eve/dist/src/evals/runner/reporters/console.js

Filesystem-first framework for durable backend AI agents that run anywhere.

import picocolors from"#compiled/picocolors/index.js";var ConsoleReporter=class{#e;#t;constructor(t){this.#e=t?.log??console.log,this.#t=picocolors.createColors(t?.color??!!process.stdout.isTTY)}onRunStart(e,t){this.#e(``),this.#e(`${this.#t.bold(this.#t.cyan(`EVALS`))} ${this.#t.bold(String(e.length))}`),this.#e(`${this.#t.dim(`target`)} ${t.kind===`local`?this.#t.green(t.url):this.#t.blue(t.url)}`),this.#e(``)}onEvalComplete(e){let{assertions:t,verdict:n,error:r}=e,i=t.filter(e=>e.severity===`gate`),a=t.filter(e=>e.severity===`soft`),o=this.#n(n),s=i.length>0?this.#r(i.filter(e=>e.passed).length,i.length):``,c=a.map(e=>this.#i(e.name,e.score)).join(`  `),l=[o,this.#t.dim(e.id),s,c].filter(Boolean).join(`  `);this.#e(l);for(let e of t){if(e.passed)continue;let t=e.message===void 0?``:`: ${e.message}`;this.#e(`  ${this.#t.red(`✗ ${e.name}${t}`)}`)}r&&this.#e(`  ${this.#t.red(r)}`)}onRunComplete(e){this.#e(``);let{passed:t,failed:n,scored:r,results:i}=e,a=i.length,o=[];t>0&&o.push(this.#t.green(`${t} passed`)),n>0&&o.push(this.#t.red(`${n} failed`)),r>0&&o.push(this.#t.yellow(`${r} scored`)),o.length===0&&o.push(this.#t.dim(`0 evals`)),this.#e(`${this.#t.bold(`Results:`)} ${o.join(`, `)} ${this.#t.dim(`(${a} total)`)}`);let s=this.#a(i);if(s.total>0){let e=this.#t.green(`${s.passed} passed`),t=s.failed>0?`, ${this.#t.red(`${s.failed} failed`)}`:``;this.#e(`${this.#t.bold(`Gates:`)} ${e}${t}`)}let c=this.#o(i);if(c.length>0){this.#e(``);for(let{name:e,avg:t,count:n}of c){let r=this.#i(e,t);this.#e(`  ${r} ${this.#t.dim(`(${n} evals)`)}`)}}let l=computeDurationMs(e.startedAt,e.completedAt);this.#e(``),this.#e(this.#t.dim(`Completed in ${formatDuration(l)}`)),this.#e(``)}#n(e){switch(e){case`passed`:return this.#t.green(`✓`);case`failed`:return this.#t.red(`✗`);case`scored`:return this.#t.yellow(`○`)}}#r(e,t){let n=`gates ${e}/${t}`;return e===t?this.#t.green(n):this.#t.red(n)}#i(e,t){let n=`${e}: ${Math.round(t*100)}%`;return t===1?this.#t.green(n):t===0?this.#t.red(n):this.#t.yellow(n)}#a(e){let t=0,n=0;for(let r of e)for(let e of gatesOf(r))e.passed?t+=1:n+=1;return{passed:t,failed:n,total:t+n}}#o(e){let t=new Map;for(let n of e)for(let e of n.assertions){if(e.severity!==`soft`)continue;let n=t.get(e.name);n?(n.sum+=e.score,n.count+=1):t.set(e.name,{sum:e.score,count:1})}return[...t.entries()].map(([e,{sum:t,count:n}])=>({name:e,avg:t/n,count:n}))}};function gatesOf(e){return e.assertions.filter(e=>e.severity===`gate`)}function computeDurationMs(e,t){return new Date(t).getTime()-new Date(e).getTime()}function formatDuration(e){return e<1e3?`${e}ms`:e<6e4?`${(e/1e3).toFixed(1)}s`:`${Math.floor(e/6e4)}m ${(e%6e4/1e3).toFixed(0)}s`}export{ConsoleReporter};