@ggui-ai/ui-gen 0.26.0 → 0.28.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (77) hide show
  1. package/dist/{adapter-BJwn74rx.d.ts → adapter-h4_V69LI.d.ts} +1 -1
  2. package/dist/adapters/base.d.ts +6 -1
  3. package/dist/adapters/base.js.map +1 -1
  4. package/dist/adapters/generation-dispatch.d.ts +2 -2
  5. package/dist/adapters/generation-dispatch.js +233 -17
  6. package/dist/adapters/generation-dispatch.js.map +1 -1
  7. package/dist/adapters/index.js +36 -14
  8. package/dist/adapters/index.js.map +1 -1
  9. package/dist/advanced/index.d.ts +4 -4
  10. package/dist/advanced/index.js +268 -28
  11. package/dist/advanced/index.js.map +1 -1
  12. package/dist/boilerplate.js +4 -6
  13. package/dist/boilerplate.js.map +1 -1
  14. package/dist/check/index.d.ts +1 -7
  15. package/dist/check/index.js +50 -4
  16. package/dist/check/index.js.map +1 -1
  17. package/dist/{create-ui-generator-N4gRXpcL.d.ts → create-ui-generator-BrZHF0cf.d.ts} +2 -2
  18. package/dist/{dispatch-BUoXRj_r.d.ts → dispatch-B07TI30d.d.ts} +1 -1
  19. package/dist/evaluation/axis-checks/index.d.ts +3 -3
  20. package/dist/evaluation/axis-checks/index.js +97 -1
  21. package/dist/evaluation/axis-checks/index.js.map +1 -1
  22. package/dist/evaluation/axis-checks/registry.d.ts +1 -1
  23. package/dist/evaluation/axis-checks/registry.js +97 -1
  24. package/dist/evaluation/axis-checks/registry.js.map +1 -1
  25. package/dist/evaluation/index.d.ts +22 -12
  26. package/dist/evaluation/index.js +138 -14
  27. package/dist/evaluation/index.js.map +1 -1
  28. package/dist/evaluation/types.d.ts +1 -1
  29. package/dist/harness/check/runtime-render/index.d.ts +3 -3
  30. package/dist/harness/index.d.ts +1 -1
  31. package/dist/harness/index.js +6 -6
  32. package/dist/harness/index.js.map +1 -1
  33. package/dist/harness/prompts.js +1 -4
  34. package/dist/harness/prompts.js.map +1 -1
  35. package/dist/harness/result-types.d.ts +14 -3
  36. package/dist/harness/result-types.js.map +1 -1
  37. package/dist/harness/runtime.d.ts +1 -1
  38. package/dist/harness/runtime.js +6 -7
  39. package/dist/harness/runtime.js.map +1 -1
  40. package/dist/harness/types-public.d.ts +1 -1
  41. package/dist/harness/types-public.js +4 -6
  42. package/dist/harness/types-public.js.map +1 -1
  43. package/dist/harness/validator-trace-sink.d.ts +1 -1
  44. package/dist/index.d.ts +6 -6
  45. package/dist/index.js +268 -28
  46. package/dist/index.js.map +1 -1
  47. package/dist/patch.d.ts +1 -1
  48. package/dist/{types-public-B7wn4ce5.d.ts → types-public-h8IMnYuh.d.ts} +4 -0
  49. package/dist/{visual-evaluator-C9L-o6ap.d.ts → visual-evaluator-C2strxqD.d.ts} +1 -1
  50. package/dist/workflows.d.ts +1 -1
  51. package/package.json +13 -13
  52. package/src/adapters/base.ts +6 -1
  53. package/src/adapters/claude/message-parser.ts +5 -6
  54. package/src/adapters/claude/sdk-tool-pins.ts +44 -0
  55. package/src/adapters/claude/sdk.ts +3 -3
  56. package/src/adapters/claude/source-capture.ts +20 -0
  57. package/src/adapters/provider-router.ts +1 -0
  58. package/src/boilerplate/system-prompt.ts +3 -2
  59. package/src/check/run-tier0.ts +61 -0
  60. package/src/create-ui-generator.ts +4 -2
  61. package/src/design-system-docs.ts +2 -1
  62. package/src/evaluation/axis-checks/checks/state-merge.ts +1 -1
  63. package/src/evaluation/axis-checks/helpers.ts +125 -0
  64. package/src/evaluation/criteria/resolve.ts +19 -1
  65. package/src/evaluation/evaluator.ts +2 -2
  66. package/src/evaluation/llm-evaluator.ts +2 -2
  67. package/src/evaluation/loop.ts +25 -4
  68. package/src/evaluation/message-parsing.ts +11 -12
  69. package/src/evaluation/types-public.ts +4 -0
  70. package/src/evaluation/visual-evaluator.ts +1 -1
  71. package/src/harness/coding/assemble-result.ts +1 -0
  72. package/src/harness/coding/dupe-break.ts +70 -0
  73. package/src/harness/coding/generate-task-runner.ts +5 -1
  74. package/src/harness/coding/run-coding-turn.ts +56 -1
  75. package/src/harness/result-types.ts +12 -0
  76. package/src/index.ts +1 -1
  77. package/src/taught-tokens.ts +5 -6
@@ -1,11 +1,11 @@
1
1
  import { GeneratorTier, UiGenerator } from '@ggui-ai/mcp-server-core';
2
2
  import { ModelRef, JsonObject } from '@ggui-ai/protocol';
3
3
  import { BehaviorFailure, PlaywrightModule } from '@ggui-ai/ui-visual-tester';
4
- import { C as CreateUiGeneratorOptions } from '../create-ui-generator-N4gRXpcL.js';
5
- import { a as RenderCheckIssue } from '../adapter-BJwn74rx.js';
4
+ import { C as CreateUiGeneratorOptions } from '../create-ui-generator-BrZHF0cf.js';
5
+ import { a as RenderCheckIssue } from '../adapter-h4_V69LI.js';
6
6
  import '../harness/runtime.js';
7
7
  import '../styling-profile-CxfvV48g.js';
8
- import '../types-public-B7wn4ce5.js';
8
+ import '../types-public-h8IMnYuh.js';
9
9
  import '../llm.js';
10
10
  import '../design-mode-DGWN_Gkm.js';
11
11
  import '../axes-CKVCzR_x.js';
@@ -21,7 +21,7 @@ import '../policy.js';
21
21
  import '../harness/result-types.js';
22
22
  import '../adapters/types.js';
23
23
  import '../prompts-BhNUuRoZ.js';
24
- import '../visual-evaluator-C9L-o6ap.js';
24
+ import '../visual-evaluator-C2strxqD.js';
25
25
  import 'puppeteer-core';
26
26
 
27
27
  /**
@@ -3,7 +3,7 @@ import * as esbuild2 from 'esbuild';
3
3
  import { build } from 'esbuild';
4
4
  import Anthropic from '@anthropic-ai/sdk';
5
5
  import { EXPANDED_FRAME, getCssTokens, expandedFrameScrimDecls, fillFitRule, expandedFramePanelRule, FORBIDDEN_IMPORT_SPECIFIERS } from '@ggui-ai/design/rendering';
6
- import { listContractGadgets, HOOK_NAME_RE, classifyUi, MODEL_REGISTRY, DANGEROUS_PATTERNS as DANGEROUS_PATTERNS$1, modelRefOfRoute, isGeneratorId, APP_GENERATION_PROFILE_BOUNDS, STDLIB_GADGETS_PACKAGE, resolveAppGadgets, anthropicRejectsForcedToolChoice, normalizeAnthropicModelId, anthropicRejectsSamplingParams, isRecord } from '@ggui-ai/protocol';
6
+ import { listContractGadgets, HOOK_NAME_RE, classifyUi, MODEL_REGISTRY, DANGEROUS_PATTERNS as DANGEROUS_PATTERNS$1, modelRefOfRoute, isGeneratorId, DEFAULT_MODEL as DEFAULT_MODEL$1, APP_GENERATION_PROFILE_BOUNDS, STDLIB_GADGETS_PACKAGE, resolveAppGadgets, anthropicRejectsForcedToolChoice, normalizeAnthropicModelId, anthropicRejectsSamplingParams, isRecord } from '@ggui-ai/protocol';
7
7
  import fs, { existsSync, mkdtempSync, writeFileSync, rmSync, readFileSync, readdirSync, statSync } from 'fs';
8
8
  import { contentHash } from '@ggui-ai/protocol/content-hash';
9
9
  import ts4 from 'typescript';
@@ -338,6 +338,21 @@ var init_raw = __esm({
338
338
  }
339
339
  });
340
340
 
341
+ // src/adapters/claude/source-capture.ts
342
+ function sourceCodeFromToolUse(name, input) {
343
+ if (typeof name !== "string") return void 0;
344
+ const field2 = name === "Write" ? "content" : name.endsWith(COMPILE_TOOL_SUFFIX) ? "code" : void 0;
345
+ if (field2 === void 0 || typeof input !== "object" || input === null) return void 0;
346
+ const value = Reflect.get(input, field2);
347
+ return typeof value === "string" && value.length > 0 ? value : void 0;
348
+ }
349
+ var COMPILE_TOOL_SUFFIX;
350
+ var init_source_capture = __esm({
351
+ "src/adapters/claude/source-capture.ts"() {
352
+ COMPILE_TOOL_SUFFIX = "__compile_component";
353
+ }
354
+ });
355
+
341
356
  // src/adapters/claude/message-parser.ts
342
357
  function createArtifacts() {
343
358
  return {
@@ -457,12 +472,9 @@ function extractSourceCode(message, artifacts) {
457
472
  const content = assistantMsg.message?.content;
458
473
  if (!Array.isArray(content)) return;
459
474
  for (const block of content) {
460
- if (block.type === "tool_use" && block.name === "Write") {
461
- const input = block.input;
462
- if (input?.content) {
463
- artifacts.sourceCode = input.content;
464
- }
465
- }
475
+ if (block.type !== "tool_use") continue;
476
+ const value = sourceCodeFromToolUse(block.name, block.input);
477
+ if (value !== void 0) artifacts.sourceCode = value;
466
478
  }
467
479
  }
468
480
  function unescapeJsonString(s) {
@@ -500,6 +512,7 @@ function extractGeneratorMeta(message, artifacts) {
500
512
  }
501
513
  var init_message_parser = __esm({
502
514
  "src/adapters/claude/message-parser.ts"() {
515
+ init_source_capture();
503
516
  }
504
517
  });
505
518
 
@@ -595,6 +608,17 @@ var init_claude_code_login = __esm({
595
608
  }
596
609
  });
597
610
 
611
+ // src/adapters/claude/sdk-tool-pins.ts
612
+ function sdkToolPins(input) {
613
+ const builtins = [...input.builtinTools ?? []];
614
+ const approved = input.allowedTools !== void 0 ? [...input.allowedTools] : Object.keys(input.mcpServers ?? {}).map((name) => `mcp__${name}`);
615
+ return { tools: builtins, settingSources: [], allowedTools: [...approved, ...builtins] };
616
+ }
617
+ var init_sdk_tool_pins = __esm({
618
+ "src/adapters/claude/sdk-tool-pins.ts"() {
619
+ }
620
+ });
621
+
598
622
  // src/adapters/claude/sdk.ts
599
623
  var sdk_exports = {};
600
624
  __export(sdk_exports, {
@@ -621,6 +645,7 @@ var init_sdk = __esm({
621
645
  init_message_parser();
622
646
  init_tool_bridge2();
623
647
  init_claude_code_login();
648
+ init_sdk_tool_pins();
624
649
  ClaudeSdkAdapter = class extends GeneratorAdapter {
625
650
  provider = "claude";
626
651
  mode = "sdk";
@@ -660,9 +685,8 @@ var init_sdk = __esm({
660
685
  systemPrompt: params.systemPrompt,
661
686
  ...mcpServers && { mcpServers },
662
687
  ...cfg.agents && { agents: cfg.agents },
663
- ...allowedTools && { allowedTools },
664
- permissionMode: "bypassPermissions",
665
- allowDangerouslySkipPermissions: true,
688
+ // ggui#1714 — no CLI built-ins unless asked for, no host settings, no permission bypass.
689
+ ...sdkToolPins({ mcpServers, allowedTools, builtinTools: cfg.builtinTools }),
666
690
  ...cfg.cliPath && { pathToClaudeCodeExecutable: cfg.cliPath },
667
691
  ...cfg.stderr && { stderr: cfg.stderr },
668
692
  // ggui#1185 — tool-less, config-less, non-bare; the model is pinned above.
@@ -2944,10 +2968,7 @@ var init_design_mode = __esm({
2944
2968
  var NOT_YET_TAUGHT_TOKENS, taughtTokenManifest;
2945
2969
  var init_taught_tokens = __esm({
2946
2970
  "src/taught-tokens.ts"() {
2947
- NOT_YET_TAUGHT_TOKENS = /* @__PURE__ */ new Set([
2948
- "--ggui-color-controlAccentOutline",
2949
- "--ggui-color-controlOutline"
2950
- ]);
2971
+ NOT_YET_TAUGHT_TOKENS = /* @__PURE__ */ new Set([]);
2951
2972
  taughtTokenManifest = consumedTokenManifest.filter(
2952
2973
  (token) => !NOT_YET_TAUGHT_TOKENS.has(token)
2953
2974
  );
@@ -3467,6 +3488,40 @@ function detectComponentNeverReturns(sourceCode) {
3467
3488
  for (const st of body.statements) if (!isFunctionLike(st)) visit(st);
3468
3489
  return returnsValue ? null : { jsxStatementLine };
3469
3490
  }
3491
+ function kitControlTags(source) {
3492
+ const out = [];
3493
+ const open = new RegExp(`<(${KIT_CONTROLS.join("|")})\\b`, "g");
3494
+ for (const m of source.matchAll(open)) {
3495
+ const start = (m.index ?? 0) + m[0].length;
3496
+ let depth = 0;
3497
+ let quote = null;
3498
+ let i = start;
3499
+ for (; i < source.length; i += 1) {
3500
+ const c = source[i];
3501
+ if (quote !== null) {
3502
+ if (c === quote && source[i - 1] !== "\\") quote = null;
3503
+ } else if (c === '"' || c === "'" || c === "`") quote = c;
3504
+ else if (c === "{") depth += 1;
3505
+ else if (c === "}") depth -= 1;
3506
+ else if (c === ">" && depth === 0) break;
3507
+ }
3508
+ out.push({ name: m[1] ?? "", attrs: source.slice(start, i), line: source.slice(0, m.index ?? 0).split("\n").length });
3509
+ }
3510
+ return out;
3511
+ }
3512
+ function inlineBorderKey(attrs, source) {
3513
+ const BORDER = /\b(border(?:Top|Right|Bottom|Left)?(?:Color|Width|Style)?)\s*:/;
3514
+ const style = attrs.match(/\bstyle=\{\{([\s\S]*)\}\}/);
3515
+ if (style !== null) {
3516
+ const hit2 = (style[1] ?? "").match(BORDER);
3517
+ if (hit2 !== null) return hit2[1] ?? null;
3518
+ }
3519
+ const named = attrs.match(/\bstyle=\{\s*([A-Za-z_$][\w$]*)\s*\}/);
3520
+ if (named === null) return null;
3521
+ const decl = source.match(new RegExp(`\\b(?:const|let)\\s+${named[1]}\\b[^=]*=\\s*\\{([^{}]*)\\}`));
3522
+ const hit = decl !== null ? (decl[1] ?? "").match(BORDER) : null;
3523
+ return hit !== null ? hit[1] ?? null : null;
3524
+ }
3470
3525
  async function runTier0Checks(sourceCode, contract, buildErrors, gadgetTypes, designMode = DEFAULT_DESIGN_MODE) {
3471
3526
  const issues = [];
3472
3527
  const lines = sourceCode.split("\n");
@@ -3864,6 +3919,20 @@ ${errorDetail}`,
3864
3919
  });
3865
3920
  }
3866
3921
  }
3922
+ for (const tag of enforceDesignVocabulary ? kitControlTags(sourceCode) : []) {
3923
+ const key = inlineBorderKey(tag.attrs, sourceCode);
3924
+ if (key === null) continue;
3925
+ issues.push({
3926
+ tier: 0,
3927
+ result: "fail",
3928
+ category: "tokens",
3929
+ subcategory: "control-inline-border",
3930
+ severity: "critical",
3931
+ description: `control-inline-border: <${tag.name}> sets an inline \`${key}\` \u2014 a kit control draws its own boundary, and an inline border paints over it`,
3932
+ fix: `Remove the border from <${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).`,
3933
+ line: tag.line
3934
+ });
3935
+ }
3867
3936
  if (!sourceCode.includes("interface Props") && !sourceCode.includes("type Props")) {
3868
3937
  issues.push({
3869
3938
  tier: 0,
@@ -4135,7 +4204,7 @@ async function runTier0(sourceCode, contract, designMode = DEFAULT_DESIGN_MODE)
4135
4204
  }
4136
4205
  return { issues, pass };
4137
4206
  }
4138
- var MANIFEST_TOKENS, RETIRED_TOKEN_SUCCESSORS, TOKEN_REF_RE, HOOK_NAME_FOR, CONTRACT_FIELD_FOR;
4207
+ var MANIFEST_TOKENS, RETIRED_TOKEN_SUCCESSORS, TOKEN_REF_RE, HOOK_NAME_FOR, CONTRACT_FIELD_FOR, KIT_CONTROLS;
4139
4208
  var init_run_tier0 = __esm({
4140
4209
  "src/check/run-tier0.ts"() {
4141
4210
  init_design_mode();
@@ -4164,6 +4233,7 @@ var init_run_tier0 = __esm({
4164
4233
  stream: "streamSpec",
4165
4234
  context: "contextSpec"
4166
4235
  };
4236
+ KIT_CONTROLS = ["Button", "Input", "Select", "TextArea", "Checkbox", "RadioGroup", "Toggle", "SearchField"];
4167
4237
  }
4168
4238
  });
4169
4239
 
@@ -4457,6 +4527,8 @@ var init_policy = __esm({
4457
4527
  function getBedrockModelId(model) {
4458
4528
  const BEDROCK_MAP = {
4459
4529
  "anthropic/claude-haiku-4-5": "us.anthropic.claude-haiku-4-5-20251001-v1:0",
4530
+ "anthropic/claude-haiku-5-5": "us.anthropic.claude-haiku-5-5",
4531
+ // listed ACTIVE in us-east-1 (ggui#1743)
4460
4532
  "anthropic/claude-sonnet-4-6": "us.anthropic.claude-sonnet-4-6",
4461
4533
  "anthropic/claude-opus-4-6": "us.anthropic.claude-opus-4-6-v1:0",
4462
4534
  // Claude 5-family (ggui#706, strings verified 2026-09-02 against
@@ -6973,10 +7045,16 @@ var dupe_break_exports = {};
6973
7045
  __export(dupe_break_exports, {
6974
7046
  SAME_EXCHANGE_BREAK_AT: () => SAME_EXCHANGE_BREAK_AT,
6975
7047
  SAME_EXCHANGE_NUDGE: () => SAME_EXCHANGE_NUDGE,
7048
+ SAME_FAILURE_BREAK_AT: () => SAME_FAILURE_BREAK_AT,
7049
+ SAME_FAILURE_NUDGE: () => SAME_FAILURE_NUDGE,
7050
+ SAME_FAILURE_NUDGE_AT: () => SAME_FAILURE_NUDGE_AT,
7051
+ clearWriteFailures: () => clearWriteFailures,
6976
7052
  computePatchFingerprint: () => computePatchFingerprint,
6977
7053
  createDupeBreakState: () => createDupeBreakState,
6978
7054
  errorClassBucket: () => errorClassBucket,
6979
- recordToolExchange: () => recordToolExchange
7055
+ failureSignatures: () => failureSignatures,
7056
+ recordToolExchange: () => recordToolExchange,
7057
+ recordWriteFailure: () => recordWriteFailure
6980
7058
  });
6981
7059
  function recordToolExchange(state, name, input, resultText) {
6982
7060
  const key = createHash("sha256").update(name).update("\0").update(JSON.stringify(input) ?? "undefined").update("\0").update(resultText).digest("hex");
@@ -6988,6 +7066,25 @@ function recordToolExchange(state, name, input, resultText) {
6988
7066
  }
6989
7067
  return state.consecutiveIdenticalExchanges;
6990
7068
  }
7069
+ function failureSignatures(resultText) {
7070
+ const norm = (text) => text.replace(/[0-9]+/g, "#").replace(/\s+/g, " ").trim();
7071
+ const violations = [...resultText.matchAll(/^\[[^\]\n]+\] [^\n]+/gm)].map((m) => norm(m[0]));
7072
+ if (violations.length > 0) return [...new Set(violations)];
7073
+ const rest = norm(resultText.replace(/^Committed [^\n]*\n/, ""));
7074
+ return rest.length > 0 ? [rest] : [];
7075
+ }
7076
+ function recordWriteFailure(state, resultText) {
7077
+ let top;
7078
+ for (const signature of failureSignatures(resultText)) {
7079
+ const repeats = (state.failureCounts.get(signature) ?? 0) + 1;
7080
+ state.failureCounts.set(signature, repeats);
7081
+ if (top === void 0 || repeats > top.repeats) top = { signature, repeats };
7082
+ }
7083
+ return top;
7084
+ }
7085
+ function clearWriteFailures(state) {
7086
+ state.failureCounts.clear();
7087
+ }
6991
7088
  function createDupeBreakState() {
6992
7089
  return {
6993
7090
  lastFailedPatchFingerprint: null,
@@ -7004,7 +7101,8 @@ function createDupeBreakState() {
7004
7101
  consecutiveBrokenApplies: 0,
7005
7102
  consecutiveNoToolCalls: 0,
7006
7103
  lastExchangeKey: null,
7007
- consecutiveIdenticalExchanges: 0
7104
+ consecutiveIdenticalExchanges: 0,
7105
+ failureCounts: /* @__PURE__ */ new Map()
7008
7106
  };
7009
7107
  }
7010
7108
  function errorClassBucket(errText) {
@@ -7029,11 +7127,14 @@ function computePatchFingerprint(changes, errText) {
7029
7127
  h.update(errBucket);
7030
7128
  return h.digest("hex");
7031
7129
  }
7032
- var SAME_EXCHANGE_BREAK_AT, SAME_EXCHANGE_NUDGE;
7130
+ var SAME_EXCHANGE_BREAK_AT, SAME_EXCHANGE_NUDGE, SAME_FAILURE_NUDGE_AT, SAME_FAILURE_BREAK_AT, SAME_FAILURE_NUDGE;
7033
7131
  var init_dupe_break = __esm({
7034
7132
  "src/harness/coding/dupe-break.ts"() {
7035
7133
  SAME_EXCHANGE_BREAK_AT = 3;
7036
7134
  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.";
7135
+ SAME_FAILURE_NUDGE_AT = 3;
7136
+ SAME_FAILURE_BREAK_AT = 5;
7137
+ 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.";
7037
7138
  }
7038
7139
  });
7039
7140
 
@@ -7425,13 +7526,13 @@ Return your evaluation using the evaluate_${criterion} tool.`;
7425
7526
  function getDefaultEvalModel(provider) {
7426
7527
  switch (provider) {
7427
7528
  case "claude":
7428
- return "claude-haiku-4-5-20251001";
7529
+ return "claude-haiku-5-5";
7429
7530
  case "openai":
7430
7531
  return "gpt-5.6-luna";
7431
7532
  case "google":
7432
7533
  return "gemini-3.5-flash-lite";
7433
7534
  default:
7434
- return "claude-haiku-4-5-20251001";
7535
+ return "claude-haiku-5-5";
7435
7536
  }
7436
7537
  }
7437
7538
  function callTokens(response) {
@@ -7891,6 +7992,7 @@ function instrumentRead(kind, m) {
7891
7992
  return { verdict: "n/a", evidence: `fill ${m.inkRatio.toFixed(3)} reported; no floor ruled` };
7892
7993
  }
7893
7994
  function readRow(row, answers, m, unansweredReason) {
7995
+ if (row.evaluation === "judge" && REQUIRES_POST_ACTION_CAPTURE.has(row.id) && m.postActionCapture !== true) return { verdict: "n/a", evidence: NO_POST_ACTION };
7894
7996
  if (row.evaluation === "judge") return majority(row.id, answers, unansweredReason);
7895
7997
  if (row.evaluation === "human") return { verdict: "n/a", evidence: `reader's column (status ${row.status})` };
7896
7998
  if (row.status !== "live") return { verdict: "n/a", evidence: `not implemented (status ${row.status})` };
@@ -7949,14 +8051,14 @@ ${text}
7949
8051
  }
7950
8052
  function buildCriteriaJudgeBlock(bank, selected, inputs = {}) {
7951
8053
  const byId = new Map(bankRows(bank).map((r) => [r.id, r]));
7952
- const asked = selected.selection.map((s) => byId.get(s.id)).filter((r) => r !== void 0 && r.evaluation === "judge");
8054
+ 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));
7953
8055
  if (asked.length === 0) return "";
7954
8056
  const lines = asked.map((r) => `- ${r.id} (${r.severity}): ${r.text} Evidence must name: ${r.evidence}`);
7955
8057
  return "## Criteria \u2014 answer each one from the FRAME only\nRead these off the screenshot. Never infer them from the request, the styling profile or any direction text above.\n" + (inputs.frame !== void 0 ? `${criteriaFrameLine(inputs.frame)}
7956
8058
  ` : "") + (inputs.propsJson !== void 0 ? criteriaPropsSection(inputs.propsJson) : "") + `${lines.join("\n")}
7957
8059
  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.`;
7958
8060
  }
7959
- var INSTRUMENT_BY_ID, INSTRUMENT_BESIDE_JUDGE, CRITERIA_PROPS_MAX_CHARS;
8061
+ var INSTRUMENT_BY_ID, INSTRUMENT_BESIDE_JUDGE, REQUIRES_POST_ACTION_CAPTURE, NO_POST_ACTION, CRITERIA_PROPS_MAX_CHARS;
7960
8062
  var init_resolve = __esm({
7961
8063
  "src/evaluation/criteria/resolve.ts"() {
7962
8064
  init_bank();
@@ -7969,6 +8071,8 @@ var init_resolve = __esm({
7969
8071
  INSTRUMENT_BESIDE_JUDGE = {
7970
8072
  "space.edge": "edge"
7971
8073
  };
8074
+ REQUIRES_POST_ACTION_CAPTURE = /* @__PURE__ */ new Set(["state.feedback"]);
8075
+ NO_POST_ACTION = "not evaluated: no post-action capture";
7972
8076
  CRITERIA_PROPS_MAX_CHARS = 2e3;
7973
8077
  }
7974
8078
  });
@@ -9506,7 +9610,7 @@ function parseCriteriaAnswers(raw) {
9506
9610
  function getDefaultVisualModel(provider) {
9507
9611
  switch (provider) {
9508
9612
  case "claude":
9509
- return "claude-haiku-4-5-20251001";
9613
+ return "claude-haiku-5-5";
9510
9614
  case "google":
9511
9615
  return "gemini-3-flash-preview";
9512
9616
  }
@@ -10140,6 +10244,7 @@ These tokens adapt automatically to any theme (light, dark, branded). **ALWAYS u
10140
10244
  | 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 |
10141
10245
  | outline | var(--ggui-color-outline) | Borders, dividers |
10142
10246
  | outlineVariant | var(--ggui-color-outlineVariant) | Subtle borders |
10247
+ | controlOutline | var(--ggui-color-controlOutline) | A control's edge \u2014 clears 3:1 on its surface; kit controls draw it themselves |
10143
10248
 
10144
10249
  **Usage pattern:**
10145
10250
  - Page/section background \u2192 \`var(--ggui-color-ground)\`
@@ -10264,7 +10369,7 @@ Use \`var(--ggui-spacing-N)\` for all padding, gap, and margin values. **Never u
10264
10369
  ### Color usage guide
10265
10370
  - **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
10266
10371
  - **primary-50/100**: Section backgrounds, highlight strips, card headers
10267
- - **primary-200/300**: Borders, dividers, focus rings, input outlines
10372
+ - **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
10268
10373
  - **primary-500/600**: Filled buttons, CTAs, badges, icon fills, active indicators \u2014 fills and borders, NOT a text color
10269
10374
  - **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\`)
10270
10375
  - **primary-700/800/900**: Headings and text on light primary backgrounds
@@ -11749,10 +11854,96 @@ function getInitialValuePropNames(contract) {
11749
11854
  }
11750
11855
  return names;
11751
11856
  }
11857
+ function objectLiteralKeysAt(src, open) {
11858
+ if (src[open] !== "{") return void 0;
11859
+ const keys = [];
11860
+ let depth = 0;
11861
+ let expectKey = true;
11862
+ for (let i = open; i < src.length; i++) {
11863
+ const c = src[i];
11864
+ if (c === '"' || c === "'" || c === "`") {
11865
+ let j = i + 1;
11866
+ while (j < src.length && src[j] !== c) j += src[j] === "\\" ? 2 : 1;
11867
+ if (depth === 1 && expectKey && /^\s*:/.test(src.slice(j + 1))) keys.push(src.slice(i + 1, j));
11868
+ i = j;
11869
+ expectKey = false;
11870
+ continue;
11871
+ }
11872
+ if (c === "{" || c === "[" || c === "(") {
11873
+ depth++;
11874
+ if (depth === 1) expectKey = true;
11875
+ continue;
11876
+ }
11877
+ if (c === "}" || c === "]" || c === ")") {
11878
+ depth--;
11879
+ if (depth === 0) return keys;
11880
+ continue;
11881
+ }
11882
+ if (depth !== 1) continue;
11883
+ if (c === ",") {
11884
+ expectKey = true;
11885
+ continue;
11886
+ }
11887
+ if (expectKey && /[A-Za-z_$]/.test(c ?? "")) {
11888
+ const m = /^([A-Za-z_$][\w$]*)\s*(:|,|\})/.exec(src.slice(i));
11889
+ if (m !== null && m[1] !== void 0) keys.push(m[1]);
11890
+ expectKey = false;
11891
+ }
11892
+ }
11893
+ return keys;
11894
+ }
11895
+ function callArgsAt(src, open) {
11896
+ const args = [];
11897
+ let depth = 0;
11898
+ let start = open + 1;
11899
+ for (let i = open; i < src.length; i++) {
11900
+ const c = src[i];
11901
+ if (c === '"' || c === "'" || c === "`") {
11902
+ let j = i + 1;
11903
+ while (j < src.length && src[j] !== c) j += src[j] === "\\" ? 2 : 1;
11904
+ i = j;
11905
+ continue;
11906
+ }
11907
+ if (c === "(" || c === "{" || c === "[") depth++;
11908
+ else if (c === ")" || c === "}" || c === "]") {
11909
+ depth--;
11910
+ if (depth === 0) {
11911
+ args.push(src.slice(start, i));
11912
+ return args;
11913
+ }
11914
+ } else if (c === "," && depth === 1) {
11915
+ args.push(src.slice(start, i));
11916
+ start = i + 1;
11917
+ }
11918
+ }
11919
+ return args;
11920
+ }
11921
+ function initialStateKeys(src, argStart, argText) {
11922
+ const lead = argText.length - argText.trimStart().length;
11923
+ let at = argStart + lead;
11924
+ let text = argText.trim();
11925
+ const lazy = /^\(\s*\)\s*=>\s*\(?\s*/.exec(text);
11926
+ if (lazy !== null) {
11927
+ at += lazy[0].length;
11928
+ text = text.slice(lazy[0].length);
11929
+ }
11930
+ if (text.startsWith("{")) return objectLiteralKeysAt(src, at) ?? [];
11931
+ const ident = /^([A-Za-z_$][\w$]*)\s*\)?\s*$/.exec(text);
11932
+ if (ident === null || ident[1] === void 0) return [];
11933
+ const decls = [...src.matchAll(new RegExp(`(?:const|let|var)\\s+${ident[1]}\\s*(?::[^=]{1,200})?=\\s*`, "g"))];
11934
+ const decl = decls.length === 1 ? decls[0] : void 0;
11935
+ if (decl === void 0) return [];
11936
+ return objectLiteralKeysAt(src, (decl.index ?? 0) + decl[0].length) ?? [];
11937
+ }
11752
11938
  function collectStateKeys(src) {
11753
11939
  const keys = /* @__PURE__ */ new Set();
11754
11940
  const varRe = /const\s*\[\s*(\w+)\s*,\s*(\w+)\s*\]\s*=\s*useState/g;
11755
11941
  for (const m of src.matchAll(varRe)) keys.add(m[1]);
11942
+ const contextRe = /const\s*\[\s*(\w+)\s*,\s*\w+\s*\]\s*=\s*useGguiContext\s*(?:<[^>]*>)?\s*\(\s*['"`](\w+)['"`]/g;
11943
+ for (const m of src.matchAll(contextRe)) {
11944
+ keys.add(m[1]);
11945
+ keys.add(m[2]);
11946
+ }
11756
11947
  const objRe = /useState(?:<[^>]*>)?\s*\(\s*\{([\s\S]*?)\}\s*\)/g;
11757
11948
  for (const m of src.matchAll(objRe)) {
11758
11949
  const keyRe = /(?:^|,)\s*(\w+)\s*:/g;
@@ -11763,6 +11954,16 @@ function collectStateKeys(src) {
11763
11954
  const keyRe = /(?:^|,)\s*(\w+)\s*:/g;
11764
11955
  for (const km of m[1].matchAll(keyRe)) keys.add(km[1]);
11765
11956
  }
11957
+ for (const m of src.matchAll(/\b(useState|useReducer)\s*(?:<[^>]*>)?\s*\(/g)) {
11958
+ const open = (m.index ?? 0) + m[0].length - 1;
11959
+ const args = callArgsAt(src, open);
11960
+ const which = m[1] === "useReducer" ? 1 : 0;
11961
+ const arg = args[which];
11962
+ if (arg === void 0) continue;
11963
+ let argStart = open + 1;
11964
+ for (let k = 0; k < which; k++) argStart += (args[k]?.length ?? 0) + 1;
11965
+ for (const key of initialStateKeys(src, argStart, arg)) keys.add(key);
11966
+ }
11766
11967
  return keys;
11767
11968
  }
11768
11969
  function mkIssue(subcategory, description, fix, result = "fail") {
@@ -12375,7 +12576,7 @@ function runNoHardcodedEntities(input) {
12375
12576
  mkIssue(
12376
12577
  "state.merge.no_hardcoded_entities",
12377
12578
  `Hardcoded entity array literal (multiple objects with "${idField}") in the component \u2014 entity data should come from state/props.`,
12378
- `Remove the literal array. Seed state from props.{entityProp} and merge updates via stream.`
12579
+ `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.`
12379
12580
  )
12380
12581
  );
12381
12582
  }
@@ -15267,13 +15468,14 @@ MANDATORY:
15267
15468
  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))\`.
15268
15469
  3. NEVER invent your own palette. The system provides primary, neutral, success, warning, error, and info \u2014 use ONLY these.
15269
15470
  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.
15471
+ 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).
15270
15472
 
15271
15473
  Token categories:
15272
15474
  - Brand fills: \`var(--ggui-color-primary-600)\`, \`var(--ggui-color-primary-50)\`
15273
15475
  - Accent text: \`var(--ggui-color-link)\` \u2014 eyebrows, taglines, links, labels; readable per mode (never a bare \`primary-<stop>\` as a text color)
15274
15476
  - Text: \`var(--ggui-color-onContainer)\`, \`var(--ggui-color-onSunken)\`
15275
15477
  - Backgrounds: \`var(--ggui-color-container)\`, \`var(--ggui-color-sunken)\`
15276
- - Borders: \`var(--ggui-color-outline)\`
15478
+ - Borders: \`var(--ggui-color-outline)\` \u2014 dividers and the edges of non-interactive boxes; a control's edge is the kit's (rule 5)
15277
15479
  - Spacing: \`var(--ggui-spacing-4)\`, \`var(--ggui-spacing-6)\`
15278
15480
  - Typography: \`var(--ggui-font-size-sm)\`, \`var(--ggui-font-weight-semibold)\`
15279
15481
  - Shadows: \`var(--ggui-shape-shadow-sm)\`, \`var(--ggui-shape-shadow-md)\`, \`var(--ggui-shape-shadow-lg)\`
@@ -15288,7 +15490,7 @@ Use the FULL primary palette throughout the component \u2014 NOT only on submit
15288
15490
  | Element | Token | Purpose |
15289
15491
  |---------|-------|---------|
15290
15492
  | Section headers, hero areas, highlight strips | \`primary-50\` / \`primary-100\` | Subtle branded backgrounds |
15291
- | Borders, dividers, focus rings, input focus | \`primary-200\` / \`primary-300\` | Branded structure |
15493
+ | Dividers, section rules, edges of non-interactive boxes | \`primary-200\` / \`primary-300\` | Branded structure \u2014 never a control's border (rule 5) |
15292
15494
  | 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 |
15293
15495
  | Icon fills, badges, active indicators | \`primary-500\` / \`primary-600\` | Core accent fill |
15294
15496
  | Buttons, CTAs, filled interactive elements | \`primary-600\` / \`primary-700\` | Primary actions |
@@ -18260,6 +18462,7 @@ async function assembleGenerationResult(input) {
18260
18462
  evalResult: telemetry.evalResult !== void 0 && telemetry.contractFeedback !== void 0 ? { ...telemetry.evalResult, contractFeedback: telemetry.contractFeedback } : telemetry.evalResult,
18261
18463
  // ggui#404 — present only when the same-exchange guard ended the run.
18262
18464
  ...telemetry.sameExchangeBreak !== void 0 ? { sameExchangeBreak: telemetry.sameExchangeBreak } : {},
18465
+ ...telemetry.sameFailureBreak !== void 0 ? { sameFailureBreak: telemetry.sameFailureBreak } : {},
18263
18466
  needsBackgroundImprovement,
18264
18467
  selfCheckPassed: telemetry.selfCheckPassed,
18265
18468
  timing,
@@ -19199,6 +19402,18 @@ Repeated error: ${errText}
19199
19402
  augmentedResultText += SAME_EXCHANGE_NUDGE;
19200
19403
  }
19201
19404
  }
19405
+ const isWrite = call.name === "apply_changes" || call.name === "rewrite";
19406
+ let repeatedFailure;
19407
+ if (ctx.dupeBreak && isWrite) {
19408
+ if (result.done) {
19409
+ clearWriteFailures(ctx.dupeBreak);
19410
+ } else {
19411
+ repeatedFailure = recordWriteFailure(ctx.dupeBreak, result.result);
19412
+ if (repeatedFailure?.repeats === SAME_FAILURE_NUDGE_AT) {
19413
+ augmentedResultText += SAME_FAILURE_NUDGE;
19414
+ }
19415
+ }
19416
+ }
19202
19417
  await codingAgent.sendToolResult([{
19203
19418
  callId: call.id,
19204
19419
  name: call.name,
@@ -19227,6 +19442,29 @@ Repeated error: ${errText}
19227
19442
  sameExchangeBreak: { tool: call.name, repeats: sameExchangeRepeats }
19228
19443
  };
19229
19444
  }
19445
+ if (repeatedFailure !== void 0 && repeatedFailure.repeats >= SAME_FAILURE_BREAK_AT) {
19446
+ const signature = repeatedFailure.signature.slice(0, 160);
19447
+ console.log(
19448
+ `[coding-agent] SAME_FAILURE_BREAK | tool=${call.name} | repeats=${repeatedFailure.repeats} | failure=${JSON.stringify(signature.slice(0, 80))} \u2014 terminating run (ggui#1791)`
19449
+ );
19450
+ return {
19451
+ control: "break",
19452
+ compiledCode: "",
19453
+ selfCheckPassed: false,
19454
+ outcome: void 0,
19455
+ phase: void 0,
19456
+ tokens,
19457
+ ...cacheTokens,
19458
+ llmMs,
19459
+ toolMs,
19460
+ lastResultText: augmentedResultText,
19461
+ lastDiffFailed: false,
19462
+ isEvalFeedback: false,
19463
+ iconNamesCache,
19464
+ preWarmedContext,
19465
+ sameFailureBreak: { signature, repeats: repeatedFailure.repeats }
19466
+ };
19467
+ }
19230
19468
  const turnDiffFailed = !!result.error && result.result.includes("DIFF");
19231
19469
  const isDiffFail = turnDiffFailed && !isPatchInvalid;
19232
19470
  const outcome = result.done ? "PASS" : isPatchInvalid ? "PATCH_INVALID" : isDiffFail ? "DIFF_FAIL" : "SELF_CHECK_FAIL";
@@ -20214,6 +20452,7 @@ function createTelemetry() {
20214
20452
  probeRepairTrigger: void 0,
20215
20453
  contractFeedback: void 0,
20216
20454
  sameExchangeBreak: void 0,
20455
+ sameFailureBreak: void 0,
20217
20456
  compiledCode: "",
20218
20457
  pairedSource: "",
20219
20458
  selfCheckPassed: false,
@@ -20269,6 +20508,7 @@ function absorbTurn(telemetry, turn) {
20269
20508
  else telemetry.counters.outcomes.selfCheckFail++;
20270
20509
  }
20271
20510
  if (turn.sameExchangeBreak !== void 0) telemetry.sameExchangeBreak = turn.sameExchangeBreak;
20511
+ if (turn.sameFailureBreak !== void 0) telemetry.sameFailureBreak = turn.sameFailureBreak;
20272
20512
  }
20273
20513
  function createGenerateTaskRunner(input) {
20274
20514
  const { session, params, classification, telemetry } = input;
@@ -21423,7 +21663,7 @@ function injectVariance(userPrompt, variance, options) {
21423
21663
  init_design_mode();
21424
21664
  init_provider_router();
21425
21665
  var DEFAULT_TIER = "default";
21426
- var DEFAULT_MODEL = "anthropic/claude-haiku-4-5";
21666
+ var DEFAULT_MODEL = DEFAULT_MODEL$1;
21427
21667
  function createUiGenerator(options = {}) {
21428
21668
  const {
21429
21669
  enableRuntimeRender = false,