@ggui-ai/ui-gen 0.28.0 → 0.29.0

package/dist/evaluation/index.d.ts

@@ -1,16 +1,16 @@
-import { r as CriteriaSelection, h as CriteriaContext, C as CriteriaAnswerCounts, s as VisionFinishReason, k as CriteriaBlock } from '../types-public-h8IMnYuh.js';
-export { A as AxisCheck, a as AxisCheckInput, t as CRITERIA, n as CanvasHostPresentation, j as CanvasJudgeRecord, l as CanvasPresentationOutcome, u as CanvasVisualSummary, v as ContractFeedbackRecord, w as CriterionCoverage, x as CriterionRunStatus, D as DEFAULT_QUALITY_CONFIG, y as EvalCategory, z as EvalCriterion, E as EvalIssue, B as EvalOutcome, e as EvalResult, q as EvalTier, H as HostInlineFrame, F as HostPresentationIgnoredReason, G as LLM_EVAL_STATIC_CRITERIA, I as Priority, Q as QualityConfig, K as QualityMode, o as VisualEvalSummary, M as buildCodingCriteriaSummary, N as criteriaFor, O as getActionableIssues, S as getCriteriaByPriority, T as getCriterionById, U as getLLMCriteria, W as isBlocked, m as matches, X as notApplicableCoverage, Y as priorityForIssue } from '../types-public-h8IMnYuh.js';
-import * as _anthropic_ai_claude_agent_sdk from '@anthropic-ai/claude-agent-sdk';
-import { McpServerConfig } from '@anthropic-ai/claude-agent-sdk';
-import { EvaluationContext, EvaluationConfig, QualityMetadata, EvaluationResult, EvaluationIssue, CriteriaAnswer } from './types.js';
-export { DEFAULT_EVAL_MAX_ROUNDS, DimensionScores, MAX_EVAL_ROUNDS_HARD_LIMIT } from './types.js';
-export { R as RunAxisChecksInput, r as runAxisChecks } from '../dispatch-B07TI30d.js';
-import { E as EdgeVerdict, a as EdgeBlock, C as CriteriaCallTokens } from '../visual-evaluator-C2strxqD.js';
-export { b as CARD_HEIGHT_EXPRESSION, c as CONTENT_HEIGHT_EXPRESSION, d as CanvasFitPolicy, e as CanvasVisualResult, f as CaptureMode, g as ChromiumProvider, h as CriteriaContextInput, i as CriteriaContextSources, j as EDGE_TOLERANCE_PX, k as EdgeOffset, H as HOST_COLOUR_PATTERN, l as HOST_FRAME_BOUNDS, J as JUDGE_GROUND_MARGIN_PX, m as JUDGE_INK_INSET_PX, n as JudgeDesignIdentity, N as NaturalCaptureOptions, S as ScreenshotAttempt, o as ScreenshotBrowser, p as ScreenshotClip, q as ScreenshotDeps, r as ScreenshotLauncher, s as ScreenshotPage, t as StoredCaptureInput, u as StoredCaptureOutcome, v as StoredCaptureVerdict, w as VISUAL_EVAL_PROMPT, x as VISUAL_JUDGE_MAX_OUTPUT_TOKENS, y as VISUAL_JUDGE_PROMPT_DIGEST, z as VISUAL_JUDGE_PROMPT_VERSION, A as VisualEvalConfig, B as VisualEvalContext, D as VisualEvalDeps, F as VisualEvalDetailed, G as VisualEvalOutcome, I as VisualEvaluationResult, V as VisualLegTokens, K as callMultimodalLLM, L as canvasBlankIssue, M as canvasFitPolicy, O as canvasOverflowIssue, P as captureScreenshot, Q as captureScreenshotDetailed, R as chromaOfCssTokens, T as criteriaContextFor, U as designTreeSha256, W as judgeDesignIdentity, X as judgeEdges, Y as judgeStoredCapture, Z as resetJudgeDesignIdentityCache, _ as resolveDesignPackageDir, $ as resolveLaunchOptions, a0 as runVisualEval, a1 as runVisualEvaluation, a2 as runVisualEvaluationDetailed, a3 as summarizeVisualResult } from '../visual-evaluator-C2strxqD.js';
-export { b as CANVAS_BREAKPOINTS, c as CANVAS_CLASSES, d as CANVAS_DESCRIPTORS, e as CANVAS_VIEWPORTS, C as CanvasClass, f as CanvasDescriptor, g as CanvasDisplayMode, a as CanvasViewport, j as canvasDisplayModeForShell, k as canvasForRendering, l as canvasForViewportWidth, m as describeCanvas } from '../design-mode-DGWN_Gkm.js';
+import { t as CriteriaSelection, h as CriteriaContext, C as CriteriaAnswerCounts, u as VisionFinishReason, k as CriteriaBlock } from '../types-public-hYlrerZh.js';
+export { A as AxisCheck, a as AxisCheckInput, v as CRITERIA, n as CanvasHostPresentation, j as CanvasJudgeRecord, l as CanvasPresentationOutcome, w as CanvasVisualSummary, x as ContractFeedbackRecord, y as CriterionCoverage, z as CriterionRunStatus, D as DEFAULT_QUALITY_CONFIG, B as EvalCategory, F as EvalCriterion, E as EvalIssue, G as EvalOutcome, e as EvalResult, q as EvalTier, H as HostInlineFrame, I as HostPresentationIgnoredReason, K as LLM_EVAL_STATIC_CRITERIA, M as Priority, Q as QualityConfig, N as QualityMode, o as VisualEvalSummary, O as buildCodingCriteriaSummary, S as criteriaFor, T as getActionableIssues, U as getCriteriaByPriority, W as getCriterionById, X as getLLMCriteria, Y as isBlocked, m as matches, Z as notApplicableCoverage, _ as priorityForIssue } from '../types-public-hYlrerZh.js';
+import { EvaluationContext, EvaluationConfig, EvaluationResult, EvaluationIssue, CriteriaAnswer } from './types.js';
+export { DEFAULT_EVAL_MAX_ROUNDS, DimensionScores, MAX_EVAL_ROUNDS_HARD_LIMIT, QualityMetadata } from './types.js';
+export { R as RunAxisChecksInput, r as runAxisChecks } from '../dispatch-8LzTddKw.js';
+import { E as EdgeVerdict, a as EdgeBlock, C as CriteriaCallTokens } from '../visual-evaluator-KskajxD5.js';
+export { b as CARD_HEIGHT_EXPRESSION, c as CONTENT_HEIGHT_EXPRESSION, d as CanvasFitPolicy, e as CanvasVisualResult, f as CaptureMode, g as ChromiumProvider, h as CriteriaContextInput, i as CriteriaContextSources, j as EDGE_TOLERANCE_PX, k as EdgeOffset, H as HOST_COLOUR_PATTERN, l as HOST_FRAME_BOUNDS, J as JUDGE_GROUND_MARGIN_PX, m as JUDGE_INK_INSET_PX, n as JudgeDesignIdentity, N as NaturalCaptureOptions, S as ScreenshotAttempt, o as ScreenshotBrowser, p as ScreenshotClip, q as ScreenshotDeps, r as ScreenshotLauncher, s as ScreenshotPage, t as StoredCaptureInput, u as StoredCaptureOutcome, v as StoredCaptureVerdict, w as VISUAL_EVAL_PROMPT, x as VISUAL_JUDGE_MAX_OUTPUT_TOKENS, y as VISUAL_JUDGE_PROMPT_DIGEST, z as VISUAL_JUDGE_PROMPT_VERSION, A as VisualEvalConfig, B as VisualEvalContext, D as VisualEvalDeps, F as VisualEvalDetailed, G as VisualEvalOutcome, I as VisualEvaluationResult, V as VisualLegTokens, K as callMultimodalLLM, L as canvasBlankIssue, M as canvasFitPolicy, O as canvasOverflowIssue, P as captureScreenshot, Q as captureScreenshotDetailed, R as chromaOfCssTokens, T as criteriaContextFor, U as designTreeSha256, W as judgeDesignIdentity, X as judgeEdges, Y as judgeStoredCapture, Z as resetJudgeDesignIdentityCache, _ as resolveDesignPackageDir, $ as resolveLaunchOptions, a0 as runVisualEval, a1 as runVisualEvaluation, a2 as runVisualEvaluationDetailed, a3 as summarizeVisualResult } from '../visual-evaluator-KskajxD5.js';
+export { a as CANVAS_BREAKPOINTS, b as CANVAS_DESCRIPTORS, c as CANVAS_VIEWPORTS, d as CanvasDescriptor, e as CanvasDisplayMode, C as CanvasViewport, h as canvasDisplayModeForShell, i as canvasForRendering, j as canvasForViewportWidth, k as describeCanvas } from '../design-mode-XyOQmzE0.js';
 import { AppTheme } from '@ggui-ai/protocol';
-import { C as CriteriaBank } from '../bank-Ctdp3FUA.js';
-export { B as BankRow, a as CriteriaScope, b as bankRows, c as criteriaBankSchema, p as parseCriteriaBank } from '../bank-Ctdp3FUA.js';
+import { C as CriteriaBank } from '../bank-LWaiAtaA.js';
+export { B as BankRow, a as CriteriaScope, b as bankRows, c as criteriaBankSchema, p as parseCriteriaBank } from '../bank-LWaiAtaA.js';
+import { MountFit } from '@ggui-ai/design/rendering';
+export { CANVAS_CLASSES, CanvasClass } from '@ggui-ai/design/rendering';
 import '../llm.js';
 import '@ggui-ai/mcp-server-core';
 import '../axes-CKVCzR_x.js';
@@ -19,73 +19,42 @@ import '../styling-profile-CxfvV48g.js';
 import 'puppeteer-core';
 import 'zod';
 
-/**
- * Options for running the evaluation loop.
- *
- * Controls the evaluate-fix-re-evaluate cycle that improves generated
- * component quality until it passes or the round limit is reached.
- */
-interface EvaluationLoopOptions {
-    /** Session ID of the generator to resume for fixes */
-    generatorSessionId: string;
-    /** Evaluation context (code, prompt, design, theme) */
-    context: EvaluationContext;
-    /** Evaluation configuration (thresholds, budgets, round limits) */
-    config: EvaluationConfig;
-    /** Progress callback for evaluating/fixing status updates */
-    onProgress?: (event: {
-        type: 'evaluating' | 'fixing';
-        round: number;
-    }) => void;
-    /** Generator context to pass when resuming the session for fixes */
-    generatorOptions?: {
-        /** Working directory for the generator session */
-        cwd?: string;
-        /** MCP servers available to the fix agent */
-        mcpServers?: Record<string, McpServerConfig>;
-        /** Tool names to approve. Absent → every tool of every server in `mcpServers` (ggui#1714). */
-        allowedTools?: string[];
-        /**
-         * CLI built-in tools to offer the fix agent (e.g. `['Write']`). Default: none.
-         * An explicit opt-in: a built-in that writes or executes acts on the host (ggui#1714).
-         */
-        builtinTools?: readonly string[];
-        /** LLM model for fix rounds */
-        model?: string;
-        /** Environment variables (includes BYOK credentials) */
-        env?: Record<string, string>;
-        /** Stderr capture callback for debugging */
-        stderr?: (data: string) => void;
-    };
+/** A channel difference at or below this is the same colour (anti-aliasing and compositing noise). */
+declare const INK_TOLERANCE = 8;
+interface PixelRegion {
+    readonly x: number;
+    readonly y: number;
+    readonly width: number;
+    readonly height: number;
 }
-/**
- * Result of the evaluation loop.
- *
- * Contains the final (possibly fixed) code, quality scores, and
- * the evaluation history from each round.
- */
-interface EvaluationLoopResult {
-    /** Final compiled code (may be updated by fix rounds) */
-    finalCode: string;
-    /** Final source code (may be updated by fix rounds) */
-    finalSourceCode?: string;
-    /** Quality metadata for the generation result */
-    qualityMetadata: QualityMetadata;
-    /** All evaluation results from each round */
-    evaluationResults: EvaluationResult[];
+interface InkExtent {
+    /** `(lastInkRow + 1) / region.height`, rounded to 3 places; `0` when nothing differs from the ground. */
+    readonly ratio: number;
+    /** The last region row (0-based from the region's top) carrying ink; `null` when there is none — the blank. */
+    readonly lastInkRow: number | null;
+    /** The region read, in image px. */
+    readonly region: PixelRegion;
+    /** The dominant colour the ink was read against, `[r, g, b]`. */
+    readonly ground: readonly [number, number, number];
+}
+/** Why a capture could not be read — reported, never mistaken for a blank. */
+interface InkUnreadable {
+    readonly reason: string;
 }
+interface DecodedImage {
+    readonly width: number;
+    readonly height: number;
+    readonly channels: 3 | 4;
+    /** Row-major, `channels` bytes per pixel, scanline filters undone. */
+    readonly pixels: Uint8Array;
+}
+/** The PNG subset the judge's browser writes: 8-bit RGB (colour type 2) or RGBA (6), not interlaced. */
+declare function decodePng(png: Uint8Array): DecodedImage | InkUnreadable;
 /**
- * Run the evaluation loop: evaluate -> fix -> re-evaluate (up to maxRounds).
- *
- * If the first evaluation passes, returns immediately.
- * If it fails, resumes the generator session with critique feedback,
- * captures the fixed code, and re-evaluates. Repeats until the score
- * passes or the round limit is reached.
- *
- * @param options - Evaluation loop configuration and callbacks
- * @returns The final code, quality metadata, and evaluation history
+ * The ink extent of a capture, read `insetPx` in from every edge — the judge's inset skips a host
+ * panel's gap, rounded corners and hairline ({@link JUDGE_INK_INSET_PX}); `0` reads the whole image.
  */
-declare function runEvaluationLoop(options: EvaluationLoopOptions): Promise<EvaluationLoopResult>;
+declare function readInkExtent(png: Uint8Array, insetPx?: number): InkExtent | InkUnreadable;
 
 /**
  * Run a single evaluation round using any LLM provider.
@@ -113,14 +82,6 @@ interface EvaluateScoreInput {
  * unit-tested directly without going through MCP protocol.
  */
 declare function computeEvaluationScore(args: EvaluateScoreInput, passThreshold: number): EvaluationResult;
-/**
- * Create the evaluation MCP server with the evaluate_score tool.
- *
- * The evaluator LLM provides qualitative scores per dimension.
- * This tool handles the arithmetic (average, pass/fail) so the LLM
- * doesn't need to do math.
- */
-declare function createEvaluationToolsServer(passThreshold?: number): _anthropic_ai_claude_agent_sdk.McpSdkServerConfigWithInstance;
 
 /**
  * System prompt for the evaluator agent.
@@ -195,43 +156,6 @@ declare function extractCompiledCode(messages: SdkMessage[]): string | undefined
  */
 declare function extractSourceCode(messages: SdkMessage[]): string | undefined;
 
-/** A channel difference at or below this is the same colour (anti-aliasing and compositing noise). */
-declare const INK_TOLERANCE = 8;
-interface PixelRegion {
-    readonly x: number;
-    readonly y: number;
-    readonly width: number;
-    readonly height: number;
-}
-interface InkExtent {
-    /** `(lastInkRow + 1) / region.height`, rounded to 3 places; `0` when nothing differs from the ground. */
-    readonly ratio: number;
-    /** The last region row (0-based from the region's top) carrying ink; `null` when there is none — the blank. */
-    readonly lastInkRow: number | null;
-    /** The region read, in image px. */
-    readonly region: PixelRegion;
-    /** The dominant colour the ink was read against, `[r, g, b]`. */
-    readonly ground: readonly [number, number, number];
-}
-/** Why a capture could not be read — reported, never mistaken for a blank. */
-interface InkUnreadable {
-    readonly reason: string;
-}
-interface DecodedImage {
-    readonly width: number;
-    readonly height: number;
-    readonly channels: 3 | 4;
-    /** Row-major, `channels` bytes per pixel, scanline filters undone. */
-    readonly pixels: Uint8Array;
-}
-/** The PNG subset the judge's browser writes: 8-bit RGB (colour type 2) or RGBA (6), not interlaced. */
-declare function decodePng(png: Uint8Array): DecodedImage | InkUnreadable;
-/**
- * The ink extent of a capture, read `insetPx` in from every edge — the judge's inset skips a host
- * panel's gap, rounded corners and hairline ({@link JUDGE_INK_INSET_PX}); `0` reads the whole image.
- */
-declare function readInkExtent(png: Uint8Array, insetPx?: number): InkExtent | InkUnreadable;
-
 /**
  * The stylesheet a visual judge renders under for an app: the app's theme
  * composed the way the runtime composes it (`composeThemeCss`, layer `page`
@@ -260,17 +184,6 @@ interface CriteriaSelectionResult {
  */
 declare function selectCriteria(bank: CriteriaBank, ctx: CriteriaContext): CriteriaSelectionResult;
 
-/**
- * From K judge answers and the capture's measurements to one `CriteriaBlock` — ggui#1436.
- *
- * The judge is asked every `evaluation: judge` row (this block IS their implementation; the row's
- * `status` is carried on the verdict as the author wrote it). An `instrument` row is answered by a
- * measurement the capture already ran when its status is `live` and a binding exists; otherwise it
- * reads `n/a` with the reason from `status`. A `human` row reads `n/a` for the judge. K-majority per
- * id; a tie or an unanswered id reads `n/a` (the CTO's A2) — a `must` at `n/a` is never a pass when
- * a gate later reads the set.
- */
-
 /** The instruments the capture runs today, by the bank ids they answer (v1 binding; a bank field may replace it). */
 declare const INSTRUMENT_BY_ID: Readonly<Record<string, 'fit' | 'fill'>>;
 /**
@@ -288,6 +201,8 @@ interface CriteriaMeasurements {
      * `null` when the probe ran and failed.
      */
     readonly edge?: EdgeVerdict | null;
+    /** #1765 — the v2 edge rule on the same capture, read beside v1: absent / `null` as for `edge`. */
+    readonly edgeV2?: EdgeVerdict | null;
     /** ggui#1711 — the capture includes a state after an action ran. Absent = no (every capture today). */
     readonly postActionCapture?: boolean;
 }
@@ -316,6 +231,8 @@ interface CriteriaJudgeFrame {
     readonly height: number;
     /** ggui#1492 — the host whose presentation framed this canvas (e.g. "a chat widget, inline, reference theme"); absent ⇒ the judge's own stand-in. */
     readonly hostLabel?: string;
+    /** How the card was composed on this frame (`resolveMountFit`, the one rule the runtime serves by); the sentence follows it. */
+    readonly fit: MountFit;
 }
 /** What the block carries besides the selected rows: the frame's box, and the props the card was rendered with. */
 interface CriteriaJudgeInputs {
@@ -401,4 +318,4 @@ interface JudgedTokens {
 /** A judged frame's spend over every leg: the scoring calls plus the report-only criteria call, when one ran. */
 declare function judgedFrameTokens(result: JudgedTokens): JudgeSpend;
 
-export { CRITERIA_SELECTOR_VERSION, CriteriaBank, CriteriaCallTokens, type CriteriaMeasurements, type CriteriaSelectionResult, type DecodedImage, EDGE_BLOCKS_EXPRESSION, EdgeBlock, type EdgeProbe, EdgeVerdict, type EvaluateScoreInput, EvaluationConfig, EvaluationContext, EvaluationIssue, type EvaluationLoopOptions, type EvaluationLoopResult, EvaluationResult, INK_TOLERANCE, INSTRUMENT_BESIDE_JUDGE, INSTRUMENT_BY_ID, type InkExtent, type InkUnreadable, type JudgeSpend, type JudgedTokens, type PixelRegion, QualityMetadata, type SdkMessage, buildCriteriaJudgeBlock, buildFixPrompt, canonicalCriteriaContext, computeEvaluationScore, createEvaluationToolsServer, criteriaSetIdFor, cssTokensForAppTheme, decodePng, extractCompiledCode, extractCompiledCodeFromMessage, extractEvalResult, extractSourceCode, extractSourceCodeFromMessage, extractToolResultTexts, getEvaluatorSystemPrompt, judgedFrameTokens, parseEdgeProbe, readInkExtent, resolveCriteriaBlock, runEvaluation, runEvaluationLoop, selectCriteria };
+export { CRITERIA_SELECTOR_VERSION, CriteriaBank, CriteriaCallTokens, type CriteriaMeasurements, type CriteriaSelectionResult, type DecodedImage, EDGE_BLOCKS_EXPRESSION, EdgeBlock, type EdgeProbe, EdgeVerdict, type EvaluateScoreInput, EvaluationConfig, EvaluationContext, EvaluationIssue, EvaluationResult, INK_TOLERANCE, INSTRUMENT_BESIDE_JUDGE, INSTRUMENT_BY_ID, type InkExtent, type InkUnreadable, type JudgeSpend, type JudgedTokens, type PixelRegion, type SdkMessage, buildCriteriaJudgeBlock, buildFixPrompt, canonicalCriteriaContext, computeEvaluationScore, criteriaSetIdFor, cssTokensForAppTheme, decodePng, extractCompiledCode, extractCompiledCodeFromMessage, extractEvalResult, extractSourceCode, extractSourceCodeFromMessage, extractToolResultTexts, getEvaluatorSystemPrompt, judgedFrameTokens, parseEdgeProbe, readInkExtent, resolveCriteriaBlock, runEvaluation, selectCriteria };