@ggui-ai/ui-gen 0.26.0 → 0.28.0

package/dist/{adapter-BJwn74rx.d.ts → adapter-h4_V69LI.d.ts}

@@ -1,7 +1,7 @@
 import { RuntimeRenderCheck } from './harness/types-public.js';
 import { GenerationRuntimeProbeCheck } from '@ggui-ai/mcp-server-core';
 import { JsonObject, DataContract } from '@ggui-ai/protocol';
-import { f as ProbeHostLoad, g as ProbePendingAffordance } from './types-public-B7wn4ce5.js';
+import { f as ProbeHostLoad, g as ProbePendingAffordance } from './types-public-h8IMnYuh.js';
 
 /**
  * The subprocess bounds of one isolated check (ggui#1380). They are

package/dist/adapters/base.d.ts

@@ -41,8 +41,13 @@ interface ClaudeSdkConfig extends AdapterConfig {
     env?: Record<string, string>;
     /** MCP server configs for Claude Agent SDK (keyed by name) */
     mcpServers?: Record<string, McpServerConfig>;
-    /** Allowed tool names for Claude Agent SDK */
+    /** Tool names to approve. Absent → every tool of every server in `mcpServers` (ggui#1714). */
     allowedTools?: string[];
+    /**
+     * CLI built-in tools to offer the model (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[];
     /** Subagent definitions for Claude Agent SDK */
     agents?: Record<string, AgentDefinition>;
     /** Custom CLI path for Docker deployments */

package/dist/adapters/base.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/adapters/base.ts"],"names":[],"mappings":";AA2FO,IAAe,mBAAf,MAAgC;AAAA,EAKrC,YAA+B,MAAA,EAA0B;AAA1B,IAAA,IAAA,CAAA,MAAA,GAAA,MAAA;AAAA,EAA2B;AAAA,EAA3B,MAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAiB/B,eAAe,cAAA,EAAgC;AAC7C,IAAA,OAAO,iBAAiB,cAAc,CAAA;AAAA,EACxC;AACF;AAQA,SAAS,iBAAiB,OAAA,EAAyB;AACjD,EAAA,MAAM,UAAA,GAAa,OAAA,CAAQ,OAAA,CAAQ,GAAG,CAAA;AACtC,EAAA,OAAO,cAAc,CAAA,GAAI,OAAA,CAAQ,KAAA,CAAM,UAAA,GAAa,CAAC,CAAA,GAAI,OAAA;AAC3D;AAKO,SAAS,cAAA,CACd,WACG,WAAA,EACM;AACT,EAAA,IAAI,MAAA,CAAO,QAAQ,OAAO,IAAA;AAC1B,EAAA,IAAI,MAAA,CAAO,YAAY,OAAO,IAAA;AAC9B,EAAA,IAAI,MAAA,CAAO,iBAAiB,OAAO,IAAA;AACnC,EAAA,OAAO,WAAA,CAAY,KAAK,CAAC,IAAA,KAAS,CAAC,CAAC,OAAA,CAAQ,GAAA,CAAI,IAAI,CAAC,CAAA;AACvD","file":"base.js","sourcesContent":["// packages/ui-gen/src/adapters/base.ts\n//\n// Abstract base class for all generator adapters.\n\nimport type { McpServerConfig, AgentDefinition } from '@anthropic-ai/claude-agent-sdk';\nimport type { JsonObject } from '@ggui-ai/protocol';\nimport type { AdapterResult, ToolDefinition, ProviderName, AdapterMode } from './types';\n\n/**\n * Core configuration shared by all adapters.\n */\nexport interface AdapterConfig {\n  /** API key for the provider (BYOK) */\n  apiKey?: string;\n  /** Use Bedrock instead of direct API (Claude only) */\n  useBedrock?: boolean;\n  /**\n   * ggui#1185 — run on the machine's own Claude Code login instead of a key.\n   * Claude SDK adapter only. The adapter strips every provider key from the\n   * spawned env and pins the run tool-less, config-less and non-bare (see\n   * `claude/claude-code-login.ts`). Opt-in; absent ⇒ key-based auth.\n   */\n  claudeCodeLogin?: boolean;\n  /**\n   * Thinking level for models with extended thinking (Gemini 3.x Flash\n   * via the Interactions API). Mirrors the API's `thinking_level` enum —\n   * `'minimal'` is the floor (the API has NO true off), then `'low'` /\n   * `'medium'` / `'high'`. Omitted → provider default, and Gemini 3.5\n   * Flash defaults thinking ON with no cap — the dominant latency cost\n   * in contract-bounded UI generation. Only the Google adapter consumes\n   * this today.\n   */\n  thinkingLevel?: 'minimal' | 'low' | 'medium' | 'high';\n}\n\n/**\n * Extended configuration for Claude Agent SDK adapter.\n * Separated from AdapterConfig to avoid polluting the base interface\n * with provider-specific concerns.\n */\nexport interface ClaudeSdkConfig extends AdapterConfig {\n  /** Working directory for SDK subprocess */\n  cwd?: string;\n  /** Environment variables for SDK subprocess */\n  env?: Record<string, string>;\n  /** MCP server configs for Claude Agent SDK (keyed by name) */\n  mcpServers?: Record<string, McpServerConfig>;\n  /** Allowed tool names for Claude Agent SDK */\n  allowedTools?: string[];\n  /** Subagent definitions for Claude Agent SDK */\n  agents?: Record<string, AgentDefinition>;\n  /** Custom CLI path for Docker deployments */\n  cliPath?: string;\n  /** stderr callback for debug logging */\n  stderr?: (data: string) => void;\n  /** Progress callback */\n  onProgress?: (event: { type: string; [key: string]: unknown }) => void;\n}\n\n/**\n * Union config type — adapters accept their specific config\n * but the registry uses this union so callers can pass any config.\n */\nexport type AnyAdapterConfig = AdapterConfig | ClaudeSdkConfig;\n\n/**\n * Parameters for a single generation run.\n */\nexport interface GenerateParams {\n  /** System prompt (HOW to build — shared across all SDKs) */\n  systemPrompt: string;\n  /** User prompt (WHAT to build — the specific request) */\n  userPrompt: string;\n  /** Model identifier (provider-native format, e.g., 'claude-sonnet-4-6') */\n  model: string;\n  /** SDK-agnostic tool definitions */\n  tools: ToolDefinition[];\n  /** Maximum agentic loop turns */\n  maxTurns: number;\n  /** Model-specific options (e.g., speed_priority for Claude fast mode) */\n  modelOptions?: JsonObject;\n}\n\n/**\n * Abstract base class for generator adapters.\n *\n * Each adapter wraps a specific provider SDK (Claude, OpenAI, Google)\n * in either raw API mode or agent SDK mode, and implements the tool-use\n * loop using the SDK's native API. All adapters produce the same\n * output format (AdapterResult).\n */\nexport abstract class GeneratorAdapter {\n  abstract readonly provider: ProviderName;\n  abstract readonly mode: AdapterMode;\n  abstract readonly displayName: string;\n\n  constructor(protected readonly config: AnyAdapterConfig) {}\n\n  /**\n   * Run UI generation with the given tools, prompt, and model.\n   */\n  abstract generate(params: GenerateParams): Promise<AdapterResult>;\n\n  /**\n   * Check if this adapter can run (SDK installed, credentials available).\n   */\n  abstract isAvailable(): boolean;\n\n  /**\n   * Map a LiteLLM-format model ID to this SDK's native format.\n   * Default: strips the provider prefix (e.g., 'anthropic/claude-sonnet-4-6' -> 'claude-sonnet-4-6').\n   * Override only if the prefix convention differs.\n   */\n  resolveModelId(litellmModelId: string): string {\n    return stripModelPrefix(litellmModelId);\n  }\n}\n\n/**\n * Strip provider prefix from LiteLLM-format model IDs.\n * 'anthropic/claude-sonnet-4-6' -> 'claude-sonnet-4-6'\n * 'openai/gpt-5.3-codex' -> 'gpt-5.3-codex'\n * 'gemini/gemini-3-flash-preview' -> 'gemini-3-flash-preview'\n */\nfunction stripModelPrefix(modelId: string): string {\n  const slashIndex = modelId.indexOf('/');\n  return slashIndex >= 0 ? modelId.slice(slashIndex + 1) : modelId;\n}\n\n/**\n * Check if credentials are available from config or environment variables.\n */\nexport function hasCredentials(\n  config: AdapterConfig,\n  ...envVarNames: string[]\n): boolean {\n  if (config.apiKey) return true;\n  if (config.useBedrock) return true;\n  if (config.claudeCodeLogin) return true;\n  return envVarNames.some((name) => !!process.env[name]);\n}\n"]}
+{"version":3,"sources":["../../src/adapters/base.ts"],"names":[],"mappings":";AAgGO,IAAe,mBAAf,MAAgC;AAAA,EAKrC,YAA+B,MAAA,EAA0B;AAA1B,IAAA,IAAA,CAAA,MAAA,GAAA,MAAA;AAAA,EAA2B;AAAA,EAA3B,MAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAiB/B,eAAe,cAAA,EAAgC;AAC7C,IAAA,OAAO,iBAAiB,cAAc,CAAA;AAAA,EACxC;AACF;AAQA,SAAS,iBAAiB,OAAA,EAAyB;AACjD,EAAA,MAAM,UAAA,GAAa,OAAA,CAAQ,OAAA,CAAQ,GAAG,CAAA;AACtC,EAAA,OAAO,cAAc,CAAA,GAAI,OAAA,CAAQ,KAAA,CAAM,UAAA,GAAa,CAAC,CAAA,GAAI,OAAA;AAC3D;AAKO,SAAS,cAAA,CACd,WACG,WAAA,EACM;AACT,EAAA,IAAI,MAAA,CAAO,QAAQ,OAAO,IAAA;AAC1B,EAAA,IAAI,MAAA,CAAO,YAAY,OAAO,IAAA;AAC9B,EAAA,IAAI,MAAA,CAAO,iBAAiB,OAAO,IAAA;AACnC,EAAA,OAAO,WAAA,CAAY,KAAK,CAAC,IAAA,KAAS,CAAC,CAAC,OAAA,CAAQ,GAAA,CAAI,IAAI,CAAC,CAAA;AACvD","file":"base.js","sourcesContent":["// packages/ui-gen/src/adapters/base.ts\n//\n// Abstract base class for all generator adapters.\n\nimport type { McpServerConfig, AgentDefinition } from '@anthropic-ai/claude-agent-sdk';\nimport type { JsonObject } from '@ggui-ai/protocol';\nimport type { AdapterResult, ToolDefinition, ProviderName, AdapterMode } from './types';\n\n/**\n * Core configuration shared by all adapters.\n */\nexport interface AdapterConfig {\n  /** API key for the provider (BYOK) */\n  apiKey?: string;\n  /** Use Bedrock instead of direct API (Claude only) */\n  useBedrock?: boolean;\n  /**\n   * ggui#1185 — run on the machine's own Claude Code login instead of a key.\n   * Claude SDK adapter only. The adapter strips every provider key from the\n   * spawned env and pins the run tool-less, config-less and non-bare (see\n   * `claude/claude-code-login.ts`). Opt-in; absent ⇒ key-based auth.\n   */\n  claudeCodeLogin?: boolean;\n  /**\n   * Thinking level for models with extended thinking (Gemini 3.x Flash\n   * via the Interactions API). Mirrors the API's `thinking_level` enum —\n   * `'minimal'` is the floor (the API has NO true off), then `'low'` /\n   * `'medium'` / `'high'`. Omitted → provider default, and Gemini 3.5\n   * Flash defaults thinking ON with no cap — the dominant latency cost\n   * in contract-bounded UI generation. Only the Google adapter consumes\n   * this today.\n   */\n  thinkingLevel?: 'minimal' | 'low' | 'medium' | 'high';\n}\n\n/**\n * Extended configuration for Claude Agent SDK adapter.\n * Separated from AdapterConfig to avoid polluting the base interface\n * with provider-specific concerns.\n */\nexport interface ClaudeSdkConfig extends AdapterConfig {\n  /** Working directory for SDK subprocess */\n  cwd?: string;\n  /** Environment variables for SDK subprocess */\n  env?: Record<string, string>;\n  /** MCP server configs for Claude Agent SDK (keyed by name) */\n  mcpServers?: Record<string, McpServerConfig>;\n  /** Tool names to approve. Absent → every tool of every server in `mcpServers` (ggui#1714). */\n  allowedTools?: string[];\n  /**\n   * CLI built-in tools to offer the model (e.g. `['Write']`). Default: none.\n   * An explicit opt-in: a built-in that writes or executes acts on the host (ggui#1714).\n   */\n  builtinTools?: readonly string[];\n  /** Subagent definitions for Claude Agent SDK */\n  agents?: Record<string, AgentDefinition>;\n  /** Custom CLI path for Docker deployments */\n  cliPath?: string;\n  /** stderr callback for debug logging */\n  stderr?: (data: string) => void;\n  /** Progress callback */\n  onProgress?: (event: { type: string; [key: string]: unknown }) => void;\n}\n\n/**\n * Union config type — adapters accept their specific config\n * but the registry uses this union so callers can pass any config.\n */\nexport type AnyAdapterConfig = AdapterConfig | ClaudeSdkConfig;\n\n/**\n * Parameters for a single generation run.\n */\nexport interface GenerateParams {\n  /** System prompt (HOW to build — shared across all SDKs) */\n  systemPrompt: string;\n  /** User prompt (WHAT to build — the specific request) */\n  userPrompt: string;\n  /** Model identifier (provider-native format, e.g., 'claude-sonnet-4-6') */\n  model: string;\n  /** SDK-agnostic tool definitions */\n  tools: ToolDefinition[];\n  /** Maximum agentic loop turns */\n  maxTurns: number;\n  /** Model-specific options (e.g., speed_priority for Claude fast mode) */\n  modelOptions?: JsonObject;\n}\n\n/**\n * Abstract base class for generator adapters.\n *\n * Each adapter wraps a specific provider SDK (Claude, OpenAI, Google)\n * in either raw API mode or agent SDK mode, and implements the tool-use\n * loop using the SDK's native API. All adapters produce the same\n * output format (AdapterResult).\n */\nexport abstract class GeneratorAdapter {\n  abstract readonly provider: ProviderName;\n  abstract readonly mode: AdapterMode;\n  abstract readonly displayName: string;\n\n  constructor(protected readonly config: AnyAdapterConfig) {}\n\n  /**\n   * Run UI generation with the given tools, prompt, and model.\n   */\n  abstract generate(params: GenerateParams): Promise<AdapterResult>;\n\n  /**\n   * Check if this adapter can run (SDK installed, credentials available).\n   */\n  abstract isAvailable(): boolean;\n\n  /**\n   * Map a LiteLLM-format model ID to this SDK's native format.\n   * Default: strips the provider prefix (e.g., 'anthropic/claude-sonnet-4-6' -> 'claude-sonnet-4-6').\n   * Override only if the prefix convention differs.\n   */\n  resolveModelId(litellmModelId: string): string {\n    return stripModelPrefix(litellmModelId);\n  }\n}\n\n/**\n * Strip provider prefix from LiteLLM-format model IDs.\n * 'anthropic/claude-sonnet-4-6' -> 'claude-sonnet-4-6'\n * 'openai/gpt-5.3-codex' -> 'gpt-5.3-codex'\n * 'gemini/gemini-3-flash-preview' -> 'gemini-3-flash-preview'\n */\nfunction stripModelPrefix(modelId: string): string {\n  const slashIndex = modelId.indexOf('/');\n  return slashIndex >= 0 ? modelId.slice(slashIndex + 1) : modelId;\n}\n\n/**\n * Check if credentials are available from config or environment variables.\n */\nexport function hasCredentials(\n  config: AdapterConfig,\n  ...envVarNames: string[]\n): boolean {\n  if (config.apiKey) return true;\n  if (config.useBedrock) return true;\n  if (config.claudeCodeLogin) return true;\n  return envVarNames.some((name) => !!process.env[name]);\n}\n"]}

package/dist/adapters/generation-dispatch.d.ts

@@ -2,7 +2,7 @@ import { EvaluationConfig } from '../evaluation/types.js';
 import { RuntimeRenderCheck } from '../harness/types-public.js';
 import { ModelRoles, GenerationResult } from '../harness/result-types.js';
 import { ProviderName, ToolDefinition } from './types.js';
-import { Q as QualityConfig, b as AgentConfig, V as VisionProvider } from '../types-public-B7wn4ce5.js';
+import { Q as QualityConfig, b as AgentConfig, V as VisionProvider } from '../types-public-h8IMnYuh.js';
 import { DataContract, JsonObject, GadgetDescriptor } from '@ggui-ai/protocol';
 import { C as CanvasClass, D as DesignMode } from '../design-mode-DGWN_Gkm.js';
 import { C as CriteriaBank } from '../bank-Ctdp3FUA.js';
@@ -12,7 +12,7 @@ import '../axes-CKVCzR_x.js';
 import '../llm.js';
 import '../types-Dh1s4K9j.js';
 import '../policy.js';
-import '../visual-evaluator-C9L-o6ap.js';
+import '../visual-evaluator-C2strxqD.js';
 import 'puppeteer-core';
 import 'zod';
 import '@ggui-ai/mcp-server-core';

package/dist/adapters/generation-dispatch.js

@@ -532,10 +532,7 @@ var init_json_schema_ts = __esm({
 var NOT_YET_TAUGHT_TOKENS, taughtTokenManifest;
 var init_taught_tokens = __esm({
   "src/taught-tokens.ts"() {
-    NOT_YET_TAUGHT_TOKENS = /* @__PURE__ */ new Set([
-      "--ggui-color-controlAccentOutline",
-      "--ggui-color-controlOutline"
-    ]);
+    NOT_YET_TAUGHT_TOKENS = /* @__PURE__ */ new Set([]);
     taughtTokenManifest = consumedTokenManifest.filter(
       (token) => !NOT_YET_TAUGHT_TOKENS.has(token)
     );
@@ -714,6 +711,8 @@ var init_claude_code_login = __esm({
 function getBedrockModelId(model) {
   const BEDROCK_MAP = {
     "anthropic/claude-haiku-4-5": "us.anthropic.claude-haiku-4-5-20251001-v1:0",
+    "anthropic/claude-haiku-5-5": "us.anthropic.claude-haiku-5-5",
+    // listed ACTIVE in us-east-1 (ggui#1743)
     "anthropic/claude-sonnet-4-6": "us.anthropic.claude-sonnet-4-6",
     "anthropic/claude-opus-4-6": "us.anthropic.claude-opus-4-6-v1:0",
     // Claude 5-family (ggui#706, strings verified 2026-09-02 against
@@ -4626,6 +4625,40 @@ function detectComponentNeverReturns(sourceCode) {
   for (const st of body.statements) if (!isFunctionLike(st)) visit(st);
   return returnsValue ? null : { jsxStatementLine };
 }
+function kitControlTags(source) {
+  const out = [];
+  const open = new RegExp(`<(${KIT_CONTROLS.join("|")})\\b`, "g");
+  for (const m of source.matchAll(open)) {
+    const start = (m.index ?? 0) + m[0].length;
+    let depth = 0;
+    let quote = null;
+    let i = start;
+    for (; i < source.length; i += 1) {
+      const c = source[i];
+      if (quote !== null) {
+        if (c === quote && source[i - 1] !== "\\") quote = null;
+      } else if (c === '"' || c === "'" || c === "`") quote = c;
+      else if (c === "{") depth += 1;
+      else if (c === "}") depth -= 1;
+      else if (c === ">" && depth === 0) break;
+    }
+    out.push({ name: m[1] ?? "", attrs: source.slice(start, i), line: source.slice(0, m.index ?? 0).split("\n").length });
+  }
+  return out;
+}
+function inlineBorderKey(attrs, source) {
+  const BORDER = /\b(border(?:Top|Right|Bottom|Left)?(?:Color|Width|Style)?)\s*:/;
+  const style = attrs.match(/\bstyle=\{\{([\s\S]*)\}\}/);
+  if (style !== null) {
+    const hit2 = (style[1] ?? "").match(BORDER);
+    if (hit2 !== null) return hit2[1] ?? null;
+  }
+  const named = attrs.match(/\bstyle=\{\s*([A-Za-z_$][\w$]*)\s*\}/);
+  if (named === null) return null;
+  const decl = source.match(new RegExp(`\\b(?:const|let)\\s+${named[1]}\\b[^=]*=\\s*\\{([^{}]*)\\}`));
+  const hit = decl !== null ? (decl[1] ?? "").match(BORDER) : null;
+  return hit !== null ? hit[1] ?? null : null;
+}
 async function runTier0Checks(sourceCode, contract, buildErrors, gadgetTypes, designMode = DEFAULT_DESIGN_MODE) {
   const issues = [];
   const lines = sourceCode.split("\n");
@@ -5023,6 +5056,20 @@ ${errorDetail}`,
       });
     }
   }
+  for (const tag of enforceDesignVocabulary ? kitControlTags(sourceCode) : []) {
+    const key = inlineBorderKey(tag.attrs, sourceCode);
+    if (key === null) continue;
+    issues.push({
+      tier: 0,
+      result: "fail",
+      category: "tokens",
+      subcategory: "control-inline-border",
+      severity: "critical",
+      description: `control-inline-border: <${tag.name}> sets an inline \`${key}\` \u2014 a kit control draws its own boundary, and an inline border paints over it`,
+      fix: `Remove the border from <${tag.name}>'s style (keep borderRadius if you need it). An outline chip is <Button variant="outline"> exactly as it comes; a control-like edge on your own element uses var(--ggui-color-controlOutline).`,
+      line: tag.line
+    });
+  }
   if (!sourceCode.includes("interface Props") && !sourceCode.includes("type Props")) {
     issues.push({
       tier: 0,
@@ -5294,7 +5341,7 @@ async function runTier0(sourceCode, contract, designMode = DEFAULT_DESIGN_MODE)
   }
   return { issues, pass };
 }
-var MANIFEST_TOKENS, RETIRED_TOKEN_SUCCESSORS, TOKEN_REF_RE, HOOK_NAME_FOR, CONTRACT_FIELD_FOR;
+var MANIFEST_TOKENS, RETIRED_TOKEN_SUCCESSORS, TOKEN_REF_RE, HOOK_NAME_FOR, CONTRACT_FIELD_FOR, KIT_CONTROLS;
 var init_run_tier0 = __esm({
   "src/check/run-tier0.ts"() {
     init_design_mode();
@@ -5323,6 +5370,7 @@ var init_run_tier0 = __esm({
       stream: "streamSpec",
       context: "contextSpec"
     };
+    KIT_CONTROLS = ["Button", "Input", "Select", "TextArea", "Checkbox", "RadioGroup", "Toggle", "SearchField"];
   }
 });
 
@@ -5649,10 +5697,16 @@ var dupe_break_exports = {};
 __export(dupe_break_exports, {
   SAME_EXCHANGE_BREAK_AT: () => SAME_EXCHANGE_BREAK_AT,
   SAME_EXCHANGE_NUDGE: () => SAME_EXCHANGE_NUDGE,
+  SAME_FAILURE_BREAK_AT: () => SAME_FAILURE_BREAK_AT,
+  SAME_FAILURE_NUDGE: () => SAME_FAILURE_NUDGE,
+  SAME_FAILURE_NUDGE_AT: () => SAME_FAILURE_NUDGE_AT,
+  clearWriteFailures: () => clearWriteFailures,
   computePatchFingerprint: () => computePatchFingerprint,
   createDupeBreakState: () => createDupeBreakState,
   errorClassBucket: () => errorClassBucket,
-  recordToolExchange: () => recordToolExchange
+  failureSignatures: () => failureSignatures,
+  recordToolExchange: () => recordToolExchange,
+  recordWriteFailure: () => recordWriteFailure
 });
 function recordToolExchange(state, name, input, resultText) {
   const key = createHash("sha256").update(name).update("\0").update(JSON.stringify(input) ?? "undefined").update("\0").update(resultText).digest("hex");
@@ -5664,6 +5718,25 @@ function recordToolExchange(state, name, input, resultText) {
   }
   return state.consecutiveIdenticalExchanges;
 }
+function failureSignatures(resultText) {
+  const norm = (text) => text.replace(/[0-9]+/g, "#").replace(/\s+/g, " ").trim();
+  const violations = [...resultText.matchAll(/^\[[^\]\n]+\] [^\n]+/gm)].map((m) => norm(m[0]));
+  if (violations.length > 0) return [...new Set(violations)];
+  const rest = norm(resultText.replace(/^Committed [^\n]*\n/, ""));
+  return rest.length > 0 ? [rest] : [];
+}
+function recordWriteFailure(state, resultText) {
+  let top;
+  for (const signature of failureSignatures(resultText)) {
+    const repeats = (state.failureCounts.get(signature) ?? 0) + 1;
+    state.failureCounts.set(signature, repeats);
+    if (top === void 0 || repeats > top.repeats) top = { signature, repeats };
+  }
+  return top;
+}
+function clearWriteFailures(state) {
+  state.failureCounts.clear();
+}
 function createDupeBreakState() {
   return {
     lastFailedPatchFingerprint: null,
@@ -5680,7 +5753,8 @@ function createDupeBreakState() {
     consecutiveBrokenApplies: 0,
     consecutiveNoToolCalls: 0,
     lastExchangeKey: null,
-    consecutiveIdenticalExchanges: 0
+    consecutiveIdenticalExchanges: 0,
+    failureCounts: /* @__PURE__ */ new Map()
   };
 }
 function errorClassBucket(errText) {
@@ -5705,11 +5779,14 @@ function computePatchFingerprint(changes, errText) {
   h.update(errBucket);
   return h.digest("hex");
 }
-var SAME_EXCHANGE_BREAK_AT, SAME_EXCHANGE_NUDGE;
+var SAME_EXCHANGE_BREAK_AT, SAME_EXCHANGE_NUDGE, SAME_FAILURE_NUDGE_AT, SAME_FAILURE_BREAK_AT, SAME_FAILURE_NUDGE;
 var init_dupe_break = __esm({
   "src/harness/coding/dupe-break.ts"() {
     SAME_EXCHANGE_BREAK_AT = 3;
     SAME_EXCHANGE_NUDGE = "\n\n[harness] This is the SAME tool call with the SAME result as your previous turn \u2014 repeating it cannot produce new information. Use what you already have and emit an apply_changes (or scoped) call now; one more identical repeat terminates the run.";
+    SAME_FAILURE_NUDGE_AT = 3;
+    SAME_FAILURE_BREAK_AT = 5;
+    SAME_FAILURE_NUDGE = "\n\n[harness] This same failure has now been reported for three of your patches. Changing other code does not clear it: re-read that check above and change exactly what it names. Two more reports of it end the run.";
   }
 });
 
@@ -6101,13 +6178,13 @@ Return your evaluation using the evaluate_${criterion} tool.`;
 function getDefaultEvalModel(provider) {
   switch (provider) {
     case "claude":
-      return "claude-haiku-4-5-20251001";
+      return "claude-haiku-5-5";
     case "openai":
       return "gpt-5.6-luna";
     case "google":
       return "gemini-3.5-flash-lite";
     default:
-      return "claude-haiku-4-5-20251001";
+      return "claude-haiku-5-5";
   }
 }
 function callTokens(response) {
@@ -6567,6 +6644,7 @@ 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})` };
@@ -6625,14 +6703,14 @@ ${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)}
 ` : "") + (inputs.propsJson !== void 0 ? criteriaPropsSection(inputs.propsJson) : "") + `${lines.join("\n")}
 Add to your JSON a top-level "criteria" array with exactly these ids: [{ "id": "<id>", "verdict": "pass" | "fail" | "n/a", "evidence": "<one sentence naming the region or element you read it from>" }]. Every row here was selected as applicable to this card, so "n/a" never means "does not apply": use it ONLY when this frame cannot show the criterion \u2014 it is then recorded as not evaluated.`;
 }
-var INSTRUMENT_BY_ID, INSTRUMENT_BESIDE_JUDGE, CRITERIA_PROPS_MAX_CHARS;
+var INSTRUMENT_BY_ID, INSTRUMENT_BESIDE_JUDGE, REQUIRES_POST_ACTION_CAPTURE, NO_POST_ACTION, CRITERIA_PROPS_MAX_CHARS;
 var init_resolve = __esm({
   "src/evaluation/criteria/resolve.ts"() {
     init_bank();
@@ -6645,6 +6723,8 @@ var init_resolve = __esm({
     INSTRUMENT_BESIDE_JUDGE = {
       "space.edge": "edge"
     };
+    REQUIRES_POST_ACTION_CAPTURE = /* @__PURE__ */ new Set(["state.feedback"]);
+    NO_POST_ACTION = "not evaluated: no post-action capture";
     CRITERIA_PROPS_MAX_CHARS = 2e3;
   }
 });
@@ -8182,7 +8262,7 @@ function parseCriteriaAnswers(raw) {
 function getDefaultVisualModel(provider) {
   switch (provider) {
     case "claude":
-      return "claude-haiku-4-5-20251001";
+      return "claude-haiku-5-5";
     case "google":
       return "gemini-3-flash-preview";
   }
@@ -9451,10 +9531,96 @@ function getInitialValuePropNames(contract) {
   }
   return names;
 }
+function objectLiteralKeysAt(src, open) {
+  if (src[open] !== "{") return void 0;
+  const keys = [];
+  let depth = 0;
+  let expectKey = true;
+  for (let i = open; i < src.length; i++) {
+    const c = src[i];
+    if (c === '"' || c === "'" || c === "`") {
+      let j = i + 1;
+      while (j < src.length && src[j] !== c) j += src[j] === "\\" ? 2 : 1;
+      if (depth === 1 && expectKey && /^\s*:/.test(src.slice(j + 1))) keys.push(src.slice(i + 1, j));
+      i = j;
+      expectKey = false;
+      continue;
+    }
+    if (c === "{" || c === "[" || c === "(") {
+      depth++;
+      if (depth === 1) expectKey = true;
+      continue;
+    }
+    if (c === "}" || c === "]" || c === ")") {
+      depth--;
+      if (depth === 0) return keys;
+      continue;
+    }
+    if (depth !== 1) continue;
+    if (c === ",") {
+      expectKey = true;
+      continue;
+    }
+    if (expectKey && /[A-Za-z_$]/.test(c ?? "")) {
+      const m = /^([A-Za-z_$][\w$]*)\s*(:|,|\})/.exec(src.slice(i));
+      if (m !== null && m[1] !== void 0) keys.push(m[1]);
+      expectKey = false;
+    }
+  }
+  return keys;
+}
+function callArgsAt(src, open) {
+  const args = [];
+  let depth = 0;
+  let start = open + 1;
+  for (let i = open; i < src.length; i++) {
+    const c = src[i];
+    if (c === '"' || c === "'" || c === "`") {
+      let j = i + 1;
+      while (j < src.length && src[j] !== c) j += src[j] === "\\" ? 2 : 1;
+      i = j;
+      continue;
+    }
+    if (c === "(" || c === "{" || c === "[") depth++;
+    else if (c === ")" || c === "}" || c === "]") {
+      depth--;
+      if (depth === 0) {
+        args.push(src.slice(start, i));
+        return args;
+      }
+    } else if (c === "," && depth === 1) {
+      args.push(src.slice(start, i));
+      start = i + 1;
+    }
+  }
+  return args;
+}
+function initialStateKeys(src, argStart, argText) {
+  const lead = argText.length - argText.trimStart().length;
+  let at = argStart + lead;
+  let text = argText.trim();
+  const lazy = /^\(\s*\)\s*=>\s*\(?\s*/.exec(text);
+  if (lazy !== null) {
+    at += lazy[0].length;
+    text = text.slice(lazy[0].length);
+  }
+  if (text.startsWith("{")) return objectLiteralKeysAt(src, at) ?? [];
+  const ident = /^([A-Za-z_$][\w$]*)\s*\)?\s*$/.exec(text);
+  if (ident === null || ident[1] === void 0) return [];
+  const decls = [...src.matchAll(new RegExp(`(?:const|let|var)\\s+${ident[1]}\\s*(?::[^=]{1,200})?=\\s*`, "g"))];
+  const decl = decls.length === 1 ? decls[0] : void 0;
+  if (decl === void 0) return [];
+  return objectLiteralKeysAt(src, (decl.index ?? 0) + decl[0].length) ?? [];
+}
 function collectStateKeys(src) {
   const keys = /* @__PURE__ */ new Set();
   const varRe = /const\s*\[\s*(\w+)\s*,\s*(\w+)\s*\]\s*=\s*useState/g;
   for (const m of src.matchAll(varRe)) keys.add(m[1]);
+  const contextRe = /const\s*\[\s*(\w+)\s*,\s*\w+\s*\]\s*=\s*useGguiContext\s*(?:<[^>]*>)?\s*\(\s*['"`](\w+)['"`]/g;
+  for (const m of src.matchAll(contextRe)) {
+    keys.add(m[1]);
+    keys.add(m[2]);
+  }
   const objRe = /useState(?:<[^>]*>)?\s*\(\s*\{([\s\S]*?)\}\s*\)/g;
   for (const m of src.matchAll(objRe)) {
     const keyRe = /(?:^|,)\s*(\w+)\s*:/g;
@@ -9465,6 +9631,16 @@ function collectStateKeys(src) {
     const keyRe = /(?:^|,)\s*(\w+)\s*:/g;
     for (const km of m[1].matchAll(keyRe)) keys.add(km[1]);
   }
+  for (const m of src.matchAll(/\b(useState|useReducer)\s*(?:<[^>]*>)?\s*\(/g)) {
+    const open = (m.index ?? 0) + m[0].length - 1;
+    const args = callArgsAt(src, open);
+    const which = m[1] === "useReducer" ? 1 : 0;
+    const arg = args[which];
+    if (arg === void 0) continue;
+    let argStart = open + 1;
+    for (let k = 0; k < which; k++) argStart += (args[k]?.length ?? 0) + 1;
+    for (const key of initialStateKeys(src, argStart, arg)) keys.add(key);
+  }
   return keys;
 }
 function mkIssue(subcategory, description, fix, result = "fail") {
@@ -10077,7 +10253,7 @@ function runNoHardcodedEntities(input) {
       mkIssue(
         "state.merge.no_hardcoded_entities",
         `Hardcoded entity array literal (multiple objects with "${idField}") in the component \u2014 entity data should come from state/props.`,
-        `Remove the literal array. Seed state from props.{entityProp} and merge updates via stream.`
+        `Remove the literal array: this data comes from ${all.filter((e) => e.idField === idField).map((e) => `props.${e.name}`).join(" / ") || "props"}. Read it there (\`?? []\` for an absent value, never a literal default), seed state from it, and merge updates via stream.`
       )
     );
   }
@@ -12969,13 +13145,14 @@ MANDATORY:
 2. NEVER use CSS gradients with custom colors. If you need a gradient: \`linear-gradient(to bottom, var(--ggui-color-primary-500), var(--ggui-color-primary-700))\`.
 3. NEVER invent your own palette. The system provides primary, neutral, success, warning, error, and info \u2014 use ONLY these.
 4. NEVER include literal fallback values in ANY token reference \u2014 colors, spacing, typography, radius, shadows alike. Write \`var(--ggui-spacing-4)\` bare, never \`var(--ggui-spacing-4, 16px)\`; \`var(--ggui-color-primary-600)\` bare, never \`var(--ggui-color-primary-600, #0284c7)\`. The runtime injects every token; a literal fallback paints the wrong value exactly when the operator's theme matters most.
+5. NEVER set an inline \`border\`, \`borderColor\`, \`borderWidth\` or \`borderStyle\` on a kit control \u2014 \`Button\`, \`Input\`, \`Select\`, \`TextArea\`, \`Checkbox\`, \`RadioGroup\`, \`Toggle\`, \`SearchField\`. The control draws its own boundary, and it clears 3:1 on its surface in every theme: an outline chip is \`<Button variant="outline">\` exactly as it comes (its edge is \`controlAccentOutline\`). Where you draw a control-like edge on your own element, use \`var(--ggui-color-controlOutline)\`. A ramp step such as \`primary-200\` as a control's border is nearly invisible (about 1.1:1).
 
 Token categories:
 - Brand fills: \`var(--ggui-color-primary-600)\`, \`var(--ggui-color-primary-50)\`
 - Accent text: \`var(--ggui-color-link)\` \u2014 eyebrows, taglines, links, labels; readable per mode (never a bare \`primary-<stop>\` as a text color)
 - Text: \`var(--ggui-color-onContainer)\`, \`var(--ggui-color-onSunken)\`
 - Backgrounds: \`var(--ggui-color-container)\`, \`var(--ggui-color-sunken)\`
-- Borders: \`var(--ggui-color-outline)\`
+- Borders: \`var(--ggui-color-outline)\` \u2014 dividers and the edges of non-interactive boxes; a control's edge is the kit's (rule 5)
 - Spacing: \`var(--ggui-spacing-4)\`, \`var(--ggui-spacing-6)\`
 - Typography: \`var(--ggui-font-size-sm)\`, \`var(--ggui-font-weight-semibold)\`
 - Shadows: \`var(--ggui-shape-shadow-sm)\`, \`var(--ggui-shape-shadow-md)\`, \`var(--ggui-shape-shadow-lg)\`
@@ -12990,7 +13167,7 @@ Use the FULL primary palette throughout the component \u2014 NOT only on submit
 | Element | Token | Purpose |
 |---------|-------|---------|
 | Section headers, hero areas, highlight strips | \`primary-50\` / \`primary-100\` | Subtle branded backgrounds |
-| Borders, dividers, focus rings, input focus | \`primary-200\` / \`primary-300\` | Branded structure |
+| Dividers, section rules, edges of non-interactive boxes | \`primary-200\` / \`primary-300\` | Branded structure \u2014 never a control's border (rule 5) |
 | Accent TEXT \u2014 eyebrows, taglines, links, labels, inline arrows | \`link\` (\`var(--ggui-color-link)\`) or \`<Text tone="emphasized">\` | The brand's READABLE ink, derived per mode to clear 4.5:1 \u2014 a bare \`primary-300\u2026600\` as a text color fails tier-0 |
 | Icon fills, badges, active indicators | \`primary-500\` / \`primary-600\` | Core accent fill |
 | Buttons, CTAs, filled interactive elements | \`primary-600\` / \`primary-700\` | Primary actions |
@@ -13663,6 +13840,7 @@ These tokens adapt automatically to any theme (light, dark, branded). **ALWAYS u
 | link | var(--ggui-color-link) | Accent TEXT \u2014 eyebrows, taglines, links, labels, inline arrows: the brand's readable ink, derived per mode to clear 4.5:1 on the container and the ground |
 | outline | var(--ggui-color-outline) | Borders, dividers |
 | outlineVariant | var(--ggui-color-outlineVariant) | Subtle borders |
+| controlOutline | var(--ggui-color-controlOutline) | A control's edge \u2014 clears 3:1 on its surface; kit controls draw it themselves |
 
 **Usage pattern:**
 - Page/section background \u2192 \`var(--ggui-color-ground)\`
@@ -13787,7 +13965,7 @@ Use \`var(--ggui-spacing-N)\` for all padding, gap, and margin values. **Never u
 ### Color usage guide
 - **Layering roles** (ground, container, elevated, sunken + their on* inks; outline): one kind of AREA each \u2014 use them for every background/text/border decision; they adapt to any theme
 - **primary-50/100**: Section backgrounds, highlight strips, card headers
-- **primary-200/300**: Borders, dividers, focus rings, input outlines
+- **primary-200/300**: Dividers and section rules \u2014 never a control's border: kit controls draw their own edge (\`controlOutline\` / \`controlAccentOutline\`), so set no inline border on them
 - **primary-500/600**: Filled buttons, CTAs, badges, icon fills, active indicators \u2014 fills and borders, NOT a text color
 - **link**: Accent TEXT \u2014 eyebrows, taglines, links, labels, inline arrows; the brand's readable ink per mode (a bare \`primary-300\u2026600\` as a text color fails tier-0: \`universal.accent_text_ink\`)
 - **primary-700/800/900**: Headings and text on light primary backgrounds
@@ -16295,6 +16473,7 @@ async function assembleGenerationResult(input) {
     evalResult: telemetry.evalResult !== void 0 && telemetry.contractFeedback !== void 0 ? { ...telemetry.evalResult, contractFeedback: telemetry.contractFeedback } : telemetry.evalResult,
     // ggui#404 — present only when the same-exchange guard ended the run.
     ...telemetry.sameExchangeBreak !== void 0 ? { sameExchangeBreak: telemetry.sameExchangeBreak } : {},
+    ...telemetry.sameFailureBreak !== void 0 ? { sameFailureBreak: telemetry.sameFailureBreak } : {},
     needsBackgroundImprovement,
     selfCheckPassed: telemetry.selfCheckPassed,
     timing,
@@ -17248,6 +17427,18 @@ Repeated error: ${errText}
       augmentedResultText += SAME_EXCHANGE_NUDGE;
     }
   }
+  const isWrite = call.name === "apply_changes" || call.name === "rewrite";
+  let repeatedFailure;
+  if (ctx.dupeBreak && isWrite) {
+    if (result.done) {
+      clearWriteFailures(ctx.dupeBreak);
+    } else {
+      repeatedFailure = recordWriteFailure(ctx.dupeBreak, result.result);
+      if (repeatedFailure?.repeats === SAME_FAILURE_NUDGE_AT) {
+        augmentedResultText += SAME_FAILURE_NUDGE;
+      }
+    }
+  }
   await codingAgent.sendToolResult([{
     callId: call.id,
     name: call.name,
@@ -17276,6 +17467,29 @@ Repeated error: ${errText}
       sameExchangeBreak: { tool: call.name, repeats: sameExchangeRepeats }
     };
   }
+  if (repeatedFailure !== void 0 && repeatedFailure.repeats >= SAME_FAILURE_BREAK_AT) {
+    const signature = repeatedFailure.signature.slice(0, 160);
+    console.log(
+      `[coding-agent] SAME_FAILURE_BREAK | tool=${call.name} | repeats=${repeatedFailure.repeats} | failure=${JSON.stringify(signature.slice(0, 80))} \u2014 terminating run (ggui#1791)`
+    );
+    return {
+      control: "break",
+      compiledCode: "",
+      selfCheckPassed: false,
+      outcome: void 0,
+      phase: void 0,
+      tokens,
+      ...cacheTokens,
+      llmMs,
+      toolMs,
+      lastResultText: augmentedResultText,
+      lastDiffFailed: false,
+      isEvalFeedback: false,
+      iconNamesCache,
+      preWarmedContext,
+      sameFailureBreak: { signature, repeats: repeatedFailure.repeats }
+    };
+  }
   const turnDiffFailed = !!result.error && result.result.includes("DIFF");
   const isDiffFail = turnDiffFailed && !isPatchInvalid;
   const outcome = result.done ? "PASS" : isPatchInvalid ? "PATCH_INVALID" : isDiffFail ? "DIFF_FAIL" : "SELF_CHECK_FAIL";
@@ -18263,6 +18477,7 @@ function createTelemetry() {
     probeRepairTrigger: void 0,
     contractFeedback: void 0,
     sameExchangeBreak: void 0,
+    sameFailureBreak: void 0,
     compiledCode: "",
     pairedSource: "",
     selfCheckPassed: false,
@@ -18318,6 +18533,7 @@ function absorbTurn(telemetry, turn) {
     else telemetry.counters.outcomes.selfCheckFail++;
   }
   if (turn.sameExchangeBreak !== void 0) telemetry.sameExchangeBreak = turn.sameExchangeBreak;
+  if (turn.sameFailureBreak !== void 0) telemetry.sameFailureBreak = turn.sameFailureBreak;
 }
 function createGenerateTaskRunner(input) {
   const { session, params, classification, telemetry } = input;