@ggui-ai/ui-gen 0.24.0 → 0.26.0

package/dist/harness/result-types.d.ts

@@ -1,13 +1,19 @@
 import { JsonObject, DataContract } from '@ggui-ai/protocol';
 export { ActionSpec, DataContract, JsonSchema, PropsSpec, StreamSpec } from '@ggui-ai/protocol';
 import { AdapterResult } from '../adapters/types.js';
-import { R as RenderingContext } from '../prompts-YU9piOgs.js';
-import { b as EvalResult } from '../types-public-Bk6d1WLN.js';
-import { D as DesignMode, C as CanvasClass } from '../design-mode-DLufhrNp.js';
+import { R as RenderingContext } from '../prompts-BhNUuRoZ.js';
+import { e as EvalResult } from '../types-public-B7wn4ce5.js';
+import { V as VisualLegTokens } from '../visual-evaluator-C9L-o6ap.js';
+import { D as DesignMode, C as CanvasClass } from '../design-mode-DGWN_Gkm.js';
 import 'zod';
+import '../llm.js';
 import '@ggui-ai/mcp-server-core';
-import '../axes-CzLEMDeB.js';
-import '../types-BOvHNG7K.js';
+import '../axes-CKVCzR_x.js';
+import '../types-Dh1s4K9j.js';
+import '../styling-profile-CxfvV48g.js';
+import '../bank-Ctdp3FUA.js';
+import '../evaluation/types.js';
+import 'puppeteer-core';
 
 /**
  * Model role — each internal task in the harness can use a different model.
@@ -49,6 +55,14 @@ interface GenerationResult extends AdapterResult {
     passesUsed: number;
     /** Three-tier evaluation result (tier 0 + LLM tier 1+2 + visual) */
     evalResult?: EvalResult;
+    /**
+     * ggui#1522 — what the IN-LOOP visual judge spent during this generation (its scoring calls, and its report-only
+     * criteria calls apart), priced at the visual agent's model. NOT inside `tokens`, which covers the coding turns
+     * and the text evaluator. And NOT the same calls as a post-generation judge's own visual pass (a benchmark
+     * report's `meta.costs.visualUsd`): a cost rule that adds both is correct, one read as the other double-counts.
+     * Absent when no in-loop visual leg ran or it reported no tokens.
+     */
+    inLoopVisualTokens?: VisualLegTokens;
     /** Set only when the same-exchange guard ended the run (ggui#404); no key otherwise. */
     sameExchangeBreak?: SameExchangeBreak;
     /** Whether background improvement should be spawned (auto-improve mode) */

package/dist/harness/result-types.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/harness/result-types.ts"],"names":[],"mappings":";AAkLA,SAAS,iBAAiB,OAAA,EAAyB;AACjD,EAAA,MAAM,UAAA,GAAa,OAAA,CAAQ,OAAA,CAAQ,GAAG,CAAA;AACtC,EAAA,OAAO,cAAc,CAAA,GAAI,OAAA,CAAQ,KAAA,CAAM,UAAA,GAAa,CAAC,CAAA,GAAI,OAAA;AAC3D;AAOO,SAAS,mBAAA,CACd,IAAA,EACA,MAAA,EACA,QAAA,EACQ;AACR,EAAA,IAAI,CAAC,QAAQ,OAAO,QAAA;AACpB,EAAA,MAAM,QAAA,GAAW,MAAA,CAAO,IAAI,CAAA,IAAK,OAAO,OAAA,IAAW,QAAA;AACnD,EAAA,OAAO,iBAAiB,QAAQ,CAAA;AAClC","file":"result-types.js","sourcesContent":["// packages/ui-gen/src/harness/result-types.ts\n//\n// Result shapes returned by the harness runtime.\n//\n// Two output types travel the dispatch surface:\n//   - AdapterResult      — raw adapter output (provider-direct path)\n//   - GenerationResult   — full harness output (run-loop, eval, breakdown)\n//\n// ModelRoles + resolveModelForRole live here too (the harness decides\n// which model to use per internal task).\n\nimport type { DataContract, JsonObject } from \"@ggui-ai/protocol\";\nimport type { AdapterResult } from \"../adapters/types.js\";\nimport type { RenderingContext } from \"../contract-context.js\";\nimport type { EvalResult } from \"../evaluation/types-public.js\";\nimport type { CanvasClass, DesignMode } from \"../design-mode.js\";\n\n// =============================================================================\n// Model Roles\n// =============================================================================\n\n/**\n * Model role — each internal task in the harness can use a different model.\n *\n * - thinking: planning, design decisions, architecture (use frontier/thinking model)\n * - coding:   writing TSX, implementing the design (use fast model)\n * - evaluation: scoring quality, finding issues (use balanced model)\n * - default:  fallback for any role not explicitly set\n */\nexport type ModelRole = \"thinking\" | \"coding\" | \"evaluation\" | \"default\";\n\n/**\n * Maps model roles to model IDs.\n * Any role not specified falls back to 'default', then to params.model.\n *\n * @example\n * // Opus plans, Haiku codes, Sonnet evaluates\n * { thinking: 'claude-opus-4-6', coding: 'claude-haiku-4-5', default: 'claude-sonnet-4-6' }\n *\n * // All Haiku (cheapest)\n * { default: 'claude-haiku-4-5' }\n *\n * // All Opus (highest quality)\n * { default: 'claude-opus-4-6' }\n */\nexport type ModelRoles = Partial<Record<ModelRole, string>>;\n\n// =============================================================================\n// Generation Result\n// =============================================================================\n\n/**\n * The same-exchange guard (ggui#404) ended this generation: the coding model\n * made one identical tool exchange (same tool, same input, same result)\n * `repeats` times, so the loop stopped instead of burning its remaining turns.\n * Present ONLY when that guard ended the run — whether anything shipped is\n * the compiled code's to say (the loop falls back to an earlier build).\n */\nexport interface SameExchangeBreak {\n  readonly tool: string;\n  readonly repeats: number;\n}\n\nexport interface GenerationResult extends AdapterResult {\n  /** Number of generation passes the harness performed */\n  passesUsed: number;\n  /** Three-tier evaluation result (tier 0 + LLM tier 1+2 + visual) */\n  evalResult?: EvalResult;\n  /** Set only when the same-exchange guard ended the run (ggui#404); no key otherwise. */\n  sameExchangeBreak?: SameExchangeBreak;\n  /** Whether background improvement should be spawned (auto-improve mode) */\n  needsBackgroundImprovement?: boolean;\n  /** Whether at least one commit passed the self-check (compile + type check + lint) */\n  selfCheckPassed?: boolean;\n  /** Which model was used for each role (for benchmark reporting) */\n  modelRolesUsed?: Record<string, string>;\n  /**\n   * The triad this generation ran under — recorded ONLY when the caller\n   * set `designMode` on the dispatch (a bench arm switch, an operator\n   * option). Absent = the constrained default path, byte-identical to\n   * the pre-`designMode` result.\n   */\n  designMode?: DesignMode;\n  /**\n   * The canvas class the `free` prompt stated — the explicit `canvas`\n   * when the caller passed one, else the shell × screen derivation the\n   * prompt itself used. Absent in constrained mode (that prompt states\n   * no canvas) unless the caller passed one explicitly.\n   */\n  canvas?: CanvasClass;\n  /** Phase-level timing breakdown (ms) */\n  timing?: Record<string, number>;\n  /** Per-turn/per-outcome breakdown for benchmark analysis.\n   *  Machine-parseable summary — phases counted by role, outcomes counted by rejection class. */\n  breakdown?: {\n    phases: { impl: number; patch: number; evalFix: number; scaffold?: number; fill?: number };\n    outcomes: { pass: number; patchInvalid: number; selfCheckFail: number; diffFail: number };\n    evalRounds: number;\n    /** Coding-turn LLM wall-time (sum of per-turn llm=Xms). Stable semantics. */\n    llmMs: number;\n    /** Eval-round LLM wall-time (parallel LLM+visual). Added 2026-04-14. */\n    evalLlmMs?: number;\n    toolMs: number;\n    /** Eval-round wall-clock (fix: was structurally 0 before 2026-04-14). */\n    evalMs: number;\n    /** Coding-only wall-time (inner loop minus eval). Added 2026-04-14. */\n    codingMs?: number;\n    /** Session setup+cleanup wall-time (outside inner loop). Added 2026-04-14. */\n    setupMs?: number;\n  };\n}\n\n// =============================================================================\n// Generation Context\n// =============================================================================\n\nexport interface GenerationContext {\n  /** Progress callback */\n  onProgress?: (event: unknown) => void;\n\n  /** Called with initial result before evaluation loop (for early delivery) */\n  onInitialResult?: (result: {\n    componentCode: string;\n    sourceCode?: string;\n  }) => void | Promise<void>;\n\n  /** Original user prompt (for evaluation context) */\n  originalPrompt: string;\n\n  /** Design context for evaluation */\n  designContext?: string;\n\n  /** Model roles — which model to use for each internal task */\n  models?: ModelRoles;\n\n  /** Rendering context — device and shell type affect layout strategy */\n  rendering?: RenderingContext;\n\n  /** Data contract — bind the component to its consumers */\n  contract?: DataContract;\n\n  /** Evaluation configuration */\n  evaluation?: EvaluationConfig;\n}\n\n// Rendering context — canonical definition lives in\n// `../contract-context.ts` (one type, one prompt-builder family).\n// Re-exported here because this subpath is the harness's type hub.\nexport type { RenderingContext } from \"../contract-context.js\";\n\n// Data contract — re-exported from protocol (canonical types)\nexport type {\n  ActionSpec,\n  DataContract,\n  JsonSchema,\n  PropsSpec,\n  StreamSpec,\n} from \"@ggui-ai/protocol\";\n\nexport interface EvaluationConfig {\n  enabled: boolean;\n  passThreshold: number;\n  maxRounds?: number;\n  model?: string;\n  provider?: \"claude\" | \"openai\" | \"google\" | \"openrouter\";\n  maxBudgetPerEval?: number;\n  maxBudgetPerFix?: number;\n  generatorOptions?: JsonObject;\n}\n\n// =============================================================================\n// Helpers\n// =============================================================================\n\n/**\n * Strip provider prefix from LiteLLM-format model IDs.\n * 'anthropic/claude-sonnet-4-6' -> 'claude-sonnet-4-6'\n */\nfunction stripModelPrefix(modelId: string): string {\n  const slashIndex = modelId.indexOf(\"/\");\n  return slashIndex >= 0 ? modelId.slice(slashIndex + 1) : modelId;\n}\n\n/**\n * Resolve the model ID for a given role.\n * Priority: role-specific → default role → params.model\n * Strips provider prefix (e.g., 'anthropic/') so the result is a native API model ID.\n */\nexport function resolveModelForRole(\n  role: ModelRole,\n  models: ModelRoles | undefined,\n  fallback: string\n): string {\n  if (!models) return fallback;\n  const resolved = models[role] ?? models.default ?? fallback;\n  return stripModelPrefix(resolved);\n}\n"]}
+{"version":3,"sources":["../../src/harness/result-types.ts"],"names":[],"mappings":";AA2LA,SAAS,iBAAiB,OAAA,EAAyB;AACjD,EAAA,MAAM,UAAA,GAAa,OAAA,CAAQ,OAAA,CAAQ,GAAG,CAAA;AACtC,EAAA,OAAO,cAAc,CAAA,GAAI,OAAA,CAAQ,KAAA,CAAM,UAAA,GAAa,CAAC,CAAA,GAAI,OAAA;AAC3D;AAOO,SAAS,mBAAA,CACd,IAAA,EACA,MAAA,EACA,QAAA,EACQ;AACR,EAAA,IAAI,CAAC,QAAQ,OAAO,QAAA;AACpB,EAAA,MAAM,QAAA,GAAW,MAAA,CAAO,IAAI,CAAA,IAAK,OAAO,OAAA,IAAW,QAAA;AACnD,EAAA,OAAO,iBAAiB,QAAQ,CAAA;AAClC","file":"result-types.js","sourcesContent":["// packages/ui-gen/src/harness/result-types.ts\n//\n// Result shapes returned by the harness runtime.\n//\n// Two output types travel the dispatch surface:\n//   - AdapterResult      — raw adapter output (provider-direct path)\n//   - GenerationResult   — full harness output (run-loop, eval, breakdown)\n//\n// ModelRoles + resolveModelForRole live here too (the harness decides\n// which model to use per internal task).\n\nimport type { DataContract, JsonObject } from \"@ggui-ai/protocol\";\nimport type { AdapterResult } from \"../adapters/types.js\";\nimport type { RenderingContext } from \"../contract-context.js\";\nimport type { EvalResult } from \"../evaluation/types-public.js\";\nimport type { VisualLegTokens } from \"../evaluation/visual-evaluator.js\";\nimport type { CanvasClass, DesignMode } from \"../design-mode.js\";\n\n// =============================================================================\n// Model Roles\n// =============================================================================\n\n/**\n * Model role — each internal task in the harness can use a different model.\n *\n * - thinking: planning, design decisions, architecture (use frontier/thinking model)\n * - coding:   writing TSX, implementing the design (use fast model)\n * - evaluation: scoring quality, finding issues (use balanced model)\n * - default:  fallback for any role not explicitly set\n */\nexport type ModelRole = \"thinking\" | \"coding\" | \"evaluation\" | \"default\";\n\n/**\n * Maps model roles to model IDs.\n * Any role not specified falls back to 'default', then to params.model.\n *\n * @example\n * // Opus plans, Haiku codes, Sonnet evaluates\n * { thinking: 'claude-opus-4-6', coding: 'claude-haiku-4-5', default: 'claude-sonnet-4-6' }\n *\n * // All Haiku (cheapest)\n * { default: 'claude-haiku-4-5' }\n *\n * // All Opus (highest quality)\n * { default: 'claude-opus-4-6' }\n */\nexport type ModelRoles = Partial<Record<ModelRole, string>>;\n\n// =============================================================================\n// Generation Result\n// =============================================================================\n\n/**\n * The same-exchange guard (ggui#404) ended this generation: the coding model\n * made one identical tool exchange (same tool, same input, same result)\n * `repeats` times, so the loop stopped instead of burning its remaining turns.\n * Present ONLY when that guard ended the run — whether anything shipped is\n * the compiled code's to say (the loop falls back to an earlier build).\n */\nexport interface SameExchangeBreak {\n  readonly tool: string;\n  readonly repeats: number;\n}\n\nexport interface GenerationResult extends AdapterResult {\n  /** Number of generation passes the harness performed */\n  passesUsed: number;\n  /** Three-tier evaluation result (tier 0 + LLM tier 1+2 + visual) */\n  evalResult?: EvalResult;\n  /**\n   * ggui#1522 — what the IN-LOOP visual judge spent during this generation (its scoring calls, and its report-only\n   * criteria calls apart), priced at the visual agent's model. NOT inside `tokens`, which covers the coding turns\n   * and the text evaluator. And NOT the same calls as a post-generation judge's own visual pass (a benchmark\n   * report's `meta.costs.visualUsd`): a cost rule that adds both is correct, one read as the other double-counts.\n   * Absent when no in-loop visual leg ran or it reported no tokens.\n   */\n  inLoopVisualTokens?: VisualLegTokens;\n  /** Set only when the same-exchange guard ended the run (ggui#404); no key otherwise. */\n  sameExchangeBreak?: SameExchangeBreak;\n  /** Whether background improvement should be spawned (auto-improve mode) */\n  needsBackgroundImprovement?: boolean;\n  /** Whether at least one commit passed the self-check (compile + type check + lint) */\n  selfCheckPassed?: boolean;\n  /** Which model was used for each role (for benchmark reporting) */\n  modelRolesUsed?: Record<string, string>;\n  /**\n   * The triad this generation ran under — recorded ONLY when the caller\n   * set `designMode` on the dispatch (a bench arm switch, an operator\n   * option). Absent = the constrained default path, byte-identical to\n   * the pre-`designMode` result.\n   */\n  designMode?: DesignMode;\n  /**\n   * The canvas class the `free` prompt stated — the explicit `canvas`\n   * when the caller passed one, else the shell × screen derivation the\n   * prompt itself used. Absent in constrained mode (that prompt states\n   * no canvas) unless the caller passed one explicitly.\n   */\n  canvas?: CanvasClass;\n  /** Phase-level timing breakdown (ms) */\n  timing?: Record<string, number>;\n  /** Per-turn/per-outcome breakdown for benchmark analysis.\n   *  Machine-parseable summary — phases counted by role, outcomes counted by rejection class. */\n  breakdown?: {\n    phases: { impl: number; patch: number; evalFix: number; scaffold?: number; fill?: number };\n    outcomes: { pass: number; patchInvalid: number; selfCheckFail: number; diffFail: number };\n    evalRounds: number;\n    /** Coding-turn LLM wall-time (sum of per-turn llm=Xms). Stable semantics. */\n    llmMs: number;\n    /** Eval-round LLM wall-time (parallel LLM+visual). Added 2026-04-14. */\n    evalLlmMs?: number;\n    toolMs: number;\n    /** Eval-round wall-clock (fix: was structurally 0 before 2026-04-14). */\n    evalMs: number;\n    /** Coding-only wall-time (inner loop minus eval). Added 2026-04-14. */\n    codingMs?: number;\n    /** Session setup+cleanup wall-time (outside inner loop). Added 2026-04-14. */\n    setupMs?: number;\n  };\n}\n\n// =============================================================================\n// Generation Context\n// =============================================================================\n\nexport interface GenerationContext {\n  /** Progress callback */\n  onProgress?: (event: unknown) => void;\n\n  /** Called with initial result before evaluation loop (for early delivery) */\n  onInitialResult?: (result: {\n    componentCode: string;\n    sourceCode?: string;\n  }) => void | Promise<void>;\n\n  /** Original user prompt (for evaluation context) */\n  originalPrompt: string;\n\n  /** Design context for evaluation */\n  designContext?: string;\n\n  /** Model roles — which model to use for each internal task */\n  models?: ModelRoles;\n\n  /** Rendering context — device and shell type affect layout strategy */\n  rendering?: RenderingContext;\n\n  /** Data contract — bind the component to its consumers */\n  contract?: DataContract;\n\n  /** Evaluation configuration */\n  evaluation?: EvaluationConfig;\n}\n\n// Rendering context — canonical definition lives in\n// `../contract-context.ts` (one type, one prompt-builder family).\n// Re-exported here because this subpath is the harness's type hub.\nexport type { RenderingContext } from \"../contract-context.js\";\n\n// Data contract — re-exported from protocol (canonical types)\nexport type {\n  ActionSpec,\n  DataContract,\n  JsonSchema,\n  PropsSpec,\n  StreamSpec,\n} from \"@ggui-ai/protocol\";\n\nexport interface EvaluationConfig {\n  enabled: boolean;\n  passThreshold: number;\n  maxRounds?: number;\n  model?: string;\n  provider?: \"claude\" | \"openai\" | \"google\" | \"openrouter\";\n  maxBudgetPerEval?: number;\n  maxBudgetPerFix?: number;\n  generatorOptions?: JsonObject;\n}\n\n// =============================================================================\n// Helpers\n// =============================================================================\n\n/**\n * Strip provider prefix from LiteLLM-format model IDs.\n * 'anthropic/claude-sonnet-4-6' -> 'claude-sonnet-4-6'\n */\nfunction stripModelPrefix(modelId: string): string {\n  const slashIndex = modelId.indexOf(\"/\");\n  return slashIndex >= 0 ? modelId.slice(slashIndex + 1) : modelId;\n}\n\n/**\n * Resolve the model ID for a given role.\n * Priority: role-specific → default role → params.model\n * Strips provider prefix (e.g., 'anthropic/') so the result is a native API model ID.\n */\nexport function resolveModelForRole(\n  role: ModelRole,\n  models: ModelRoles | undefined,\n  fallback: string\n): string {\n  if (!models) return fallback;\n  const resolved = models[role] ?? models.default ?? fallback;\n  return stripModelPrefix(resolved);\n}\n"]}

package/dist/harness/runtime.d.ts

@@ -1,13 +1,14 @@
 import { G as GenerationProfileInput } from '../styling-profile-CxfvV48g.js';
-import { Q as QualityConfig } from '../types-public-Bk6d1WLN.js';
+import { Q as QualityConfig, b as AgentConfig } from '../types-public-B7wn4ce5.js';
 import { DataContract, JsonObject, GadgetDescriptor } from '@ggui-ai/protocol';
-import { A as AgentConfig } from '../llm-router-DwwUK1bz.js';
-import { C as CanvasClass, D as DesignMode } from '../design-mode-DLufhrNp.js';
-export { g as generateBoilerplate } from '../generate-9IeNT4ZY.js';
-import '@ggui-ai/mcp-server-core';
-import '../axes-CzLEMDeB.js';
-import '../types-BOvHNG7K.js';
+import { C as CriteriaBank } from '../bank-Ctdp3FUA.js';
+import { C as CanvasClass, D as DesignMode } from '../design-mode-DGWN_Gkm.js';
+export { g as generateBoilerplate } from '../generate-tSO_shQ0.js';
 import '../llm.js';
+import '@ggui-ai/mcp-server-core';
+import '../axes-CKVCzR_x.js';
+import '../types-Dh1s4K9j.js';
+import 'zod';
 
 type Provider = "anthropic" | "openai" | "google" | "openrouter";
 /** Agent config — provider + model pair. */
@@ -73,6 +74,19 @@ interface SingleComponentParams {
          * default tokens. Never part of the prompt or the identity.
          */
         cssTokens?: string;
+        /** ggui#1436 — the visual criteria bank (data, parsed by `parseCriteriaBank`); absent = no block, today's judge. */
+        criteriaBank?: CriteriaBank;
+        /** ggui#1436 — the caller's item kind for the selector (a bootstrap mint's `hello`); absent = not gated by kind. */
+        criteriaKind?: string;
+        /**
+         * ggui#1542 — which of the VISUAL leg's own findings may start another eval round when the quality mode is `fast`.
+         * `'critical'` (the default, today's behaviour): only a critical finding, which the harness makes a blocking fail.
+         * `'major'`: the visual JUDGE's major findings (tier-2 warns) too. The fit measurement's own warn-level overflow is
+         * not among them (fit runs with or without a judge, so it is its own feedback source), and neither are the text
+         * evaluator's warns. The other quality modes already act on every warn, so the setting changes nothing there. The
+         * fix turn is fed as any continuing round is (fails first, then warns).
+         */
+        actOn?: "critical" | "major";
     };
     onProgress?: (event: unknown) => void;
     onInitialResult?: (result: {