hack/tools/scripts/llm-key-hunter/results/_cache/content/alexalexalex222_frontend-design-loop-mcp.json

{"repo": "alexalexalex222/frontend-design-loop-mcp", "files": {"config/config.yaml": {"c65f7caa09db9c4ce04c2ed0872d401e291c525a": {"t": 1786068509.9842818, "n": 8612, "h": "3e01614e4f697d3b", "c": "# TITAN-4-DESIGN Dataset Factory Configuration\n\nmodels:\n  # Stage 1: Planner (generates UI_SPEC)\n  planner:\n    provider: vertex\n    model: deepseek-ai/deepseek-v3.2-maas\n    publishable: true\n    max_tokens: 65000\n    temperature: 0.7\n\n  # Stage 2: UI Generators (generate code candidates)\n  # Note: Thinking models need extra tokens for <think> blocks + code\n  # INCREASED to 32k for FULL landing pages (hero + features + pricing + FAQ + footer)\n  # Kimi K2 supports 32k output, MiniMax M2 supports similar\n  ui_generators:\n    # Multi-temperature spread for diversity.\n    #\n    # NOTE: Candidate count per task = sum(variants for each entry).\n    # Current spread:\n    # - Kimi: 4 temps × 1 variant = 4 candidates\n    # - MiniMax: 4 temps × 1 variant = 4 candidates\n    # Total = 8 candidates per task (keeps throughput sane while still spanning temps).\n\n    # === Kimi K2: conservative → balanced → creative → bold ===\n    - provider: vertex\n      model: moonshotai/kimi-k2-thinking-maas\n      publishable: true\n      variants: 1\n      max_tokens: 65000\n      temperature: 0.70\n\n    - provider: vertex\n      model: moonshotai/kimi-k2-thinking-maas\n      publishable: true\n      variants: 1\n      max_tokens: 65000\n      temperature: 0.85\n\n    - provider: vertex\n      model: moonshotai/kimi-k2-thinking-maas\n      publishable: true\n      variants: 1\n      max_tokens: 65000\n      temperature: 1.00\n\n    - provider: vertex\n      model: moonshotai/kimi-k2-thinking-maas\n      publishable: true\n      variants: 1\n      max_tokens: 65000\n      temperature: 1.15\n\n    # === MiniMax M2: include the same 4 temp points ===\n    - provider: vertex\n      model: minimaxai/minimax-m2-maas\n      publishable: true\n      variants: 1\n      max_tokens: 65000\n      temperature: 0.70\n\n    - provider: vertex\n      model: minimaxai/minimax-m2-maas\n      publishable: true\n      variants: 1\n      max_tokens: 65000\n      temperature: 0.85\n\n    - provider: vertex\n      model: minimaxai/minimax-m2-maas\n      publishable: true\n      variants: 1\n      max_tokens: 65000\n      temperature: 1.00\n\n    - provider: vertex\n      model: minimaxai/minimax-m2-maas\n      publishable: true\n      variants: 1\n      max_tokens: 65000\n      temperature: 1.15\n    # NOTE: GLM-4.7 and MiniMax M2.1 are NOT on Vertex AI MaaS yet\n    # Only available: Kimi K2, MiniMax M2, DeepSeek V3.2\n\n  # Stage 3: Patcher/Fixer (fixes build errors)\n  patcher:\n    # Vertex-only: do NOT use OpenRouter (no credits).\n    provider: vertex\n    model: deepseek-ai/deepseek-v3.2-maas\n    publishable: true\n    max_tokens: 65000\n    temperature: 0.3\n\n  # Vision Judge (scores screenshots - NOT used for training)\n  # Uses Gemini native API for multimodal vision capabilities\n  vision_judge:\n    provider: gemini\n    # Prefer the stable Gemini 3 Flash name; the code will fall back to preview if not available\n    # in this Vertex project.\n    model: gemini-3.0-flash\n    publishable: false\n    max_tokens: 65000\n    temperature: 0.7\n    force_adc: true\n\n  # Refinement Loop Models (quality improvement based on vision feedback)\n  # These are DIFFERENT from patcher - patcher fixes build errors,\n  # refiners improve visual/UX quality based on judge feedback.\n\n  # Refine Reasoner: Plans targeted fixes from judge feedback\n  refine_reasoner:\n    provider: vertex\n    model: deepseek-ai/deepseek-v3.2-maas\n    publishable: true\n    max_tokens: 65000\n    temperature: 0.5\n\n  # Refine Coder: Applies visual/UX fixes (needs strong coding model)\n  refine_coder:\n    provider: vertex\n    model: moonshotai/kimi-k2-thinking-maas\n    publishable: true\n    max_tokens: 65000  # Match ui_generators for full rewrites\n    temperature: 0.7\n\n  # Megamind models (avoid homogeneity: do not fall back to planner for every sub-reasoner)\n  megamind_bold:\n    provider: vertex\n    model: moonshotai/kimi-k2-thinking-maas\n    publishable: true\n    max_tokens: 65000\n    temperature: 0.9\n\n  megamind_minimal:\n    provider: vertex\n    model: minimaxai/minimax-m2-maas\n    publishable: true\n    max_tokens: 65000\n    temperature: 0.8\n\n  megamind_safe:\n    provider: vertex\n    model: deepseek-ai/deepseek-v3.2-maas\n    publishable: true\n    max_tokens: 65000\n    temperature: 0.7\n\n  megamind_synthesizer:\n    provider: vertex\n    model: deepseek-ai/deepseek-v3.2-maas\n    publishable: true\n    max_tokens: 65000\n    temperature: 0.7\n\npipeline:\n  # Scoring threshold (0-10)\n  vision_score_threshold: 8.0\n\n  # Allow generator temps > 1.0 (hard-clamped in uigen to this cap).\n  generator_temp_cap: 1.3\n\n  # === WINNER SELECTION (Fix E) ===\n  # Blend judge score (0-10) with section creativity (0-1) so that distinctive\n  # candidates can win when judge scores are close.\n  selection:\n    # Creativity is the north star: pick the most distinctive candidate once shippable is satisfied.\n    # \"creativity_first\" ignores weights for ordering, but keep weights for logging/compat.\n    mode: creativity_first\n    judge_weight: 0.4\n    creativity_weight: 0.6\n\n  # === CREATIVITY GATE (north star) ===\n  # Keep in observe mode by default (enforce=false) so throughput isn't destroyed while we tune prompts.\n  creativity_gate:\n    enabled: true\n    min_avg: 0.7\n    enforce: false\n\n  # === MEGAMIND 3-PASS REASONING ===\n  # When enabled, uses 3 parallel sub-reasoners (bold/minimal/safe) to generate plans,\n  # then synthesizes them into a unified best-of-3 plan before UI generation.\n  # This increases plan diversity and quality at the cost of 4x planner API calls.\n  megamind_enabled: true\n  megamind_v2_enabled: true\n\n  # === STYLE ROUTING + EVOL (prevent monoculture) ===\n  style_routing_enabled: true\n  style_gates_enabled: true\n  style_gates_enforce: false\n  evol_enabled: true\n  evol_passes: 2\n\n  # === CREATIVE DIRECTOR MODE ===\n  # When enabled, replaces numeric scoring with qualitative feedback.\n  # This mode is more generous with creative risk-taking and focuses\n  # on production readiness rather than aesthetic preferences.\n  # Refinement is guided by specific feedback (missing_for_production)\n  # rather than score thresholds.\n  creative_director_mode: false\n\n  # === REFINEMENT LOOP THRESHOLDS ===\n  # Only used when creative_director_mode is false\n  refinement_enabled: true\n  # Don't over-polish: high thresholds tend to normalize creativity.\n  refine_pass2_threshold: 7.0   # Refine if score < 7.0\n  refine_pass3_threshold: 8.0   # Refine again if score < 8.0\n  max_refine_passes: 2\n\n  # === REFINEMENT SKIP POLICY (Fix F) ===\n  # Refinement tends to normalize distinctive layouts. If deterministic gates pass\n  # AND section creativity is high, skip refinement to preserve creative intent.\n  refinement:\n    skip_for_high_creativity: true\n    creativity_skip_threshold: 0.7\n\n  # === BROKEN VISION GATE (blank/error pages) ===\n  broken_vision_gate_enabled: true\n  broken_vision_gate_min_confidence: 0.85\n\n  # Max fix attempts per candidate\n  max_fix_rounds: 2\n\n  # Whether to attempt polish loop for low-scoring candidates\n  polish_loop_enabled: true\n\n  # Tasks per niche (7 = 6 page types + 1 edit)\n  tasks_per_niche: 7\n\n  # Total niches to generate\n  total_niches: 100\n\n  # Timeouts (milliseconds)\n  model_timeout_ms: 120000\n  build_timeout_ms: 240000\n  render_timeout_ms: 90000\n\n  # === DETERMINISTIC QUALITY GATES (AXE + LIGHTHOUSE) ===\n  # Runs measurable validation before subjective vision judging.\n  deterministic_gates_enabled: true\n  # Start in \"observe\" mode; flip true once thresholds are tuned.\n  deterministic_gates_enforce: false\n\n  axe_gate_enabled: true\n  axe_fail_impacts: [\"critical\"]\n  axe_timeout_ms: 60000\n\n  lighthouse_gate_enabled: true\n  lighthouse_preset: desktop\n  lighthouse_timeout_ms: 180000\n  lighthouse_min_scores:\n    performance: 0.35\n    accessibility: 0.70\n    best_practices: 0.70\n    seo: 0.60\n\nbudget:\n  # Concurrent requests per provider\n  concurrency_vertex: 5\n\n  # Rate limits (requests per minute)\n  requests_per_min_vertex: 60\n\n  # Optional limits\n  max_total_tasks: null  # null = run all tasks\n  stop_after_usd: null   # null = no budget limit\n\nexport:\n  # Number of niches to hold out for generalization testing\n  holdout_niches: 12\n\n  # Validation split (within non-holdout data)\n  validation_split: 0.08\n\n  # Specific niche IDs to hold out (optional, overrides random selection)\n  holdout_niche_ids: []\n\n# Optional GCS backup\ngcs:\n  bucket: null  # Set to bucket name to enable\n  prefix: frontend-design-loop-mcp-outputs\n  upload_interval_tasks: 50\n\n# Vertex AI configuration\nvertex:\n  endpoint_template: \"https://{region}-aiplatform.googleapis.com/v1/projects/{project}/locations/{region}/endpoints/openapi/chat/completions\"\n"}}}}