@ggui-ai/ui-gen 0.24.0 → 0.26.0

package/dist/harness/index.d.ts

@@ -1,7 +1,7 @@
 import { Harness } from './types-public.js';
 export { CheckLeg, CheckResult, CreateHarnessInput, DefaultTaskRunner, HarnessConstructionContext, HarnessId, HarnessMeta, HarnessName, HarnessOverrides, HarnessRevision, HowLeg, IterationRecord, LLMEvaluator, PatchFn, Phase, PhaseRunResult, PlannerDecision, PlannerFn, ProcessLeg, PromptBuilder, RetryPolicy, RunCheckInput, RunHarnessInput, RunHarnessReason, RunHarnessResult, RunWorkflowInput, Task, TaskContext, TaskRunner, TierCheck, WhatLeg, Workflow, WorkflowRunResult, createHarness, runCheck, runHarness, runWorkflow } from './types-public.js';
 export { applyLineRanges, defaultApplyPatch } from '../patch.js';
-export { A as AgentConfig, L as LLMResponse, P as ProviderRetryInfo, c as callLLM, a as createAgent } from '../llm-router-DwwUK1bz.js';
+export { b as AgentConfig, L as LLMResponse, P as ProviderRetryInfo, c as callLLM, d as createAgent } from '../types-public-B7wn4ce5.js';
 export { WORKFLOWS, WorkflowId, pickWorkflow } from '../workflows.js';
 export { LlmTraceEvent, LlmTraceKind, LlmTraceProvider, LlmTraceSink, getLlmTraceSink, setLlmTraceSink } from './llm-trace-sink.js';
 export { ValidatorTraceEvent, ValidatorTraceSink, getValidatorTraceSink, setValidatorTraceSink } from './validator-trace-sink.js';
@@ -9,12 +9,11 @@ export { ProcessMode } from '../policy.js';
 export { computeHarnessId, computeHarnessName, hashClassification } from '../hash.js';
 import '@ggui-ai/protocol';
 import '../styling-profile-CxfvV48g.js';
-import '../axes-CzLEMDeB.js';
-import '../types-public-Bk6d1WLN.js';
-import '@ggui-ai/mcp-server-core';
-import '../design-mode-DLufhrNp.js';
-import '../types-BOvHNG7K.js';
+import '../axes-CKVCzR_x.js';
 import '../llm.js';
+import '../types-Dh1s4K9j.js';
+import '../design-mode-DGWN_Gkm.js';
+import '@ggui-ai/mcp-server-core';
 
 /**
  * Fingerprint a materialized Harness. `h.id` is already a deterministic

package/dist/harness/index.js

@@ -582,7 +582,9 @@ async function runCheck(input) {
     classification: harness.classification,
     designMode: harness.designMode,
     // ggui#1117: the canvas the harness was built for, when it has one — never a default.
-    ...harness.canvas !== void 0 ? { canvas: harness.canvas } : {}
+    ...harness.canvas !== void 0 ? { canvas: harness.canvas } : {},
+    // ggui#1285: the props the probe and judge rendered with — the fixture-echo check reads them.
+    ...input.fixtureProps !== void 0 ? { fixtureProps: input.fixtureProps } : {}
   };
   const axisRun = runGatedAxisChecks(check.axisChecks, axisInput);
   issues.push(...axisRun.issues);
@@ -1464,6 +1466,7 @@ function generateBoilerplate(_userPrompt, contract, shellType, screen, composedS
   }
   const actionTypeAliases = [];
   const actionHookCalls = [];
+  let hasOneShotAction = false;
   const actionsMap = contract?.actionSpec ?? {};
   for (const [key, entry] of Object.entries(actionsMap)) {
     const label = entry.label ?? key;
@@ -1484,7 +1487,20 @@ type ${typeName} = ${tsType};`
     const callSig = tsType === "void" ? "() => void \u2014 fire and forget" : `(data: ${tsType}) => void`;
     const toolHint = tool ? ` \u2192 nextStep: ${tool}` : "";
     const onceHint = " \u2014 if this is meant ONCE, disable its control after it fires";
-    actionHookCalls.push(`  const ${key} = useAction<${typeName}>('${key}'); // ${callSig}${toolHint}${onceHint}`);
+    if (entry.oneShot === true) {
+      hasOneShotAction = true;
+      actionHookCalls.push(
+        `  const ${key} = useAction<${typeName}>('${key}'); // ${callSig}${toolHint} \u2014 declared oneShot: the runtime drops a second dispatch`
+      );
+      actionHookCalls.push(
+        `  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`
+      );
+    } else {
+      actionHookCalls.push(`  const ${key} = useAction<${typeName}>('${key}'); // ${callSig}${toolHint}${onceHint}`);
+    }
+    actionHookCalls.push(
+      `  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`
+    );
   }
   const streamChannels = contract?.streamSpec ?? {};
   const streamChannelEntries = Object.entries(streamChannels);
@@ -1582,6 +1598,8 @@ ${hookLines.join("\n")}
   const hasAnyWireFromWire = hasActions || hasStream || hasContext;
   const wireHooks = [];
   if (hasActions) wireHooks.push("useAction");
+  if (hasOneShotAction) wireHooks.push("useActionSpent");
+  if (hasActions) wireHooks.push("useActionPending");
   if (hasStream) wireHooks.push("useStream");
   if (hasContext) wireHooks.push("useGguiContext");
   const wireImport = hasAnyWireFromWire ? `import { ${wireHooks.join(", ")} } from '@ggui-ai/wire';
@@ -1632,6 +1650,13 @@ ${body}
     AXIS_SECTIONS: composedSections ?? ""
   }, designMode);
 }
+var NOT_YET_TAUGHT_TOKENS = /* @__PURE__ */ new Set([
+  "--ggui-color-controlAccentOutline",
+  "--ggui-color-controlOutline"
+]);
+var taughtTokenManifest = consumedTokenManifest.filter(
+  (token) => !NOT_YET_TAUGHT_TOKENS.has(token)
+);
 
 // src/boilerplate/hard-sections.ts
 var PROTOCOL_NOTES = `## Protocol Notes
@@ -1818,21 +1843,17 @@ Rules:
 2. Layout and styling are universal. Colors, spacing, structure \u2014 these are the reusable part.
 3. Default values come from the current request \u2014 so the component renders correctly standalone.
 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"\`).
-5. Props interface must be typed and exported.`;
+5. Props interface must be typed and exported.
+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.`;
 var TERMINAL_ACTIONS = `## Actions the user means once
 
 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.
 
-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".
-
-**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.
+**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.
 
-When you do guard one:
-- Hold a small piece of state for it and set it in the handler: \`const [submitted, setSubmitted] = useState(false)\`.
-- \`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.
-- A card re-rendered with new props starts fresh; the state above lives with the component, not with the request.
+**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.
 
-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.`;
+**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.`;
 var FRAME_SIZING = `## Sizing \u2014 the frame owns the height and the width
 
 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:
@@ -1845,7 +1866,7 @@ The width works the same way, and the mistake is the mirror image: **your outerm
 
 - Never put \`maxWidth\` (or a \`max-w-*\` class) on the element you return at the top. On a wide canvas that turns your card into a narrow strip with the frame's ground on either side of it, and on a narrow one it does nothing \u2014 the same composition reading differently at two sizes.
 - DO cap the reading measure inside: a column of prose is easiest to read at 60\u201370 characters, so wrap the TEXT in a \`<Container maxWidth="sm">\` (or your own \`max-width\`) and leave the root filling. A hero band, a toolbar, a table or an image then still uses the whole width, which is what makes a wide card look composed rather than padded.
-- Centre that inner column with the container's own centring, not by shrinking the root.
+- Centre that inner column with the container's own centring, not by shrinking the root \u2014 and give the card ONE column: every block of the reading flow (a heading, its text, and the chips or buttons that answer it) sits in a \`<Container>\` of the same \`maxWidth\`, so their left edges line up even where a full-width band runs between them.
 - **They are two different elements.** Even when the whole card is one column of text, the root still fills and the column sits INSIDE it \u2014 \`<Box padding="lg"><Container maxWidth="sm">\u2026</Container></Box>\`, never a root that is itself the container. A centred card is centring inside a filled frame, not a shrunken frame.`;
 var COMPONENT_STRUCTURE = `## Component Structure
 
@@ -1922,7 +1943,7 @@ function colorVocabularyFromManifest(manifest) {
 function renderRamp(r) {
   return `\`${r.family}-${r.steps.join("/")}\`${r.bare ? ` (plus bare \`${r.family}\`)` : ""}`;
 }
-function renderFreeColorRule(manifest = consumedTokenManifest) {
+function renderFreeColorRule(manifest = taughtTokenManifest) {
   const { ramps, roles } = colorVocabularyFromManifest(manifest);
   const brand = ramps.filter((r) => r.family === "primary" || r.family === "neutral");
   const state = ramps.filter((r) => r.family !== "primary" && r.family !== "neutral");
@@ -1988,7 +2009,7 @@ With raw elements YOU own the semantics \u2014 nothing is accessible by construc
 3. **Images have \`alt\`** (\`alt=""\` for decorative). Icon-only buttons need \`aria-label\`; decorative icons and emoji are \`aria-hidden="true"\`.
 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.
 5. **Headings nest**: one \`<h1>\` per screen, \`<h2>\` for sections, \`<h3>\` below \u2014 never skipped or inverted.
-6. **Busy state is announced**: \`<button disabled={isLoading} aria-busy={isLoading}>{isLoading ? 'Submitting\u2026' : 'Submit'}</button>\`.
+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.
 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.
 8. **Focus is visible** \u2014 style \`:focus-visible\` with a token color; never \`outline: none\` without a replacement.`;
 var FREE_QUALITY_CHECKLIST = `## Quality Checklist (verify before returning)
@@ -2012,7 +2033,7 @@ var TOKEN_GROUPS = [
   { label: "Radius (optional)", prefixes: ["--ggui-shape-radius-"], note: "" },
   { label: "Shadow (optional)", prefixes: ["--ggui-shape-shadow-"], note: "" }
 ];
-function renderTokenVocabulary(manifest = consumedTokenManifest) {
+function renderTokenVocabulary(manifest = taughtTokenManifest) {
   const sorted = [...manifest].sort();
   const claimed = /* @__PURE__ */ new Set();
   const lines = [
@@ -2532,7 +2553,7 @@ The design-system primitives are accessible by construction \u2014 they emit the
 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.
 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.
 5. **Headings nest** \u2014 one \`<Heading level={1}>\` per screen, \`level={2}\` for sections, \`level={3}\` for subsections. Never skip or invert levels.
-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>\`.
+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.
 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.
 
 ## Design System Usage (CRITICAL)
@@ -2775,6 +2796,10 @@ var APPLY_CHANGES_TOOL_SCOPED = {
   }
 };
 
+// src/harness-tool-names.ts
+var HARNESS_CODING_TOOLS = [APPLY_CHANGES_TOOL];
+var HARNESS_SCOPED_TOOLS = [APPLY_CHANGES_TOOL_SCOPED];
+
 // src/patch.ts
 function applyLineRanges(sourceBefore, rawChanges) {
   if (rawChanges.length === 0) {
@@ -3099,8 +3124,8 @@ function createHarness(input) {
   const baseWhat = {
     boilerplate,
     fragments: whatFragments,
-    codingTools: [APPLY_CHANGES_TOOL],
-    scopedTools: [APPLY_CHANGES_TOOL_SCOPED],
+    codingTools: [...HARNESS_CODING_TOOLS],
+    scopedTools: [...HARNESS_SCOPED_TOOLS],
     applyPatch: defaultPatchFn,
     // Registered gadget catalog — drives the system prompt's gadget
     // table + the boilerplate's direct-import emission. Omitted when no
@@ -3631,6 +3656,26 @@ async function collectLoginTurn(stream, opts) {
 }
 
 // src/harness/llm-router.ts
+function servedModelOf(name) {
+  return typeof name === "string" && name.length > 0 ? { servedModel: name } : {};
+}
+function anthropicFinishReason(stopReason) {
+  if (stopReason === null || stopReason === void 0) return void 0;
+  if (stopReason === "end_turn" || stopReason === "stop_sequence") return "stop";
+  if (stopReason === "max_tokens") return "length";
+  if (stopReason === "refusal") return "content-filter";
+  return "other";
+}
+function googleFinishReason(finishReason) {
+  if (finishReason === null || finishReason === void 0) return void 0;
+  if (finishReason === "STOP") return "stop";
+  if (finishReason === "MAX_TOKENS") return "length";
+  if (["SAFETY", "RECITATION", "BLOCKLIST", "PROHIBITED_CONTENT", "SPII", "IMAGE_SAFETY"].includes(finishReason)) return "content-filter";
+  return "other";
+}
+function visionFinish(finishReason) {
+  return finishReason !== void 0 ? { finishReason } : {};
+}
 function addStrictSchemaConstraints(schema) {
   const result = { ...schema };
   if (result.type === "object" && result.properties) {
@@ -3981,7 +4026,9 @@ var AnthropicAgent = class extends LLMAgent {
       return {
         text,
         inputTokens: response.usage.input_tokens,
-        outputTokens: response.usage.output_tokens
+        outputTokens: response.usage.output_tokens,
+        ...visionFinish(anthropicFinishReason(response.stop_reason)),
+        ...servedModelOf(response.model)
       };
     } catch (e) {
       const endedAt = Date.now();
@@ -4652,7 +4699,9 @@ var GoogleAgent = class extends LLMAgent {
         text,
         // ggui#1186 (third site): `promptTokenCount` includes `cachedContentTokenCount` — report the non-cached input.
         inputTokens: splitCacheInclusiveUsage(usage?.promptTokenCount ?? 0, usage?.candidatesTokenCount ?? 0, usage?.cachedContentTokenCount ?? 0).tokens.input,
-        outputTokens: usage?.candidatesTokenCount ?? 0
+        outputTokens: usage?.candidatesTokenCount ?? 0,
+        ...visionFinish(googleFinishReason(response.candidates?.[0]?.finishReason)),
+        ...servedModelOf(response.modelVersion)
       };
       const endedAt = Date.now();
       emitLlmTraceEvent({