OpenAI proxy request failed: ${f.message}`),new Error(`Failed to format output: ${f.message}`,{cause:f})}}function G(s){let e=`${s}_POOL`,t=process.env[e];if(!t||typeof t!="string")return{picked:null,count:0,dispose:()=>{}};let r=t.split(/[,\n]+/).map(i=>i.trim()).filter(Boolean);if(r.length===0)return{picked:null,count:0,dispose:()=>{}};let o=r[Math.floor(Math.random()*r.length)],n=process.env[s];return process.env[s]=o,a.debug(`[auth-pool] ${s}: picked 1 of ${r.length} from pool (***${o.slice(-4)})`),{picked:o,count:r.length,dispose(){n===void 0?delete process.env[s]:process.env[s]=n}}}import{timeline as Y,Timeline as jo,WORKFLOW_GRAPH_LOG_MARKER_PREFIX as Fo}from"@zibby/agent-workflow";import{copyFileSync as At,existsSync as Pe,lstatSync as bt,mkdirSync as Fe,rmSync as wt,symlinkSync as Tt,unlinkSync as xt}from"node:fs";import{join as z}from"node:path";import{homedir as It}from"node:os";import{randomBytes as $t}from"node:crypto";var vt=["cli-config.json","config.json","auth.json","argv.json"];function Ye(s){return!(!s||typeof s!="string"||process.env.ZIBBY_CURSOR_USE_GLOBAL_MCP==="1"||process.env.ZIBBY_CURSOR_USE_GLOBAL_MCP==="true")}function ze(s){let e=z(s||process.cwd(),".zibby","tmp");Fe(e,{recursive:!0});let t=`${process.pid}-${Date.now()}-${$t(4).toString("hex")}`,r=z(e,`cursor-agent-home-${t}`),o=z(r,".cursor");Fe(o,{recursive:!0});let n=It(),i=z(n,".cursor");if(Pe(i))for(let l of vt){let c=z(i,l);if(Pe(c))try{At(c,z(o,l))}catch{}}if(process.platform==="darwin"){let l=z(n,"Library");if(Pe(l))try{Tt(l,z(r,"Library"))}catch{}}return r}function He(s){if(!(!s||typeof s!="string"))try{let e=z(s,"Library");try{bt(e).isSymbolicLink()&&xt(e)}catch{}wt(s,{recursive:!0,force:!0})}catch{}}var ge=class extends Ct{constructor(){super("cursor","Cursor (CLI)",100)}canHandle(e){let t=[j(ae(),".local","bin","cursor-agent"),j(ae(),".cursor","bin","cursor-agent"),"/usr/local/bin/cursor-agent","/usr/local/bin/agent","/Applications/Cursor.app/Contents/Resources/app/bin/cursor","agent","cursor-agent"];for(let r of t)try{if(r.startsWith("/")){Ve(r,Qe.X_OK);let o=
|
|
42
|
+
Response: ${JSON.stringify(m)}`,{cause:f}):new Error(`Failed to format Cursor output: ${m?.error?.message||"Unknown error"}`,{cause:f})}throw a.error(`\u274C OpenAI proxy request failed: ${f.message}`),new Error(`Failed to format output: ${f.message}`,{cause:f})}}function G(s){let e=`${s}_POOL`,t=process.env[e];if(!t||typeof t!="string")return{picked:null,count:0,dispose:()=>{}};let r=t.split(/[,\n]+/).map(i=>i.trim()).filter(Boolean);if(r.length===0)return{picked:null,count:0,dispose:()=>{}};let o=r[Math.floor(Math.random()*r.length)],n=process.env[s];return process.env[s]=o,a.debug(`[auth-pool] ${s}: picked 1 of ${r.length} from pool (***${o.slice(-4)})`),{picked:o,count:r.length,dispose(){n===void 0?delete process.env[s]:process.env[s]=n}}}import{timeline as Y,Timeline as jo,WORKFLOW_GRAPH_LOG_MARKER_PREFIX as Fo}from"@zibby/agent-workflow";import{copyFileSync as At,existsSync as Pe,lstatSync as bt,mkdirSync as Fe,rmSync as wt,symlinkSync as Tt,unlinkSync as xt}from"node:fs";import{join as z}from"node:path";import{homedir as It}from"node:os";import{randomBytes as $t}from"node:crypto";var vt=["cli-config.json","config.json","auth.json","argv.json"];function Ye(s){return!(!s||typeof s!="string"||process.env.ZIBBY_CURSOR_USE_GLOBAL_MCP==="1"||process.env.ZIBBY_CURSOR_USE_GLOBAL_MCP==="true")}function ze(s){let e=z(s||process.cwd(),".zibby","tmp");Fe(e,{recursive:!0});let t=`${process.pid}-${Date.now()}-${$t(4).toString("hex")}`,r=z(e,`cursor-agent-home-${t}`),o=z(r,".cursor");Fe(o,{recursive:!0});let n=It(),i=z(n,".cursor");if(Pe(i))for(let l of vt){let c=z(i,l);if(Pe(c))try{At(c,z(o,l))}catch{}}if(process.platform==="darwin"){let l=z(n,"Library");if(Pe(l))try{Tt(l,z(r,"Library"))}catch{}}return r}function He(s){if(!(!s||typeof s!="string"))try{let e=z(s,"Library");try{bt(e).isSymbolicLink()&&xt(e)}catch{}wt(s,{recursive:!0,force:!0})}catch{}}var ge=class extends Ct{constructor(){super("cursor","Cursor (CLI)",100)}canHandle(e){let t=[j(ae(),".local","bin","cursor-agent"),j(ae(),".cursor","bin","cursor-agent"),"/usr/local/bin/cursor-agent","/usr/local/bin/agent","/Applications/Cursor.app/Contents/Resources/app/bin/cursor","agent","cursor-agent"];for(let r of t)try{if(r.startsWith("/")){Ve(r,Qe.X_OK);let o=W(`"${r}" --version 2>&1`,{encoding:"utf-8",timeout:3e3,stdio:"pipe"});if(o&&o.length>0)return a.debug(`[Cursor] Found agent at: ${r} (version: ${o.trim().slice(0,50)})`),!0}else{let o=W(`which ${r}`,{encoding:"utf-8",timeout:2e3,stdio:"pipe"}).trim();if(!o)continue;let n=W(`${r} --version 2>&1`,{encoding:"utf-8",timeout:3e3,stdio:"pipe"});if(n&&n.length>0)return a.debug(`[Cursor] Found '${r}' in PATH at ${o} (version: ${n.trim().slice(0,50)})`),!0}}catch{continue}return a.warn("[Cursor] \u274C Cursor Agent CLI not found or not working. Run: agent --version"),!1}async invoke(e,t={}){let{workspace:r=process.cwd(),print:o=!1,schema:n=null,skills:i=null,sessionPath:l=null,nodeName:c=null,timeout:p=de.CURSOR_AGENT_DEFAULT,config:f={},signal:_=null}=t,m=f?.agent?.strictMode||!1,g=t.model??f?.agent?.cursor?.model??B.CURSOR,S=G("CURSOR_API_KEY");try{a.debug(`[Cursor] Invoking (model: ${g}, timeout: ${p/1e3}s, skills: ${JSON.stringify(i)})`);let h=(this._setupMcpConfig(l,r,f,i,c)||{}).isolatedMcpHome??null,I=[j(ae(),".local","bin","cursor-agent"),j(ae(),".cursor","bin","cursor-agent"),"/usr/local/bin/cursor-agent","/usr/local/bin/agent","/Applications/Cursor.app/Contents/Resources/app/bin/cursor","agent","cursor-agent"],T=null;for(let w of I)try{if(w.startsWith("/"))Ve(w,Qe.X_OK),W(`"${w}" --version 2>&1`,{encoding:"utf-8",timeout:3e3,stdio:"pipe"});else{if(!W(`which ${w}`,{encoding:"utf-8",timeout:2e3}).trim())throw new Error("not in PATH");W(`${w} --version 2>&1`,{encoding:"utf-8",timeout:3e3,stdio:"pipe"})}T=w,a.debug(`[Agent] Using binary: ${w}`);break}catch(u){a.debug(`[Agent] Binary '${w}' check failed: ${u.message}`);continue}if(!T)throw new Error(`Cursor Agent CLI not found or not working.
|
|
43
43
|
|
|
44
44
|
Checked paths:
|
|
45
45
|
${I.map(w=>` - ${w}`).join(`
|
|
@@ -55,7 +55,7 @@ Test with: agent --version`);let $=new V(e),y=null;if(n){let w=`zibby-result-${D
|
|
|
55
55
|
\u25C6 Model: ${g||"auto"}${C}
|
|
56
56
|
`);let k=(await import("chalk")).default,O=$.getUserPrompt();console.log(`
|
|
57
57
|
${k.bold("Prompt sent to LLM:")}`),console.log(k.dim("\u2500".repeat(60))),console.log(k.dim(O)),console.log(k.dim("\u2500".repeat(60)));let E=$.build(),L=$.getStats(),b=["--print","--force","--approve-mcps","--output-format","stream-json","--stream-partial-output","--model",g||"auto"];if(process.env.CURSOR_API_KEY&&b.push("--api-key",process.env.CURSOR_API_KEY),b.push(E),a.debug(`[Agent] User prompt: ${L.userPromptLength} chars, System instructions: ${L.systemInstructionsLength} chars (${L.instructionCount} blocks), Full: ${L.fullPromptLength} chars, model: ${g||"auto"}`),a.debug(`[Agent] Workspace: ${r}`),process.env.LOG_LEVEL==="debug"||process.env.ZIBBY_LOG_CURSOR_CLI==="1")try{console.log(`\u{1F527} Cursor CLI --model ${g||"auto"} (from .zibby.config.js agent.cursor.model)
|
|
58
|
-
`)}catch{}let P,U=null;try{let w=l||(process.env.ZIBBY_SESSION_PATH?String(process.env.ZIBBY_SESSION_PATH).trim():null);P=await this._spawnWithStreaming(T,b,r,p,null,w,h,_)}catch(w){U=w}let v=P?.stdout||"";if(n){let w=typeof n.parse=="function",u=null,x=!!(y&&le(y));if(y&&a.info(`[Agent] Result file: ${x?"present":"missing"} at ${y}`),x)try{let d=
|
|
58
|
+
`)}catch{}let P,U=null;try{let w=l||(process.env.ZIBBY_SESSION_PATH?String(process.env.ZIBBY_SESSION_PATH).trim():null);P=await this._spawnWithStreaming(T,b,r,p,null,w,h,_)}catch(w){U=w}let v=P?.stdout||"";if(n){let w=typeof n.parse=="function",u=null,x=!!(y&&le(y));if(y&&a.info(`[Agent] Result file: ${x?"present":"missing"} at ${y}`),x)try{let d=Ze(y,"utf-8").trim();u=JSON.parse(d),a.info(`[Agent] Parsed JSON from result file OK (${d.length} chars) \u2192 object ready for validation`),U&&a.debug("[Agent] Agent exited non-zero but result file was written \u2014 recovering")}catch(d){a.warn(`\u26A0\uFE0F [Agent] Result file exists on disk but is not valid JSON: ${d.message}`)}else if(U)a.warn(`[Agent] Result file missing at ${y} (agent process error \u2014 may still recover if strictMode repairs)`);else throw a.error(`\u274C [Agent] Result file was never created at ${y}`),new Error(`Agent did not write required result file at ${y}`);if(u&&w)try{let d=n.parse(u);return a.info("\u2705 [Agent] Zod validation passed for structured result file"),m&&a.debug("[Agent] strictMode enabled but not needed \u2014 agent wrote valid file"),{raw:v,structured:d}}catch(d){let D=d.issues?d.issues.map(R=>` - ${R.path.join(".")}: ${R.message}`).join(`
|
|
59
59
|
`):d.message?.slice(0,400);if(a.error(`\u274C [Agent] Zod validation FAILED - invalid output schema:
|
|
60
60
|
${D}`),a.error(`\u{1F4C4} [Agent] Invalid JSON written to file:
|
|
61
61
|
${JSON.stringify(u,null,2).slice(0,500)}`),!m)throw new Error(`Agent output failed schema validation:
|
|
@@ -63,8 +63,8 @@ ${D}
|
|
|
63
63
|
|
|
64
64
|
The agent wrote invalid data that doesn't match the required outputSchema.
|
|
65
65
|
Enable strictMode in .zibby.config.js for automatic repair.`,{cause:d})}else{if(u)return a.info("\u2705 [Agent] File-based output extracted (no Zod parse fn) \u2014 accepting as structured"),m&&a.debug("[Agent] strictMode enabled but not needed \u2014 agent wrote valid file"),{raw:v,structured:u};x&&a.error("\u274C [Agent] Result file exists but produced no in-memory JSON (parse failed earlier)")}if(m&&!U){let d=P.parsedText,D=u?JSON.stringify(u):d;a.info(`[Agent] strictMode: calling OpenAI proxy to fix structured output (${D.length} chars in)`);try{let R=await je(D,n);if(w){let J=n.parse(R.structured);return a.info("\u2705 [Agent] Proxy output passed Zod validation"),{raw:v,structured:J}}return{raw:v,...R}}catch(R){if(a.warn(`\u26A0\uFE0F [Agent] strictMode proxy failed: ${R.message}`),u)return a.warn("[Agent] Using agent's original result file as fallback"),{raw:v,structured:u}}}if(U)throw U;let A=x?u==null?"file existed but JSON.parse failed \u2014 see WARN log above":w?"JSON was valid but Zod validation failed \u2014 see WARN log above":"no structured object after read (unexpected)":"file never appeared (agent may not have run Write tool to the path above)";throw a.error(`\u274C [Agent] No validated structured output: ${A}`),a.error("\u{1F4A1} Tip: Set strictMode=true in .zibby.config.js for OpenAI proxy fallback"),new Error(`Agent did not produce a valid result file at ${y}. Enable strictMode for proxy fallback.`)}if(U)throw U;return this._extractFinalResult(v)||P?.parsedText||v}finally{S.dispose()}}_extractFinalResult(e){if(!e)return null;let t=e.split(`
|
|
66
|
-
`),r=null;for(let o of t){let n=o.trim();if(n)try{let i=JSON.parse(n);if(i.type==="assistant"&&i.message?.content){let l=i.message.content;if(Array.isArray(l)){let c=l.filter(p=>p.type==="text"&&p.text).map(p=>p.text).join("");c&&(r=c)}else typeof l=="string"&&l&&(r=l)}}catch{}}return r?.trim()||null}_setupMcpConfig(e,t,r,o=null,n=null){let i=r?.headless,l=j(ae(),".cursor"),c=j(l,"mcp.json"),p={};if(le(c))try{p=JSON.parse(
|
|
67
|
-
`;if(Ye(e)){let h=ze(t||process.cwd()),I=j(h,".cursor","mcp.json");return
|
|
66
|
+
`),r=null;for(let o of t){let n=o.trim();if(n)try{let i=JSON.parse(n);if(i.type==="assistant"&&i.message?.content){let l=i.message.content;if(Array.isArray(l)){let c=l.filter(p=>p.type==="text"&&p.text).map(p=>p.text).join("");c&&(r=c)}else typeof l=="string"&&l&&(r=l)}}catch{}}return r?.trim()||null}_setupMcpConfig(e,t,r,o=null,n=null){let i=r?.headless,l=j(ae(),".cursor"),c=j(l,"mcp.json"),p={};if(le(c))try{p=JSON.parse(Ze(c,"utf-8"))}catch{}let f=p.mcpServers||{},_=r?.paths?.output||kt,m=j(t||process.cwd(),_,Pt),g=Array.isArray(o)?o.map(h=>Xe(h)).filter(Boolean):[...Rt()].map(([,h])=>h),S=new Set;for(let h of g)typeof h.resolve=="function"&&(S.has(h.serverName)||(S.add(h.serverName),this._ensureSkillConfigured(f,h,e,m,n,i)));if(e){let h=Xe("browser");h&&typeof h.resolve=="function"&&!S.has(h.serverName)&&this._ensureSkillConfigured(f,h,e,m,"execute_live",i)}if(Object.keys(f).length===0)return a.debug("[MCP] No MCP servers configured - agent will run without tool access"),{isolatedMcpHome:null};let N=`${JSON.stringify({mcpServers:f},null,2)}
|
|
67
|
+
`;if(Ye(e)){let h=ze(t||process.cwd()),I=j(h,".cursor","mcp.json");return We(I,N,"utf8"),a.debug(`[MCP] Isolated cursor-agent HOME (session-scoped mcp.json): ${h} | servers: ${Object.keys(f).join(", ")}`),{isolatedMcpHome:h}}return le(l)||qe(l,{recursive:!0}),We(c,N,"utf8"),a.debug(`[MCP] Global ~/.cursor/mcp.json | servers: ${Object.keys(f).join(", ")}`),{isolatedMcpHome:null}}_ensureSkillConfigured(e,t,r,o,n=null,i){let l=t.cursorKey||t.serverName,c=e[l]?l:e[t.serverName]?t.serverName:null;if(c&&r){let f=typeof t.resolve=="function"?t.resolve({sessionPath:r,nodeName:n,headless:i}):null;f?.args?e[c].args=f.args:e[c].args=(e[c].args||[]).map(g=>g.startsWith("--output-dir=")?`--output-dir=${r}`:g);let _=f?.env||{},m=t.sessionEnvKey?{[t.sessionEnvKey]:o}:{};e[c].env={...e[c].env||{},..._,...m},a.debug(`[MCP] Updated ${c} session \u2192 ${r}`);return}if(c)return;let p=t.resolve({sessionPath:r,nodeName:n,headless:i});p&&(e[l]={...p,...t.sessionEnvKey&&{env:{...p.env||{},[t.sessionEnvKey]:o}}},a.debug(`[MCP] Configured ${l}`))}_spawnWithStreaming(e,t,r,o,n=null,i=null,l=null,c=null){return new Promise((p,f)=>{let _=Date.now(),m="",g="",S=Date.now(),N=0,h=!1,I=null,T=!1,$=!1,y=null;if(i)try{y=j(Dt(String(i)),Nt)}catch{y=null}let M=!1,C=()=>{M||(M=!0,He(l))},k={...process.env};l&&(k.HOME=l,process.platform==="win32"&&(k.USERPROFILE=l),a.debug(`[Agent] cursor-agent HOME=${l} (isolated MCP config)`));let O=Mt(e,t,{cwd:r,shell:!1,stdio:["pipe","pipe","pipe"],env:k,...c&&{signal:c}});a.debug(`[Agent] PID: ${O.pid}`);let E=null,L=()=>{h||$||(h=!0,I="aborted",a.warn("\u{1F6D1} Abort signal \u2014 terminating cursor-agent (SIGKILL in 5s if SIGTERM trapped)"),E=setTimeout(()=>{if(!$){a.warn("\u{1F6D1} SIGTERM not honored \u2014 sending SIGKILL");try{O.kill("SIGKILL")}catch{}}},5e3))};c&&(c.aborted?L():c.addEventListener("abort",L,{once:!0})),O.stdin.on("error",u=>{u.code!=="EPIPE"&&a.warn(`[Agent] stdin error: ${u.message}`)}),O.stdout.on("error",u=>{u.code!=="EPIPE"&&a.warn(`[Agent] stdout error: ${u.message}`)}),O.stderr.on("error",u=>{u.code!=="EPIPE"&&a.warn(`[Agent] stderr error: ${u.message}`)}),n?(O.stdin.write(n,u=>{u&&u.code!=="EPIPE"&&a.warn(`[Agent] Failed to write to stdin: ${u.message}`),O.stdin.end()}),a.debug(`[Agent] Prompt also piped to stdin (${n.length} chars)`)):O.stdin.end();let b=null;y&&(b=setInterval(()=>{if(!(h||$))try{if(le(y)){h=!0,I="aborted";try{Lt(y)}catch{}a.warn("\u{1F6D1} External stop requested \u2014 terminating Cursor agent (and MCP browser session)"),O.kill("SIGTERM"),setTimeout(()=>{O.killed||O.kill("SIGKILL")},2e3)}}catch{}},600));let P=new Set,U=new Date(_).toISOString().replace(/\.\d+Z$/,""),v=setInterval(()=>{let u=Math.round((Date.now()-_)/1e3),x=Math.round((Date.now()-S)/1e3),A=[];try{let D=Math.ceil(u/60)+1,R=W(`find "${r}" -type f -mmin -${D} -not -path '*/node_modules/*' -not -path '*/.git/*' -not -path '*/target/*' 2>/dev/null | head -20`,{encoding:"utf-8",timeout:5e3}).trim();if(R)for(let J of R.split(`
|
|
68
68
|
`)){let X=J.replace(`${r}/`,"");P.has(X)||(P.add(X),A.push(X))}}catch{}let d="";A.length>0&&(d=` | \u{1F4C1} new: ${A.map(R=>R.split("/").pop()).join(", ")}`),P.size>0&&(d+=` | \u{1F4E6} total: ${P.size} files`),a.debug(`\u{1F493} [Agent] Running for ${u}s | ${N} lines output${d}`),N===0&&u>=30&&P.size===0&&(u<35&&a.warn(`\u26A0\uFE0F [Agent] No output after ${u}s \u2014 agent may be stuck. Check your CURSOR_API_KEY.`),u>=60&&(h=!0,I=I||"stall",a.error(`\u274C [Agent] No response after ${u}s \u2014 killing. Verify CURSOR_API_KEY is valid and agent CLI works: agent --version`),O.kill("SIGTERM"),setTimeout(()=>{O.killed||O.kill("SIGKILL")},3e3)))},3e4),K=setTimeout(()=>{h=!0,I=I||"timeout";let u=Math.round((Date.now()-_)/1e3);a.error(`\u23F1\uFE0F [Agent] Timeout after ${u}s \u2014 killing process (PID: ${O.pid})`),m.trim()&&a.warn(`\u{1F4E4} [Agent] Partial output (${m.length} chars) before timeout:
|
|
69
69
|
${m.slice(-2e3)}`),O.kill("SIGTERM"),setTimeout(()=>{O.killed||O.kill("SIGKILL")},5e3)},o),w=new q;w.onToolCall=(u,x)=>{let A=u,d=x;if(u==="mcpToolCall"&&x?.name)A=x.name.replace(/^mcp_+[^_]+_+/,""),A.includes("-")&&A.split("-")[0]===A.split("-")[1]&&(A=A.split("-")[0]),d=x.args??x.input??x;else{if(u==="readToolCall"||u==="editToolCall"||u==="writeToolCall")return;(u.startsWith("mcp__")||u.includes("ToolCall"))&&(A=u.replace(/^mcp_+[^_]+_+/,"").replace(/ToolCall$/,""))}if(A.includes("memory")?Y.stepMemory(`Tool: ${A}`):Y.stepTool(`Tool: ${A}`),d!=null&&typeof d=="object"&&Object.keys(d).length>0&&!$){let R=JSON.stringify(d),J=R.length>100?`${R.substring(0,100)}...`:R;console.log(` Input: ${J}`)}},O.stdout.on("data",u=>{let x=u.toString();m+=x,S=Date.now(),T||(T=!0);let A=w.processChunk(x);A&&!$&&process.stdout.write(A);let d=x.split(`
|
|
70
70
|
`).filter(D=>D.trim());N+=d.length}),O.stderr.on("data",u=>{let x=u.toString();g+=x,S=Date.now(),T||(T=!0);let A=x.split(`
|
|
@@ -75,20 +75,20 @@ Stderr: ${g.slice(-1e3)}`:""}${m.trim()?`
|
|
|
75
75
|
Stdout (last 500 chars): ${m.slice(-500)}`:""}`));return}let d=w.getResult(),D=d?JSON.stringify(d,null,2):w.getRawText()||m||"";p({stdout:m||g||"",parsedText:D})}),O.on("error",u=>{if(C(),clearTimeout(K),clearInterval(v),b&&clearInterval(b),E&&clearTimeout(E),c&&!c.aborted)try{c.removeEventListener("abort",L)}catch{}f(new Error(`Cursor Agent spawn error: ${u.message}
|
|
76
76
|
Binary: ${e}
|
|
77
77
|
This usually means the binary is not in PATH. Try:
|
|
78
|
-
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && source ~/.zshrc`))})})}};import{AgentStrategy as Ut,getSkill as Kt}from"@zibby/agent-workflow";import{query as Bt}from"@anthropic-ai/claude-agent-sdk";import{zodToJsonSchema as Jt}from"zod-to-json-schema";import{writeFileSync as Ne,mkdirSync as Re,existsSync as et,readFileSync as tt}from"fs";import{homedir as ot}from"os";import{join as ee}from"path";var he=class extends Ut{constructor(){super("claude","Claude (Anthropic API)",50)}canHandle(e){let t=!!process.env.ANTHROPIC_API_KEY||!!process.env.ANTHROPIC_API_KEY_POOL,r=!!process.env.CLAUDE_CODE_OAUTH_TOKEN||!!process.env.CLAUDE_CODE_OAUTH_TOKEN_POOL,o=!!process.env.ANTHROPIC_AUTH_TOKEN||!!process.env.ANTHROPIC_AUTH_TOKEN_POOL,n=t||r||o;return n||a.debug("ClaudeAgentStrategy: no credentials. Set one of ANTHROPIC_API_KEY / CLAUDE_CODE_OAUTH_TOKEN / ANTHROPIC_AUTH_TOKEN (or a *_POOL variant for multi-token rotation)."),n}async invoke(e,t={}){let{model:r,workspace:o=process.cwd(),schema:n=null,images:i=[],skills:l=null,sessionPath:c=null,nodeName:p=null,timeout:f,config:_={},signal:m=null}=t,g=r;(!g||g==="auto")&&(a.debug(`Model is '${g||"undefined"}', using default: ${B.CLAUDE}`),g=B.CLAUDE);let S=Ce[g]||g;Ce[g]&&g!==S&&a.debug(`Mapped model: ${g} \u2192 ${S}`);let N=[G("ANTHROPIC_API_KEY"),G("CLAUDE_CODE_OAUTH_TOKEN"),G("ANTHROPIC_AUTH_TOKEN")];try{a.debug(`Invoking Claude Agent SDK with model: ${S}, skills: ${JSON.stringify(l)}`);let h=process.env.CLAUDE_CODE_OAUTH_TOKEN,I=process.env.ANTHROPIC_API_KEY,T=process.env.ANTHROPIC_AUTH_TOKEN;process.env.CLAUDE_CODE_CREDENTIALS_JSON?Ft(process.env.CLAUDE_CODE_CREDENTIALS_JSON):h&&jt(h,process.env.__ZIBBY_CLAUDE_PLAN);let $=process.env.__ZIBBY_CLAUDE_PLAN,y=$?$.charAt(0).toUpperCase()+$.slice(1):null,M;h?M=y?`Claude ${y} plan \xB7 ***${h.slice(-4)}`:`Subscription \xB7 ***${h.slice(-4)}`:I?M=`API key \xB7 ***${I.slice(-4)}`:T?M=`Bearer \xB7 ***${T.slice(-4)}`:M="NONE \u2014 request will fail";let C=(await import("chalk")).default;console.log(""),console.log(` ${C.cyan("\u25C6")} ${C.bold(S)}`),console.log(` ${C.dim("Auth: ")}${M}`),console.log(` ${C.dim("Workdir: ")}${o}`),console.log(""),console.log(C.bold("Prompt sent to LLM:")),console.log(C.dim("\u2500".repeat(60))),console.log(C.dim(e)),console.log(C.dim("\u2500".repeat(60)));let{allowedTools:k,mcpServers:O}=this._resolveSkills(l,{sessionPath:c,workspace:o,nodeName:p});try{let E={cwd:o,allowedTools:k,
|
|
78
|
+
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && source ~/.zshrc`))})})}};import{AgentStrategy as Ut,getSkill as Kt}from"@zibby/agent-workflow";import{query as Bt}from"@anthropic-ai/claude-agent-sdk";import{zodToJsonSchema as Jt}from"zod-to-json-schema";import{writeFileSync as Ne,mkdirSync as Re,existsSync as et,readFileSync as tt}from"fs";import{homedir as ot}from"os";import{join as ee}from"path";var he=class extends Ut{constructor(){super("claude","Claude (Anthropic API)",50)}canHandle(e){let t=!!process.env.ANTHROPIC_API_KEY||!!process.env.ANTHROPIC_API_KEY_POOL,r=!!process.env.CLAUDE_CODE_OAUTH_TOKEN||!!process.env.CLAUDE_CODE_OAUTH_TOKEN_POOL,o=!!process.env.ANTHROPIC_AUTH_TOKEN||!!process.env.ANTHROPIC_AUTH_TOKEN_POOL,n=t||r||o;return n||a.debug("ClaudeAgentStrategy: no credentials. Set one of ANTHROPIC_API_KEY / CLAUDE_CODE_OAUTH_TOKEN / ANTHROPIC_AUTH_TOKEN (or a *_POOL variant for multi-token rotation)."),n}async invoke(e,t={}){let{model:r,workspace:o=process.cwd(),schema:n=null,images:i=[],skills:l=null,sessionPath:c=null,nodeName:p=null,timeout:f,config:_={},signal:m=null}=t,g=r;(!g||g==="auto")&&(a.debug(`Model is '${g||"undefined"}', using default: ${B.CLAUDE}`),g=B.CLAUDE);let S=Ce[g]||g;Ce[g]&&g!==S&&a.debug(`Mapped model: ${g} \u2192 ${S}`);let N=[G("ANTHROPIC_API_KEY"),G("CLAUDE_CODE_OAUTH_TOKEN"),G("ANTHROPIC_AUTH_TOKEN")];try{a.debug(`Invoking Claude Agent SDK with model: ${S}, skills: ${JSON.stringify(l)}`);let h=process.env.CLAUDE_CODE_OAUTH_TOKEN,I=process.env.ANTHROPIC_API_KEY,T=process.env.ANTHROPIC_AUTH_TOKEN;process.env.CLAUDE_CODE_CREDENTIALS_JSON?Ft(process.env.CLAUDE_CODE_CREDENTIALS_JSON):h&&jt(h,process.env.__ZIBBY_CLAUDE_PLAN);let $=process.env.__ZIBBY_CLAUDE_PLAN,y=$?$.charAt(0).toUpperCase()+$.slice(1):null,M;h?M=y?`Claude ${y} plan \xB7 ***${h.slice(-4)}`:`Subscription \xB7 ***${h.slice(-4)}`:I?M=`API key \xB7 ***${I.slice(-4)}`:T?M=`Bearer \xB7 ***${T.slice(-4)}`:M="NONE \u2014 request will fail";let C=(await import("chalk")).default;console.log(""),console.log(` ${C.cyan("\u25C6")} ${C.bold(S)}`),console.log(` ${C.dim("Auth: ")}${M}`),console.log(` ${C.dim("Workdir: ")}${o}`),console.log(""),console.log(C.bold("Prompt sent to LLM:")),console.log(C.dim("\u2500".repeat(60))),console.log(C.dim(e)),console.log(C.dim("\u2500".repeat(60)));let{allowedTools:k,mcpServers:O}=this._resolveSkills(l,{sessionPath:c,workspace:o,nodeName:p});try{let E={cwd:o,allowedTools:k,settings:{permissions:{allow:["Bash(*)","Read(*)","Write(*)","Edit(*)","MultiEdit(*)","NotebookEdit(*)","Grep(*)","Glob(*)","WebFetch(*)","WebSearch(*)","Task(*)","TodoWrite(*)","Skill(*)"]}},model:S,...Object.keys(O).length>0&&{mcpServers:O}};if(n){let x=typeof n.parse=="function"?Jt(n,{target:"openApi3"}):n;E.outputFormat={type:"json_schema",schema:x},a.debug("Structured output enforced via SDK outputFormat")}a.debug(`Agent SDK options: ${JSON.stringify({cwd:E.cwd,toolCount:k.length,permissionMode:E.permissionMode,model:E.model,hasOutputFormat:!!E.outputFormat})}`);let L="",b=0,P=[];a.debug("Starting Claude Agent SDK query stream");let U;try{U=Bt({prompt:e,options:E})}catch(u){throw a.error(`Failed to initialize Claude Agent SDK: ${u.message}`),u}let v=null,K=0,w=3;if(m?.aborted){let u=new Error("Aborted via signal");throw u.name="AbortError",u}try{for await(let u of U){if(m?.aborted){let A=new Error("Aborted via signal");throw A.name="AbortError",A}if(P.push(u),u.type==="error"||u.error){let A=u.error?.message||u.error||u.message||"Unknown API error";try{a.error("SDK error event raw payload",{fullMessage:JSON.stringify(u,null,2)})}catch{}throw new Error(typeof A=="string"?A:JSON.stringify(A))}let x=JSON.stringify(u.message?.content||u.text||"").slice(0,200);if(x===v){if(K++,K>=w){let A=(u.message?.content?.[0]?.text||u.text||"unknown").slice(0,100);throw new Error(`API stuck in loop (${K}x repeated): ${A}`)}}else v=x,K=1;if(u.type==="assistant"||u.constructor?.name==="AssistantMessage"){let A=u.message?.content||u.content||[];for(let d of A)if(d.type==="thinking"&&d.thinking)console.log(`${d.thinking.substring(0,200)}${d.thinking.length>200?"...":""}`);else if(d.type==="text"&&d.text)L+=d.text,d.text.length<500?console.log(`${d.text}`):console.log(`${d.text.substring(0,200)}... (${d.text.length} chars)`);else if(d.type==="tool_use"){b++,d.name.includes("memory")?Y.stepMemory(`Tool: ${d.name}`):Y.stepTool(`Tool: ${d.name}`);let R=JSON.stringify(d.input).substring(0,100);console.log(` Input: ${R}${JSON.stringify(d.input).length>100?"...":""}`)}}else if(!(u.type==="user"&&u.tool_use_result)){if(u.type==="result"||u.constructor?.name==="ResultMessage"){let A=u.result||u.text||u.content||L;try{Gt({message:u,sessionPath:c,nodeName:p,model:g,toolCallCount:b})}catch(d){a.debug(`usage.json write failed: ${d.message}`)}if(n){if(u.structured_output){a.debug("Using SDK native structured_output");let D=typeof n.parse=="function"?n.parse(u.structured_output):u.structured_output;return{raw:A,structured:D}}if(A){let d=this._extractJson(A,n);if(d)return{raw:A,structured:d}}a.warn(`Could not extract structured output \u2014 returning raw text (${(A||"").length} chars)`)}return A||""}}}if(a.warn(`Agent SDK ended without result. Collected ${P.length} messages`),L.length>0)return a.debug("Returning accumulated text from messages"),L;throw new Error("Claude Agent SDK query ended without result")}catch(u){let x={message:u?.message,name:u?.name,code:u?.code,cause:u?.cause?.message||u?.cause,stack:u?.stack?.split(`
|
|
79
79
|
`).slice(0,5).join(`
|
|
80
80
|
`),stringified:(()=>{try{return JSON.stringify(u,Object.getOwnPropertyNames(Object(u)))}catch{return String(u)}})()};throw a.error(`Error during query stream: ${u?.message||u}`,x),u}}catch(E){throw a.error("Claude Agent SDK call failed",{message:E?.message,name:E?.name,code:E?.code,cause:E?.cause?.message||E?.cause}),E}}finally{for(let h of N)h.dispose()}}_resolveSkills(e,t){let r=["Read","Write","Edit","Bash","Grep","Glob"];if(e===null)return a.debug("No skills \u2014 pure LLM mode"),{allowedTools:[],mcpServers:{}};if(!Array.isArray(e)||e.length===0)return a.debug("Default IDE skills for code generation"),{allowedTools:[...r],mcpServers:{}};let o=[...r],n={};for(let i of e){let l=Kt(i);if(!l){a.warn(`Unknown skill "${i}" \u2014 skipping`);continue}if(l.allowedTools&&o.push(...l.allowedTools),typeof l.resolve=="function"){let c=l.resolve(t);c&&(n[l.serverName]=c,a.debug(`MCP: ${l.serverName} \u2192 ${c.command} ${c.args[0]}`))}}return{allowedTools:o,mcpServers:n}}_extractJson(e,t){let r=[()=>{if(e.includes("===JSON_START===")){let o=e.indexOf("===JSON_START===")+16,n=e.indexOf("===JSON_END===");return e.substring(o,n).trim()}},()=>e.match(/```json\s*\n([\s\S]*?)\n```/)?.[1]?.trim(),()=>{if(!e.startsWith("{"))return e.match(/```\s*\n([\s\S]*?)\n```/)?.[1]?.trim()},()=>e.trim(),()=>{let o=e.indexOf("{"),n=e.lastIndexOf("}");if(o!==-1&&n>o)return e.substring(o,n+1)}];for(let o of r)try{let n=o();if(!n)continue;let i=JSON.parse(n);if(typeof i!="object"||i===null)continue;return typeof t.parse=="function"?t.parse(i):i}catch{}return null}};function Gt({message:s,sessionPath:e,nodeName:t,model:r,toolCallCount:o}){if(!e||!t)return;let n=s?.usage||s?.message?.usage||{},i=ce(n.input_tokens,n.inputTokens),l=ce(n.output_tokens,n.outputTokens),c=ce(n.cache_read_input_tokens,n.cacheReadInputTokens),p=ce(n.cache_creation_input_tokens,n.cacheCreationInputTokens),f=ce(s?.duration_ms,s?.durationMs),_={model:r||null,input_tokens:i,output_tokens:l,cache_read_tokens:c,cache_creation_tokens:p,tool_call_count:typeof o=="number"?o:null,duration_ms:f,captured_at:new Date().toISOString()},m=ee(e,t);try{Re(m,{recursive:!0})}catch{}Ne(ee(m,"usage.json"),JSON.stringify(_,null,2),"utf-8")}function ce(...s){for(let e of s){if(typeof e=="number"&&Number.isFinite(e))return e;if(typeof e=="string"&&e.trim()!==""&&!isNaN(Number(e)))return Number(e)}return null}function jt(s,e){let t=ee(ot(),".claude"),r=ee(t,".credentials.json"),o={claudeAiOauth:{accessToken:s,...e?{subscriptionType:String(e).toLowerCase()}:{}}};try{if(et(r)&&JSON.parse(tt(r,"utf-8"))?.claudeAiOauth?.accessToken===s)return;Re(t,{recursive:!0}),Ne(r,JSON.stringify(o,null,2),"utf-8"),a.debug(`Wrote Claude Code OAuth credentials to ${r}`)}catch(n){a.warn(`Could not write ${r}: ${n.message} \u2014 Claude SDK may fail to authenticate.`)}}function Ft(s){let e=ee(ot(),".claude"),t=ee(e,".credentials.json"),r;try{r=JSON.parse(s)}catch(o){a.warn(`CLAUDE_CODE_CREDENTIALS_JSON not valid JSON: ${o.message}`);return}if(!r?.claudeAiOauth?.accessToken){a.warn("CLAUDE_CODE_CREDENTIALS_JSON missing claudeAiOauth.accessToken \u2014 skipping write.");return}try{if(et(t)){let o=JSON.parse(tt(t,"utf-8"));if(o?.claudeAiOauth?.accessToken===r.claudeAiOauth.accessToken&&o?.claudeAiOauth?.expiresAt===r.claudeAiOauth.expiresAt)return}Re(e,{recursive:!0}),Ne(t,JSON.stringify(r,null,2),"utf-8"),a.debug(`Wrote full Claude Code credentials blob to ${t}`)}catch(o){a.warn(`Could not write ${t}: ${o.message} \u2014 Claude SDK may fail to authenticate.`)}}import{AgentStrategy as Yt,getSkill as zt}from"@zibby/agent-workflow";import{execSync as Ht}from"node:child_process";import{zodToJsonSchema as Xt}from"zod-to-json-schema";var ye=class extends Yt{constructor(){super("codex","Codex (OpenAI)",75)}canHandle(e){if(!!!(process.env.OPENAI_API_KEY||process.env.OPENAI_API_KEY_POOL||process.env.CODEX_API_KEY||process.env.CODEX_API_KEY_POOL))return a.debug("CodexAgentStrategy: OPENAI_API_KEY or CODEX_API_KEY not set"),!1;try{return Ht("codex --version",{encoding:"utf-8",timeout:5e3,stdio:"pipe"}),!0}catch{return a.warn("[Codex] codex CLI not found. Install: npm install -g @openai/codex"),!1}}async invoke(e,t={}){let{model:r,workspace:o=process.cwd(),schema:n=null,skills:i=null,sessionPath:l=null,nodeName:c=null,timeout:p,config:f={},signal:_=null}=t,m=[G("OPENAI_API_KEY"),G("CODEX_API_KEY")];try{let{Codex:g}=await import("@openai/codex-sdk"),S=r;(!S||S==="auto")&&(a.debug(`Model is '${S||"undefined"}', using default: ${B.CODEX}`),S=B.CODEX);let N=ke[S]||S;ke[S]&&S!==N&&a.debug(`Mapped model: ${S} \u2192 ${N}`),a.debug(`Invoking Codex SDK with model: ${N}, skills: ${JSON.stringify(i)}`);let h=process.env.CODEX_API_KEY||process.env.OPENAI_API_KEY;h&&!process.env.CODEX_API_KEY&&(process.env.CODEX_API_KEY=h);let I=h?` | key: ***${h.slice(-4)}`:" | key: not set";console.log(`
|
|
81
81
|
\u25C6 Model: ${N}${I}
|
|
82
82
|
`);let T=(await import("chalk")).default;console.log(`
|
|
83
|
-
${T.bold("Prompt sent to LLM:")}`),console.log(T.dim("\u2500".repeat(60))),console.log(T.dim(e)),console.log(T.dim("\u2500".repeat(60)));let $=this._resolveSkillsToMcp(i,{sessionPath:l,workspace:o,nodeName:c}),y={};Object.keys($).length>0&&(y.mcp_servers=$,a.debug(`[Codex] MCP servers: ${Object.keys($).join(", ")}`));let C=new g({...Object.keys(y).length>0&&{config:y}}).startThread({workingDirectory:o,skipGitRepoCheck:!0,approvalPolicy:"never",sandboxMode:"danger-full-access",networkAccessEnabled:!0}),k=n&&typeof n.parse=="function",O={};if(n)try{let E=k?Xt(n,{target:"openAi"}):n;O.outputSchema=E,a.debug("Structured output via SDK outputSchema")}catch(E){a.warn(`[Codex] Schema conversion failed, will extract from text: ${E.message}`)}if(_?.aborted){let E=new Error("Aborted via signal");throw E.name="AbortError",E}try{let{events:E}=await C.runStreamed(e,O),L=0,b="";for await(let P of E){if(_?.aborted){let v=new Error("Aborted via signal");throw v.name="AbortError",v}let U=P.type;if(U==="item.completed"){let v=P.item,K=v?.type;if(K==="mcp_tool_call"){L++;let w=`${v.server}/${v.tool}`;if(Y.stepTool(`Tool: ${w}`),v.arguments){let u=JSON.stringify(v.arguments),x=u.length>100?`${u.substring(0,100)}...`:u;console.log(` Input: ${x}`)}}else if(K==="tool_call"||K==="function_call"||K==="command_execution"){L++;let w=v.name||v.tool||v.command||"unknown";Y.stepTool(`Tool: ${w}`)}else K==="agent_message"&&(b=v.text||"",b.length<500?console.log(b):console.log(`${b.substring(0,200)}... (${b.length} chars)`))}else U==="turn.completed"?a.debug(`[Codex] Turn completed. Usage: ${JSON.stringify(P.usage||{})}`):a.debug(`[Codex] Event: ${U} ${JSON.stringify(P).slice(0,300)}`)}if(a.debug(`[Codex] Last agent message (${b.length} chars): ${b.slice(0,500)}`),n){if(!b)throw new Error("Codex agent returned no response");let P=JSON.parse(b),U=k?n.parse(P):P;return a.debug("\u2705 [Codex] Structured output validated"),{raw:b,structured:U}}return b||""}catch(E){let L=E.message||String(E);throw a.error(`\u274C [Codex] SDK call failed: ${L}`),L.includes("exited with code")&&(a.error("\u{1F4A1} [Codex] Verify: codex --version && echo $OPENAI_API_KEY"),a.error("\u{1F4A1} [Codex] If codex is missing: npm install -g @openai/codex")),E}}finally{for(let g of m)g.dispose()}}_resolveSkillsToMcp(e,t={}){if(!Array.isArray(e)||e.length===0)return{};let r={};for(let o of e){let n=zt(o);if(!n){a.warn(`[Codex] Unknown skill "${o}" \u2014 skipping`);continue}if(typeof n.resolve!="function")continue;let i=n.resolve(t);if(!i)continue;let l=n.serverName||o,c={command:i.command};i.args?.length&&(c.args=i.args),i.env&&Object.keys(i.env).length>0&&(c.env=i.env),r[l]=c,a.debug(`[Codex] MCP: ${l} \u2192 ${i.command} ${(i.args||[]).join(" ")}`)}return r}};import{AgentStrategy as
|
|
84
|
-
${d}`,"schema_instruction","append");let D=`zibby-result-${Date.now()}.json`,R=
|
|
83
|
+
${T.bold("Prompt sent to LLM:")}`),console.log(T.dim("\u2500".repeat(60))),console.log(T.dim(e)),console.log(T.dim("\u2500".repeat(60)));let $=this._resolveSkillsToMcp(i,{sessionPath:l,workspace:o,nodeName:c}),y={};Object.keys($).length>0&&(y.mcp_servers=$,a.debug(`[Codex] MCP servers: ${Object.keys($).join(", ")}`));let C=new g({...Object.keys(y).length>0&&{config:y}}).startThread({workingDirectory:o,skipGitRepoCheck:!0,approvalPolicy:"never",sandboxMode:"danger-full-access",networkAccessEnabled:!0}),k=n&&typeof n.parse=="function",O={};if(n)try{let E=k?Xt(n,{target:"openAi"}):n;O.outputSchema=E,a.debug("Structured output via SDK outputSchema")}catch(E){a.warn(`[Codex] Schema conversion failed, will extract from text: ${E.message}`)}if(_?.aborted){let E=new Error("Aborted via signal");throw E.name="AbortError",E}try{let{events:E}=await C.runStreamed(e,O),L=0,b="";for await(let P of E){if(_?.aborted){let v=new Error("Aborted via signal");throw v.name="AbortError",v}let U=P.type;if(U==="item.completed"){let v=P.item,K=v?.type;if(K==="mcp_tool_call"){L++;let w=`${v.server}/${v.tool}`;if(Y.stepTool(`Tool: ${w}`),v.arguments){let u=JSON.stringify(v.arguments),x=u.length>100?`${u.substring(0,100)}...`:u;console.log(` Input: ${x}`)}}else if(K==="tool_call"||K==="function_call"||K==="command_execution"){L++;let w=v.name||v.tool||v.command||"unknown";Y.stepTool(`Tool: ${w}`)}else K==="agent_message"&&(b=v.text||"",b.length<500?console.log(b):console.log(`${b.substring(0,200)}... (${b.length} chars)`))}else U==="turn.completed"?a.debug(`[Codex] Turn completed. Usage: ${JSON.stringify(P.usage||{})}`):a.debug(`[Codex] Event: ${U} ${JSON.stringify(P).slice(0,300)}`)}if(a.debug(`[Codex] Last agent message (${b.length} chars): ${b.slice(0,500)}`),n){if(!b)throw new Error("Codex agent returned no response");let P=JSON.parse(b),U=k?n.parse(P):P;return a.debug("\u2705 [Codex] Structured output validated"),{raw:b,structured:U}}return b||""}catch(E){let L=E.message||String(E);throw a.error(`\u274C [Codex] SDK call failed: ${L}`),L.includes("exited with code")&&(a.error("\u{1F4A1} [Codex] Verify: codex --version && echo $OPENAI_API_KEY"),a.error("\u{1F4A1} [Codex] If codex is missing: npm install -g @openai/codex")),E}}finally{for(let g of m)g.dispose()}}_resolveSkillsToMcp(e,t={}){if(!Array.isArray(e)||e.length===0)return{};let r={};for(let o of e){let n=zt(o);if(!n){a.warn(`[Codex] Unknown skill "${o}" \u2014 skipping`);continue}if(typeof n.resolve!="function")continue;let i=n.resolve(t);if(!i)continue;let l=n.serverName||o,c={command:i.command};i.args?.length&&(c.args=i.args),i.env&&Object.keys(i.env).length>0&&(c.env=i.env),r[l]=c,a.debug(`[Codex] MCP: ${l} \u2192 ${i.command} ${(i.args||[]).join(" ")}`)}return r}};import{AgentStrategy as Wt,getSkill as Zt}from"@zibby/agent-workflow";import{execSync as qt,spawn as Vt}from"node:child_process";import{zodToJsonSchema as Qt}from"zod-to-json-schema";import{existsSync as rt,mkdirSync as nt,readFileSync as st,rmSync as eo,writeFileSync as it}from"node:fs";import{join as Z}from"node:path";function to(s){if(!s)return null;let e=String(s),t=e.match(/```(?:json)?\s*([\s\S]*?)```/i);if(t?.[1])try{return JSON.parse(t[1].trim())}catch{}let r=e.indexOf("{");if(r<0)return null;let o=0,n=!1,i=!1,l=-1;for(let c=r;c<e.length;c++){let p=e[c];if(n){i?i=!1:p==="\\"?i=!0:p==='"'&&(n=!1);continue}if(p==='"'){n=!0;continue}if(p==="{"){o===0&&(l=c),o+=1;continue}if(p==="}"){if(o===0)continue;if(o-=1,o===0&&l>=0){let f=e.slice(l,c+1);try{return JSON.parse(f)}catch{l=-1}}}}return null}function oo(s){let e=String(s||"").trim();if(!e)return null;try{return JSON.parse(e)}catch{return to(e)}}function ro(s){try{let e=JSON.parse(s);if(typeof e=="string")return e;if(typeof e?.response=="string")return e.response;if(typeof e?.text=="string")return e.text;if(typeof e?.output=="string")return e.output;if(Array.isArray(e?.candidates)&&e.candidates.length>0){let t=e.candidates[0];if(typeof t?.content=="string")return t.content;if(Array.isArray(t?.content?.parts)){let r=t.content.parts.map(o=>typeof o?.text=="string"?o.text:"").join("");if(r.trim())return r}}}catch{}return s}var _e=class extends Wt{constructor(){super("gemini","Gemini (Google)",70)}canHandle(e){if(!!!(process.env.GEMINI_API_KEY||process.env.GEMINI_API_KEY_POOL||process.env.GOOGLE_API_KEY||process.env.GOOGLE_API_KEY_POOL))return a.debug("GeminiAgentStrategy: GEMINI_API_KEY or GOOGLE_API_KEY not set"),!1;try{return qt("gemini --version",{encoding:"utf-8",timeout:5e3,stdio:"pipe"}),!0}catch{return a.warn("[Gemini] gemini CLI not found. Install: npm install -g @google/gemini-cli"),!1}}async invoke(e,t={}){let{model:r,workspace:o=process.cwd(),schema:n=null,skills:i=null,sessionPath:l=null,nodeName:c=null,timeout:p=600*1e3,signal:f=null}=t,_=[G("GEMINI_API_KEY"),G("GOOGLE_API_KEY")];try{let m=r;(!m||m==="auto")&&(m=B.GEMINI);let g=Ge[m]||m,S=String(process.env.GEMINI_API_KEY||"").trim(),N=String(process.env.GOOGLE_API_KEY||"").trim(),h=this._resolveSkillsToMcp(i,{sessionPath:l,workspace:o,nodeName:c}),I=Object.keys(h).length>0,T=new V(e),$=n&&typeof n.parse=="function",y=null;if(n){let d;try{let D=$?Qt(n,{target:"openAi"}):n;d=JSON.stringify(D,null,2)}catch{d="{}"}if(I){T.addSystemInstruction(`Write valid JSON that matches this schema:
|
|
84
|
+
${d}`,"schema_instruction","append");let D=`zibby-result-${Date.now()}.json`,R=Z(o,".zibby","tmp");y=Z(R,D),nt(R,{recursive:!0}),T.addStructuredOutput(n,y)}else T.addSystemInstruction(`Return ONLY valid JSON (no markdown, no commentary) that matches this schema:
|
|
85
85
|
${d}`,"json_instruction","append")}let M=T.build(),C=T.getUserPrompt(),k=T.getStats(),O=String(process.env.GEMINI_API_KEY||process.env.GOOGLE_API_KEY||"").trim(),E=O?` | key: ***${O.slice(-4)}`:" | key: not set";console.log(`
|
|
86
86
|
\u25C6 Model: ${g||"auto"}${E}
|
|
87
87
|
`);let L=(await import("chalk")).default;console.log(`
|
|
88
88
|
${L.bold("Prompt sent to LLM:")}`),console.log(L.dim("\u2500".repeat(60))),console.log(L.dim(C)),console.log(L.dim("\u2500".repeat(60)));let b=this._createGeminiConfigDir(o,h),P=["--output-format","json"];g&&g!=="auto"&&P.push("--model",g);let U=Object.keys(h);if(U.length>0){P.push("--approval-mode","yolo");for(let d of U)P.push("--allowed-mcp-server-names",d);a.info(`[Gemini] Enabling MCP servers: ${U.join(", ")}`)}else i&&i.length>0&&a.warn(`[Gemini] Skills requested but no MCP servers configured: ${i.join(", ")}`);P.push("-p",M);let v={...process.env,GEMINI_CLI_HOME:b};S?(v.GEMINI_API_KEY=S,delete v.GOOGLE_API_KEY):N&&(v.GOOGLE_API_KEY=N,delete v.GEMINI_API_KEY),a.debug(`[Gemini] Command: gemini ${P.slice(0,8).join(" ")}... (${P.length} total args)`),a.debug(`[Gemini] User prompt: ${k.userPromptLength} chars, System instructions: ${k.systemInstructionsLength} chars (${k.instructionCount} blocks), Full: ${k.fullPromptLength} chars`),a.debug(`[Gemini] Config home: ${b}`),a.debug(`[Gemini] GEMINI_CLI_HOME env: ${v.GEMINI_CLI_HOME}`);let K="",w=null;try{K=await new Promise((D,R)=>{let J=Vt("gemini",P,{cwd:o,env:v,stdio:["ignore","pipe","pipe"],...f&&{signal:f}}),X="",Ke="",Ie=!1,se=null,$e=!1,Be=setTimeout(()=>{try{J.kill("SIGTERM")}catch{}},p),pe=()=>{$e||Ie||($e=!0,se=setTimeout(()=>{if(!Ie)try{J.kill("SIGKILL")}catch{}},5e3))};f&&(f.aborted?pe():f.addEventListener("abort",pe,{once:!0})),J.stdout.on("data",H=>{X+=H.toString()}),J.stderr.on("data",H=>{Ke+=H.toString()}),J.on("error",H=>{if(clearTimeout(Be),se&&clearTimeout(se),f&&!f.aborted)try{f.removeEventListener("abort",pe)}catch{}R(H)}),J.on("close",H=>{if(Ie=!0,clearTimeout(Be),se&&clearTimeout(se),f&&!f.aborted)try{f.removeEventListener("abort",pe)}catch{}if($e){let Je=new Error("Aborted via signal");return Je.name="AbortError",R(Je)}if(H===0)return D(X.trim());R(new Error(`gemini failed with code ${H}: ${(Ke||X).trim()}`))})})}catch(d){w=d}finally{try{eo(b,{recursive:!0,force:!0})}catch{}}let u=ro(K).trim();if(!n){if(w)throw w;return u}if(y){let d=rt(y);if(a.info(`[Gemini] Result file: ${d?"present":"missing"} at ${y}`),d)try{let D=st(y,"utf-8").trim(),R=JSON.parse(D),J=$?n.parse(R):R;return a.info("[Gemini] Structured output recovered from result file"),{raw:u,structured:J}}catch(D){a.warn(`[Gemini] Result file parse/validation failed: ${D.message}`)}else w||a.warn("[Gemini] Result file missing; falling back to stream-parsed JSON")}let x=null;if(n){let d=new q;d.zodSchema=n,d.processChunk(u),d.flush(),x=d.getResult()}if(a.info(`[Gemini] Raw stdout length: ${K.length} chars`),a.info(`[Gemini] Extracted text length: ${u.length} chars`),a.info(`[Gemini] StreamParser result: ${x?"extracted":"null"}`),x||(u.length<2e3?a.info(`[Gemini] Raw text preview:
|
|
89
89
|
${u}`):a.info(`[Gemini] Raw text preview (first 1000 chars):
|
|
90
|
-
${u.slice(0,1e3)}`),x=oo(u)),!x)throw w||(a.error("[Gemini] Failed to extract valid JSON from output"),a.error("\u{1F4A1} Tip: Set strictMode=true in .zibby.config.js for OpenAI proxy fallback"),new Error("Gemini did not return valid JSON for structured output. Enable strictMode for proxy fallback."));let A=$?n.parse(x):x;return{raw:u,structured:A}}finally{for(let m of _)m.dispose()}}_resolveSkillsToMcp(e,t={}){if(!Array.isArray(e)||e.length===0)return{};let r={};for(let o of e){let n=
|
|
91
|
-
`,"utf-8");let f=
|
|
90
|
+
${u.slice(0,1e3)}`),x=oo(u)),!x)throw w||(a.error("[Gemini] Failed to extract valid JSON from output"),a.error("\u{1F4A1} Tip: Set strictMode=true in .zibby.config.js for OpenAI proxy fallback"),new Error("Gemini did not return valid JSON for structured output. Enable strictMode for proxy fallback."));let A=$?n.parse(x):x;return{raw:u,structured:A}}finally{for(let m of _)m.dispose()}}_resolveSkillsToMcp(e,t={}){if(!Array.isArray(e)||e.length===0)return{};let r={};for(let o of e){let n=Zt(o);if(!n||typeof n.resolve!="function")continue;let i=n.resolve(t);if(!i)continue;let l=n.cursorKey||n.serverName||o,c={command:i.command};i.args?.length&&(c.args=i.args),i.env&&Object.keys(i.env).length>0&&(c.env=i.env),i.cwd&&(c.cwd=i.cwd),r[l]=c}return r}_createGeminiConfigDir(e,t){let r=`${Date.now()}-${Math.random().toString(16).slice(2,10)}`,o=Z(e||process.cwd(),".zibby","tmp",`gemini-home-${r}`),n=Z(o,".gemini");nt(n,{recursive:!0});let i=Z(n,"settings.json"),l={},c=Z(process.env.HOME||"",".gemini","settings.json");if(rt(c))try{l=JSON.parse(st(c,"utf-8"))}catch{l={}}let p={...l,mcpServers:{...l.mcpServers&&typeof l.mcpServers=="object"?l.mcpServers:{},...t||{}}};it(i,`${JSON.stringify(p,null,2)}
|
|
91
|
+
`,"utf-8");let f=Z(e||process.cwd(),".zibby","tmp","gemini-settings-debug.json");try{it(f,`${JSON.stringify(p,null,2)}
|
|
92
92
|
`,"utf-8")}catch{}return a.debug(`[Gemini] Created isolated config with ${Object.keys(p.mcpServers||{}).length} MCP servers`),a.debug(`[Gemini] MCP servers: ${JSON.stringify(Object.keys(p.mcpServers||{}),null,2)}`),o}};import{AgentStrategy as yo,getSkill as we}from"@zibby/agent-workflow";var ue=class{formatTools(e){throw new Error("formatTools() must be implemented")}hasToolCalls(e){throw new Error("hasToolCalls() must be implemented")}parseToolCalls(e){throw new Error("parseToolCalls() must be implemented")}getTextContent(e){throw new Error("getTextContent() must be implemented")}buildAssistantMessage(e){throw new Error("buildAssistantMessage() must be implemented")}buildToolResultMessage(e,t){throw new Error("buildToolResultMessage() must be implemented")}injectToolsIntoBody(e,t){throw new Error("injectToolsIntoBody() must be implemented")}};var te=class extends ue{formatTools(e){return e.map(t=>({type:"function",function:{name:t.name,description:t.description,parameters:t.parameters||t.input_schema||{type:"object",properties:{}}}}))}hasToolCalls(e){let t=e.choices?.[0]?.message;return!!(t?.tool_calls&&t.tool_calls.length>0)}parseToolCalls(e){return(e.choices?.[0]?.message?.tool_calls||[]).map(r=>({id:r.id,name:r.function.name,args:JSON.parse(r.function.arguments||"{}")}))}getTextContent(e){return e.choices?.[0]?.message?.content||""}buildAssistantMessage(e){return e.choices?.[0]?.message}buildToolResultMessage(e,t){return{role:"tool",tool_call_id:e,content:typeof t=="string"?t:JSON.stringify(t)}}injectToolsIntoBody(e,t){return t.length>0&&(e.tools=t),e}};var fe=class{async fetchCompletion(e,t,r={}){throw new Error("fetchCompletion() must be implemented")}async fetchStreamingCompletion(e,t,r={}){throw new Error("fetchStreamingCompletion() must be implemented")}};function Se(s){return Buffer.byteLength(JSON.stringify(s),"utf8")}function Ee(s,e){let t=String(s||"");if(t.length<=e)return t;let r=Math.max(0,e-28);return`${t.slice(0,r)}
|
|
93
93
|
|
|
94
94
|
[truncated for size budget]`}function Me(s,e=0){if(!s||typeof s!="object"||e>8)return s;if(Array.isArray(s))return s.map(r=>Me(r,e+1));let t={};for(let[r,o]of Object.entries(s))r==="description"||r==="title"||r==="examples"||r==="default"||(t[r]=Me(o,e+1));return t}function no(s=[]){return s.map(e=>({...e,function:{...e.function,description:Ee(e.function?.description||"",180),parameters:Me(e.function?.parameters||{type:"object",properties:{}})}}))}function lt(s){let e=new Set;for(let o of s)if(o.role==="assistant"&&Array.isArray(o.tool_calls))for(let n of o.tool_calls)e.add(n.id);let t=s.filter(o=>o.role==="tool"?e.has(o.tool_call_id):!0),r=new Set;for(let o of t)o.role==="tool"&&r.add(o.tool_call_id);return t.map(o=>{if(o.role!=="assistant"||!Array.isArray(o.tool_calls)||o.tool_calls.every(c=>r.has(c.id)))return o;let{tool_calls:i,...l}=o;return{...l,content:l.content||""}})}function Le(s,e={}){let t=e.maxBytes||49e3,r=e.systemMaxChars||12e3,o={...s,messages:Array.isArray(s.messages)?[...s.messages]:[],tools:Array.isArray(s.tools)?no(s.tools):s.tools};o.messages.length>0&&o.messages[0]?.role==="system"&&(o.messages[0]={...o.messages[0],content:Ee(o.messages[0].content,r)});let n=!1;for(;Se(o)>t&&o.messages.length>2;)o.messages.splice(1,1),n=!0;if(n&&(o.messages=lt(o.messages)),Se(o)>t&&o.messages.length>0&&(o.messages[0]={...o.messages[0],content:Ee(o.messages[0].content,6e3)},n=!0),Se(o)>t){let i=o.messages.find(c=>c.role==="system")||o.messages[0],l=o.messages.slice(-2);o.messages=lt([i,...l].filter(Boolean).map((c,p)=>({...c,content:Ee(c.content,p===0?4e3:8e3)}))),n=!0}return{body:o,meta:{bytes:Se(o),trimmed:n,maxBytes:t,messageCount:o.messages.length}}}var at=s=>Buffer.byteLength(JSON.stringify(s),"utf8"),oe=class extends fe{async fetchCompletion(e,t,r={}){let o=at(e),{body:n,meta:i}=Le(e,r.payloadCompaction);r.onBudget?.({streaming:!1,beforeBytes:o,meta:i});let l=this.#e(r),c=`${t.baseUrl}${r.chatCompletionsPath||"/v1/chat/completions"}`,p=await fetch(c,{method:"POST",headers:t.headers,body:JSON.stringify(n),signal:l});if(!p.ok){let f=await p.text();throw p.status===401||p.status===403?new Error("Session expired. Run `zibby login` to re-authenticate."):new Error(`Proxy error ${p.status}: ${f}`)}return p.json()}async fetchStreamingCompletion(e,t,r={}){let o={...e,stream:!0},n=at(o),{body:i,meta:l}=Le(o,r.payloadCompaction);r.onBudget?.({streaming:!0,beforeBytes:n,meta:l});let c=this.#e(r),p=`${t.baseUrl}${r.chatCompletionsPath||"/v1/chat/completions"}`,f=await fetch(p,{method:"POST",headers:t.headers,body:JSON.stringify(i),signal:c});if(!f.ok){let h=await f.text();throw f.status===401||f.status===403?new Error("Session expired. Run `zibby login` to re-authenticate."):new Error(`Proxy error ${f.status}: ${h}`)}let _=f.body.getReader(),m=new TextDecoder,g="",S="",N=new Map;for(;;){let{done:h,value:I}=await _.read();if(h)break;g+=m.decode(I,{stream:!0});let T=g.split(`
|
|
@@ -8,6 +8,7 @@
|
|
|
8
8
|
import { WorkflowAgent, WorkflowGraph } from '@zibby/core';
|
|
9
9
|
import {
|
|
10
10
|
preflightNode,
|
|
11
|
+
cacheReplayNode,
|
|
11
12
|
executeLiveNode,
|
|
12
13
|
generateScriptNode,
|
|
13
14
|
} from './nodes/index.mjs';
|
|
@@ -20,9 +21,10 @@ export class BrowserTestAutomationAgent extends WorkflowAgent {
|
|
|
20
21
|
graph.setStateSchema(browserTestAutomationStateSchema);
|
|
21
22
|
|
|
22
23
|
graph.addNode('preflight', preflightNode);
|
|
24
|
+
graph.addNode('cache_replay', cacheReplayNode);
|
|
23
25
|
graph.addNode('execute_live', executeLiveNode);
|
|
24
26
|
graph.addNode('generate_script', generateScriptNode);
|
|
25
|
-
|
|
27
|
+
|
|
26
28
|
graph.setEntryPoint('preflight');
|
|
27
29
|
|
|
28
30
|
// Short-circuit when preflight produced nothing usable. Triggered when:
|
|
@@ -35,7 +37,17 @@ export class BrowserTestAutomationAgent extends WorkflowAgent {
|
|
|
35
37
|
// ~30s before failing — bad UX and bad bill.
|
|
36
38
|
graph.addConditionalEdges('preflight', (state) => {
|
|
37
39
|
const assertions = state.preflight?.assertions || [];
|
|
38
|
-
return assertions.length > 0 ? '
|
|
40
|
+
return assertions.length > 0 ? 'cache_replay' : 'END';
|
|
41
|
+
});
|
|
42
|
+
|
|
43
|
+
// Lever-#2 fork: cache_replay attempted a Playwright-only replay of
|
|
44
|
+
// a prior successful action sequence. On hit it side-wrote
|
|
45
|
+
// state.execute_live with synthesized output, so we can skip
|
|
46
|
+
// execute_live and jump straight to generate_script — zero LLM
|
|
47
|
+
// tokens. On miss / replay failure / cold cache, fall through to
|
|
48
|
+
// the normal LLM-driven execute_live path.
|
|
49
|
+
graph.addConditionalEdges('cache_replay', (state) => {
|
|
50
|
+
return state.cache_replay?.hit === true ? 'generate_script' : 'execute_live';
|
|
39
51
|
});
|
|
40
52
|
|
|
41
53
|
graph.addConditionalEdges('execute_live', (state) => {
|
|
@@ -0,0 +1,213 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* cache_replay node — lever-#2 read path inside the workflow.
|
|
3
|
+
*
|
|
4
|
+
* Sits between `preflight` and `execute_live` in the graph. Tries to
|
|
5
|
+
* replay a prior successful run's action sequence via Playwright
|
|
6
|
+
* directly, completely skipping the LLM. On a clean cache hit it
|
|
7
|
+
* populates `state.execute_live` with the result so downstream
|
|
8
|
+
* `generate_script` works exactly as if execute_live had run.
|
|
9
|
+
*
|
|
10
|
+
* Conditional edge after this node:
|
|
11
|
+
* - state.cache_replay.hit === true → skip execute_live → generate_script
|
|
12
|
+
* - state.cache_replay.hit === false → execute_live (LLM-driven path)
|
|
13
|
+
*
|
|
14
|
+
* Not user-configurable per-spec — the cache key derivation handles
|
|
15
|
+
* staleness (page fingerprint drift invalidates) and replay failures
|
|
16
|
+
* fall through cleanly to the LLM path.
|
|
17
|
+
*/
|
|
18
|
+
|
|
19
|
+
import { z } from '@zibby/core';
|
|
20
|
+
import { chromium } from 'playwright';
|
|
21
|
+
import { spawn } from 'child_process';
|
|
22
|
+
import { extractDomain, replayActions } from '@zibby/ui-memory';
|
|
23
|
+
import { join } from 'path';
|
|
24
|
+
|
|
25
|
+
const REPLAY_TIMEOUT_MS = 60_000;
|
|
26
|
+
|
|
27
|
+
export const cacheReplayNode = {
|
|
28
|
+
name: 'cache_replay',
|
|
29
|
+
skills: [],
|
|
30
|
+
timeout: 90000,
|
|
31
|
+
outputSchema: z.object({
|
|
32
|
+
hit: z.boolean(),
|
|
33
|
+
elapsed_ms: z.number().nullish(),
|
|
34
|
+
executed: z.number().nullish(),
|
|
35
|
+
total: z.number().nullish(),
|
|
36
|
+
cache_key: z.string().nullish(),
|
|
37
|
+
error: z.string().nullish(),
|
|
38
|
+
// When hit, we also write a synthesized execute_live block so the
|
|
39
|
+
// downstream generate_script node sees what it expects.
|
|
40
|
+
execute_live_synthesized: z.boolean().nullish(),
|
|
41
|
+
}),
|
|
42
|
+
|
|
43
|
+
execute: async (context) => {
|
|
44
|
+
// graph.js builds nodeContext as `{ state, invokeAgent, _coreInvokeAgent,
|
|
45
|
+
// ...state.getAll() }`. So `context.testSpec` works (spread) AND
|
|
46
|
+
// `context.state.get('testSpec')` works (instance). Reading from the
|
|
47
|
+
// spread is the natural shape — `context.state` is reserved for the
|
|
48
|
+
// .set(key, value) side-write below.
|
|
49
|
+
const cwd = context.cwd || context.workspace || process.cwd();
|
|
50
|
+
const testSpec = context.testSpec || '';
|
|
51
|
+
const specPath = context.specPath || '';
|
|
52
|
+
|
|
53
|
+
// Derive domain from the spec text (no DOM access yet — pure parse).
|
|
54
|
+
const domain = extractDomainFromSpec(testSpec);
|
|
55
|
+
if (!domain) {
|
|
56
|
+
return { hit: false, error: 'cannot derive domain from spec' };
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
// Cache key requires page_fingerprint, which is page-state-dependent
|
|
60
|
+
// and only available AFTER navigation. We compute a key WITHOUT
|
|
61
|
+
// fingerprint first and look up by (domain, spec_path) prefix —
|
|
62
|
+
// the persister wrote spec_path too. If we find a candidate, we
|
|
63
|
+
// use its stored fingerprint to compute the full key and verify.
|
|
64
|
+
//
|
|
65
|
+
// Lookup order:
|
|
66
|
+
// 1. Exact (domain, spec_path) match in action_cache.
|
|
67
|
+
// 2. If found, use its actions for replay attempt.
|
|
68
|
+
// 3. On replay success: signal hit, populate state.execute_live.
|
|
69
|
+
// 4. On replay failure (or cache miss): hit=false, fall back to LLM.
|
|
70
|
+
const cached = await lookupCacheByDomainAndSpec({ cwd, domain, specPath });
|
|
71
|
+
if (!cached) {
|
|
72
|
+
return { hit: false, error: 'no cached actions for this spec' };
|
|
73
|
+
}
|
|
74
|
+
|
|
75
|
+
// Run the replay in a freshly-launched Playwright browser. Cleanly
|
|
76
|
+
// independent from the @zibby/mcp-browser path execute_live uses.
|
|
77
|
+
const t0 = Date.now();
|
|
78
|
+
const browser = await chromium.launch({ headless: true });
|
|
79
|
+
const page = await browser.newPage();
|
|
80
|
+
let replayResult;
|
|
81
|
+
try {
|
|
82
|
+
replayResult = await Promise.race([
|
|
83
|
+
replayActions({
|
|
84
|
+
actions: cached.actions,
|
|
85
|
+
page,
|
|
86
|
+
log: (m) => console.log(`[cache_replay] ${m}`),
|
|
87
|
+
}),
|
|
88
|
+
new Promise((_, reject) =>
|
|
89
|
+
setTimeout(() => reject(new Error('replay timeout')), REPLAY_TIMEOUT_MS),
|
|
90
|
+
),
|
|
91
|
+
]);
|
|
92
|
+
} catch (err) {
|
|
93
|
+
replayResult = { success: false, error: err.message, executed: 0, total: cached.actions.length };
|
|
94
|
+
}
|
|
95
|
+
const finalUrl = page.url();
|
|
96
|
+
await browser.close().catch(() => {});
|
|
97
|
+
const elapsedMs = Date.now() - t0;
|
|
98
|
+
|
|
99
|
+
if (!replayResult.success) {
|
|
100
|
+
// Increment failure_count so we can drop chronic misses later.
|
|
101
|
+
await incrementCacheFailure({ cwd, cacheKey: cached.cache_key });
|
|
102
|
+
return {
|
|
103
|
+
hit: false,
|
|
104
|
+
elapsed_ms: elapsedMs,
|
|
105
|
+
executed: replayResult.executed,
|
|
106
|
+
total: replayResult.total,
|
|
107
|
+
cache_key: cached.cache_key,
|
|
108
|
+
error: replayResult.error,
|
|
109
|
+
};
|
|
110
|
+
}
|
|
111
|
+
|
|
112
|
+
// HIT path. Side-write the synthesized execute_live output via
|
|
113
|
+
// context.state.set so downstream generate_script reads the same
|
|
114
|
+
// shape it expects (actions[], finalUrl, …). The customExecute
|
|
115
|
+
// return-value lands in state.cache_replay; the execute_live slot
|
|
116
|
+
// has to be populated separately.
|
|
117
|
+
if (typeof context.state?.set === 'function') {
|
|
118
|
+
context.state.set('execute_live', {
|
|
119
|
+
success: true,
|
|
120
|
+
steps: cached.actions.map((a) => a.description),
|
|
121
|
+
actions: cached.actions,
|
|
122
|
+
assertions: [],
|
|
123
|
+
finalUrl,
|
|
124
|
+
browserClosed: true,
|
|
125
|
+
notes: 'cache_replay hit — actions replayed via Playwright, no LLM',
|
|
126
|
+
});
|
|
127
|
+
}
|
|
128
|
+
|
|
129
|
+
return {
|
|
130
|
+
hit: true,
|
|
131
|
+
elapsed_ms: elapsedMs,
|
|
132
|
+
executed: replayResult.executed,
|
|
133
|
+
total: replayResult.total,
|
|
134
|
+
cache_key: cached.cache_key,
|
|
135
|
+
execute_live_synthesized: true,
|
|
136
|
+
};
|
|
137
|
+
},
|
|
138
|
+
};
|
|
139
|
+
|
|
140
|
+
// ─── helpers ────────────────────────────────────────────────────────────
|
|
141
|
+
|
|
142
|
+
function extractDomainFromSpec(spec) {
|
|
143
|
+
if (!spec) return null;
|
|
144
|
+
// Find the first http(s) URL in the spec and run it through the
|
|
145
|
+
// SAME `extractDomain` the persister uses, so the cache-key lookup
|
|
146
|
+
// matches what was actually written (notably: `www.` is stripped).
|
|
147
|
+
const m = String(spec).match(/https?:\/\/[^\s"'<>]+/);
|
|
148
|
+
if (!m) return null;
|
|
149
|
+
return extractDomain(m[0]);
|
|
150
|
+
}
|
|
151
|
+
|
|
152
|
+
/**
|
|
153
|
+
* Find a cached row by (domain, spec_path). Picks the row with
|
|
154
|
+
* highest success_count if multiple match.
|
|
155
|
+
* Uses dolt via subprocess (matching the rest of the codebase's
|
|
156
|
+
* Dolt-access pattern).
|
|
157
|
+
*/
|
|
158
|
+
async function lookupCacheByDomainAndSpec({ cwd, domain, specPath }) {
|
|
159
|
+
const dbDir = join(cwd, '.zibby', 'memory');
|
|
160
|
+
const safeDomain = escapeSql(domain);
|
|
161
|
+
const safeSpec = escapeSql(specPath);
|
|
162
|
+
const sql = `SELECT cache_key, actions_json, page_fingerprint
|
|
163
|
+
FROM action_cache
|
|
164
|
+
WHERE domain = ${safeDomain} AND spec_path = ${safeSpec}
|
|
165
|
+
ORDER BY success_count DESC, last_used_at DESC
|
|
166
|
+
LIMIT 1`;
|
|
167
|
+
const rows = await runDoltJson(dbDir, sql);
|
|
168
|
+
if (!rows || rows.length === 0) return null;
|
|
169
|
+
try {
|
|
170
|
+
const actions = JSON.parse(rows[0].actions_json);
|
|
171
|
+
return { cache_key: rows[0].cache_key, actions, fingerprint: rows[0].page_fingerprint };
|
|
172
|
+
} catch {
|
|
173
|
+
return null;
|
|
174
|
+
}
|
|
175
|
+
}
|
|
176
|
+
|
|
177
|
+
async function incrementCacheFailure({ cwd, cacheKey }) {
|
|
178
|
+
const dbDir = join(cwd, '.zibby', 'memory');
|
|
179
|
+
const sql = `UPDATE action_cache
|
|
180
|
+
SET failure_count = failure_count + 1, last_replay_status = 'replay-failed'
|
|
181
|
+
WHERE cache_key = ${escapeSql(cacheKey)}`;
|
|
182
|
+
await runDoltExec(dbDir, sql).catch(() => { /* non-fatal */ });
|
|
183
|
+
}
|
|
184
|
+
|
|
185
|
+
function escapeSql(v) {
|
|
186
|
+
if (v == null) return 'NULL';
|
|
187
|
+
return `'${String(v).replace(/'/g, "''")}'`;
|
|
188
|
+
}
|
|
189
|
+
|
|
190
|
+
function runDoltJson(dir, sql) {
|
|
191
|
+
return new Promise((resolve) => {
|
|
192
|
+
const child = spawn('dolt', ['sql', '-r', 'json', '-q', sql], { cwd: dir });
|
|
193
|
+
let out = '';
|
|
194
|
+
child.stdout.on('data', (d) => { out += d; });
|
|
195
|
+
child.on('close', () => {
|
|
196
|
+
try {
|
|
197
|
+
const parsed = JSON.parse(out);
|
|
198
|
+
resolve(parsed.rows || []);
|
|
199
|
+
} catch {
|
|
200
|
+
resolve([]);
|
|
201
|
+
}
|
|
202
|
+
});
|
|
203
|
+
child.on('error', () => resolve([]));
|
|
204
|
+
});
|
|
205
|
+
}
|
|
206
|
+
|
|
207
|
+
function runDoltExec(dir, sql) {
|
|
208
|
+
return new Promise((resolve, reject) => {
|
|
209
|
+
const child = spawn('dolt', ['sql', '-q', sql], { cwd: dir });
|
|
210
|
+
child.on('close', (code) => (code === 0 ? resolve() : reject(new Error(`dolt exit ${code}`))));
|
|
211
|
+
child.on('error', reject);
|
|
212
|
+
});
|
|
213
|
+
}
|
package/package.json
CHANGED
|
@@ -8,6 +8,7 @@
|
|
|
8
8
|
import { WorkflowAgent, WorkflowGraph } from '@zibby/core';
|
|
9
9
|
import {
|
|
10
10
|
preflightNode,
|
|
11
|
+
cacheReplayNode,
|
|
11
12
|
executeLiveNode,
|
|
12
13
|
generateScriptNode,
|
|
13
14
|
} from './nodes/index.mjs';
|
|
@@ -20,9 +21,10 @@ export class BrowserTestAutomationAgent extends WorkflowAgent {
|
|
|
20
21
|
graph.setStateSchema(browserTestAutomationStateSchema);
|
|
21
22
|
|
|
22
23
|
graph.addNode('preflight', preflightNode);
|
|
24
|
+
graph.addNode('cache_replay', cacheReplayNode);
|
|
23
25
|
graph.addNode('execute_live', executeLiveNode);
|
|
24
26
|
graph.addNode('generate_script', generateScriptNode);
|
|
25
|
-
|
|
27
|
+
|
|
26
28
|
graph.setEntryPoint('preflight');
|
|
27
29
|
|
|
28
30
|
// Short-circuit when preflight produced nothing usable. Triggered when:
|
|
@@ -35,7 +37,17 @@ export class BrowserTestAutomationAgent extends WorkflowAgent {
|
|
|
35
37
|
// ~30s before failing — bad UX and bad bill.
|
|
36
38
|
graph.addConditionalEdges('preflight', (state) => {
|
|
37
39
|
const assertions = state.preflight?.assertions || [];
|
|
38
|
-
return assertions.length > 0 ? '
|
|
40
|
+
return assertions.length > 0 ? 'cache_replay' : 'END';
|
|
41
|
+
});
|
|
42
|
+
|
|
43
|
+
// Lever-#2 fork: cache_replay attempted a Playwright-only replay of
|
|
44
|
+
// a prior successful action sequence. On hit it side-wrote
|
|
45
|
+
// state.execute_live with synthesized output, so we can skip
|
|
46
|
+
// execute_live and jump straight to generate_script — zero LLM
|
|
47
|
+
// tokens. On miss / replay failure / cold cache, fall through to
|
|
48
|
+
// the normal LLM-driven execute_live path.
|
|
49
|
+
graph.addConditionalEdges('cache_replay', (state) => {
|
|
50
|
+
return state.cache_replay?.hit === true ? 'generate_script' : 'execute_live';
|
|
39
51
|
});
|
|
40
52
|
|
|
41
53
|
graph.addConditionalEdges('execute_live', (state) => {
|