@outputai/evals 0.1.4 → 0.1.5

package/dist/aggregate.d.ts

@@ -0,0 +1,2 @@
+import type { EvaluatorResult, Verdict } from './schemas.js';
+export declare function aggregateCaseVerdict(evaluatorResults: EvaluatorResult[]): Verdict;

package/dist/aggregate.js

@@ -0,0 +1,11 @@
+import { VERDICT, CRITICALITY } from './schemas.js';
+export function aggregateCaseVerdict(evaluatorResults) {
+    const required = evaluatorResults.filter(r => r.criticality === CRITICALITY.REQUIRED);
+    if (required.some(r => r.verdict === VERDICT.FAIL)) {
+        return VERDICT.FAIL;
+    }
+    if (required.some(r => r.verdict === VERDICT.PARTIAL)) {
+        return VERDICT.PARTIAL;
+    }
+    return VERDICT.PASS;
+}

package/dist/aggregate.spec.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/aggregate.spec.js

@@ -0,0 +1,48 @@
+import { describe, it, expect } from 'vitest';
+import { aggregateCaseVerdict } from './aggregate.js';
+const makeResult = (verdict, criticality = 'required') => ({
+    name: 'test_eval',
+    verdict,
+    criticality
+});
+describe('aggregateCaseVerdict', () => {
+    it('returns pass when all required evaluators pass', () => {
+        expect(aggregateCaseVerdict([
+            makeResult('pass'),
+            makeResult('pass')
+        ])).toBe('pass');
+    });
+    it('returns fail when any required evaluator fails', () => {
+        expect(aggregateCaseVerdict([
+            makeResult('pass'),
+            makeResult('fail')
+        ])).toBe('fail');
+    });
+    it('returns partial when any required evaluator is partial and none fail', () => {
+        expect(aggregateCaseVerdict([
+            makeResult('pass'),
+            makeResult('partial')
+        ])).toBe('partial');
+    });
+    it('returns fail over partial when both exist', () => {
+        expect(aggregateCaseVerdict([
+            makeResult('partial'),
+            makeResult('fail')
+        ])).toBe('fail');
+    });
+    it('ignores informational evaluators for verdict', () => {
+        expect(aggregateCaseVerdict([
+            makeResult('pass'),
+            makeResult('fail', 'informational')
+        ])).toBe('pass');
+    });
+    it('returns pass when only informational evaluators fail', () => {
+        expect(aggregateCaseVerdict([
+            makeResult('fail', 'informational'),
+            makeResult('partial', 'informational')
+        ])).toBe('pass');
+    });
+    it('returns pass for empty evaluator list', () => {
+        expect(aggregateCaseVerdict([])).toBe('pass');
+    });
+});

package/dist/eval_workflow.d.ts

@@ -0,0 +1,20 @@
+import type { InterpretConfig, EvalOutput, Dataset, Criticality } from './schemas.js';
+export interface EvalDefinition {
+    evaluator: Function;
+    criticality?: Criticality;
+    interpret: InterpretConfig;
+}
+export interface EvalWorkflowConfig {
+    concurrency?: number;
+}
+export type EvalWorkflowInput = {
+    datasets: Dataset[];
+};
+export type EvalWorkflowFn = (input: EvalWorkflowInput, context: unknown) => Promise<EvalOutput>;
+export interface EvalWorkflowOptions {
+    name: string;
+    evals: EvalDefinition[];
+    fn?: EvalWorkflowFn;
+    config?: EvalWorkflowConfig;
+}
+export declare function evalWorkflow({ name, evals, fn, config }: EvalWorkflowOptions): unknown;

package/dist/eval_workflow.js

@@ -0,0 +1,88 @@
+import { workflow, executeInParallel, z } from '@outputai/core';
+import { getMetadata } from '@outputai/core/sdk_utils';
+import { VERDICT, CRITICALITY, DatasetSchema, EvalOutputSchema } from './schemas.js';
+import { interpretResult } from './interpret.js';
+import { aggregateCaseVerdict } from './aggregate.js';
+export function evalWorkflow({ name, evals, fn, config = {} }) {
+    const concurrency = config.concurrency ?? 10;
+    const evalDefs = evals.map(def => {
+        const meta = getMetadata(def.evaluator);
+        if (!meta) {
+            throw new Error('Evaluator passed to evalWorkflow is missing metadata. Ensure it was created with evaluator().');
+        }
+        return {
+            name: meta.name,
+            criticality: def.criticality ?? CRITICALITY.REQUIRED,
+            interpret: def.interpret
+        };
+    });
+    const defaultFn = async function (input) {
+        const { invokeEvaluator } = this;
+        const jobs = input.datasets.map((dataset) => async () => {
+            const evaluatorResults = [];
+            for (const evalDef of evalDefs) {
+                const { evals: perEvalTruth, ...globalTruth } = dataset.ground_truth ?? {};
+                const evalTruth = perEvalTruth?.[evalDef.name] ?? {};
+                const mergedTruth = { ...globalTruth, ...evalTruth };
+                const result = await invokeEvaluator(evalDef.name, {
+                    input: dataset.input,
+                    output: dataset.last_output?.output,
+                    ground_truth: mergedTruth
+                });
+                const verdict = interpretResult(result, evalDef.interpret);
+                evaluatorResults.push({
+                    name: evalDef.name,
+                    verdict,
+                    criticality: evalDef.criticality,
+                    confidence: result.confidence,
+                    reasoning: result.reasoning,
+                    feedback: result.feedback?.length ? result.feedback : undefined
+                });
+            }
+            const caseVerdict = aggregateCaseVerdict(evaluatorResults);
+            return {
+                datasetName: dataset.name,
+                verdict: caseVerdict,
+                evaluators: evaluatorResults
+            };
+        });
+        const results = await executeInParallel({ jobs, concurrency });
+        const cases = results.map((r, i) => {
+            if (r.ok) {
+                return r.result;
+            }
+            const datasetName = input.datasets[i]?.name ?? `dataset_${i}`;
+            return {
+                datasetName,
+                verdict: VERDICT.FAIL,
+                evaluators: [{
+                        name: '_error',
+                        verdict: VERDICT.FAIL,
+                        criticality: CRITICALITY.REQUIRED,
+                        reasoning: r.error instanceof Error ? r.error.message : String(r.error)
+                    }]
+            };
+        });
+        const passed = cases.filter(c => c.verdict === VERDICT.PASS).length;
+        const partial = cases.filter(c => c.verdict === VERDICT.PARTIAL).length;
+        const failed = cases.filter(c => c.verdict === VERDICT.FAIL).length;
+        const total = cases.length;
+        return {
+            cases,
+            summary: {
+                total,
+                passed,
+                partial,
+                failed,
+                acceptableRate: total > 0 ? (passed + partial) / total : 0
+            }
+        };
+    };
+    return workflow({
+        name,
+        description: `Eval workflow for ${name}`,
+        inputSchema: z.object({ datasets: z.array(DatasetSchema) }),
+        outputSchema: EvalOutputSchema,
+        fn: fn ?? defaultFn
+    });
+}

package/dist/index.d.ts

@@ -0,0 +1,13 @@
+export { evalWorkflow } from './eval_workflow.js';
+export type { EvalDefinition, EvalWorkflowConfig, EvalWorkflowInput, EvalWorkflowFn, EvalWorkflowOptions } from './eval_workflow.js';
+export { Verdict } from './verdict.js';
+export { verify } from './verify.js';
+export type { CheckContext } from './verify.js';
+export { judgeVerdict, judgeScore, judgeBoolean, judgeLabel } from './judge.js';
+export type { JudgeArgs } from './judge.js';
+export { interpretResult } from './interpret.js';
+export { aggregateCaseVerdict } from './aggregate.js';
+export { renderEvalOutput, computeExitCode } from './render.js';
+export { getEvalWorkflowName, isEvalWorkflow, getParentWorkflowName } from './naming.js';
+export { VERDICT, CRITICALITY, VerdictSchema, CriticalitySchema, InterpretConfigSchema, EvaluatorResultSchema, EvalCaseSchema, EvalOutputSchema, LastOutputSchema, LastEvalSchema, DatasetSchema } from './schemas.js';
+export type { Verdict as VerdictType, Criticality, InterpretConfig, EvaluatorResult, EvalCase, EvalSummary, EvalOutput, GroundTruth, LastOutput, LastEval, Dataset } from './schemas.js';

package/dist/index.js

@@ -0,0 +1,9 @@
+export { evalWorkflow } from './eval_workflow.js';
+export { Verdict } from './verdict.js';
+export { verify } from './verify.js';
+export { judgeVerdict, judgeScore, judgeBoolean, judgeLabel } from './judge.js';
+export { interpretResult } from './interpret.js';
+export { aggregateCaseVerdict } from './aggregate.js';
+export { renderEvalOutput, computeExitCode } from './render.js';
+export { getEvalWorkflowName, isEvalWorkflow, getParentWorkflowName } from './naming.js';
+export { VERDICT, CRITICALITY, VerdictSchema, CriticalitySchema, InterpretConfigSchema, EvaluatorResultSchema, EvalCaseSchema, EvalOutputSchema, LastOutputSchema, LastEvalSchema, DatasetSchema } from './schemas.js';

package/dist/index_sandbox.d.ts

@@ -0,0 +1,16 @@
+export { evalWorkflow } from './eval_workflow.js';
+export type { EvalDefinition, EvalWorkflowConfig, EvalWorkflowInput, EvalWorkflowFn, EvalWorkflowOptions } from './eval_workflow.js';
+export { Verdict } from './verdict.js';
+export { verify } from './verify.js';
+export type { CheckContext } from './verify.js';
+export type { JudgeArgs } from './judge.js';
+export { interpretResult } from './interpret.js';
+export { aggregateCaseVerdict } from './aggregate.js';
+export { renderEvalOutput, computeExitCode } from './render.js';
+export { getEvalWorkflowName, isEvalWorkflow, getParentWorkflowName } from './naming.js';
+export { VERDICT, CRITICALITY, VerdictSchema, CriticalitySchema, InterpretConfigSchema, EvaluatorResultSchema, EvalCaseSchema, EvalOutputSchema, LastOutputSchema, LastEvalSchema, DatasetSchema } from './schemas.js';
+export type { Verdict as VerdictType, Criticality, InterpretConfig, EvaluatorResult, EvalCase, EvalSummary, EvalOutput, GroundTruth, LastOutput, LastEval, Dataset } from './schemas.js';
+export declare const judgeVerdict: (...args: unknown[]) => never;
+export declare const judgeScore: (...args: unknown[]) => never;
+export declare const judgeBoolean: (...args: unknown[]) => never;
+export declare const judgeLabel: (...args: unknown[]) => never;

package/dist/index_sandbox.js

@@ -0,0 +1,51 @@
+/*
+ * HACK: Export workflow-safe function signatures independently.
+ *
+ * Problem:
+ * Eval workflow files pass evaluator functions as config values to evalWorkflow():
+ *
+ *   export default evalWorkflow({ evals: [{ evaluator: evaluateTopic, ... }] });
+ *
+ * The webpack rewriter (workflow_rewriter/index.mjs) is supposed to strip evaluator
+ * imports from the AST via collectTargetImports(), which calls path.remove() on each
+ * matched import declaration. However, after collecting imports, the rewriter checks
+ * if rewriteFnBodies() performed any rewrites. For eval workflows — which have no
+ * fn body, only a config object — rewriteFnBodies() returns false. When that happens,
+ * the rewriter returns the original source string unchanged (lines 46-48), discarding
+ * the import stripping that collectTargetImports() already applied to the AST.
+ *
+ * This means webpack follows the full import chain: evaluators.js → @outputai/evals
+ * → judge.js → @outputai/llm → node:zlib — which fails because Node.js built-ins
+ * can't be bundled into Temporal's deterministic workflow bundle.
+ *
+ * This file provides an alternative entry point for @outputai/evals that excludes
+ * the real judge functions (breaking the chain to @outputai/llm) and replaces them
+ * with no-op stubs. The stubs satisfy webpack's named-export resolution. Judge
+ * functions are never called inside the workflow bundle — they execute as Temporal
+ * activities at runtime where the real entry point (index.js) is used.
+ *
+ * Remove when:
+ * The workflow rewriter in sdk/core/src/worker/webpack_loaders/workflow_rewriter/index.mjs
+ * is fixed to generate output from the modified AST even when rewriteFnBodies() returns
+ * false — i.e. when collectTargetImports() stripped imports but no fn bodies needed
+ * rewriting. This also requires injecting metadata stubs for the stripped evaluator
+ * references, since evalWorkflow() calls getMetadata(def.evaluator) at module init time
+ * and the identifiers would otherwise be undefined. Once the rewriter properly strips the
+ * evaluator → judge → @outputai/llm chain, this file and the "output-workflow-bundle" export
+ * condition in package.json can be deleted.
+ */
+export { evalWorkflow } from './eval_workflow.js';
+export { Verdict } from './verdict.js';
+export { verify } from './verify.js';
+export { interpretResult } from './interpret.js';
+export { aggregateCaseVerdict } from './aggregate.js';
+export { renderEvalOutput, computeExitCode } from './render.js';
+export { getEvalWorkflowName, isEvalWorkflow, getParentWorkflowName } from './naming.js';
+export { VERDICT, CRITICALITY, VerdictSchema, CriticalitySchema, InterpretConfigSchema, EvaluatorResultSchema, EvalCaseSchema, EvalOutputSchema, LastOutputSchema, LastEvalSchema, DatasetSchema } from './schemas.js';
+const sandboxStub = () => {
+    throw new Error('Judge functions are not available in the Temporal workflow bundle');
+};
+export const judgeVerdict = sandboxStub;
+export const judgeScore = sandboxStub;
+export const judgeBoolean = sandboxStub;
+export const judgeLabel = sandboxStub;

package/dist/interpret.d.ts

@@ -0,0 +1,3 @@
+import type { EvaluationResult } from '@outputai/core';
+import type { InterpretConfig, Verdict } from './schemas.js';
+export declare const interpretResult: (result: EvaluationResult, config: InterpretConfig) => Verdict;

package/dist/interpret.js

@@ -0,0 +1,40 @@
+import { VERDICT, VerdictSchema } from './schemas.js';
+const VALID_VERDICTS = new Set(VerdictSchema.options);
+const interpretVerdict = (value) => VALID_VERDICTS.has(value) ? value : VERDICT.FAIL;
+const interpretBoolean = (value) => value === true ? VERDICT.PASS : VERDICT.FAIL;
+const interpretNumber = (value, config) => {
+    if (typeof value !== 'number' || Number.isNaN(value)) {
+        return VERDICT.FAIL;
+    }
+    if (value >= config.pass) {
+        return VERDICT.PASS;
+    }
+    if (config.partial !== undefined && config.partial !== null && value >= config.partial) {
+        return VERDICT.PARTIAL;
+    }
+    return VERDICT.FAIL;
+};
+const interpretString = (value, config) => {
+    if (config.pass.includes(value)) {
+        return VERDICT.PASS;
+    }
+    if (config.partial?.includes(value)) {
+        return VERDICT.PARTIAL;
+    }
+    return VERDICT.FAIL;
+};
+export const interpretResult = (result, config) => {
+    const { value } = result;
+    switch (config.type) {
+        case 'verdict':
+            return interpretVerdict(value);
+        case 'boolean':
+            return interpretBoolean(value);
+        case 'number':
+            return interpretNumber(value, config);
+        case 'string':
+            return interpretString(value, config);
+        default:
+            return VERDICT.FAIL;
+    }
+};

package/dist/interpret.spec.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/interpret.spec.js

@@ -0,0 +1,86 @@
+import { describe, it, expect } from 'vitest';
+import { interpretResult } from './interpret.js';
+const makeResult = (value) => ({
+    value,
+    confidence: 1.0,
+    feedback: [],
+    dimensions: []
+});
+describe('interpretResult', () => {
+    describe('verdict config', () => {
+        const config = { type: 'verdict' };
+        it('passes through valid verdict values', () => {
+            expect(interpretResult(makeResult('pass'), config)).toBe('pass');
+            expect(interpretResult(makeResult('partial'), config)).toBe('partial');
+            expect(interpretResult(makeResult('fail'), config)).toBe('fail');
+        });
+        it('returns fail for invalid verdict values', () => {
+            expect(interpretResult(makeResult('maybe'), config)).toBe('fail');
+            expect(interpretResult(makeResult(''), config)).toBe('fail');
+            expect(interpretResult(makeResult(null), config)).toBe('fail');
+            expect(interpretResult(makeResult(42), config)).toBe('fail');
+        });
+    });
+    describe('boolean config', () => {
+        const config = { type: 'boolean' };
+        it('returns pass for true', () => {
+            expect(interpretResult(makeResult(true), config)).toBe('pass');
+        });
+        it('returns fail for false', () => {
+            expect(interpretResult(makeResult(false), config)).toBe('fail');
+        });
+        it('returns fail for truthy non-boolean values', () => {
+            expect(interpretResult(makeResult(1), config)).toBe('fail');
+            expect(interpretResult(makeResult('true'), config)).toBe('fail');
+        });
+    });
+    describe('number config', () => {
+        it('returns pass when value >= pass threshold', () => {
+            const config = { type: 'number', pass: 0.8 };
+            expect(interpretResult(makeResult(0.9), config)).toBe('pass');
+            expect(interpretResult(makeResult(0.8), config)).toBe('pass');
+            expect(interpretResult(makeResult(1.0), config)).toBe('pass');
+        });
+        it('returns fail when value < pass threshold with no partial', () => {
+            const config = { type: 'number', pass: 0.8 };
+            expect(interpretResult(makeResult(0.7), config)).toBe('fail');
+            expect(interpretResult(makeResult(0), config)).toBe('fail');
+        });
+        it('returns partial when value between partial and pass thresholds', () => {
+            const config = { type: 'number', pass: 0.8, partial: 0.5 };
+            expect(interpretResult(makeResult(0.6), config)).toBe('partial');
+            expect(interpretResult(makeResult(0.5), config)).toBe('partial');
+        });
+        it('returns fail when value < partial threshold', () => {
+            const config = { type: 'number', pass: 0.8, partial: 0.5 };
+            expect(interpretResult(makeResult(0.4), config)).toBe('fail');
+        });
+        it('returns fail for non-number values', () => {
+            const config = { type: 'number', pass: 0.8 };
+            expect(interpretResult(makeResult(null), config)).toBe('fail');
+            expect(interpretResult(makeResult(undefined), config)).toBe('fail');
+            expect(interpretResult(makeResult('high'), config)).toBe('fail');
+        });
+    });
+    describe('string config', () => {
+        it('returns pass for matching pass values', () => {
+            const config = { type: 'string', pass: ['good', 'great'] };
+            expect(interpretResult(makeResult('good'), config)).toBe('pass');
+            expect(interpretResult(makeResult('great'), config)).toBe('pass');
+        });
+        it('returns partial for matching partial values', () => {
+            const config = { type: 'string', pass: ['good'], partial: ['ok', 'decent'] };
+            expect(interpretResult(makeResult('ok'), config)).toBe('partial');
+            expect(interpretResult(makeResult('decent'), config)).toBe('partial');
+        });
+        it('returns fail for non-matching values', () => {
+            const config = { type: 'string', pass: ['good'], partial: ['ok'] };
+            expect(interpretResult(makeResult('bad'), config)).toBe('fail');
+            expect(interpretResult(makeResult(''), config)).toBe('fail');
+        });
+        it('returns fail when no partial config and value not in pass', () => {
+            const config = { type: 'string', pass: ['good'] };
+            expect(interpretResult(makeResult('ok'), config)).toBe('fail');
+        });
+    });
+});

package/dist/judge.d.ts

@@ -0,0 +1,10 @@
+import { EvaluationVerdictResult, EvaluationNumberResult, EvaluationBooleanResult, EvaluationStringResult, z } from '@outputai/core';
+export type JudgeArgs = {
+    prompt: string;
+    variables?: Record<string, string | number | boolean>;
+    schema?: z.ZodType;
+};
+export declare function judgeVerdict({ prompt, variables, schema }: JudgeArgs): Promise<EvaluationVerdictResult>;
+export declare function judgeScore({ prompt, variables, schema }: JudgeArgs): Promise<EvaluationNumberResult>;
+export declare function judgeBoolean({ prompt, variables, schema }: JudgeArgs): Promise<EvaluationBooleanResult>;
+export declare function judgeLabel({ prompt, variables, schema }: JudgeArgs): Promise<EvaluationStringResult>;

package/dist/judge.js

@@ -0,0 +1,31 @@
+import { EvaluationVerdictResult, EvaluationNumberResult, EvaluationBooleanResult, EvaluationStringResult, z } from '@outputai/core';
+import { resolveInvocationDir } from '@outputai/core/sdk_utils';
+import { generateText, Output } from '@outputai/llm';
+const verdictSchema = z.object({ verdict: z.enum(['pass', 'partial', 'fail']), reasoning: z.string() });
+const scoreSchema = z.object({ score: z.number(), reasoning: z.string() });
+const booleanSchema = z.object({ result: z.boolean(), reasoning: z.string() });
+const labelSchema = z.object({ label: z.string(), reasoning: z.string() });
+export async function judgeVerdict({ prompt, variables, schema = verdictSchema }) {
+    const promptDir = resolveInvocationDir();
+    const response = await generateText({ prompt, variables, promptDir, output: Output.object({ schema }) });
+    const result = response.output;
+    return new EvaluationVerdictResult({ value: result.verdict, confidence: 0.9, reasoning: result.reasoning });
+}
+export async function judgeScore({ prompt, variables, schema = scoreSchema }) {
+    const promptDir = resolveInvocationDir();
+    const response = await generateText({ prompt, variables, promptDir, output: Output.object({ schema }) });
+    const result = response.output;
+    return new EvaluationNumberResult({ value: result.score, confidence: 0.9, reasoning: result.reasoning });
+}
+export async function judgeBoolean({ prompt, variables, schema = booleanSchema }) {
+    const promptDir = resolveInvocationDir();
+    const response = await generateText({ prompt, variables, promptDir, output: Output.object({ schema }) });
+    const result = response.output;
+    return new EvaluationBooleanResult({ value: result.result, confidence: 0.9, reasoning: result.reasoning });
+}
+export async function judgeLabel({ prompt, variables, schema = labelSchema }) {
+    const promptDir = resolveInvocationDir();
+    const response = await generateText({ prompt, variables, promptDir, output: Output.object({ schema }) });
+    const result = response.output;
+    return new EvaluationStringResult({ value: result.label, confidence: 0.9, reasoning: result.reasoning });
+}

package/dist/naming.d.ts

@@ -0,0 +1,3 @@
+export declare const getEvalWorkflowName: (workflowName: string) => string;
+export declare const isEvalWorkflow: (name: string) => boolean;
+export declare const getParentWorkflowName: (evalName: string) => string;

package/dist/naming.js

@@ -0,0 +1,4 @@
+const EVAL_SUFFIX = '_eval';
+export const getEvalWorkflowName = (workflowName) => `${workflowName}${EVAL_SUFFIX}`;
+export const isEvalWorkflow = (name) => name.endsWith(EVAL_SUFFIX);
+export const getParentWorkflowName = (evalName) => evalName.slice(0, -EVAL_SUFFIX.length);

package/dist/naming.spec.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/naming.spec.js

@@ -0,0 +1,30 @@
+import { describe, it, expect } from 'vitest';
+import { getEvalWorkflowName, isEvalWorkflow, getParentWorkflowName } from './naming.js';
+describe('naming', () => {
+    describe('getEvalWorkflowName', () => {
+        it('appends _eval suffix', () => {
+            expect(getEvalWorkflowName('simple')).toBe('simple_eval');
+            expect(getEvalWorkflowName('my_workflow')).toBe('my_workflow_eval');
+        });
+    });
+    describe('isEvalWorkflow', () => {
+        it('returns true for names ending with _eval', () => {
+            expect(isEvalWorkflow('simple_eval')).toBe(true);
+            expect(isEvalWorkflow('my_workflow_eval')).toBe(true);
+        });
+        it('returns false for regular workflow names', () => {
+            expect(isEvalWorkflow('simple')).toBe(false);
+            expect(isEvalWorkflow('eval_runner')).toBe(false);
+            expect(isEvalWorkflow('evaluation')).toBe(false);
+        });
+    });
+    describe('getParentWorkflowName', () => {
+        it('strips the _eval suffix', () => {
+            expect(getParentWorkflowName('simple_eval')).toBe('simple');
+            expect(getParentWorkflowName('my_workflow_eval')).toBe('my_workflow');
+        });
+        it('returns empty string for bare _eval', () => {
+            expect(getParentWorkflowName('_eval')).toBe('');
+        });
+    });
+});

package/dist/render.d.ts

@@ -0,0 +1,3 @@
+import type { EvalOutput } from './schemas.js';
+export declare function renderEvalOutput(evalOutput: EvalOutput, evalName?: string): string;
+export declare function computeExitCode(evalOutput: EvalOutput): number;

package/dist/render.js

@@ -0,0 +1,75 @@
+import { VERDICT, CRITICALITY } from './schemas.js';
+const COLORS = {
+    reset: '\x1b[0m',
+    green: '\x1b[32m',
+    yellow: '\x1b[33m',
+    red: '\x1b[31m',
+    dim: '\x1b[2m',
+    bold: '\x1b[1m',
+    cyan: '\x1b[36m'
+};
+const VERDICT_STYLE = {
+    [VERDICT.PASS]: { label: 'PASS', color: COLORS.green },
+    [VERDICT.PARTIAL]: { label: 'PARTIAL', color: COLORS.yellow },
+    [VERDICT.FAIL]: { label: 'FAIL', color: COLORS.red }
+};
+function colorize(text, color) {
+    return `${color}${text}${COLORS.reset}`;
+}
+function padWithDots(label, width = 30) {
+    const dotCount = Math.max(2, width - label.length);
+    return ` ${'.'.repeat(dotCount)} `;
+}
+function renderEvaluator(evaluator) {
+    const lines = [];
+    const style = VERDICT_STYLE[evaluator.verdict];
+    const prefix = evaluator.criticality === CRITICALITY.INFORMATIONAL ?
+        `${COLORS.dim}(info)${COLORS.reset} ` :
+        '';
+    lines.push(`    ${prefix}${evaluator.name}${padWithDots(evaluator.name, 24)}${colorize(evaluator.verdict, style.color)}`);
+    if (evaluator.reasoning && evaluator.verdict !== VERDICT.PASS) {
+        lines.push(`      ${COLORS.dim}-> ${evaluator.reasoning}${COLORS.reset}`);
+    }
+    if (evaluator.feedback) {
+        for (const fb of evaluator.feedback) {
+            const item = fb;
+            if (item.issue) {
+                lines.push(`      ${COLORS.dim}-> ${item.issue}${COLORS.reset}`);
+            }
+        }
+    }
+    return lines;
+}
+function renderCase(evalCase) {
+    const caseStyle = VERDICT_STYLE[evalCase.verdict];
+    const lines = [
+        `  ${evalCase.datasetName}${padWithDots(evalCase.datasetName)}${colorize(caseStyle.label, caseStyle.color)}`
+    ];
+    evalCase.evaluators.forEach(evaluator => lines.push(...renderEvaluator(evaluator)));
+    lines.push('');
+    return lines;
+}
+export function renderEvalOutput(evalOutput, evalName) {
+    const lines = [];
+    if (evalName) {
+        lines.push(colorize(evalName, COLORS.bold));
+        lines.push('');
+    }
+    evalOutput.cases.forEach(evalCase => lines.push(...renderCase(evalCase)));
+    const { summary } = evalOutput;
+    const parts = [];
+    if (summary.passed > 0) {
+        parts.push(colorize(`${summary.passed} passed`, COLORS.green));
+    }
+    if (summary.partial > 0) {
+        parts.push(colorize(`${summary.partial} partial`, COLORS.yellow));
+    }
+    if (summary.failed > 0) {
+        parts.push(colorize(`${summary.failed} failed`, COLORS.red));
+    }
+    lines.push(`${parts.join(', ')} (${Math.round(summary.acceptableRate * 100)}% acceptable)`);
+    return lines.join('\n');
+}
+export function computeExitCode(evalOutput) {
+    return evalOutput.cases.some(c => c.verdict === VERDICT.FAIL) ? 1 : 0;
+}

package/dist/render.spec.d.ts

@@ -0,0 +1 @@
+export {};