@ggui-ai/ui-gen 0.24.0 → 0.25.0

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 (98) hide show
  1. package/dist/{adapter-Du7P_5Fx.d.ts → adapter-Cg_rkpGq.d.ts} +1 -1
  2. package/dist/adapters/generation-dispatch.d.ts +27 -5
  3. package/dist/adapters/generation-dispatch.js +909 -53
  4. package/dist/adapters/generation-dispatch.js.map +1 -1
  5. package/dist/adapters/index.js +3 -1
  6. package/dist/adapters/index.js.map +1 -1
  7. package/dist/advanced/index.d.ts +12 -10
  8. package/dist/advanced/index.js +917 -98
  9. package/dist/advanced/index.js.map +1 -1
  10. package/dist/{axes-CzLEMDeB.d.ts → axes-CKVCzR_x.d.ts} +1 -1
  11. package/dist/bank-CT4eEkta.d.ts +118 -0
  12. package/dist/boilerplate.js +24 -12
  13. package/dist/boilerplate.js.map +1 -1
  14. package/dist/check/index.d.ts +11 -4
  15. package/dist/check/index.js +47 -1
  16. package/dist/check/index.js.map +1 -1
  17. package/dist/classifier/index.d.ts +2 -2
  18. package/dist/compose.d.ts +2 -2
  19. package/dist/{create-ui-generator-GIhzx08x.d.ts → create-ui-generator-DFnnArrv.d.ts} +11 -3
  20. package/dist/{dispatch-BnAH56p4.d.ts → dispatch-CStl9eAw.d.ts} +5 -3
  21. package/dist/evaluation/axis-checks/index.d.ts +5 -5
  22. package/dist/evaluation/axis-checks/index.js +90 -2
  23. package/dist/evaluation/axis-checks/index.js.map +1 -1
  24. package/dist/evaluation/axis-checks/registry.d.ts +3 -3
  25. package/dist/evaluation/axis-checks/registry.js +88 -1
  26. package/dist/evaluation/axis-checks/registry.js.map +1 -1
  27. package/dist/evaluation/index.d.ts +300 -23
  28. package/dist/evaluation/index.js +746 -30
  29. package/dist/evaluation/index.js.map +1 -1
  30. package/dist/evaluation/types.d.ts +9 -1
  31. package/dist/evaluation/types.js.map +1 -1
  32. package/dist/fragments/index.d.ts +3 -3
  33. package/dist/harness/check/runtime-render/index.d.ts +5 -5
  34. package/dist/harness/index.d.ts +4 -4
  35. package/dist/harness/index.js +33 -15
  36. package/dist/harness/index.js.map +1 -1
  37. package/dist/harness/result-types.d.ts +3 -3
  38. package/dist/harness/runtime.d.ts +10 -4
  39. package/dist/harness/runtime.js +26 -14
  40. package/dist/harness/runtime.js.map +1 -1
  41. package/dist/harness/types-public.d.ts +3 -3
  42. package/dist/harness/types-public.js +33 -15
  43. package/dist/harness/types-public.js.map +1 -1
  44. package/dist/harness/validator-trace-sink.d.ts +3 -3
  45. package/dist/hash.d.ts +1 -1
  46. package/dist/index.d.ts +25 -9
  47. package/dist/index.js +926 -101
  48. package/dist/index.js.map +1 -1
  49. package/dist/{llm-router-DwwUK1bz.d.ts → llm-router-nfO7wPec.d.ts} +3 -1
  50. package/dist/patch.d.ts +3 -3
  51. package/dist/policy.d.ts +1 -1
  52. package/dist/tools/get-primitives-ts.d.ts +1 -1
  53. package/dist/tools/get-primitives-ts.js +1 -1
  54. package/dist/tools/get-primitives-ts.js.map +1 -1
  55. package/dist/{types-BOvHNG7K.d.ts → types-Dh1s4K9j.d.ts} +1 -1
  56. package/dist/{types-public-Bk6d1WLN.d.ts → types-public-BJTLpzWn.d.ts} +134 -7
  57. package/dist/validation/index.d.ts +1 -1
  58. package/dist/validation/index.js +1 -1
  59. package/dist/validation/index.js.map +1 -1
  60. package/dist/workflows.d.ts +3 -3
  61. package/package.json +12 -12
  62. package/src/adapters/generation-dispatch.ts +36 -2
  63. package/src/adapters/tools.ts +2 -1
  64. package/src/boilerplate/free-design-prompt.ts +1 -1
  65. package/src/boilerplate/generate.ts +26 -1
  66. package/src/boilerplate/hard-sections.ts +13 -13
  67. package/src/boilerplate/system-prompt.ts +1 -1
  68. package/src/check/index.ts +1 -0
  69. package/src/check/sample-props.ts +63 -0
  70. package/src/coding-agent/tools.ts +4 -1
  71. package/src/create-harness.ts +3 -3
  72. package/src/create-ui-generator.ts +20 -1
  73. package/src/evaluation/__fixtures__/png.ts +100 -0
  74. package/src/evaluation/axis-checks/checks/state-merge.ts +1 -1
  75. package/src/evaluation/axis-checks/checks/universal.ts +111 -0
  76. package/src/evaluation/axis-checks/dispatch.ts +4 -1
  77. package/src/evaluation/criteria/bank.ts +85 -0
  78. package/src/evaluation/criteria/context.ts +52 -0
  79. package/src/evaluation/criteria/resolve.ts +159 -0
  80. package/src/evaluation/criteria/select.ts +79 -0
  81. package/src/evaluation/index.ts +23 -0
  82. package/src/evaluation/ink-extent.ts +200 -0
  83. package/src/evaluation/theme-css.ts +12 -5
  84. package/src/evaluation/types-public.ts +130 -3
  85. package/src/evaluation/types.ts +9 -0
  86. package/src/evaluation/visual-evaluator.ts +636 -30
  87. package/src/harness/coding/generate-task-runner.ts +3 -0
  88. package/src/harness/coding/run-coding-turn.ts +7 -3
  89. package/src/harness/coding/run-eval-round.ts +19 -0
  90. package/src/harness/runtime.ts +5 -0
  91. package/src/harness-tool-names.ts +38 -0
  92. package/src/index.ts +12 -0
  93. package/src/run-check.ts +2 -0
  94. package/src/tools/get-primitives-ts.ts +1 -1
  95. package/src/tools/get-wire.ts +1 -1
  96. package/src/tools/render-check.ts +0 -60
  97. package/src/tools.ts +11 -1
  98. package/src/validation/primitives.ts +1 -1
@@ -1,10 +1,10 @@
1
1
  import * as _ggui_ai_protocol from '@ggui-ai/protocol';
2
2
  import { JsonValue, DataContract, JsonObject, GadgetDescriptor } from '@ggui-ai/protocol';
3
3
  import { G as GenerationProfileInput } from '../styling-profile-CxfvV48g.js';
4
- import { C as Classification } from '../axes-CzLEMDeB.js';
5
- import { E as EvalIssue, R as RuntimeProbeStatus, P as ProbeHostLoad, A as AxisCheck, c as EvalTier } from '../types-public-Bk6d1WLN.js';
4
+ import { C as Classification } from '../axes-CKVCzR_x.js';
5
+ import { E as EvalIssue, R as RuntimeProbeStatus, P as ProbeHostLoad, A as AxisCheck, c as EvalTier } from '../types-public-BJTLpzWn.js';
6
6
  import { LLMToolDef } from '../llm.js';
7
- import { H as HarnessFragment, a as CacheTier } from '../types-BOvHNG7K.js';
7
+ import { H as HarnessFragment, a as CacheTier } from '../types-Dh1s4K9j.js';
8
8
  import { ProcessMode, HarnessPolicy } from '../policy.js';
9
9
  import { D as DesignMode, C as CanvasClass } from '../design-mode-DLufhrNp.js';
10
10
  import '@ggui-ai/mcp-server-core';
@@ -427,7 +427,9 @@ async function runCheck(input) {
427
427
  classification: harness.classification,
428
428
  designMode: harness.designMode,
429
429
  // ggui#1117: the canvas the harness was built for, when it has one — never a default.
430
- ...harness.canvas !== void 0 ? { canvas: harness.canvas } : {}
430
+ ...harness.canvas !== void 0 ? { canvas: harness.canvas } : {},
431
+ // ggui#1285: the props the probe and judge rendered with — the fixture-echo check reads them.
432
+ ...input.fixtureProps !== void 0 ? { fixtureProps: input.fixtureProps } : {}
431
433
  };
432
434
  const axisRun = runGatedAxisChecks(check.axisChecks, axisInput);
433
435
  issues.push(...axisRun.issues);
@@ -1295,6 +1297,7 @@ function generateBoilerplate(_userPrompt, contract, shellType, screen, composedS
1295
1297
  }
1296
1298
  const actionTypeAliases = [];
1297
1299
  const actionHookCalls = [];
1300
+ let hasOneShotAction = false;
1298
1301
  const actionsMap = contract?.actionSpec ?? {};
1299
1302
  for (const [key, entry] of Object.entries(actionsMap)) {
1300
1303
  const label = entry.label ?? key;
@@ -1315,7 +1318,20 @@ type ${typeName} = ${tsType};`
1315
1318
  const callSig = tsType === "void" ? "() => void \u2014 fire and forget" : `(data: ${tsType}) => void`;
1316
1319
  const toolHint = tool ? ` \u2192 nextStep: ${tool}` : "";
1317
1320
  const onceHint = " \u2014 if this is meant ONCE, disable its control after it fires";
1318
- actionHookCalls.push(` const ${key} = useAction<${typeName}>('${key}'); // ${callSig}${toolHint}${onceHint}`);
1321
+ if (entry.oneShot === true) {
1322
+ hasOneShotAction = true;
1323
+ actionHookCalls.push(
1324
+ ` const ${key} = useAction<${typeName}>('${key}'); // ${callSig}${toolHint} \u2014 declared oneShot: the runtime drops a second dispatch`
1325
+ );
1326
+ actionHookCalls.push(
1327
+ ` const ${key}Spent = useActionSpent('${key}'); // true once fired, also after a reload \u2014 render the control from THIS: disabled={${key}Spent} and a spent label`
1328
+ );
1329
+ } else {
1330
+ actionHookCalls.push(` const ${key} = useAction<${typeName}>('${key}'); // ${callSig}${toolHint}${onceHint}`);
1331
+ }
1332
+ actionHookCalls.push(
1333
+ ` const ${key}Pending = useActionPending('${key}'); // true from the tap until the agent's answer repaints the card \u2014 render EVERY control that dispatches '${key}' (a Button, a Select's onChange, an Input's Enter) from THIS: disabled={${key}Pending} aria-busy={${key}Pending}, and a working label on a button; no other control carries its aria-busy`
1334
+ );
1319
1335
  }
1320
1336
  const streamChannels = contract?.streamSpec ?? {};
1321
1337
  const streamChannelEntries = Object.entries(streamChannels);
@@ -1413,6 +1429,8 @@ ${hookLines.join("\n")}
1413
1429
  const hasAnyWireFromWire = hasActions || hasStream || hasContext;
1414
1430
  const wireHooks = [];
1415
1431
  if (hasActions) wireHooks.push("useAction");
1432
+ if (hasOneShotAction) wireHooks.push("useActionSpent");
1433
+ if (hasActions) wireHooks.push("useActionPending");
1416
1434
  if (hasStream) wireHooks.push("useStream");
1417
1435
  if (hasContext) wireHooks.push("useGguiContext");
1418
1436
  const wireImport = hasAnyWireFromWire ? `import { ${wireHooks.join(", ")} } from '@ggui-ai/wire';
@@ -1649,21 +1667,17 @@ Rules:
1649
1667
  2. Layout and styling are universal. Colors, spacing, structure \u2014 these are the reusable part.
1650
1668
  3. Default values come from the current request \u2014 so the component renders correctly standalone.
1651
1669
  4. Copy comes from the props. NEVER invent display text the contract does not carry \u2014 no eyebrow / kicker / overline (a \`<Text caps>\` label), no helper line, no status text, no tagline. **Do NOT add a caps label above the heading** \u2014 not \`<Text caps>LET'S BEGIN</Text>\`, not \`WELCOME\`, not a section kicker like \`GETTING STARTED\` or \`OVERVIEW\` \u2014 UNLESS a prop supplies that exact text (\`<Text caps>{props.section}</Text>\`). When the request asks for a heading and a line, render the heading and the line and NOTHING above the heading. A heading's fallback is the request's own words, never a slogan of your own. Text the agent cannot change is chatter on every render, and a caps eyebrow you wrote is the most common form of it. An action's control shows the contract's \`label\` VERBATIM \u2014 every character, including \`&\`, \`/\`, \`\u2013\` and quotes \u2014 never paraphrased, shortened or re-cased (\`"Confirm & Schedule"\` is not \`"Confirm Schedule"\`).
1652
- 5. Props interface must be typed and exported.`;
1670
+ 5. Props interface must be typed and exported.
1671
+ 6. Keys on a mapped list come from the item's own identity: \`key={item.id}\` when the contract gives an id; when the items carry none (KPI tiles \`{label, value}\`, forecast days \`{day, high}\`), key by the field that names each one \u2014 \`key={item.label}\`, \`key={day.day}\` \u2014 never by the map index: a reorder or a stream merge re-keys by index and re-mounts the wrong rows, and the self-check refuses an index key.`;
1653
1672
  var TERMINAL_ACTIONS = `## Actions the user means once
1654
1673
 
1655
1674
  A second press is a mistake only when it would make a SECOND THING HAPPEN IN THE WORLD \u2014 a second booking, a second charge, a second order, a second application sent. If pressing again would only write the SAME STATE AGAIN \u2014 saving, updating, renaming, setting a preference, toggling \u2014 nothing is duplicated and the control stays armed.
1656
1675
 
1657
- Guard a control only where you have evidence, not a feeling about the verb: the request itself names a one-time commitment \u2014 "submit the application", "place the order", "confirm the booking".
1658
-
1659
- **When in doubt, leave it armed.** A control the user cannot press a second time has lost them something that worked \u2014 save, edit, save again is ordinary. A control left armed is only what every interface does by default.
1676
+ **The contract says which actions are once.** An action the contract declares \`oneShot: true\` (in the Contract block above) is guarded by the runtime \u2014 a second dispatch is dropped \u2014 and its spent state is DATA you read, never state you hold: render its control from \`const spent = useActionSpent('<name>')\`, \`disabled={spent}\`, and say so in the label or beside it ("Submitted", "Request sent") \u2014 a control that is disabled without explanation reads as broken. The hook's answer survives a reload and a second tab; \`useState(false)\` does not, so a reloaded card built on local state re-arms a control that already fired.
1660
1677
 
1661
- When you do guard one:
1662
- - Hold a small piece of state for it and set it in the handler: \`const [submitted, setSubmitted] = useState(false)\`.
1663
- - \`disabled={submitted}\` on the control that fires it, and say so in the label or beside it ("Submitted", "Request sent") \u2014 a control that is disabled without explanation reads as broken.
1664
- - A card re-rendered with new props starts fresh; the state above lives with the component, not with the request.
1678
+ **An action the contract does not declare \`oneShot\` stays armed.** Never guard one from a feeling about the verb \u2014 "confirm", "submit", "place" alone are not evidence. Guard it only where the request itself names a one-time commitment the contract failed to declare ("submit the application", "place the order"), and then with local state as the fallback: \`const [submitted, setSubmitted] = useState(false)\` set in the handler, \`disabled={submitted}\` with a label \u2014 knowing that this state lives with the component, not with the request, and starts fresh on a re-render with new props.
1665
1679
 
1666
- Nothing in the wire DECLARES one-shot-ness yet, which is why the test is what a second press would DO \u2014 not how final the word sounds.`;
1680
+ **When in doubt, leave it armed.** A control the user cannot press a second time has lost them something that worked \u2014 save, edit, save again is ordinary. A control left armed is only what every interface does by default.`;
1667
1681
  var FRAME_SIZING = `## Sizing \u2014 the frame owns the height and the width
1668
1682
 
1669
1683
  The host sizes the iframe; the component never sizes itself to the viewport. NEVER write \`100vh\` / \`100dvh\` / \`100svh\` (or an \`h-screen\` / \`min-h-screen\` class) as a height on any element:
@@ -1819,7 +1833,7 @@ With raw elements YOU own the semantics \u2014 nothing is accessible by construc
1819
1833
  3. **Images have \`alt\`** (\`alt=""\` for decorative). Icon-only buttons need \`aria-label\`; decorative icons and emoji are \`aria-hidden="true"\`.
1820
1834
  4. **Live & streaming data** (a \`useStream\` \`.latest\` value, counters, clocks, flashing prices) sits in an \`aria-live="polite"\` (or \`role="status"\`) region so screen readers announce the change.
1821
1835
  5. **Headings nest**: one \`<h1>\` per screen, \`<h2>\` for sections, \`<h3>\` below \u2014 never skipped or inverted.
1822
- 6. **Busy state is announced**: \`<button disabled={isLoading} aria-busy={isLoading}>{isLoading ? 'Submitting\u2026' : 'Submit'}</button>\`.
1836
+ 6. **Busy state is announced**: every control that dispatches an action (a button, a select whose change dispatches, an input that sends on Enter) stays pending from \`useActionPending('<action>')\` until the agent's answer repaints the card \u2014 \`<button disabled={sendPending} aria-busy={sendPending}>{sendPending ? 'Sending\u2026' : 'Send'}</button>\`. Only a control that dispatches the action carries its busy look; while it is pending, other controls of the same record may be disabled \u2014 nothing more (controls of other records, creation controls and the action's own cancel stay enabled); busy state the card owns (\`isLoading\`) the same way.
1823
1837
  7. **Stateful controls announce state**: \`role="checkbox"\` + \`aria-checked\` on a toggleable row, \`aria-pressed\` on a toggle button, \`aria-selected\` on the chosen item, \`aria-expanded\` on an open/close affordance. Styling (strikethrough, color, a check mark) is never the only carrier of state.
1824
1838
  8. **Focus is visible** \u2014 style \`:focus-visible\` with a token color; never \`outline: none\` without a replacement.`;
1825
1839
  var FREE_QUALITY_CHECKLIST = `## Quality Checklist (verify before returning)
@@ -2363,7 +2377,7 @@ The design-system primitives are accessible by construction \u2014 they emit the
2363
2377
  3. **Icons are decorative by default** \u2014 \`<Icon name="check" />\` is hidden from screen readers, which is correct for an icon sitting next to text. Add \`aria-label\` ONLY for a standalone, meaning-bearing icon with no adjacent text. Icon-only \`Button\`s still need \`aria-label\` on the **Button** itself.
2364
2378
  4. **Live & streaming data** \u2014 wrap any region whose content updates on its own (a \`useStream\` \`.latest\` value, a live clock, an "N new" counter, a flashing price) in an element with \`aria-live="polite"\` so screen readers announce the change.
2365
2379
  5. **Headings nest** \u2014 one \`<Heading level={1}>\` per screen, \`level={2}\` for sections, \`level={3}\` for subsections. Never skip or invert levels.
2366
- 6. **Buttons** \u2014 descriptive text content; icon-only buttons need \`aria-label\`. Announce busy state: \`<Button disabled={isLoading} aria-busy={isLoading}>{isLoading ? 'Submitting\u2026' : 'Submit'}</Button>\`.
2380
+ 6. **Buttons** \u2014 descriptive text content; icon-only buttons need \`aria-label\`. Announce busy state. EVERY control that dispatches an action \u2014 a Button, a Select whose change dispatches, an Input that sends on Enter \u2014 stays pending from that action's \`useActionPending\` until the agent's answer repaints the card: with \`const sendPending = useActionPending('send');\`, write \`<Button disabled={sendPending} aria-busy={sendPending} onClick={() => send({ text })}>{sendPending ? 'Sending\u2026' : 'Send'}</Button>\`, and give a dispatching Select or Input the same \`disabled={sendPending} aria-busy={sendPending}\`. Only a control that dispatches the action carries its \`aria-busy\` and a working label. While it is pending, other controls of the SAME record may be \`disabled\` \u2014 nothing more; controls of other records, creation controls and the action's own cancel stay enabled. Busy state the card owns (a local \`isLoading\`) is announced the same way.
2367
2381
  7. **Stateful controls announce their state.** Toggleable / selectable / expandable things use the primitive that carries the semantics (\`Checkbox\`, \`Toggle\`, \`Tabs\`, \`Accordion\`) \u2014 never a styled \`Box\`/\`Icon\` that only *looks* checked. When you hand-roll one with \`as={Clickable}\`, \`Clickable\` adds button + keyboard semantics but NOT state \u2014 add it yourself: \`role="checkbox"\` + \`aria-checked={done}\` on a toggleable row, \`aria-pressed\` on a toggle button, \`aria-selected\` on the chosen item, \`aria-expanded\` on an open/close affordance. Styling (strikethrough, color, a check icon) must never be the only carrier of state.
2368
2382
 
2369
2383
  ## Design System Usage (CRITICAL)
@@ -2606,6 +2620,10 @@ var APPLY_CHANGES_TOOL_SCOPED = {
2606
2620
  }
2607
2621
  };
2608
2622
 
2623
+ // src/harness-tool-names.ts
2624
+ var HARNESS_CODING_TOOLS = [APPLY_CHANGES_TOOL];
2625
+ var HARNESS_SCOPED_TOOLS = [APPLY_CHANGES_TOOL_SCOPED];
2626
+
2609
2627
  // src/patch.ts
2610
2628
  function applyLineRanges(sourceBefore, rawChanges) {
2611
2629
  if (rawChanges.length === 0) {
@@ -2810,8 +2828,8 @@ function createHarness(input) {
2810
2828
  const baseWhat = {
2811
2829
  boilerplate,
2812
2830
  fragments: whatFragments,
2813
- codingTools: [APPLY_CHANGES_TOOL],
2814
- scopedTools: [APPLY_CHANGES_TOOL_SCOPED],
2831
+ codingTools: [...HARNESS_CODING_TOOLS],
2832
+ scopedTools: [...HARNESS_SCOPED_TOOLS],
2815
2833
  applyPatch: defaultPatchFn,
2816
2834
  // Registered gadget catalog — drives the system prompt's gadget
2817
2835
  // table + the boilerplate's direct-import emission. Omitted when no