@ggui-ai/ui-gen 0.26.0 → 0.27.0

package/dist/evaluation/axis-checks/registry.d.ts

@@ -1,4 +1,4 @@
-import { A as AxisCheck } from '../../types-public-B7wn4ce5.js';
+import { A as AxisCheck } from '../../types-public-h8IMnYuh.js';
 import '@ggui-ai/protocol';
 import '../../llm.js';
 import '@ggui-ai/mcp-server-core';

package/dist/evaluation/index.d.ts

@@ -1,12 +1,12 @@
-import { r as CriteriaSelection, h as CriteriaContext, C as CriteriaAnswerCounts, s as VisionFinishReason, k as CriteriaBlock } from '../types-public-B7wn4ce5.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-B7wn4ce5.js';
+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-BUoXRj_r.js';
-import { E as EdgeVerdict, a as EdgeBlock, C as CriteriaCallTokens } from '../visual-evaluator-C9L-o6ap.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-C9L-o6ap.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 { AppTheme } from '@ggui-ai/protocol';
 import { C as CriteriaBank } from '../bank-Ctdp3FUA.js';
@@ -43,8 +43,13 @@ interface EvaluationLoopOptions {
         cwd?: string;
         /** MCP servers available to the fix agent */
         mcpServers?: Record<string, McpServerConfig>;
-        /** Tools the fix agent is allowed to call */
+        /** 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) */
@@ -167,13 +172,14 @@ declare function extractEvalResult(messages: SdkMessage[]): EvaluationResult | u
  */
 declare function extractCompiledCodeFromMessage(message: SdkMessage): string | undefined;
 /**
- * Extract source code from an assistant's Write tool_use message.
+ * Extract source code from an assistant message's tool_use blocks.
  *
- * Looks for `tool_use` blocks with `name === 'Write'` and returns
- * the `input.content` string. If multiple Write calls exist, returns
- * the last one (the final version).
+ * Reads the bridged compile tool (`*__compile_component`, `input.code`)
+ * and the CLI's `Write` (`input.content`, offered only by opt-in), by the
+ * one rule in `adapters/claude/source-capture.ts` (ggui#1714). If several
+ * calls carry source, returns the last one (the final version).
  *
- * Used by: loop.ts, generator.ts
+ * Used by: loop.ts
  */
 declare function extractSourceCodeFromMessage(message: SdkMessage): string | undefined;
 /**
@@ -183,7 +189,7 @@ declare function extractSourceCodeFromMessage(message: SdkMessage): string | und
  */
 declare function extractCompiledCode(messages: SdkMessage[]): string | undefined;
 /**
- * Scan all messages for the last sourceCode from Write tool_use.
+ * Scan all messages for the last sourceCode a tool_use carried.
  *
  * Convenience wrapper for tests and one-shot extraction.
  */
@@ -282,6 +288,8 @@ interface CriteriaMeasurements {
      * `null` when the probe ran and failed.
      */
     readonly edge?: EdgeVerdict | null;
+    /** ggui#1711 — the capture includes a state after an action ran. Absent = no (every capture today). */
+    readonly postActionCapture?: boolean;
 }
 declare function resolveCriteriaBlock(args: {
     readonly bank: CriteriaBank;
@@ -314,6 +322,8 @@ interface CriteriaJudgeInputs {
     readonly frame?: CriteriaJudgeFrame;
     /** JSON text of the props the card was rendered with; bounded by {@link CRITERIA_PROPS_MAX_CHARS}. */
     readonly propsJson?: string;
+    /** ggui#1711 — the capture includes a post-action state; absent = no, and rows needing one are not asked. */
+    readonly postActionCapture?: boolean;
 }
 /** The judge's prompt block for the selected `judge` rows; empty when none is selected. */
 declare function buildCriteriaJudgeBlock(bank: CriteriaBank, selected: CriteriaSelectionResult, inputs?: CriteriaJudgeInputs): string;

package/dist/evaluation/index.js

@@ -415,6 +415,13 @@ function buildCodingCriteriaSummary(designMode = "constrained") {
   }
   return lines.join("\n");
 }
+
+// src/adapters/claude/sdk-tool-pins.ts
+function sdkToolPins(input) {
+  const builtins = [...input.builtinTools ?? []];
+  const approved = input.allowedTools !== void 0 ? [...input.allowedTools] : Object.keys(input.mcpServers ?? {}).map((name) => `mcp__${name}`);
+  return { tools: builtins, settingSources: [], allowedTools: [...approved, ...builtins] };
+}
 function createAnthropicClient(rawKey) {
   if (rawKey === void 0) {
     return new Anthropic({
@@ -2347,6 +2354,16 @@ function getDefaultEvalModel(provider) {
   }
 }
 
+// src/adapters/claude/source-capture.ts
+var COMPILE_TOOL_SUFFIX = "__compile_component";
+function sourceCodeFromToolUse(name, input) {
+  if (typeof name !== "string") return void 0;
+  const field2 = name === "Write" ? "content" : name.endsWith(COMPILE_TOOL_SUFFIX) ? "code" : void 0;
+  if (field2 === void 0 || typeof input !== "object" || input === null) return void 0;
+  const value = Reflect.get(input, field2);
+  return typeof value === "string" && value.length > 0 ? value : void 0;
+}
+
 // src/evaluation/message-parsing.ts
 function extractToolResultTexts(message) {
   if (message.type !== "user") return [];
@@ -2416,12 +2433,9 @@ function extractSourceCodeFromMessage(message) {
   const content = betaMessage?.content;
   if (!Array.isArray(content)) return void 0;
   for (const block of content) {
-    if (block.type === "tool_use" && block.name === "Write") {
-      const input = block.input;
-      if (input?.content) {
-        sourceCode = input.content;
-      }
-    }
+    if (block.type !== "tool_use") continue;
+    const value = sourceCodeFromToolUse(block.name, block.input);
+    if (value !== void 0) sourceCode = value;
   }
   return sourceCode;
 }
@@ -2540,6 +2554,12 @@ async function runEvaluationLoop(options) {
   let currentSourceCode = context.sourceCode;
   let round2 = 0;
   const maxRounds = Math.min(config.maxRounds ?? 3, MAX_EVAL_ROUNDS_HARD_LIMIT);
+  const offersATool = Object.keys(generatorOptions?.mcpServers ?? {}).length > 0 || (generatorOptions?.builtinTools?.length ?? 0) > 0;
+  if (maxRounds > 1 && !offersATool) {
+    throw new Error(
+      "runEvaluationLoop: a fix round needs a tool that can return code \u2014 pass generatorOptions.mcpServers (e.g. a server with compile_component) or opt into a built-in with generatorOptions.builtinTools (e.g. ['Write'])."
+    );
+  }
   while (round2 < maxRounds) {
     round2++;
     onProgress?.({ type: "evaluating", round: round2 });
@@ -2579,12 +2599,15 @@ async function runEvaluationLoop(options) {
         maxTurns: 15,
         maxBudgetUsd: config.maxBudgetPerFix,
         env,
-        permissionMode: "bypassPermissions",
-        allowDangerouslySkipPermissions: true,
         ...cliPath && { pathToClaudeCodeExecutable: cliPath },
         ...generatorOptions?.cwd && { cwd: generatorOptions.cwd },
         ...generatorOptions?.mcpServers && { mcpServers: generatorOptions.mcpServers },
-        ...generatorOptions?.allowedTools && { allowedTools: generatorOptions.allowedTools },
+        // ggui#1714 — no CLI built-ins unless asked for, no host settings, no permission bypass.
+        ...sdkToolPins({
+          mcpServers: generatorOptions?.mcpServers,
+          allowedTools: generatorOptions?.allowedTools,
+          builtinTools: generatorOptions?.builtinTools
+        }),
         ...generatorOptions?.model && { model: generatorOptions.model },
         ...generatorOptions?.stderr && { stderr: generatorOptions.stderr }
       }
@@ -4719,6 +4742,8 @@ var INSTRUMENT_BY_ID = {
 var INSTRUMENT_BESIDE_JUDGE = {
   "space.edge": "edge"
 };
+var REQUIRES_POST_ACTION_CAPTURE = /* @__PURE__ */ new Set(["state.feedback"]);
+var NO_POST_ACTION = "not evaluated: no post-action capture";
 function besideRead(kind, m) {
   if (m.edge === void 0) return { verdict: "n/a", evidence: "not measured: this capture had no page to read" };
   if (m.edge === null) return { verdict: "n/a", evidence: "edge probe failed on this capture" };
@@ -4744,6 +4769,7 @@ function instrumentRead(kind, m) {
   return { verdict: "n/a", evidence: `fill ${m.inkRatio.toFixed(3)} reported; no floor ruled` };
 }
 function readRow(row, answers, m, unansweredReason) {
+  if (row.evaluation === "judge" && REQUIRES_POST_ACTION_CAPTURE.has(row.id) && m.postActionCapture !== true) return { verdict: "n/a", evidence: NO_POST_ACTION };
   if (row.evaluation === "judge") return majority(row.id, answers, unansweredReason);
   if (row.evaluation === "human") return { verdict: "n/a", evidence: `reader's column (status ${row.status})` };
   if (row.status !== "live") return { verdict: "n/a", evidence: `not implemented (status ${row.status})` };
@@ -4803,7 +4829,7 @@ ${text}
 }
 function buildCriteriaJudgeBlock(bank, selected, inputs = {}) {
   const byId = new Map(bankRows(bank).map((r) => [r.id, r]));
-  const asked = selected.selection.map((s) => byId.get(s.id)).filter((r) => r !== void 0 && r.evaluation === "judge");
+  const asked = selected.selection.map((s) => byId.get(s.id)).filter((r) => r !== void 0 && r.evaluation === "judge").filter((r) => inputs.postActionCapture === true || !REQUIRES_POST_ACTION_CAPTURE.has(r.id));
   if (asked.length === 0) return "";
   const lines = asked.map((r) => `- ${r.id} (${r.severity}): ${r.text} Evidence must name: ${r.evidence}`);
   return "## Criteria \u2014 answer each one from the FRAME only\nRead these off the screenshot. Never infer them from the request, the styling profile or any direction text above.\n" + (inputs.frame !== void 0 ? `${criteriaFrameLine(inputs.frame)}