evolve-sdk 0.0.52.dev638__py3-none-any.whl → 0.0.52.dev656__py3-none-any.whl

bridge/dist/bridge.bundle.cjs

@@ -1233,7 +1233,7 @@ PY`,{timeoutMs:1e4})).exitCode!==0)throw new Error("Managed secret egress proxy
 `);d=P.pop()??"";for(let L of P){if(!L.trim())continue;let z=f(L);if(this.captureDroidSession(L,z),this.sessionLogger?.writeEventParsed(L,z),t?.onStdout?.(L+`
 `),z&&t?.onContent)for(let j of z)t.onContent(j)}},m=C=>{t?.onStderr?.(C)},v=await a.commands.spawn(c,{cwd:this.workingDir,timeoutMs:n,envs:u,onStdout:p,onStderr:m}),b=this.beginOperation("run",v,t,"run_start");if(this.hasRun=!0,i)return this.watchBackgroundOperation(b,"run",v,t,a),{sandboxId:a.sandboxId,sessionId:this.managedBrowserSession?.sessionId,browser:this.browserResponseInfo(),runId:l,exitCode:0,stdout:`Background process started with ID ${v.processId}`,stderr:""};let S;try{S=await v.wait()}catch(C){throw this.interruptedOperations.delete(b),this.finalizeOperation(b,t,"run_failed","error"),C}if(this.interruptedOperations.delete(b)||S.exitCode===130?this.finalizeOperation(b,t,"run_interrupted","interrupted"):S.exitCode===0?this.finalizeOperation(b,t,"run_complete","idle"):this.finalizeOperation(b,t,"run_failed","error"),d.trim()){let C=f(d);if(this.captureDroidSession(d,C),this.sessionLogger?.writeEventParsed(d,C),t?.onStdout?.(d+`
 `),C&&t?.onContent)for(let P of C)t.onContent(P)}await this.writeDroidSessionState(a),this.sessionLogger&&!i&&await this.flushSessionLoggerWithTimeout();let T;if(this.storage&&!i&&S.exitCode===0)try{T=await e8t(a,this.storage,this.agentConfig.type,this.workingDir,{tag:this.sessionTag,model:this.agentConfig.model||this.registry.defaultModel,workspaceMode:this.options.workspaceMode||"knowledge",comment:s,parentId:this.lastCheckpointId}),this.lastCheckpointId=T.id}catch(C){console.warn(`[Evolve] Auto-checkpoint failed: ${C.message}`)}return{sandboxId:a.sandboxId,sessionId:this.managedBrowserSession?.sessionId,browser:this.browserResponseInfo(),runId:l,exitCode:S.exitCode,stdout:S.stdout,stderr:S.stderr,checkpoint:T}}async executeCommand(e,t={},r){let{timeoutMs:n=K3,background:i=!1}=t;if(this.activeCommand)throw new Error("Agent is already running. Call interrupt(), wait for the active/background command to finish, or create a new Evolve instance.");let s=await this.getSandbox(r);this.credentialsSealed||await this.ensureProviderRuntimeToken(),this.lastRunTimestamp=Date.now(),this.requiresPreRunDashboardIngest()&&(this.ensureSessionLogger(s),this.sessionLogger?.writePrompt(e),await this.flushSessionLoggerWithTimeout());let o="",a="",c=m=>{o+=m,r?.onStdout?.(m)},l=m=>{a+=m,r?.onStderr?.(m)},u=this.credentialsSealed?{}:this.buildProviderRuntimeProcessEnvs(),d=await s.commands.spawn(e,{cwd:this.workingDir,timeoutMs:n,envs:Object.keys(u).length>0?u:void 0,onStdout:c,onStderr:l}),f=this.beginOperation("command",d,r,"command_start");if(i)return this.watchBackgroundOperation(f,"command",d,r),{sandboxId:s.sandboxId,sessionId:this.managedBrowserSession?.sessionId,browser:this.browserResponseInfo(),exitCode:0,stdout:`Background process started with ID ${d.processId}`,stderr:""};let p;try{p=await d.wait()}catch(m){throw this.interruptedOperations.delete(f),this.finalizeOperation(f,r,"command_failed","error"),m}return this.interruptedOperations.delete(f)||p.exitCode===130?this.finalizeOperation(f,r,"command_interrupted","interrupted"):p.exitCode===0?this.finalizeOperation(f,r,"command_complete","idle"):this.finalizeOperation(f,r,"command_failed","error"),{sandboxId:s.sandboxId,sessionId:this.managedBrowserSession?.sessionId,browser:this.browserResponseInfo(),exitCode:p.exitCode,stdout:o||p.stdout||"",stderr:a||p.stderr||""}}async sealCredentials(){if(!this.credentialsSealed){if(!this.sandbox)throw new Error("No active sandbox. Call run() or executeCommand() first.");if(this.activeCommand)throw new Error("Cannot seal credentials while a process is running.");if(this.agentConfig.isDirectMode&&!this.agentConfig.externalGateway)throw new Error("Credential sealing requires gateway mode; direct provider credentials cannot be revoked by Evolve");if([this.options.secrets&&Object.keys(this.options.secrets).length>0,this.options.sandboxCreateOptions?.envs&&Object.keys(this.options.sandboxCreateOptions.envs).length>0,this.options.managedSecrets,this.options.integrations,this.options.mcpServers&&Object.keys(this.options.mcpServers).length>0,this.options.managedBrowser,this.options.browserCredentials].some(Boolean))throw new Error("Credential sealing requires a credential-minimal sandbox: omit secrets, managed secrets, integrations, MCP servers, and browser credentials");if(this.agentConfig.externalGateway){await this.agentConfig.externalGateway.revoke(),this.credentialsSealed=!0;return}if(!this.providerRuntimeToken||!this.options.providerRouting)throw new Error("Cannot seal credentials: no active provider runtime token to revoke. Without a gateway-issued runtime token the sandbox may hold credentials Evolve cannot revoke, so sealing would be a false guarantee.");await this.closeProviderRuntimeToken(!0),this.credentialsSealed=!0}}isSealed(){return this.credentialsSealed}async collectArtifacts(e){if(!this.credentialsSealed)throw new Error("Call sealCredentials() before collecting artifacts.");if(!this.sandbox)throw new Error("No active sandbox.");if(this.activeCommand)throw new Error("Cannot collect artifacts while a process is running.");return cvn(this.sandbox,this.workingDir,e)}async uploadContext(e){let t=await this.getSandbox();await this.uploadContextFiles(t,e)}async uploadFiles(e){let t=await this.getSandbox();await this.uploadWorkspaceFiles(t,e)}async uploadFileFromPath(e,t){let r=await this.getSandbox(),n=e.startsWith("/")?e:`${this.workingDir}/${e}`,i=n.substring(0,n.lastIndexOf("/"));if(i&&await r.commands.run(`mkdir -p ${i}`,{timeoutMs:3e4}),r.files.writeFromPath){await r.files.writeFromPath(n,t);return}let{readFile:s}=await import("fs/promises");await r.files.write(n,await s(t))}async getOutputFiles(e=!1){let t=await this.getSandbox(),r=`${this.workingDir}/output`,n=e?"":"-maxdepth 1",i=(await t.commands.run(`find ${r} ${n} -type f -exec stat -c '%n|%Z' {} \\; 2>/dev/null || true`,{timeoutMs:3e4})).stdout.split(`
-`).filter(Boolean),s=this.lastRunTimestamp?Math.floor(this.lastRunTimestamp/1e3)-2:0,o=[],a=`${r}/`;for(let f of i){let[p,m]=f.split("|");if(!p||!m)continue;let v=parseInt(m,10);s>0&&v<s||o.push(p)}let c={},l=await Promise.all(o.map(async f=>{try{let p=await t.files.read(f);return{relativePath:f.startsWith(a)?f.slice(a.length):f.split("/").pop()||f,content:p}}catch{return null}}));for(let f of l)f&&(c[f.relativePath]=f.content);if(!this.zodSchema&&!this.jsonSchema)return{files:c,data:null};let u=c["result.json"];if(!u)return{files:c,data:null,error:"Schema provided but agent did not create output/result.json"};let d=typeof u=="string"?u:new TextDecoder().decode(u);try{let f=JSON.parse(d);if(this.zodSchema){let p=this.zodSchema.safeParse(f);return p.success?{files:c,data:p.data}:{files:c,data:null,error:`Schema validation failed: ${p.error.message}`,rawData:d}}if(this.compiledValidator){if(this.compiledValidator(f))return{files:c,data:f};{let p=this.compiledValidator.errors?.map(m=>`${m.instancePath} ${m.message}`).join(", ")||"Unknown validation error";return{files:c,data:null,error:`Schema validation failed: ${p}`,rawData:d}}}return{files:c,data:null}}catch(f){return{files:c,data:null,error:`Failed to parse result.json: ${f.message}`,rawData:d}}}async checkpoint(e){if(!this.storage)throw new Error("Storage not configured. Call .withStorage().");if(!this.sandbox)throw new Error("No active sandbox. Call run() first.");let t=await e8t(this.sandbox,this.storage,this.agentConfig.type,this.workingDir,{tag:this.sessionTag,model:this.agentConfig.model||this.registry.defaultModel,workspaceMode:this.options.workspaceMode||"knowledge",comment:e?.comment,parentId:this.lastCheckpointId});return this.lastCheckpointId=t.id,t}getSession(){return this.sandbox?.sandboxId||this.options.sandboxId||null}async setSession(e){if(this.credentialsSealed)throw new Error("Agent credentials are sealed and this Evolve instance cannot switch sandboxes");if(this.options.managedSecrets)throw new Error("setSession() cannot be used with managed secrets; managed secret grants are sandbox-scoped.");if(this.options.sandboxCreateOptions?.user)throw new Error("setSession() cannot be used with a configured sandbox user; the user cannot be enforced on an existing sandbox.");if(this.activeCommand&&!await this.interrupt())throw new Error("Cannot switch session while an active process is running and could not be interrupted.");await this.rotateSession(),await this.closeProviderRuntimeToken(),await this.closeManagedSecretRuntimeToken(),await this.closeManagedBrowserSession(),this.options.sandboxId=e,this.sandbox=void 0,this.interruptedOperations.clear(),this.invalidateActiveOperation(),this.sandboxState="ready",this.agentState="idle",this.hasRun=!0,this.lastCheckpointId=void 0}async pause(e){this.sandbox&&(this.activeCommand&&await this.interrupt(e),await this.sandbox.pause(),this.sandboxState="paused",this.agentState="idle",this.emitLifecycle(e,"sandbox_pause"))}async resume(e){this.sandbox&&this.options.sandboxProvider&&(this.sandbox=await this.options.sandboxProvider.connect(this.sandbox.sandboxId),this.sandboxState="ready",this.agentState="idle",this.emitLifecycle(e,"sandbox_resume"))}async interrupt(e){if(!this.activeCommand&&!this.activeProcessId)return!1;let t=this.activeOperationId,r=this.activeOperationKind,n=!1;try{this.activeCommand?n=await this.activeCommand.kill():this.sandbox&&this.activeProcessId&&(n=await this.sandbox.commands.kill(this.activeProcessId))}catch{n=!1}if(!n)return this.sandboxState="running",this.agentState="running",!1;t!==null&&this.interruptedOperations.add(t),this.invalidateActiveOperation(),this.sandboxState="ready",this.agentState="interrupted";let i=r==="run"?"run_interrupted":"command_interrupted";return this.emitLifecycle(e,i),n}status(){let e=this.browserRuntimeInfo();return{sandboxId:this.getSession(),sandbox:this.sandboxState,agent:this.agentState,activeProcessId:this.activeProcessId,hasRun:this.hasRun,timestamp:new Date().toISOString(),...e?{browser:e}:{}}}async kill(e){await this.rotateSession();let t;try{this.activeCommand&&await this.interrupt(e),this.sandbox&&(await this.sandbox.kill(),this.sandbox=void 0)}catch(r){t=r}finally{await this.closeProviderRuntimeToken(),await this.closeManagedSecretRuntimeToken(),await this.closeManagedBrowserSession()}if(t)throw t;this.options.sandboxId=void 0,this.interruptedOperations.clear(),this.invalidateActiveOperation(),this.sandboxState="stopped",this.agentState="idle",this.hasRun=!1,this.lastCheckpointId=void 0,this.emitLifecycle(e,"sandbox_killed")}async getHost(e){return(await this.getSandbox()).getHost(e)}getAgentType(){return this.agentConfig.type}getSessionTag(){return!this.sandbox&&!this.sessionLogger?null:this.sessionTag}getSessionTimestamp(){return this.sessionLogger?.getTimestamp()||null}async flushObservability(){await this.sessionLogger?.flush()}async rotateSession(){let e=!!this.sessionLogger;this.sessionLogger&&(await this.sessionLogger.close(),this.sessionLogger=void 0),e&&(this.previousSessionTag=this.sessionTag),this.sessionTag=g8t(this.options.sessionTagPrefix||"evolve")}async fetchSpend(e){if(this.agentConfig.isDirectMode)throw new Error("Cost tracking requires gateway mode (set EVOLVE_API_KEY).");let t=this.agentConfig.apiKey;if(!t)throw new Error("Cost tracking requires an API key.");let r=process.env.EVOLVE_DASHBOARD_URL||aa,n=await fetch(`${r}/api/sessions/spend?${e}`,{headers:{Authorization:`Bearer ${t}`},signal:AbortSignal.timeout(1e4)});if(!n.ok){let i=await n.text().catch(()=>"");throw new Error(`Spend query failed (${n.status}): ${i}`)}return n}resolveSpendTag(){if(this.sandbox||this.sessionLogger)return this.sessionTag;if(this.previousSessionTag)return this.previousSessionTag;throw new Error("No session to query. Call run() first.")}normalizeRunCost(e,t){return{...e,asOf:e.asOf??t.asOf,isComplete:e.isComplete??t.isComplete,truncated:e.truncated??t.truncated}}normalizeSessionCost(e){let t={asOf:e.asOf,isComplete:e.isComplete,truncated:e.truncated};return{...e,runs:e.runs.map(r=>this.normalizeRunCost(r,t))}}async getSessionCost(){let e=this.resolveSpendTag(),t=new URLSearchParams({tag:e}),r=await(await this.fetchSpend(t)).json();return this.normalizeSessionCost(r)}async getRunCost(e){let t=this.resolveSpendTag();if("runId"in e){let s=new URLSearchParams({tag:t,runId:e.runId}),o=await(await this.fetchSpend(s)).json();return this.normalizeRunCost(o,{asOf:new Date().toISOString(),isComplete:!1,truncated:!1})}let r=await this.getSessionCost(),n=e.index>0?e.index-1:r.runs.length+e.index,i=r.runs[n];if(!i)throw new Error(`Run index ${e.index} out of range. Session has ${r.runs.length} run(s).`);return i}};function hvn(e){return(e||process.env.EVOLVE_DASHBOARD_URL||aa).replace(/\/$/,"")}function mvn(e){let t=e||process.env.EVOLVE_API_KEY;if(!t)throw new Error("Browser profiles require EVOLVE_API_KEY or an explicit apiKey");return t}async function gvn(e){return await e.text().catch(()=>"")}async function b8t(e,t,r={}){let n=await fetch(`${hvn(e?.dashboardUrl)}${t}`,{...r,headers:{Authorization:`Bearer ${mvn(e?.apiKey)}`,accept:"application/json",...r.body?{"content-type":"application/json"}:{},...r.headers||{}}});if(!n.ok)throw new Error(`Browser profiles request failed (${n.status}): ${await gvn(n)}`);return n.status===204?{}:await n.json()}var yvn=class{constructor(e={}){this.config=e}toMetadata(e){return{id:e.id,profile:e.profile,createdAt:e.createdAt||e.created_at||"",updatedAt:e.updatedAt||e.updated_at||"",lastUsedAt:e.lastUsedAt??e.last_used_at??null}}async list(){return{profiles:(await b8t(this.config,"/api/browser-profiles")).profiles.map(e=>this.toMetadata(e))}}async delete(e){return await b8t(this.config,"/api/browser-profiles",{method:"DELETE",body:JSON.stringify(e)})}};function vvn(e={}){return new yvn(e)}var Lw=class extends I8t.EventEmitter{config={};agent;fallbackSandboxState="stopped";fallbackAgentState="idle";fallbackHasRun=!1;constructor(){super()}on(e,t){return super.on(e,t)}off(e,t){return super.off(e,t)}emit(e,...t){return super.emit(e,...t)}withAgent(e){return e&&(U8t(e),this.config.agent=e,this._cachedGatewayOverrides=null),this}withSandbox(e){return this.config.sandbox=e,this}withSandboxCreateOptions(e){return this.config.sandboxCreateOptions={...e,envs:e.envs?{...e.envs}:void 0,metadata:e.metadata?{...e.metadata}:void 0},this}withWorkingDirectory(e){return this.config.workingDirectory=e,this}withWorkspaceMode(e){return this.config.workspaceMode=e,this}withSecrets(e){return this.config.secrets={...this.config.secrets,...e},this}withManagedSecrets(e){return this.config.managedSecrets=e,this}withSession(e){return this.config.sandboxId=e,this.agent||(this.fallbackSandboxState="ready",this.fallbackAgentState="idle",this.fallbackHasRun=!0),this}withSystemPrompt(e){return this.config.systemPrompt=e,this}withContext(e){return this.config.context={...this.config.context,...e},this}withFiles(e){return this.config.files={...this.config.files,...e},this}withMcpServers(e){return this.config.mcpServers={...this.config.mcpServers,...e},this}withBrowser(e={provider:"agent-browser",remote:!0}){return e===!1?delete this.config.browser:this.config.browser=e,this}withBrowserCredentials(e={}){return this.config.browserCredentials=e,this}withPlugins(e){let t=Array.isArray(e)?e:[e];return this.config.plugins=[...this.config.plugins??[],...t],this}withSkills(e){return this.config.skills=e,this}withSchema(e,t){return this.config.schema=e,t&&(dK(e)?console.warn("[Evolve] schemaOptions ignored for Zod schemas - use .passthrough(), .strip(), z.coerce instead"):this.config.schemaOptions=t),this}withSessionTagPrefix(e){return this.config.sessionTagPrefix=e,this}withObservability(e){return this.config.observability={...this.config.observability,...e},this}withIntegrations(e){return this.config.integrations=e,this}withStorage(e){return this.config.storage=e||{},this}static integrations=Qyn;static browserCredentials=R0n;static browserProfiles=vvn;static managedSecrets=J0n;async initializeAgent(){let e=WJt(this.config.agent),t=rR(e.type);if(this.config.integrations&&e.isDirectMode)throw new Error("withIntegrations() is available only in gateway mode with EVOLVE_API_KEY");if(this.config.managedSecrets&&e.isDirectMode)throw new Error("withManagedSecrets() is available only in gateway mode with EVOLVE_API_KEY");if(this.config.managedSecrets&&this.config.sandboxId)throw new Error("withManagedSecrets() cannot be used with withSession(); managed secret grants are sandbox-scoped.");let r=tvn(this.config.managedSecrets,t);if(r&&this.config.sandbox&&!L8t(this.config.sandbox))throw new Error("withManagedSecrets() requires an Evolve-managed sandbox; remove withSandbox() or pass a managed sandbox provider.");let n=this.config.sandbox??(r?await vxe(e.apiKey):await vyn()),i={},s,o,a=this.config.skills,c;if(this.config.browser){let d=_0n(this.config.browser);if(c=d,d.provider==="browser-use"){if(e.isDirectMode)throw new Error('withBrowser("browser-use") requires gateway mode. Use apiKey/EVOLVE_API_KEY instead of providerApiKey/direct mode.');i=Rxe(e.apiKey)}if((d.provider==="actionbook"||d.provider==="agent-browser")&&(a=I0n(d.provider,a),d.managed)){if(e.isDirectMode)throw new Error(`Managed browser provider "${d.provider}" requires gateway mode. Use apiKey/EVOLVE_API_KEY instead of providerApiKey/direct mode.`);s=d.provider==="actionbook"?Uyn:Fyn,o={provider:d.provider,apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa,...d.profile?{profile:d.profile}:{}}}if(d.profile&&!d.managed)throw new Error('withBrowser({ profile }) requires managed remote browser mode. Use .withBrowser({ profile }) or .withBrowser({ provider: "agent-browser", remote: true, profile }).')}if(this.config.browserCredentials!==void 0){if(this.config.sandboxId)throw new Error("withBrowserCredentials() cannot be used with withSession(); browser login tokens are run-scoped.");if(e.isDirectMode)throw new Error("withBrowserCredentials() requires gateway mode. Use apiKey/EVOLVE_API_KEY instead of providerApiKey/direct mode.");if(!c||c.provider!=="agent-browser"||!c.managed)throw new Error('withBrowserCredentials() requires .withBrowser() or .withBrowser({ provider: "agent-browser", remote: true }).');if(this.config.mcpServers?.["browser-login"])throw new Error('withBrowserCredentials() reserves the "browser-login" MCP server name.')}let l=this.config.storage!==void 0?V8t(this.config.storage,!e.isDirectMode,aa,e.isDirectMode?void 0:e.apiKey):void 0,u={sandboxProvider:n,sandboxCreateOptions:this.config.sandboxCreateOptions,secrets:this.config.secrets,managedSecrets:r?{secrets:r,apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa}:void 0,sandboxId:this.config.sandboxId,workingDirectory:this.config.workingDirectory,workspaceMode:this.config.workspaceMode,systemPrompt:this.config.systemPrompt,context:this.config.context,files:this.config.files,mcpServers:{...i,...this.config.mcpServers},browserPrompt:s,managedBrowser:o,browserCredentials:this.config.browserCredentials!==void 0?{apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa,config:this.config.browserCredentials}:void 0,plugins:this.config.plugins,skills:a,schema:this.config.schema,schemaOptions:this.config.schemaOptions,sessionTagPrefix:this.config.sessionTagPrefix,observability:this.config.observability,integrations:this.config.integrations?{...this.config.integrations,apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa}:void 0,providerRouting:e.isDirectMode?void 0:{apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa},storage:l};this.agent=new pvn(e,u)}createStreamCallbacks(){let e=this.listenerCount("stdout")>0,t=this.listenerCount("stderr")>0,r=this.listenerCount("content")>0,n=this.listenerCount("lifecycle")>0;return{onStdout:e?i=>this.emit("stdout",i):void 0,onStderr:t?i=>this.emit("stderr",i):void 0,onContent:r?i=>this.emit("content",i):void 0,onLifecycle:n?i=>this.emit("lifecycle",i):void 0}}emitLifecycleFromStatus(e){if(this.listenerCount("lifecycle")===0)return;let t=this.status();this.emit("lifecycle",{sandboxId:t.sandboxId,sandbox:t.sandbox,agent:t.agent,timestamp:new Date().toISOString(),reason:e,...t.browser?{browser:t.browser}:{}})}async run({prompt:e,timeoutMs:t,background:r,from:n,checkpointComment:i}){if(n&&this.config.sandboxId)throw new Error("Cannot use 'from' with 'withSession()' \u2014 restore requires a fresh sandbox.");this.agent||await this.initializeAgent();let s=this.createStreamCallbacks();return this.agent.run({prompt:e,timeoutMs:t,background:r,from:n,checkpointComment:i},s)}async executeCommand(e,t={}){this.agent||await this.initializeAgent();let r=this.createStreamCallbacks();return this.agent.executeCommand(e,t,r)}async prepareSandbox(){this.agent||await this.initializeAgent();let e=this.createStreamCallbacks();return(await this.agent.getSandbox(e)).sandboxId}async sealCredentials(){if(!this.agent)throw new Error("Agent not initialized. Call run() or executeCommand() first.");await this.agent.sealCredentials()}isSealed(){return this.agent?.isSealed()??!1}async collectArtifacts(e){if(!this.agent)throw new Error("Agent not initialized. Call run() or executeCommand() first.");return this.agent.collectArtifacts(e)}async interrupt(){if(!this.agent)return!1;let e=this.createStreamCallbacks();return this.agent.interrupt(e)}async uploadContext(e){return this.agent||await this.initializeAgent(),this.agent.uploadContext(e)}async uploadFiles(e){return this.agent||await this.initializeAgent(),this.agent.uploadFiles(e)}async uploadFileFromPath(e,t){return this.agent||await this.initializeAgent(),this.agent.uploadFileFromPath(e,t)}async getOutputFiles(e=!1){if(!this.agent)throw new Error("Agent not initialized. Call run() first.");return this.agent.getOutputFiles(e)}async checkpoint(e){if(!this.agent)throw new Error("Agent not initialized. Call run() first.");return this.agent.checkpoint(e)}_cachedGatewayOverrides=null;resolveGatewayOverrides(){if(this._cachedGatewayOverrides!==null)return this._cachedGatewayOverrides||void 0;try{let e=WJt(this.config.agent);if(!e.isDirectMode)return this._cachedGatewayOverrides={gatewayUrl:aa,gatewayApiKey:e.apiKey},this._cachedGatewayOverrides}catch{}this._cachedGatewayOverrides=void 0}async listCheckpoints(e){if(this.config.storage===void 0)throw new Error("Storage not configured. Call .withStorage().");return o8t(this.config.storage,this.resolveGatewayOverrides()||{}).listCheckpoints(e)}storage(){if(this.config.storage===void 0)throw new Error("Storage not configured. Call .withStorage().");return o8t(this.config.storage,this.resolveGatewayOverrides()||{})}getSession(){return this.agent?this.agent.getSession():this.config.sandboxId??null}async setSession(e){if(this.config.managedSecrets)throw new Error("setSession() cannot be used with withManagedSecrets(); managed secret grants are sandbox-scoped.");this.agent?await this.agent.setSession(e):(this.fallbackSandboxState="ready",this.fallbackAgentState="idle",this.fallbackHasRun=!0),this.config.sandboxId=e}status(){return this.agent?this.agent.status():{sandboxId:this.config.sandboxId??null,sandbox:this.fallbackSandboxState,agent:this.fallbackAgentState,activeProcessId:null,hasRun:this.fallbackHasRun,timestamp:new Date().toISOString()}}async pause(){if(this.agent){let e=this.createStreamCallbacks();await this.agent.pause(e);return}this.config.sandboxId&&this.config.sandbox&&(await(await this.config.sandbox.connect(this.config.sandboxId)).pause(),this.fallbackSandboxState="paused",this.fallbackAgentState="idle",this.emitLifecycleFromStatus("sandbox_pause"))}async resume(){if(this.agent){let e=this.createStreamCallbacks();await this.agent.resume(e);return}this.config.sandboxId&&this.config.sandbox&&(await this.config.sandbox.connect(this.config.sandboxId),this.fallbackSandboxState="ready",this.fallbackAgentState="idle",this.emitLifecycleFromStatus("sandbox_resume"))}async kill(){if(this.agent){let e=this.createStreamCallbacks();await this.agent.kill(e),this.config.sandboxId=void 0,this.fallbackSandboxState="stopped",this.fallbackAgentState="idle",this.fallbackHasRun=!1;return}this.config.sandboxId&&this.config.sandbox&&(await(await this.config.sandbox.connect(this.config.sandboxId)).kill(),this.fallbackSandboxState="stopped",this.fallbackAgentState="idle",this.fallbackHasRun=!1,this.emitLifecycleFromStatus("sandbox_killed"),this.config.sandboxId=void 0)}async getHost(e){return this.agent||await this.initializeAgent(),this.agent.getHost(e)}getSessionTag(){return this.agent?.getSessionTag()||null}getSessionTimestamp(){return this.agent?.getSessionTimestamp()||null}async flushObservability(){await this.agent?.flushObservability()}async getSessionCost(){if(!this.agent)throw new Error("No agent initialized. Call run() first.");return this.agent.getSessionCost()}async getRunCost(e){if(!this.agent)throw new Error("No agent initialized. Call run() first.");return this.agent.getRunCost(e)}},bvn=class{permits;queue=[];constructor(e){if(e<1)throw new Error("Semaphore max must be >= 1");this.permits=e}async use(e){await this.acquire();try{return await e()}finally{this.release()}}acquire(){return this.permits>0?(this.permits--,Promise.resolve()):new Promise(e=>this.queue.push(e))}release(){this.queue.length>0?this.queue.shift()():this.permits++}},D3=Symbol.for("evolve.SwarmResult"),S8t=class c5t extends Array{get success(){return this.filter(t=>t.status==="success")}get filtered(){return this.filter(t=>t.status==="filtered")}get error(){return this.filter(t=>t.status==="error")}static from(t){let r=new c5t;return r.push(...t),r}},Ejn=class aK{config;semaphore;constructor(t={}){let r=t.concurrency??Dxe;if(r>vK)throw new Error(`concurrency=${r} exceeds max ${vK}. For higher parallelism, scale horizontally with multiple processes.`);this.config={agent:t.agent,sandbox:t.sandbox,tag:t.tag??"swarm",concurrency:r,timeoutMs:t.timeoutMs??K3,workspaceMode:t.workspaceMode??"knowledge",retry:t.retry,mcpServers:t.mcpServers,skills:t.skills,integrations:t.integrations},this.semaphore=new bvn(this.config.concurrency)}async map(t){let{items:r,prompt:n,bestOf:i,verify:s}=t,o=t.retry??this.config.retry,a=this.generateOperationId(),c=t.timeoutMs??this.config.timeoutMs;if(i&&s)throw new Error("map() cannot use both bestOf and verify options simultaneously");let l=await Promise.all(r.map((u,d)=>i?this.executeMapItemWithBestOf(u,n,d,a,t,c,o):s?this.executeMapItemWithVerify(u,n,d,a,t,c,o):o?q0(f=>this.executeMapItem(u,n,d,a,t,c,f),o,d):this.executeMapItem(u,n,d,a,t,c)));return S8t.from(l)}async filter(t){let{items:r,prompt:n,condition:i,verify:s}=t,o=t.retry??this.config.retry,a=this.generateOperationId(),c=t.timeoutMs??this.config.timeoutMs,l=await Promise.all(r.map((d,f)=>s?this.executeFilterItemWithVerify(d,n,f,a,t,c,o):o?q0(p=>this.executeFilterItem(d,n,f,a,t,c,p),o,f):this.executeFilterItem(d,n,f,a,t,c))),u=[];for(let d of l)if(d.status==="error")u.push(d);else if(d.data!==null)try{i(d.data)?u.push(d):u.push({...d,status:"filtered"})}catch(f){u.push({...d,status:"error",data:null,error:`Condition function threw: ${f.message}`})}return S8t.from(u)}async reduce(t){let{items:r,prompt:n,verify:i}=t,s=t.retry??this.config.retry,o=this.generateOperationId(),a=t.timeoutMs??this.config.timeoutMs,c=[],l=[];r.forEach((P,L)=>{c.push(this.getFiles(P)),l.push(this.getIndex(P,L))});let u={};c.forEach((P,L)=>{Object.entries(P).forEach(([z,j])=>{u[`item_${l[L]}/${z}`]=j})});let d=sxe(u),f=Fw(Gyn,{fileTree:d}),p=t.systemPrompt?`${f}
+`).filter(Boolean),s=this.lastRunTimestamp?Math.floor(this.lastRunTimestamp/1e3)-2:0,o=[],a=`${r}/`;for(let f of i){let[p,m]=f.split("|");if(!p||!m)continue;let v=parseInt(m,10);s>0&&v<s||o.push(p)}let c={},l=await Promise.all(o.map(async f=>{try{let p=await t.files.read(f);return{relativePath:f.startsWith(a)?f.slice(a.length):f.split("/").pop()||f,content:p}}catch{return null}}));for(let f of l)f&&(c[f.relativePath]=f.content);if(!this.zodSchema&&!this.jsonSchema)return{files:c,data:null};let u=c["result.json"];if(!u)return{files:c,data:null,error:"Schema provided but agent did not create output/result.json"};let d=typeof u=="string"?u:new TextDecoder().decode(u);try{let f=JSON.parse(d);if(this.zodSchema){let p=this.zodSchema.safeParse(f);return p.success?{files:c,data:p.data}:{files:c,data:null,error:`Schema validation failed: ${p.error.message}`,rawData:d}}if(this.compiledValidator){if(this.compiledValidator(f))return{files:c,data:f};{let p=this.compiledValidator.errors?.map(m=>`${m.instancePath} ${m.message}`).join(", ")||"Unknown validation error";return{files:c,data:null,error:`Schema validation failed: ${p}`,rawData:d}}}return{files:c,data:null}}catch(f){return{files:c,data:null,error:`Failed to parse result.json: ${f.message}`,rawData:d}}}async checkpoint(e){if(!this.storage)throw new Error("Storage not configured. Call .withStorage().");if(!this.sandbox)throw new Error("No active sandbox. Call run() first.");let t=await e8t(this.sandbox,this.storage,this.agentConfig.type,this.workingDir,{tag:this.sessionTag,model:this.agentConfig.model||this.registry.defaultModel,workspaceMode:this.options.workspaceMode||"knowledge",comment:e?.comment,parentId:this.lastCheckpointId});return this.lastCheckpointId=t.id,t}getSession(){return this.sandbox?.sandboxId||this.options.sandboxId||null}async setSession(e){if(this.credentialsSealed)throw new Error("Agent credentials are sealed and this Evolve instance cannot switch sandboxes");if(this.options.managedSecrets)throw new Error("setSession() cannot be used with managed secrets; managed secret grants are sandbox-scoped.");if(this.options.sandboxCreateOptions?.user)throw new Error("setSession() cannot be used with a configured sandbox user; the user cannot be enforced on an existing sandbox.");if(this.activeCommand&&!await this.interrupt())throw new Error("Cannot switch session while an active process is running and could not be interrupted.");await this.rotateSession(),await this.closeProviderRuntimeToken(),await this.closeManagedSecretRuntimeToken(),await this.closeManagedBrowserSession(),this.options.sandboxId=e,this.sandbox=void 0,this.interruptedOperations.clear(),this.invalidateActiveOperation(),this.sandboxState="ready",this.agentState="idle",this.hasRun=!0,this.lastCheckpointId=void 0}async pause(e){this.sandbox&&(this.activeCommand&&await this.interrupt(e),await this.sandbox.pause(),this.sandboxState="paused",this.agentState="idle",this.emitLifecycle(e,"sandbox_pause"))}async resume(e){this.sandbox&&this.options.sandboxProvider&&(this.sandbox=await this.options.sandboxProvider.connect(this.sandbox.sandboxId),this.sandboxState="ready",this.agentState="idle",this.emitLifecycle(e,"sandbox_resume"))}async interrupt(e){if(!this.activeCommand&&!this.activeProcessId)return!1;let t=this.activeOperationId,r=this.activeOperationKind,n=!1;try{this.activeCommand?n=await this.activeCommand.kill():this.sandbox&&this.activeProcessId&&(n=await this.sandbox.commands.kill(this.activeProcessId))}catch{n=!1}if(!n)return this.sandboxState="running",this.agentState="running",!1;t!==null&&this.interruptedOperations.add(t),this.invalidateActiveOperation(),this.sandboxState="ready",this.agentState="interrupted";let i=r==="run"?"run_interrupted":"command_interrupted";return this.emitLifecycle(e,i),n}status(){let e=this.browserRuntimeInfo();return{sandboxId:this.getSession(),sandbox:this.sandboxState,agent:this.agentState,activeProcessId:this.activeProcessId,hasRun:this.hasRun,timestamp:new Date().toISOString(),...e?{browser:e}:{}}}async kill(e){await this.rotateSession();let t;try{this.activeCommand&&await this.interrupt(e),this.sandbox&&(await this.sandbox.kill(),this.sandbox=void 0)}catch(r){t=r}finally{await this.closeProviderRuntimeToken(),await this.closeManagedSecretRuntimeToken(),await this.closeManagedBrowserSession()}if(t)throw t;this.options.sandboxId=void 0,this.interruptedOperations.clear(),this.invalidateActiveOperation(),this.sandboxState="stopped",this.agentState="idle",this.hasRun=!1,this.lastCheckpointId=void 0,this.emitLifecycle(e,"sandbox_killed")}async getHost(e){return(await this.getSandbox()).getHost(e)}getAgentType(){return this.agentConfig.type}getSessionTag(){return!this.sandbox&&!this.sessionLogger?null:this.sessionTag}getSessionTimestamp(){return this.sessionLogger?.getTimestamp()||null}async flushObservability(){await this.sessionLogger?.flush()}async rotateSession(){let e=!!this.sessionLogger;this.sessionLogger&&(await this.sessionLogger.close(),this.sessionLogger=void 0),e&&(this.previousSessionTag=this.sessionTag),this.sessionTag=g8t(this.options.sessionTagPrefix||"evolve")}async fetchSpend(e){if(this.agentConfig.isDirectMode)throw new Error("Cost tracking requires gateway mode (set EVOLVE_API_KEY).");let t=this.agentConfig.apiKey;if(!t)throw new Error("Cost tracking requires an API key.");let r=process.env.EVOLVE_DASHBOARD_URL||aa,n=await fetch(`${r}/api/sessions/spend?${e}`,{headers:{Authorization:`Bearer ${t}`},signal:AbortSignal.timeout(1e4)});if(!n.ok){let i=await n.text().catch(()=>"");throw new Error(`Spend query failed (${n.status}): ${i}`)}return n}resolveSpendTag(){if(this.sandbox||this.sessionLogger)return this.sessionTag;if(this.previousSessionTag)return this.previousSessionTag;throw new Error("No session to query. Call run() first.")}normalizeRunCost(e,t){return{...e,asOf:e.asOf??t.asOf,isComplete:e.isComplete??t.isComplete,truncated:e.truncated??t.truncated}}normalizeSessionCost(e){let t={asOf:e.asOf,isComplete:e.isComplete,truncated:e.truncated};return{...e,runs:e.runs.map(r=>this.normalizeRunCost(r,t))}}async getSessionCost(){let e=this.resolveSpendTag(),t=new URLSearchParams({tag:e}),r=await(await this.fetchSpend(t)).json();return this.normalizeSessionCost(r)}async getRunCost(e){let t=this.resolveSpendTag();if("runId"in e){let s=new URLSearchParams({tag:t,runId:e.runId}),o=await(await this.fetchSpend(s)).json();return this.normalizeRunCost(o,{asOf:new Date().toISOString(),isComplete:!1,truncated:!1})}let r=await this.getSessionCost(),n=e.index>0?e.index-1:r.runs.length+e.index,i=r.runs[n];if(!i)throw new Error(`Run index ${e.index} out of range. Session has ${r.runs.length} run(s).`);return i}};function hvn(e){return(e||process.env.EVOLVE_DASHBOARD_URL||aa).replace(/\/$/,"")}function mvn(e){let t=e||process.env.EVOLVE_API_KEY;if(!t)throw new Error("Browser profiles require EVOLVE_API_KEY or an explicit apiKey");return t}async function gvn(e){return await e.text().catch(()=>"")}async function b8t(e,t,r={}){let n=await fetch(`${hvn(e?.dashboardUrl)}${t}`,{...r,headers:{Authorization:`Bearer ${mvn(e?.apiKey)}`,accept:"application/json",...r.body?{"content-type":"application/json"}:{},...r.headers||{}}});if(!n.ok)throw new Error(`Browser profiles request failed (${n.status}): ${await gvn(n)}`);return n.status===204?{}:await n.json()}var yvn=class{constructor(e={}){this.config=e}toMetadata(e){return{id:e.id,profile:e.profile,createdAt:e.createdAt||e.created_at||"",updatedAt:e.updatedAt||e.updated_at||"",lastUsedAt:e.lastUsedAt??e.last_used_at??null}}async list(){return{profiles:(await b8t(this.config,"/api/browser-profiles")).profiles.map(e=>this.toMetadata(e))}}async delete(e){return await b8t(this.config,"/api/browser-profiles",{method:"DELETE",body:JSON.stringify(e)})}};function vvn(e={}){return new yvn(e)}var Lw=class extends I8t.EventEmitter{config={};agent;fallbackSandboxState="stopped";fallbackAgentState="idle";fallbackHasRun=!1;constructor(){super()}on(e,t){return super.on(e,t)}off(e,t){return super.off(e,t)}emit(e,...t){return super.emit(e,...t)}withAgent(e){return e&&(U8t(e),this.config.agent=e,this._cachedGatewayOverrides=null),this}withSandbox(e){return this.config.sandbox=e,this}withSandboxCreateOptions(e){return this.config.sandboxCreateOptions={...e,envs:e.envs?{...e.envs}:void 0,metadata:e.metadata?{...e.metadata}:void 0},this}withWorkingDirectory(e){return this.config.workingDirectory=e,this}withWorkspaceMode(e){return this.config.workspaceMode=e,this}withSecrets(e){return this.config.secrets={...this.config.secrets,...e},this}withManagedSecrets(e){return this.config.managedSecrets=e,this}withSession(e){return this.config.sandboxId=e,this.agent||(this.fallbackSandboxState="ready",this.fallbackAgentState="idle",this.fallbackHasRun=!0),this}withSystemPrompt(e){return this.config.systemPrompt=e,this}withContext(e){return this.config.context={...this.config.context,...e},this}withFiles(e){return this.config.files={...this.config.files,...e},this}withMcpServers(e){return this.config.mcpServers={...this.config.mcpServers,...e},this}withBrowser(e={provider:"agent-browser",remote:!0}){return e===!1?delete this.config.browser:this.config.browser=e,this}withBrowserCredentials(e={}){return this.config.browserCredentials=e,this}withPlugins(e){let t=Array.isArray(e)?e:[e];return this.config.plugins=[...this.config.plugins??[],...t],this}withSkills(e){return this.config.skills=e,this}withSchema(e,t){return this.config.schema=e,t&&(dK(e)?console.warn("[Evolve] schemaOptions ignored for Zod schemas - use .passthrough(), .strip(), z.coerce instead"):this.config.schemaOptions=t),this}withSessionTagPrefix(e){return this.config.sessionTagPrefix=e,this}withObservability(e){return this.config.observability={...this.config.observability,...e},this}withIntegrations(e){return this.config.integrations=e,this}withStorage(e){return this.config.storage=e||{},this}static integrations=Qyn;static browserCredentials=R0n;static browserProfiles=vvn;static managedSecrets=J0n;async initializeAgent(){let e=WJt(this.config.agent),t=rR(e.type);if(this.config.integrations&&e.isDirectMode)throw new Error("withIntegrations() is available only in gateway mode with EVOLVE_API_KEY");if(this.config.managedSecrets&&e.isDirectMode)throw new Error("withManagedSecrets() is available only in gateway mode with EVOLVE_API_KEY");if(this.config.managedSecrets&&this.config.sandboxId)throw new Error("withManagedSecrets() cannot be used with withSession(); managed secret grants are sandbox-scoped.");let r=tvn(this.config.managedSecrets,t);if(r&&this.config.sandbox&&!L8t(this.config.sandbox))throw new Error("withManagedSecrets() requires an Evolve-managed sandbox; remove withSandbox() or pass a managed sandbox provider.");let n=this.config.sandbox??(r?await vxe(e.apiKey):await vyn()),i={},s,o,a=this.config.skills,c;if(this.config.browser){let d=_0n(this.config.browser);if(c=d,d.provider==="browser-use"){if(e.isDirectMode)throw new Error('withBrowser("browser-use") requires gateway mode. Use apiKey/EVOLVE_API_KEY instead of providerApiKey/direct mode.');i=Rxe(e.apiKey)}if((d.provider==="actionbook"||d.provider==="agent-browser")&&(a=I0n(d.provider,a),d.managed)){if(e.isDirectMode)throw new Error(`Managed browser provider "${d.provider}" requires gateway mode. Use apiKey/EVOLVE_API_KEY instead of providerApiKey/direct mode.`);s=d.provider==="actionbook"?Uyn:Fyn,o={provider:d.provider,apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa,...d.profile?{profile:d.profile}:{}}}if(d.profile&&!d.managed)throw new Error('withBrowser({ profile }) requires managed remote browser mode. Use .withBrowser({ profile }) or .withBrowser({ provider: "agent-browser", remote: true, profile }).')}if(this.config.browserCredentials!==void 0){if(this.config.sandboxId)throw new Error("withBrowserCredentials() cannot be used with withSession(); browser login tokens are run-scoped.");if(e.isDirectMode)throw new Error("withBrowserCredentials() requires gateway mode. Use apiKey/EVOLVE_API_KEY instead of providerApiKey/direct mode.");if(!c||c.provider!=="agent-browser"||!c.managed)throw new Error('withBrowserCredentials() requires .withBrowser() or .withBrowser({ provider: "agent-browser", remote: true }).');if(this.config.mcpServers?.["browser-login"])throw new Error('withBrowserCredentials() reserves the "browser-login" MCP server name.')}let l=this.config.storage!==void 0?V8t(this.config.storage,!e.isDirectMode,aa,e.isDirectMode?void 0:e.apiKey):void 0,u={sandboxProvider:n,sandboxCreateOptions:this.config.sandboxCreateOptions,secrets:this.config.secrets,managedSecrets:r?{secrets:r,apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa}:void 0,sandboxId:this.config.sandboxId,workingDirectory:this.config.workingDirectory,workspaceMode:this.config.workspaceMode,systemPrompt:this.config.systemPrompt,context:this.config.context,files:this.config.files,mcpServers:{...i,...this.config.mcpServers},browserPrompt:s,managedBrowser:o,browserCredentials:this.config.browserCredentials!==void 0?{apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa,config:this.config.browserCredentials}:void 0,plugins:this.config.plugins,skills:a,schema:this.config.schema,schemaOptions:this.config.schemaOptions,sessionTagPrefix:this.config.sessionTagPrefix,observability:this.config.observability,integrations:this.config.integrations?{...this.config.integrations,apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa}:void 0,providerRouting:e.isDirectMode?void 0:{apiKey:e.apiKey,dashboardUrl:process.env.EVOLVE_DASHBOARD_URL||aa},storage:l};this.agent=new pvn(e,u)}createStreamCallbacks(){let e=this.listenerCount("stdout")>0,t=this.listenerCount("stderr")>0,r=this.listenerCount("content")>0,n=this.listenerCount("lifecycle")>0;return{onStdout:e?i=>this.emit("stdout",i):void 0,onStderr:t?i=>this.emit("stderr",i):void 0,onContent:r?i=>this.emit("content",i):void 0,onLifecycle:n?i=>this.emit("lifecycle",i):void 0}}emitLifecycleFromStatus(e){if(this.listenerCount("lifecycle")===0)return;let t=this.status();this.emit("lifecycle",{sandboxId:t.sandboxId,sandbox:t.sandbox,agent:t.agent,timestamp:new Date().toISOString(),reason:e,...t.browser?{browser:t.browser}:{}})}async run({prompt:e,timeoutMs:t,background:r,from:n,checkpointComment:i}){if(n&&this.config.sandboxId)throw new Error("Cannot use 'from' with 'withSession()' \u2014 restore requires a fresh sandbox.");this.agent||await this.initializeAgent();let s=this.createStreamCallbacks();return this.agent.run({prompt:e,timeoutMs:t,background:r,from:n,checkpointComment:i},s)}async executeCommand(e,t={}){this.agent||await this.initializeAgent();let r=this.createStreamCallbacks();return this.agent.executeCommand(e,t,r)}async prepareSandbox(){this.agent||await this.initializeAgent();let e=this.createStreamCallbacks();return(await this.agent.getSandbox(e)).sandboxId}async sealCredentials(){if(!this.agent)throw new Error("Agent not initialized. Call run() or executeCommand() first.");await this.agent.sealCredentials()}isSealed(){return this.agent?.isSealed()??!1}async collectArtifacts(e){if(!this.agent)throw new Error("Agent not initialized. Call run() or executeCommand() first.");return this.agent.collectArtifacts(e)}async interrupt(){if(!this.agent)return!1;let e=this.createStreamCallbacks();return this.agent.interrupt(e)}async uploadContext(e){return this.agent||await this.initializeAgent(),this.agent.uploadContext(e)}async uploadFiles(e){return this.agent||await this.initializeAgent(),this.agent.uploadFiles(e)}async uploadFileFromPath(e,t){return this.agent||await this.initializeAgent(),this.agent.uploadFileFromPath(e,t)}async getOutputFiles(e=!1){if(!this.agent)throw new Error("Agent not initialized. Call run() first.");return this.agent.getOutputFiles(e)}async checkpoint(e){if(!this.agent)throw new Error("Agent not initialized. Call run() first.");return this.agent.checkpoint(e)}_cachedGatewayOverrides=null;resolveGatewayOverrides(){if(this._cachedGatewayOverrides!==null)return this._cachedGatewayOverrides||void 0;try{let e=WJt(this.config.agent);if(!e.isDirectMode)return this._cachedGatewayOverrides={gatewayUrl:aa,gatewayApiKey:e.apiKey},this._cachedGatewayOverrides}catch{}this._cachedGatewayOverrides=void 0}async listCheckpoints(e){if(this.config.storage===void 0)throw new Error("Storage not configured. Call .withStorage().");return o8t(this.config.storage,this.resolveGatewayOverrides()||{}).listCheckpoints(e)}storage(){if(this.config.storage===void 0)throw new Error("Storage not configured. Call .withStorage().");return o8t(this.config.storage,this.resolveGatewayOverrides()||{})}getSession(){return this.agent?this.agent.getSession():this.config.sandboxId??null}async setSession(e){if(this.config.managedSecrets)throw new Error("setSession() cannot be used with withManagedSecrets(); managed secret grants are sandbox-scoped.");this.agent?await this.agent.setSession(e):(this.fallbackSandboxState="ready",this.fallbackAgentState="idle",this.fallbackHasRun=!0),this.config.sandboxId=e}status(){return this.agent?this.agent.status():{sandboxId:this.config.sandboxId??null,sandbox:this.fallbackSandboxState,agent:this.fallbackAgentState,activeProcessId:null,hasRun:this.fallbackHasRun,timestamp:new Date().toISOString()}}async pause(){if(this.agent){let e=this.createStreamCallbacks();await this.agent.pause(e);return}this.config.sandboxId&&this.config.sandbox&&(await(await this.config.sandbox.connect(this.config.sandboxId)).pause(),this.fallbackSandboxState="paused",this.fallbackAgentState="idle",this.emitLifecycleFromStatus("sandbox_pause"))}async resume(){if(this.agent){let e=this.createStreamCallbacks();await this.agent.resume(e);return}this.config.sandboxId&&this.config.sandbox&&(await this.config.sandbox.connect(this.config.sandboxId),this.fallbackSandboxState="ready",this.fallbackAgentState="idle",this.emitLifecycleFromStatus("sandbox_resume"))}async kill(){if(this.agent){let e=this.createStreamCallbacks();await this.agent.kill(e),this.config.sandboxId=void 0,this.fallbackSandboxState="stopped",this.fallbackAgentState="idle",this.fallbackHasRun=!1;return}this.config.sandboxId&&this.config.sandbox&&(await(await this.config.sandbox.connect(this.config.sandboxId)).kill(),this.fallbackSandboxState="stopped",this.fallbackAgentState="idle",this.fallbackHasRun=!1,this.emitLifecycleFromStatus("sandbox_killed"),this.config.sandboxId=void 0)}async getHost(e){return this.agent||await this.initializeAgent(),this.agent.getHost(e)}getSessionTag(){return this.agent?.getSessionTag()||null}getSessionTimestamp(){return this.agent?.getSessionTimestamp()||null}async flushObservability(){await this.agent?.flushObservability()}async getSessionCost(){if(!this.agent)throw new Error("No agent initialized. Call run() first.");return this.agent.getSessionCost()}async getRunCost(e){if(!this.agent)throw new Error("No agent initialized. Call run() first.");return this.agent.getRunCost(e)}},bvn=class{permits;queue=[];constructor(e){if(e<1)throw new Error("Semaphore max must be >= 1");this.permits=e}async use(e){await this.acquire();try{return await e()}finally{this.release()}}acquire(){return this.permits>0?(this.permits--,Promise.resolve()):new Promise(e=>this.queue.push(e))}release(){this.queue.length>0?this.queue.shift()():this.permits++}},D3=Symbol.for("evolve.SwarmResult"),S8t=class c5t extends Array{get success(){return this.filter(t=>t.status==="success")}get filtered(){return this.filter(t=>t.status==="filtered")}get error(){return this.filter(t=>t.status==="error")}static from(t){let r=new c5t;return r.push(...t),r}},Ajn=class aK{config;semaphore;constructor(t={}){let r=t.concurrency??Dxe;if(r>vK)throw new Error(`concurrency=${r} exceeds max ${vK}. For higher parallelism, scale horizontally with multiple processes.`);this.config={agent:t.agent,sandbox:t.sandbox,tag:t.tag??"swarm",concurrency:r,timeoutMs:t.timeoutMs??K3,workspaceMode:t.workspaceMode??"knowledge",retry:t.retry,mcpServers:t.mcpServers,skills:t.skills,integrations:t.integrations},this.semaphore=new bvn(this.config.concurrency)}async map(t){let{items:r,prompt:n,bestOf:i,verify:s}=t,o=t.retry??this.config.retry,a=this.generateOperationId(),c=t.timeoutMs??this.config.timeoutMs;if(i&&s)throw new Error("map() cannot use both bestOf and verify options simultaneously");let l=await Promise.all(r.map((u,d)=>i?this.executeMapItemWithBestOf(u,n,d,a,t,c,o):s?this.executeMapItemWithVerify(u,n,d,a,t,c,o):o?q0(f=>this.executeMapItem(u,n,d,a,t,c,f),o,d):this.executeMapItem(u,n,d,a,t,c)));return S8t.from(l)}async filter(t){let{items:r,prompt:n,condition:i,verify:s}=t,o=t.retry??this.config.retry,a=this.generateOperationId(),c=t.timeoutMs??this.config.timeoutMs,l=await Promise.all(r.map((d,f)=>s?this.executeFilterItemWithVerify(d,n,f,a,t,c,o):o?q0(p=>this.executeFilterItem(d,n,f,a,t,c,p),o,f):this.executeFilterItem(d,n,f,a,t,c))),u=[];for(let d of l)if(d.status==="error")u.push(d);else if(d.data!==null)try{i(d.data)?u.push(d):u.push({...d,status:"filtered"})}catch(f){u.push({...d,status:"error",data:null,error:`Condition function threw: ${f.message}`})}return S8t.from(u)}async reduce(t){let{items:r,prompt:n,verify:i}=t,s=t.retry??this.config.retry,o=this.generateOperationId(),a=t.timeoutMs??this.config.timeoutMs,c=[],l=[];r.forEach((P,L)=>{c.push(this.getFiles(P)),l.push(this.getIndex(P,L))});let u={};c.forEach((P,L)=>{Object.entries(P).forEach(([z,j])=>{u[`item_${l[L]}/${z}`]=j})});let d=sxe(u),f=Fw(Gyn,{fileTree:d}),p=t.systemPrompt?`${f}
 
 ${t.systemPrompt}`:f,m=(P,L,z)=>({operationId:o,operation:"reduce",tag:P.tag,sandboxId:P.sandboxId,swarmName:this.config.tag,operationName:t.name,inputCount:r.length,inputIndices:l,errorRetry:L,verifyRetry:z,...this.pipelineContextToMeta(t._pipelineContext)}),v=t.mcpServers??this.config.mcpServers,b=t.skills??this.config.skills,S=t.integrations??this.config.integrations,T=async(P,L,z,j)=>{let U=await this.semaphore.use(()=>this.execute(u,P,{systemPrompt:p,schema:t.schema,schemaOptions:t.schemaOptions,agent:t.agent,mcpServers:v,skills:b,integrations:S,tagPrefix:L,timeoutMs:a,observability:{swarmName:this.config.tag,operationName:t.name,operationId:o,operation:"reduce",role:"worker",errorRetry:z,verifyRetry:j&&j>1?j-1:void 0,...this.pipelineContextToObservability(t._pipelineContext)}})),X=m(U,z,j&&j>1?j-1:void 0);return U.error?{status:"error",data:null,files:U.files,meta:X,error:U.error,rawData:U.rawData}:{status:"success",data:U.data,files:U.files,meta:X}},C=`${this.config.tag}-reduce`;return i?this.runWithVerification((P,L,z)=>T(P,L,void 0,z),{originalPrompt:n,inputFiles:u,verifyConfig:i,timeoutMs:a,systemPrompt:p,schema:t.schema,mcpServers:v,skills:b,integrations:S,operationId:o,baseTag:C,retry:s,operation:"reduce",_pipelineContext:t._pipelineContext}):s?q0(P=>{let L=P>1?P-1:void 0,z=L?`${C}-er${L}`:C;return T(n,z,L)},s):T(n,C)}async bestOf(t){let{item:r,prompt:n,config:i}=t,s=t.retry??this.config.retry,o=i.n??i.taskAgents?.length;if(o===void 0)throw new Error("bestOf requires n or taskAgents");if(o<2)throw new Error("bestOf requires n >= 2");let a=this.generateOperationId(),c=t.timeoutMs??this.config.timeoutMs,l=this.getFiles(r),u=i.mcpServers??this.config.mcpServers,d=i.judgeMcpServers??i.mcpServers??this.config.mcpServers,f=i.skills??this.config.skills,p=i.judgeSkills??i.skills??this.config.skills,m=i.integrations??this.config.integrations,v=i.judgeIntegrations??i.integrations??this.config.integrations,b=await Promise.all(Array.from({length:o},async(L,z)=>{let j=s?await q0(U=>this.executeBestOfCandidate({inputFiles:l,prompt:n,candidateIndex:z,operationId:a,config:i,systemPrompt:t.systemPrompt,schema:t.schema,schemaOptions:t.schemaOptions,mcpServers:u,skills:f,integrations:m,timeoutMs:c,attempt:U}),s,z):await this.executeBestOfCandidate({inputFiles:l,prompt:n,candidateIndex:z,operationId:a,config:i,systemPrompt:t.systemPrompt,schema:t.schema,schemaOptions:t.schemaOptions,mcpServers:u,skills:f,integrations:m,timeoutMs:c});return i.onCandidateComplete?.(0,z,j.status==="success"?"success":"error"),j})),S=s?await q0(L=>this.executeBestOfJudge({inputFiles:l,taskPrompt:n,candidates:b,config:i,timeoutMs:c,operationId:a,systemPrompt:t.systemPrompt,schema:t.schema,schemaOptions:t.schemaOptions,mcpServers:d,skills:p,integrations:v,attempt:L}),{...s,retryOn:void 0},0):await this.executeBestOfJudge({inputFiles:l,taskPrompt:n,candidates:b,config:i,timeoutMs:c,operationId:a,systemPrompt:t.systemPrompt,schema:t.schema,schemaOptions:t.schemaOptions,mcpServers:d,skills:p,integrations:v}),T=b.findIndex(L=>L.status==="success"),C=S.decision?.winner??(T>=0?T:0);i.onJudgeComplete?.(0,C,S.decision?.reasoning??"Judge failed to provide reasoning");let P={operationId:a,operation:"bestof-judge",tag:S.tag,sandboxId:S.sandboxId,swarmName:this.config.tag,operationName:t.name,candidateCount:o};return{winner:b[C]??b[0],winnerIndex:C,judgeReasoning:S.decision?.reasoning??"Judge failed to provide reasoning",judgeMeta:P,candidates:b}}async execute(t,r,n){let i=null,s="",o=n.tagPrefix,a={},c=null,l,u,d={...this.config.agent,...n.agent};try{i=new Lw().withAgent(d).withSandbox(this.config.sandbox).withWorkspaceMode(this.config.workspaceMode).withSessionTagPrefix(n.tagPrefix),n.schema&&i.withSchema(n.schema,n.schemaOptions),n.systemPrompt&&i.withSystemPrompt(n.systemPrompt),n.mcpServers&&i.withMcpServers(n.mcpServers),n.skills?.length&&i.withSkills(n.skills),n.integrations&&i.withIntegrations(n.integrations),n.observability&&i.withObservability(n.observability),Object.keys(t).length>0&&i.withContext(t);let f=await i.run({prompt:r,timeoutMs:n.timeoutMs});s=f.sandboxId,o=i.getSessionTag()??n.tagPrefix;let p=null;try{p=await i.getOutputFiles(!0),a=p.files}catch{}f.exitCode!==0?l=`Agent exited with code ${f.exitCode}`:n.schema?p?(c=p.data,p.error&&(l=p.error),p.rawData&&(u=p.rawData)):l="Failed to read output files from sandbox":c=a}catch(f){if(l=f.message,i){o=i.getSessionTag()??n.tagPrefix;try{a=(await i.getOutputFiles(!0)).files}catch{}}}finally{i&&await i.kill().catch(()=>{})}return{files:a,data:c,tag:o,sandboxId:s,error:l,rawData:u}}async executeMapItem(t,r,n,i,s,o,a=1){let c=this.getFiles(t),l=a>1?`${this.config.tag}-map-${n}-er${a-1}`:`${this.config.tag}-map-${n}`,u=this.evaluatePrompt(r,c,n);if(u instanceof Error)return this.buildErrorResult(`Prompt function threw: ${u.message}`,{operationId:i,operation:"map",tag:l,sandboxId:"",itemIndex:n});let d=s.mcpServers??this.config.mcpServers,f=s.skills??this.config.skills,p=s.integrations??this.config.integrations,m=await this.semaphore.use(()=>this.execute(c,u,{systemPrompt:s.systemPrompt,schema:s.schema,schemaOptions:s.schemaOptions,agent:s.agent,mcpServers:d,skills:f,integrations:p,tagPrefix:l,timeoutMs:o,observability:{swarmName:this.config.tag,operationName:s.name,operationId:i,operation:"map",itemIndex:n,role:"worker",errorRetry:a>1?a-1:void 0,...this.pipelineContextToObservability(s._pipelineContext)}})),v={operationId:i,operation:"map",tag:m.tag,sandboxId:m.sandboxId,swarmName:this.config.tag,operationName:s.name,itemIndex:n,errorRetry:a>1?a-1:void 0,...this.pipelineContextToMeta(s._pipelineContext)};return this.buildResult(m,v)}async executeMapItemWithVerify(t,r,n,i,s,o,a){let c=this.getFiles(t),l=`${this.config.tag}-map-${n}`,u=s.verify,d=s.mcpServers??this.config.mcpServers,f=s.skills??this.config.skills,p=s.integrations??this.config.integrations,m=this.evaluatePrompt(r,c,n);if(m instanceof Error)return this.buildErrorResult(`Prompt function threw: ${m.message}`,{operationId:i,operation:"map",tag:l,sandboxId:"",itemIndex:n});let v=async(b,S,T)=>{let C=await this.semaphore.use(()=>this.execute(c,b,{systemPrompt:s.systemPrompt,schema:s.schema,schemaOptions:s.schemaOptions,agent:s.agent,mcpServers:d,skills:f,integrations:p,tagPrefix:S,timeoutMs:o,observability:{swarmName:this.config.tag,operationName:s.name,operationId:i,operation:"map",itemIndex:n,role:"worker",verifyRetry:T&&T>1?T-1:void 0,...this.pipelineContextToObservability(s._pipelineContext)}})),P={operationId:i,operation:"map",tag:C.tag,sandboxId:C.sandboxId,swarmName:this.config.tag,operationName:s.name,itemIndex:n,verifyRetry:T&&T>1?T-1:void 0,...this.pipelineContextToMeta(s._pipelineContext)};return this.buildResult(C,P)};return this.runWithVerification(v,{originalPrompt:m,inputFiles:c,verifyConfig:u,timeoutMs:o,systemPrompt:s.systemPrompt,schema:s.schema,mcpServers:d,skills:f,integrations:p,operationId:i,baseTag:l,retry:a,itemIndex:n,operation:"map",_pipelineContext:s._pipelineContext})}async executeMapItemWithBestOf(t,r,n,i,s,o,a){let c=this.getFiles(t),l=`${this.config.tag}-map-${n}`,u=this.evaluatePrompt(r,c,n);if(u instanceof Error)return this.buildErrorResult(`Prompt function threw: ${u.message}`,{operationId:i,operation:"map",tag:l,sandboxId:"",itemIndex:n});let d=s.bestOf,f=d.n??d.taskAgents?.length;if(f===void 0||f<2)return this.buildErrorResult("bestOf requires n >= 2 or taskAgents with at least 2 elements",{operationId:i,operation:"map",tag:l,sandboxId:"",itemIndex:n});let p=s.mcpServers??this.config.mcpServers,m=d.mcpServers??p,v=d.judgeMcpServers??d.mcpServers??p,b=s.skills??this.config.skills,S=d.skills??b,T=d.judgeSkills??d.skills??b,C=s.integrations??this.config.integrations,P=d.integrations??C,L=d.judgeIntegrations??d.integrations??C,z=await Promise.all(Array.from({length:f},async(We,de)=>{let O=a?await q0($=>this.executeBestOfCandidate({inputFiles:c,prompt:u,candidateIndex:de,operationId:i,config:d,systemPrompt:s.systemPrompt,schema:s.schema,schemaOptions:s.schemaOptions,mcpServers:m,skills:S,integrations:P,timeoutMs:o,parentIndex:n,attempt:$,_pipelineContext:s._pipelineContext}),a,de):await this.executeBestOfCandidate({inputFiles:c,prompt:u,candidateIndex:de,operationId:i,config:d,systemPrompt:s.systemPrompt,schema:s.schema,schemaOptions:s.schemaOptions,mcpServers:m,skills:S,integrations:P,timeoutMs:o,parentIndex:n,_pipelineContext:s._pipelineContext});return d.onCandidateComplete?.(n,de,O.status==="success"?"success":"error"),O})),j=a?await q0(We=>this.executeBestOfJudge({inputFiles:c,taskPrompt:u,candidates:z,config:d,timeoutMs:o,operationId:i,systemPrompt:s.systemPrompt,schema:s.schema,schemaOptions:s.schemaOptions,mcpServers:v,skills:T,integrations:L,parentIndex:n,attempt:We,_pipelineContext:s._pipelineContext}),{...a,retryOn:void 0},0):await this.executeBestOfJudge({inputFiles:c,taskPrompt:u,candidates:z,config:d,timeoutMs:o,operationId:i,systemPrompt:s.systemPrompt,schema:s.schema,schemaOptions:s.schemaOptions,mcpServers:v,skills:T,integrations:L,parentIndex:n,_pipelineContext:s._pipelineContext}),U=z.findIndex(We=>We.status==="success"),X=j.decision?.winner??(U>=0?U:0),pe=z[X]??z[0];d.onJudgeComplete?.(n,X,j.decision?.reasoning??"Judge failed to provide reasoning");let V={operationId:i,operation:"bestof-judge",tag:j.tag,sandboxId:j.sandboxId,swarmName:this.config.tag,operationName:s.name,candidateCount:f,...this.pipelineContextToMeta(s._pipelineContext)};return{...pe,meta:{...pe.meta,operation:"map",swarmName:this.config.tag,operationName:s.name,itemIndex:n,...this.pipelineContextToMeta(s._pipelineContext)},bestOf:{winnerIndex:X,judgeReasoning:j.decision?.reasoning??"Judge failed to provide reasoning",judgeMeta:V,candidates:z}}}async executeFilterItem(t,r,n,i,s,o,a=1){let c=this.getFiles(t),l=a>1?`${this.config.tag}-filter-${n}-er${a-1}`:`${this.config.tag}-filter-${n}`,u=s.mcpServers??this.config.mcpServers,d=s.skills??this.config.skills,f=s.integrations??this.config.integrations,p=await this.semaphore.use(()=>this.execute(c,r,{systemPrompt:s.systemPrompt,schema:s.schema,schemaOptions:s.schemaOptions,agent:s.agent,mcpServers:u,skills:d,integrations:f,tagPrefix:l,timeoutMs:o,observability:{swarmName:this.config.tag,operationName:s.name,operationId:i,operation:"filter",itemIndex:n,role:"worker",errorRetry:a>1?a-1:void 0,...this.pipelineContextToObservability(s._pipelineContext)}})),m={operationId:i,operation:"filter",tag:p.tag,sandboxId:p.sandboxId,swarmName:this.config.tag,operationName:s.name,itemIndex:n,errorRetry:a>1?a-1:void 0,...this.pipelineContextToMeta(s._pipelineContext)};return this.buildResult(p,m,c)}async executeFilterItemWithVerify(t,r,n,i,s,o,a){let c=this.getFiles(t),l=`${this.config.tag}-filter-${n}`,u=s.verify,d=s.mcpServers??this.config.mcpServers,f=s.skills??this.config.skills,p=s.integrations??this.config.integrations,m=async(v,b,S)=>{let T=await this.semaphore.use(()=>this.execute(c,v,{systemPrompt:s.systemPrompt,schema:s.schema,schemaOptions:s.schemaOptions,agent:s.agent,mcpServers:d,skills:f,integrations:p,tagPrefix:b,timeoutMs:o,observability:{swarmName:this.config.tag,operationName:s.name,operationId:i,operation:"filter",itemIndex:n,role:"worker",verifyRetry:S&&S>1?S-1:void 0,...this.pipelineContextToObservability(s._pipelineContext)}})),C={operationId:i,operation:"filter",tag:T.tag,sandboxId:T.sandboxId,swarmName:this.config.tag,operationName:s.name,itemIndex:n,verifyRetry:S&&S>1?S-1:void 0,...this.pipelineContextToMeta(s._pipelineContext)};return this.buildResult(T,C,c)};return this.runWithVerification(m,{originalPrompt:r,inputFiles:c,verifyConfig:u,timeoutMs:o,systemPrompt:s.systemPrompt,schema:s.schema,mcpServers:d,skills:f,integrations:p,operationId:i,baseTag:l,retry:a,itemIndex:n,operation:"filter",_pipelineContext:s._pipelineContext})}async executeBestOfCandidate(t){let{inputFiles:r,prompt:n,candidateIndex:i,operationId:s,config:o,timeoutMs:a,parentIndex:c,attempt:l=1,_pipelineContext:u}=t,d=c!==void 0?`${this.config.tag}-map-${c}-bestof-cand-${i}`:`${this.config.tag}-bestof-cand-${i}`,f=l>1?`${d}-er${l-1}`:d,p=await this.semaphore.use(()=>this.execute(r,n,{systemPrompt:t.systemPrompt,schema:t.schema,schemaOptions:t.schemaOptions,agent:o.taskAgents?.[i],mcpServers:t.mcpServers,skills:t.skills,integrations:t.integrations,tagPrefix:f,timeoutMs:a,observability:{swarmName:this.config.tag,operationId:s,operation:"map",itemIndex:c,role:"candidate",candidateIndex:i,errorRetry:l>1?l-1:void 0,...this.pipelineContextToObservability(u)}})),m={operationId:s,operation:"bestof-cand",tag:p.tag,sandboxId:p.sandboxId,swarmName:this.config.tag,itemIndex:i,errorRetry:l>1?l-1:void 0,candidateIndex:i,...this.pipelineContextToMeta(u)};return this.buildResult(p,m)}buildJudgeContext(t){let r=this.buildEvaluatorContext({inputFiles:t.inputFiles,taskPrompt:t.taskPrompt,systemPrompt:t.systemPrompt,schema:t.schema});return t.candidates.forEach((n,i)=>{n.status==="error"&&(r[`candidate_${i}/_failed.txt`]=`STATUS: FAILED
 

evolve/__init__.py

@@ -33,60 +33,76 @@ from .config import (
     HostedClientConfig,
 )
 from .hosted import (
-    ActiveBenchmark,
-    Benchmark,
-    BenchmarkImport,
-    ImportFailure,
-    BenchmarkImportFailure,
-    BenchmarkImportPage,
-    BenchmarkVersion,
-    BenchmarksClient,
+    ActiveDataset,
+    Agent,
+    AgentArm,
+    AgentCapability,
+    AgentInfo,
+    AgentPage,
+    AgentResult,
+    AgentsClient,
+    ApiKey,
+    AttemptPhase,
+    AuthClient,
+    AuthStatus,
     CapabilityDocument,
-    HarnessCapability,
-    HarnessModel,
-    HostedEvolve,
-    HOSTED_ERROR_CODES,
-    HostedErrorCode,
-    ProviderCapability,
-    StatusVocabulary,
-    UpstreamStatus,
-    is_hosted_error_code,
-    meta,
-    ComparisonAggregate,
-    ComparisonCell,
-    ComparisonCoverage,
-    ComparisonTaskRow,
-    CustomHarness,
-    CustomHarnessesClient,
+    CompareCell,
+    CompareCoverage,
+    CompareJobAggregate,
+    CompareResponse,
+    CompareTaskRow,
+    Dataset,
+    DatasetImport,
+    DatasetImportFailure,
+    DatasetImportPage,
+    DatasetPage,
+    DatasetRef,
+    DatasetSelector,
+    DatasetVersion,
+    DatasetsClient,
+    EvalSandboxProvider,
     EvolveAPIError,
     EvolveDigestMismatchError,
     EvolveIncompleteDownloadError,
-    BenchmarkPage,
-    CustomHarnessPage,
+    ExceptionInfo,
+    HostedEvolve,
+    HOSTED_ERROR_CODES,
+    HostedErrorCode,
+    ImportTaskFailure,
+    ImportWarning,
     Job,
-    JobAgent,
-    JobComparison,
     JobCounts,
     JobEvent,
     JobFailure,
     JobPage,
+    JobStatus,
+    JobTaskRollup,
+    JobTaskRollupPage,
     JobsClient,
-    ModelUsage,
+    ModelInfo,
     NoActiveVersionError,
-    RegradeFilter,
-    RegradeJob,
-    RegradePage,
-    RegradeResultsPage,
-    RegradeResult,
+    ManagedProviderCapability,
+    ProviderCapability,
+    SourceJob,
+    SpendSource,
+    StatusVocabulary,
+    StopResponse,
     Task,
     TaskPage,
     TaskProviderVerdict,
+    TimingInfo,
+    TraceEvent,
+    TraceEventPage,
     Trial,
-    TrialDetail,
     TrialPage,
+    TrialStatus,
     TrialTally,
-    TrialTraceEvent,
-    TrialTracePage,
+    TrialsClient,
+    UpstreamStatus,
+    VerifierEnvironmentMode,
+    VerifierResult,
+    is_hosted_error_code,
+    meta,
 )
 from .results import (
     AgentResponse,
@@ -235,22 +251,22 @@ def browser_profiles(config: Optional[BrowserProfilesClientConfig] = None) -> Br
     return BrowserProfilesClient(config or BrowserProfilesClientConfig())
 
 
-def benchmarks(config: Optional[HostedClientConfig] = None) -> BenchmarksClient:
-    """Create a standalone hosted-evals benchmarks client (shared catalog).
+def datasets(config: Optional[HostedClientConfig] = None) -> DatasetsClient:
+    """Create a standalone hosted-evals datasets client (shared catalog).
 
     Uses EVOLVE_API_KEY unless HostedClientConfig(api_key=...) is provided.
     """
-    return BenchmarksClient(config)
+    return DatasetsClient(config)
 
 
-def custom_harnesses(config: Optional[HostedClientConfig] = None) -> CustomHarnessesClient:
-    """Create a standalone hosted-evals custom harnesses client.
+def agents(config: Optional[HostedClientConfig] = None) -> AgentsClient:
+    """Create a standalone hosted-evals registered-agents client.
 
-    Register a private harness once, then name it in agents[].harness
+    Register a private agent once, then name it in job agents[].name
     exactly like a built-in. Uses EVOLVE_API_KEY unless
     HostedClientConfig(api_key=...) is provided.
     """
-    return CustomHarnessesClient(config)
+    return AgentsClient(config)
 
 
 def hosted(config: Optional[HostedClientConfig] = None) -> HostedEvolve:
@@ -261,14 +277,14 @@ def hosted(config: Optional[HostedClientConfig] = None) -> HostedEvolve:
     shift key apart that do completely different things is a trap. ``hosted()``
     says which half of the SDK you are reaching for.
 
-    The three clients underneath are built lazily, so ``await hosted().meta()``
+    The four clients underneath are built lazily, so ``await hosted().meta()``
     works with no API key configured::
 
         from evolve import hosted
 
         client = hosted()
-        doc = await client.meta()          # public, no key needed
-        job = await client.jobs.run(...)   # needs EVOLVE_API_KEY
+        doc = await client.meta()            # public, no key needed
+        job = await client.jobs.start(...)   # needs EVOLVE_API_KEY
     """
     return HostedEvolve(config)
 
@@ -283,6 +299,25 @@ def jobs(config: Optional[HostedClientConfig] = None) -> JobsClient:
     return JobsClient(config)
 
 
+def trials(config: Optional[HostedClientConfig] = None) -> TrialsClient:
+    """Create a standalone hosted-evals trials client.
+
+    A trial id is globally addressable — no method takes a job id. Uses
+    EVOLVE_API_KEY unless HostedClientConfig(api_key=...) is provided.
+    """
+    return TrialsClient(config)
+
+
+def auth(config: Optional[HostedClientConfig] = None) -> AuthClient:
+    """Create a standalone hosted-evals auth client (caller identity).
+
+    Wave-gated: until the server's wave lands, status() reports the route's
+    not-found as the API error it is. Uses EVOLVE_API_KEY unless
+    HostedClientConfig(api_key=...) is provided.
+    """
+    return AuthClient(config)
+
+
 def managed_secrets(config: Optional[ManagedSecretsClientConfig] = None) -> ManagedSecretsClient:
     """Create a standalone managed secrets client.
 
@@ -326,7 +361,7 @@ async def list_checkpoints(
         await store.close()
 
 
-__version__ = '0.0.52.dev638'
+__version__ = '0.0.52.dev656'
 
 __all__ = [
     # Main classes
@@ -393,60 +428,78 @@ __all__ = [
     'browser_profiles',
     'managed_secrets',
 
-    # Hosted evals (benchmarks + custom harnesses + jobs)
+    # Hosted evals (datasets + agents + jobs + trials + auth)
     'HostedClientConfig',
-    'BenchmarksClient',
-    'CustomHarnessesClient',
+    'DatasetsClient',
+    'AgentsClient',
     'JobsClient',
+    'TrialsClient',
+    'AuthClient',
     'EvolveAPIError',
     'EvolveDigestMismatchError',
     'EvolveIncompleteDownloadError',
     'NoActiveVersionError',
-    'Benchmark',
-    'ActiveBenchmark',
-    'BenchmarkImport',
-    'ImportFailure',
-    'BenchmarkImportFailure',
-    'BenchmarkVersion',
+    'Dataset',
+    'ActiveDataset',
+    'DatasetVersion',
+    'DatasetRef',
+    'DatasetSelector',
+    'DatasetImport',
+    'DatasetImportFailure',
+    'ImportTaskFailure',
+    'ImportWarning',
     'Task',
     'TaskProviderVerdict',
-    'JobAgent',
+    'AgentArm',
     'Job',
     'JobCounts',
     'JobFailure',
+    'JobStatus',
+    'TrialStatus',
+    'EvalSandboxProvider',
+    'SpendSource',
+    'VerifierEnvironmentMode',
+    'AttemptPhase',
+    'JobTaskRollup',
+    'SourceJob',
     'TrialTally',
     'JobEvent',
-    'JobComparison',
-    'ComparisonAggregate',
-    'ComparisonCell',
-    'ComparisonCoverage',
-    'ComparisonTaskRow',
-    'RegradeJob',
-    'RegradeResult',
-    'RegradeFilter',
-    'RegradeResultsPage',
-    'ModelUsage',
-    'RegradePage',
+    'CompareResponse',
+    'CompareJobAggregate',
+    'CompareCell',
+    'CompareCoverage',
+    'CompareTaskRow',
     'Trial',
-    'TrialDetail',
-    'TrialTraceEvent',
-    'TrialTracePage',
+    'TimingInfo',
+    'ModelInfo',
+    'AgentInfo',
+    'AgentResult',
+    'VerifierResult',
+    'ExceptionInfo',
+    'StopResponse',
+    'TraceEvent',
+    'TraceEventPage',
     'JobPage',
     'TrialPage',
-    'BenchmarkPage',
-    'CustomHarnessPage',
+    'JobTaskRollupPage',
+    'DatasetPage',
+    'AgentPage',
     'TaskPage',
-    'CustomHarness',
-    'BenchmarkImportPage',
-    'benchmarks',
-    'custom_harnesses',
+    'Agent',
+    'DatasetImportPage',
+    'ApiKey',
+    'AuthStatus',
+    'datasets',
+    'agents',
     'jobs',
+    'trials',
+    'auth',
     'hosted',
     'meta',
     'HostedEvolve',
     'CapabilityDocument',
-    'HarnessCapability',
-    'HarnessModel',
+    'AgentCapability',
+    'ManagedProviderCapability',
     'ProviderCapability',
     'StatusVocabulary',
     'UpstreamStatus',

evolve/config.py

@@ -82,7 +82,7 @@ class BrowserCredentialsConfig:
 
 @dataclass
 class HostedClientConfig:
-    """Standalone hosted evals client configuration (benchmarks/jobs).
+    """Standalone hosted evals client configuration (datasets/jobs).
 
     Args:
         api_key: Evolve API key override (default: EVOLVE_API_KEY)