@ggui-ai/ui-gen 0.24.0 → 0.26.0

package/dist/evaluation/types.d.ts

@@ -1,3 +1,11 @@
+import { C as CriteriaAnswerCounts } from '../types-public-B7wn4ce5.js';
+import '@ggui-ai/protocol';
+import '../llm.js';
+import '@ggui-ai/mcp-server-core';
+import '../design-mode-DGWN_Gkm.js';
+import '../axes-CKVCzR_x.js';
+import '../types-Dh1s4K9j.js';
+
 /**
  * Per-dimension quality scores from the evaluator (0-100 each).
  */
@@ -25,10 +33,23 @@ interface EvaluationIssue {
     severity: 'critical' | 'major' | 'minor';
     /** Suggested fix for the agent to apply */
     fix: string;
+    /**
+     * Who produced it, on the visual leg (ggui#1545): `judge` for the model's own findings, `instrument` for a
+     * deterministic measurement (fit, overflow, blank). Absent on issues from other evaluators. The loop's
+     * `visualEvaluation.actOn` acts on `judge` findings only, so a new instrument check never joins the treatment
+     * by accident: it has to say it is a judge finding, and absence reads as "not acted on".
+     */
+    origin?: 'judge' | 'instrument';
 }
 /**
  * Result of a single evaluation round.
  */
+/** One judge answer to a criterion (ggui#1436), as parsed — before the K-majority. */
+interface CriteriaAnswer {
+    id: string;
+    verdict: 'pass' | 'fail' | 'n/a';
+    evidence: string;
+}
 interface EvaluationResult {
     /** Whether the component passed the quality threshold */
     passed: boolean;
@@ -44,6 +65,10 @@ interface EvaluationResult {
     inputTokens?: number;
     /** Output tokens consumed by the evaluation LLM call */
     outputTokens?: number;
+    /** ggui#1436 — the judge's criteria answers, when the prompt asked for them; never part of the score. */
+    criteriaAnswers?: CriteriaAnswer[];
+    /** ggui#1687 — what the answer's criteria array held and what the parser dropped; set whenever the answer parsed. */
+    criteriaCounts?: CriteriaAnswerCounts;
 }
 /**
  * Quality metadata attached to every generated component
@@ -103,4 +128,4 @@ interface EvaluationContext {
     themeTokens: string;
 }
 
-export { DEFAULT_EVAL_MAX_ROUNDS, type DimensionScores, type EvaluationConfig, type EvaluationContext, type EvaluationIssue, type EvaluationResult, MAX_EVAL_ROUNDS_HARD_LIMIT, type QualityMetadata };
+export { type CriteriaAnswer, DEFAULT_EVAL_MAX_ROUNDS, type DimensionScores, type EvaluationConfig, type EvaluationContext, type EvaluationIssue, type EvaluationResult, MAX_EVAL_ROUNDS_HARD_LIMIT, type QualityMetadata };

package/dist/evaluation/types.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/evaluation/types.ts"],"names":[],"mappings":";AAyEO,IAAM,0BAAA,GAA6B;AAGnC,IAAM,uBAAA,GAA0B","file":"types.js","sourcesContent":["// packages/ui-gen/src/evaluation/types.ts\n\n/**\n * Per-dimension quality scores from the evaluator (0-100 each).\n */\nexport interface DimensionScores {\n  /** How well the component fulfills the original prompt requirements */\n  completeness: number;\n  /** Visual quality: layout, typography, spacing, color usage */\n  visualPolish: number;\n  /** Interaction quality: hover states, transitions, form validation */\n  interactivity: number;\n  /** Accessibility: ARIA labels, keyboard navigation, contrast ratios */\n  accessibility: number;\n  /** Code quality: TypeScript types, clean structure, no anti-patterns */\n  codeQuality: number;\n}\n\n/**\n * A specific issue found during evaluation.\n */\nexport interface EvaluationIssue {\n  /** Which quality dimension this issue belongs to */\n  dimension: string;\n  /** Human-readable description of what is wrong */\n  description: string;\n  /** Severity level: critical blocks shipping, major degrades quality, minor is a polish issue */\n  severity: 'critical' | 'major' | 'minor';\n  /** Suggested fix for the agent to apply */\n  fix: string;\n}\n\n/**\n * Result of a single evaluation round.\n */\nexport interface EvaluationResult {\n  /** Whether the component passed the quality threshold */\n  passed: boolean;\n  /** Weighted average score across all dimensions (0-100) */\n  finalScore: number;\n  /** Per-dimension score breakdown */\n  dimensions: DimensionScores;\n  /** Specific issues identified in this round */\n  issues: EvaluationIssue[];\n  /** Free-form critique text from the evaluator LLM */\n  critique?: string;\n  /** Input tokens consumed by the evaluation LLM call */\n  inputTokens?: number;\n  /** Output tokens consumed by the evaluation LLM call */\n  outputTokens?: number;\n}\n\n/**\n * Quality metadata attached to every generated component\n */\nexport interface QualityMetadata {\n  /** Number of evaluation rounds run (0 = strict/no eval) */\n  evaluationRounds: number;\n  /** Final average score 0-100 */\n  finalScore: number;\n  /** Per-dimension breakdown */\n  dimensions: DimensionScores;\n  /** Whether the component passed (score >= threshold) */\n  passed: boolean;\n  /** Model used for evaluation */\n  evaluatorModel: string;\n  /** Total evaluation loop time in ms */\n  evaluationTimeMs: number;\n  /** S3 keys for screenshots (creative only) */\n  screenshots?: string[];\n}\n\n/** Hard ceiling for evaluation rounds — never exceed this regardless of config */\nexport const MAX_EVAL_ROUNDS_HARD_LIMIT = 10;\n\n/** Default (soft) max rounds for the balanced strategy */\nexport const DEFAULT_EVAL_MAX_ROUNDS = 10;\n\n/**\n * Configuration for the evaluation loop\n */\nexport interface EvaluationConfig {\n  /** Whether evaluation is enabled */\n  enabled: boolean;\n  /** Model to use for evaluation (default: provider's default) */\n  model?: string;\n  /** LLM provider for evaluation (default: same as generation provider) */\n  provider?: 'claude' | 'openai' | 'google' | 'openrouter';\n  /** Minimum average score to pass (default: 70) */\n  passThreshold: number;\n  /** Maximum evaluation rounds (soft max: 3, hard limit: 5) */\n  maxRounds?: number;\n  /** Budget cap per evaluation round in USD */\n  maxBudgetPerEval?: number;\n  /** Budget cap per fix round in USD */\n  maxBudgetPerFix?: number;\n}\n\n/**\n * Context passed to the evaluator\n */\nexport interface EvaluationContext {\n  /** TSX source code written by the generator */\n  sourceCode: string;\n  /** Compiled JavaScript output */\n  compiledCode: string;\n  /** Original user prompt */\n  originalPrompt: string;\n  /** DESIGN.md content (if present) */\n  designContext?: string;\n  /** CSS variable tokens for the app theme */\n  themeTokens: string;\n}\n"]}
+{"version":3,"sources":["../../src/evaluation/types.ts"],"names":[],"mappings":";AA4FO,IAAM,0BAAA,GAA6B;AAGnC,IAAM,uBAAA,GAA0B","file":"types.js","sourcesContent":["import type { CriteriaAnswerCounts } from './types-public.js';\n// packages/ui-gen/src/evaluation/types.ts\n\n/**\n * Per-dimension quality scores from the evaluator (0-100 each).\n */\nexport interface DimensionScores {\n  /** How well the component fulfills the original prompt requirements */\n  completeness: number;\n  /** Visual quality: layout, typography, spacing, color usage */\n  visualPolish: number;\n  /** Interaction quality: hover states, transitions, form validation */\n  interactivity: number;\n  /** Accessibility: ARIA labels, keyboard navigation, contrast ratios */\n  accessibility: number;\n  /** Code quality: TypeScript types, clean structure, no anti-patterns */\n  codeQuality: number;\n}\n\n/**\n * A specific issue found during evaluation.\n */\nexport interface EvaluationIssue {\n  /** Which quality dimension this issue belongs to */\n  dimension: string;\n  /** Human-readable description of what is wrong */\n  description: string;\n  /** Severity level: critical blocks shipping, major degrades quality, minor is a polish issue */\n  severity: 'critical' | 'major' | 'minor';\n  /** Suggested fix for the agent to apply */\n  fix: string;\n  /**\n   * Who produced it, on the visual leg (ggui#1545): `judge` for the model's own findings, `instrument` for a\n   * deterministic measurement (fit, overflow, blank). Absent on issues from other evaluators. The loop's\n   * `visualEvaluation.actOn` acts on `judge` findings only, so a new instrument check never joins the treatment\n   * by accident: it has to say it is a judge finding, and absence reads as \"not acted on\".\n   */\n  origin?: 'judge' | 'instrument';\n}\n\n/**\n * Result of a single evaluation round.\n */\n/** One judge answer to a criterion (ggui#1436), as parsed — before the K-majority. */\nexport interface CriteriaAnswer {\n  id: string;\n  verdict: 'pass' | 'fail' | 'n/a';\n  evidence: string;\n}\n\nexport interface EvaluationResult {\n  /** Whether the component passed the quality threshold */\n  passed: boolean;\n  /** Weighted average score across all dimensions (0-100) */\n  finalScore: number;\n  /** Per-dimension score breakdown */\n  dimensions: DimensionScores;\n  /** Specific issues identified in this round */\n  issues: EvaluationIssue[];\n  /** Free-form critique text from the evaluator LLM */\n  critique?: string;\n  /** Input tokens consumed by the evaluation LLM call */\n  inputTokens?: number;\n  /** Output tokens consumed by the evaluation LLM call */\n  outputTokens?: number;\n  /** ggui#1436 — the judge's criteria answers, when the prompt asked for them; never part of the score. */\n  criteriaAnswers?: CriteriaAnswer[];\n  /** ggui#1687 — what the answer's criteria array held and what the parser dropped; set whenever the answer parsed. */\n  criteriaCounts?: CriteriaAnswerCounts;\n}\n\n/**\n * Quality metadata attached to every generated component\n */\nexport interface QualityMetadata {\n  /** Number of evaluation rounds run (0 = strict/no eval) */\n  evaluationRounds: number;\n  /** Final average score 0-100 */\n  finalScore: number;\n  /** Per-dimension breakdown */\n  dimensions: DimensionScores;\n  /** Whether the component passed (score >= threshold) */\n  passed: boolean;\n  /** Model used for evaluation */\n  evaluatorModel: string;\n  /** Total evaluation loop time in ms */\n  evaluationTimeMs: number;\n  /** S3 keys for screenshots (creative only) */\n  screenshots?: string[];\n}\n\n/** Hard ceiling for evaluation rounds — never exceed this regardless of config */\nexport const MAX_EVAL_ROUNDS_HARD_LIMIT = 10;\n\n/** Default (soft) max rounds for the balanced strategy */\nexport const DEFAULT_EVAL_MAX_ROUNDS = 10;\n\n/**\n * Configuration for the evaluation loop\n */\nexport interface EvaluationConfig {\n  /** Whether evaluation is enabled */\n  enabled: boolean;\n  /** Model to use for evaluation (default: provider's default) */\n  model?: string;\n  /** LLM provider for evaluation (default: same as generation provider) */\n  provider?: 'claude' | 'openai' | 'google' | 'openrouter';\n  /** Minimum average score to pass (default: 70) */\n  passThreshold: number;\n  /** Maximum evaluation rounds (soft max: 3, hard limit: 5) */\n  maxRounds?: number;\n  /** Budget cap per evaluation round in USD */\n  maxBudgetPerEval?: number;\n  /** Budget cap per fix round in USD */\n  maxBudgetPerFix?: number;\n}\n\n/**\n * Context passed to the evaluator\n */\nexport interface EvaluationContext {\n  /** TSX source code written by the generator */\n  sourceCode: string;\n  /** Compiled JavaScript output */\n  compiledCode: string;\n  /** Original user prompt */\n  originalPrompt: string;\n  /** DESIGN.md content (if present) */\n  designContext?: string;\n  /** CSS variable tokens for the app theme */\n  themeTokens: string;\n}\n"]}

package/dist/fragments/index.d.ts

@@ -1,6 +1,6 @@
-import { A as AxisKey, H as HarnessFragment } from '../types-BOvHNG7K.js';
-export { a as CacheTier, C as ComposedHarness } from '../types-BOvHNG7K.js';
-import '../axes-CzLEMDeB.js';
+import { A as AxisKey, H as HarnessFragment } from '../types-Dh1s4K9j.js';
+export { a as CacheTier, C as ComposedHarness } from '../types-Dh1s4K9j.js';
+import '../axes-CKVCzR_x.js';
 
 declare const FRAGMENT_REGISTRY: Record<AxisKey, Record<string, HarnessFragment>>;
 declare function lookupFragment(axis: AxisKey, value: string): HarnessFragment | undefined;

package/dist/{generate-9IeNT4ZY.d.ts → generate-tSO_shQ0.d.ts}

@@ -1,5 +1,5 @@
 import { DataContract, GadgetDescriptor } from '@ggui-ai/protocol';
-import { D as DesignMode } from './design-mode-DLufhrNp.js';
+import { D as DesignMode } from './design-mode-DGWN_Gkm.js';
 
 /** Shell layout modes supported by the boilerplate templates. */
 type ShellType = "chat" | "fullscreen" | "spatial";

package/dist/harness/check/runtime-render/index.d.ts

@@ -1,15 +1,15 @@
-import { b as RenderCheckHostBounds } from '../../../adapter-Du7P_5Fx.js';
-export { D as DEFAULT_RUNTIME_RENDER_CHECK, c as RENDER_CHECK_KINDS, d as RenderCheckHostOptions, a as RenderCheckIssue, e as RenderCheckKind, f as RenderCheckResult, R as RuntimeRenderProbeConfig, g as classifyRenderCrashFix, h as createRuntimeRenderCheck, i as isRecoverableRenderCrash, p as parseRuntimeSubcategory, r as runRenderCheck } from '../../../adapter-Du7P_5Fx.js';
+import { b as RenderCheckHostBounds } from '../../../adapter-BJwn74rx.js';
+export { D as DEFAULT_RUNTIME_RENDER_CHECK, c as RENDER_CHECK_KINDS, d as RenderCheckHostOptions, a as RenderCheckIssue, e as RenderCheckKind, f as RenderCheckResult, R as RuntimeRenderProbeConfig, g as classifyRenderCrashFix, h as createRuntimeRenderCheck, i as isRecoverableRenderCrash, p as parseRuntimeSubcategory, r as runRenderCheck } from '../../../adapter-BJwn74rx.js';
 import { WireConfig } from '@ggui-ai/wire';
 import { JsonObject, DataContract } from '@ggui-ai/protocol';
 import '../../types-public.js';
 import '../../../styling-profile-CxfvV48g.js';
-import '../../../axes-CzLEMDeB.js';
-import '../../../types-public-Bk6d1WLN.js';
-import '@ggui-ai/mcp-server-core';
-import '../../../design-mode-DLufhrNp.js';
-import '../../../types-BOvHNG7K.js';
+import '../../../axes-CKVCzR_x.js';
+import '../../../types-public-B7wn4ce5.js';
 import '../../../llm.js';
+import '@ggui-ai/mcp-server-core';
+import '../../../design-mode-DGWN_Gkm.js';
+import '../../../types-Dh1s4K9j.js';
 import '../../../policy.js';
 
 interface ActionFiredEvent {