@ggui-ai/ui-gen 0.15.0 → 0.16.0

package/dist/workflows.d.ts

@@ -1,7 +1,8 @@
 import { C as Classification } from './axes-CzLEMDeB.js';
 import { Workflow } from './harness/types-public.js';
 import '@ggui-ai/protocol';
-import './types-public-CkNWEov9.js';
+import './types-public-D24JLJgl.js';
+import './design-mode-BHEE7flj.js';
 import './types-BOvHNG7K.js';
 import './llm.js';
 import './policy.js';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ggui-ai/ui-gen",
-  "version": "0.15.0",
+  "version": "0.16.0",
   "description": "UI generation harness for ggui — the LLM system prompt, multi-turn coding loop, and evaluation engine that turn a DataContract into a React component. Implements the UiGenerator contract from @ggui-ai/mcp-server-core; consumed by @ggui-ai/mcp-server.",
   "license": "Apache-2.0",
   "keywords": [
@@ -188,9 +188,9 @@
     "puppeteer-core": "^24.40.0 || ^25.0.0",
     "react": "^18.0.0 || ^19.0.0",
     "react-dom": "^18.0.0 || ^19.0.0",
-    "@ggui-ai/mcp-server-core": "0.15.0",
-    "@ggui-ai/protocol": "0.15.0",
-    "@ggui-ai/project-config": "0.15.0"
+    "@ggui-ai/project-config": "0.16.0",
+    "@ggui-ai/mcp-server-core": "0.16.0",
+    "@ggui-ai/protocol": "0.16.0"
   },
   "peerDependenciesMeta": {
     "@google/adk": {
@@ -229,11 +229,11 @@
     "openai": "^7.9.0",
     "typescript": "^5.0.0",
     "zod": "^4.3.6",
-    "@ggui-ai/gadgets": "0.15.0",
-    "@ggui-ai/sandbox": "0.15.0",
-    "@ggui-ai/ui-visual-tester": "0.15.0",
-    "@ggui-ai/wire": "0.15.0",
-    "@ggui-ai/design": "0.15.0"
+    "@ggui-ai/gadgets": "0.16.0",
+    "@ggui-ai/design": "0.16.0",
+    "@ggui-ai/sandbox": "0.16.0",
+    "@ggui-ai/ui-visual-tester": "0.16.0",
+    "@ggui-ai/wire": "0.16.0"
   },
   "devDependencies": {
     "@google/adk": "^2.0.0",
@@ -248,9 +248,9 @@
     "react-dom": "^19.0.0",
     "tsup": "^8.5.0",
     "vitest": "^3.2.6",
-    "@ggui-ai/mcp-server-core": "0.15.0",
-    "@ggui-ai/project-config": "0.15.0",
-    "@ggui-ai/protocol": "0.15.0"
+    "@ggui-ai/mcp-server-core": "0.16.0",
+    "@ggui-ai/project-config": "0.16.0",
+    "@ggui-ai/protocol": "0.16.0"
   },
   "engines": {
     "node": ">=22.0.0"

package/src/adapters/generation-dispatch.ts

@@ -51,13 +51,14 @@ import type {
   GenerationResult,
   RenderingContext,
 } from "../harness/result-types.js";
-import type { AdapterResult, ProviderName, ToolDefinition } from "./types.js";
+import type { ProviderName, ToolDefinition } from "./types.js";
 import type { QualityConfig } from "../evaluation/types-public.js";
 import type { AgentConfig } from "../harness/llm-router.js";
 import {
   type GadgetDescriptor,
   type JsonObject,
 } from "@ggui-ai/protocol";
+import { canvasForRendering, type CanvasClass, type DesignMode } from "../design-mode.js";
 
 export interface GenerationDispatchParams {
   provider: ProviderName;
@@ -93,6 +94,13 @@ export interface GenerationDispatchParams {
     passThreshold?: number;
     sampleProps?: JsonObject;
     viewport?: { width: number; height: number };
+    /**
+     * Per-canvas judging — see `VisualEvalConfig.canvases`. When set, the
+     * eval round screenshots + judges once per class and the harness
+     * stamps `EvalResult.visual` with the PNG-free summary. Unset =
+     * today's single screenshot at `viewport`.
+     */
+    canvases?: readonly CanvasClass[];
   };
   /** Quality config controlling eval tiers and improvement behavior */
   qualityConfig?: QualityConfig;
@@ -163,6 +171,15 @@ export interface GenerationDispatchParams {
    * a no-op.
    */
   onRetry?: AgentConfig["onRetry"];
+  /**
+   * Which triad to materialise — see `CreateHarnessInput.designMode`.
+   * Reaches the system prompt, the boilerplate, the tier-0 gate, the
+   * axis checks and the evaluator criteria through the harness. Absent
+   * means `constrained` (byte-identical to the pre-`designMode` path).
+   */
+  designMode?: DesignMode;
+  /** Rendering canvas class for the `free` prompt — see `CreateHarnessInput.canvas`. */
+  canvas?: CanvasClass;
 }
 
 /**
@@ -179,7 +196,7 @@ function mapProviderForLLMRouter(provider: ProviderName): "anthropic" | "openai"
  */
 export async function dispatchGeneration(
   params: GenerationDispatchParams,
-): Promise<AdapterResult | GenerationResult> {
+): Promise<GenerationResult> {
   // NOTE: userPrompt arrives already enriched with rendering context + contract
   // (including examples) from the caller (runner.ts / generator.ts).
   // The contract are injected via injectContracts() which calls propsSpecToTypeScript()
@@ -274,6 +291,9 @@ export async function dispatchGeneration(
     ...(params.gadgetTypes !== undefined
       ? { gadgetTypes: params.gadgetTypes }
       : {}),
+    // Design mode + canvas — the harness carries them to every leg.
+    ...(params.designMode !== undefined ? { designMode: params.designMode } : {}),
+    ...(params.canvas !== undefined ? { canvas: params.canvas } : {}),
     // Policy is pre-resolved at the dispatch boundary so `createHarness`
     // stays free of experiment plumbing. `resolveHarnessPolicy` applies
     // the `GGUI_POLICY_PROFILE` experimental-profile layer and hands the
@@ -293,7 +313,7 @@ export async function dispatchGeneration(
     // prompt's gadget catalog renders a `Type:` line per third-party
     // gadget (the LLM sees a wrapper's real call shape it otherwise
     // can't know).
-    systemPromptBuilder: ({ userRequest, shellType, screen, axisDelta }) =>
+    systemPromptBuilder: ({ userRequest, shellType, screen, axisDelta, designMode, canvas }) =>
       buildSystemPromptWithFunnel(
         userRequest,
         shellType,
@@ -301,6 +321,8 @@ export async function dispatchGeneration(
         axisDelta,
         params.appGadgets,
         params.gadgetTypes,
+        designMode,
+        canvas,
       ),
     // Pre-filtered axis-check registry. The `matches()` predicate
     // selects the checks relevant to this generation's axis vector.
@@ -513,5 +535,36 @@ export async function dispatchGeneration(
   // pairedSource via the workflow output) — guaranteed to pair with
   // telemetry.compiledCode when runResult.ok is true.
   const finalSource = runResult.finalSource ?? "";
-  return assembleGenerationResult({ session, telemetry, source: finalSource });
+  // `assembleGenerationResult` is async — the spread below must see the
+  // RESOLVED result, never the promise (a spread promise is `{}`: every
+  // consumer downstream lost `compiledCode` / `tokens` on 2026-09-09).
+  const result: GenerationResult = await assembleGenerationResult({ session, telemetry, source: finalSource });
+  // Record the arm this generation ran under (bench reports read it per
+  // cell). The canvas is the one the free prompt stated: explicit when
+  // passed, else the same shell × screen derivation the prompt used.
+  // Nothing is stamped on the default (constrained, no option) path.
+  return withArmRecord(result, params);
+}
+
+/**
+ * Stamp the arm record (`designMode` / `canvas`) onto a RESOLVED
+ * {@link GenerationResult}. The input must be the resolved result, never a
+ * promise — a spread promise is `{}` and every consumer downstream loses
+ * `compiledCode` / `tokens` (2026-09-09; gate:
+ * `generation-dispatch.resolved-result.test.ts`). Nothing is stamped on the
+ * default (constrained, no option) path, so a default run's result is
+ * byte-identical to the input.
+ */
+function withArmRecord(
+  result: GenerationResult,
+  params: Pick<GenerationDispatchParams, "designMode" | "canvas" | "shellType" | "screen">,
+): GenerationResult {
+  const recordedCanvas: CanvasClass | undefined =
+    params.canvas ??
+    (params.designMode === "free" ? canvasForRendering(params.shellType, params.screen) : undefined);
+  return {
+    ...result,
+    ...(params.designMode !== undefined ? { designMode: params.designMode } : {}),
+    ...(recordedCanvas !== undefined ? { canvas: recordedCanvas } : {}),
+  };
 }

package/src/adapters/provider-router.ts

@@ -8,7 +8,7 @@
 // the typed `LlmRoute` (slice #43) guarantees `route.model` is
 // wire-canonical at construction time, so there is no LiteLLM prefix
 // to strip at this boundary. `getBedrockModelId` stays — the
-// "anthropic route + bedrock IAM" mixed-mode escape hatch (cloud-pod
+// "anthropic route + bedrock IAM" mixed-mode escape hatch (a hosted deployment's
 // legacy) still needs the wire-canonical anthropic id → cross-region
 // profile id upcast.
 
@@ -18,7 +18,7 @@ import type { LlmRoute } from '@ggui-ai/protocol';
  * Map a wire-canonical Anthropic model id (or already-bedrock id) to
  * the AWS Bedrock cross-region inference profile id. Used by the
  * mixed-mode `anthropic` route + bedrock-IAM fallback path in
- * {@link resolveRoute} (cloud-pod legacy) and the analogous branch in
+ * {@link resolveRoute} (a hosted deployment's legacy) and the analogous branch in
  * `harness/llm-router.ts::AnthropicAgent.resolveModel`.
  *
  * Operators picking the bedrock route explicitly (slice #43 Phase 4)
@@ -41,13 +41,13 @@ export function getBedrockModelId(model: string): string {
     'anthropic/claude-fable-5': 'us.anthropic.claude-fable-5',
   };
   // Normalize the Bedrock dot-form opt-in (e.g. `anthropic.claude-haiku-4-5`,
-  // produced when cloud-pod's `resolvePoolRoute` strips a `bedrock/` opt-in
+  // produced when a route resolver strips a `bedrock/` opt-in
   // prefix) to the Anthropic slash form. Both spellings should resolve to
   // the same cross-region inference profile id — without this, the dot
   // form would skip the BEDROCK_MAP hit, fall through the `^anthropic\/`
   // strip (which is a no-op for the dot form), and re-prepend
   // `us.anthropic.` to produce a double-`anthropic.` id that Bedrock
-  // rejects with 400 (cloud e2e regression 2026-05-25, bedrock-iam.spec).
+  // rejects with 400 (regression seen live on 2026-05-25).
   const normalized = model.replace(/^anthropic\./, 'anthropic/');
   if (BEDROCK_MAP[normalized]) return BEDROCK_MAP[normalized];
   if (normalized.startsWith('us.anthropic.') || normalized.startsWith('arn:')) {
@@ -131,7 +131,7 @@ function clearSiblings(
  *       2. Without `apiKey` AND explicit `env.CLAUDE_CODE_USE_BEDROCK === '1'`
  *          (or no `env.ANTHROPIC_API_KEY`) → mixed-mode bedrock fallback
  *          (model upcast to cross-region profile via `getBedrockModelId`).
- *          Cloud-pod legacy escape hatch; the OSS Phase 4 strict-fail
+ *          A hosted deployment's legacy escape hatch; the Phase 4 strict-fail
  *          prevents end-users from reaching this branch implicitly.
  *       3. Otherwise → direct Anthropic API with whatever
  *          `ANTHROPIC_API_KEY` is already in `env`.
@@ -225,7 +225,8 @@ export function resolveRoute(input: RoutingInput): RoutingDecision {
         };
       }
       // Mixed-mode legacy: anthropic route + bedrock IAM fallback.
-      // Cloud-pod escape hatch when `ANTHROPIC_API_KEY` isn't seeded.
+      // Escape hatch for a production deployment that authenticates via Bedrock
+      // IAM and never sets `ANTHROPIC_API_KEY`.
       const explicitBedrock = env.CLAUDE_CODE_USE_BEDROCK === '1';
       const noApiKey = !env.ANTHROPIC_API_KEY;
       if (explicitBedrock || noApiKey) {

package/src/adapters/tools.ts

@@ -461,7 +461,7 @@ export async function runSelfChecks(
         check: 'hex-color',
         line: lineNum,
         message: `Hardcoded color "${hexMatch[0]}" — must use design tokens`,
-        fix: `Replace with var(--ggui-color-*, ${hexMatch[0]})`,
+        fix: `Replace "${hexMatch[0]}" with a primitive variant OR a bare token reference such as var(--ggui-color-primary-500) — never a literal fallback inside var()`,
       });
     }
 
@@ -472,7 +472,7 @@ export async function runSelfChecks(
         check: 'hardcoded-color-fn',
         line: lineNum,
         message: `Hardcoded ${colorFnMatch[1]}() — use design tokens instead`,
-        fix: `Replace with semantic tokens: var(--ggui-color-surface), var(--ggui-color-onSurface), var(--ggui-color-outline), etc.`,
+        fix: `Replace with the layering roles: var(--ggui-color-container), var(--ggui-color-onContainer), var(--ggui-color-outline), etc.`,
       });
     }
 
@@ -482,7 +482,7 @@ export async function runSelfChecks(
         check: 'prefer-semantic',
         line: lineNum,
         message: `neutral-50 for background — prefer semantic token for dark-theme compatibility`,
-        fix: `Replace with var(--ggui-color-surface) or var(--ggui-color-surfaceVariant)`,
+        fix: `Replace with var(--ggui-color-container) or var(--ggui-color-sunken)`,
       });
     }
     if (/--ggui-color-neutral-900[^0-9]/.test(line) && /color/i.test(line) && !/background/i.test(line)) {
@@ -490,7 +490,7 @@ export async function runSelfChecks(
         check: 'prefer-semantic',
         line: lineNum,
         message: `neutral-900 for text color — prefer semantic token for dark-theme compatibility`,
-        fix: `Replace with var(--ggui-color-onSurface) or var(--ggui-color-onSurfaceVariant)`,
+        fix: `Replace with var(--ggui-color-onContainer) or var(--ggui-color-onSunken)`,
       });
     }
 
@@ -501,7 +501,7 @@ export async function runSelfChecks(
         check: 'raw-pixels',
         line: lineNum,
         message: `Raw pixel value in spacing — must use design tokens`,
-        fix: `Replace with var(--ggui-spacing-*, fallback)`,
+        fix: `Replace with a bare var(--ggui-spacing-*) token — no literal fallback — or the scale name on the primitive prop (gap="md", padding="lg")`,
       });
     }
 

package/src/advanced/generator.ts

@@ -1,9 +1,9 @@
 /**
- * Advanced UI generator factory (`ui-gen-advanced-opus-4-7`).
+ * Advanced UI generator factory (`ui-gen-advanced`).
  *
  * See `./index.ts` for the design narrative. This file owns the
  * iterative two-stage loop and the {@link UiGenerator} adapter that
- * registers under the slug `ui-gen-advanced-opus-4-7`.
+ * registers under the slug `ui-gen-advanced`.
  *
  * Loop structure (max 3 iterations, threshold 0.8, always-persist):
  *
@@ -26,10 +26,10 @@
  * spinning up Chromium for code we already know is broken) AND when
  * the contract has no `actionSpec` entries (no behaviour to verify).
  *
- * Generator identity is baked in — slug `ui-gen-advanced-opus-4-7`,
- * tier `advanced`, model `opus-4-7`. Operators can NOT override slug
- * via this factory; build a different factory if you want a different
- * tier/model pairing.
+ * Generator identity is baked in — slug `ui-gen-advanced`, tier
+ * `advanced`; the declared model is `ADVANCED_GENERATOR_MODEL` and is not
+ * part of the slug (ggui#923). Operators can NOT override the slug via
+ * this factory; build a different factory for a different tier.
  */
 import type {
   GenerationMetadata,
@@ -39,6 +39,7 @@ import type {
   UiGenerator,
 } from '@ggui-ai/mcp-server-core';
 import type { DataContract, JsonObject } from '@ggui-ai/protocol';
+import type { ModelRef } from '@ggui-ai/protocol';
 import type {
   BehaviorFailure,
   PlaywrightModule,
@@ -61,9 +62,9 @@ import {
 } from './feedback.js';
 
 /** Generator identity — slug, tier, and model. */
-export const ADVANCED_GENERATOR_SLUG = 'ui-gen-advanced-opus-4-7' as const;
+export const ADVANCED_GENERATOR_SLUG = 'ui-gen-advanced' as const;
 export const ADVANCED_GENERATOR_TIER: GeneratorTier = 'advanced';
-export const ADVANCED_GENERATOR_MODEL = 'opus-4-7' as const;
+export const ADVANCED_GENERATOR_MODEL: ModelRef = 'anthropic/claude-opus-4-7';
 
 /**
  * Re-exported under a friendlier name for callers wiring deploys.
@@ -162,7 +163,7 @@ const DEFAULT_SLOW_TIMEOUT_MS = 5000;
 
 /**
  * Create the advanced generator. Returns a {@link UiGenerator}
- * registrable under the slug `ui-gen-advanced-opus-4-7`.
+ * registrable under the slug `ui-gen-advanced`.
  *
  * @example
  * import { chromium } from 'playwright-core';
@@ -189,7 +190,7 @@ export function createAdvancedUiGenerator(
     options.innerGenerator ??
     createUiGenerator({
       tier: 'advanced',
-      model: 'opus-4-7',
+      model: ADVANCED_GENERATOR_MODEL,
       ...(options.maxAttempts !== undefined
         ? { maxAttempts: options.maxAttempts }
         : {}),

package/src/advanced/index.ts

@@ -1,5 +1,5 @@
 /**
- * Advanced UI generator — `ui-gen-advanced-opus-4-7`.
+ * Advanced UI generator — `ui-gen-advanced`.
  *
  * Iterative two-stage validator-feedback loop. Sibling to the default
  * generator at `@ggui-ai/ui-gen/create-ui-generator.ts`. The default
@@ -27,9 +27,10 @@
  * Missing the dep throws on every `generate()` call (not at factory
  * time) so a config file dropping the registration is observable.
  *
- * Default OSS install does NOT ship Playwright. Cloud operators opt
- * into this generator by deploying `ggui-protocol-pod-advanced`, the
- * pod image variant that bakes Playwright + Chromium.
+ * The default install does NOT ship Playwright. Operators opt into this
+ * generator by installing `playwright-core` + Chromium and registering
+ * `createAdvancedUiGenerator({ playwright })` in the generator registry
+ * they hand to the server.
  */
 export {
   createAdvancedUiGenerator,