- design.ts: Add actual readline confirmation before overwriting
  DESIGN.md, with --force flag support to skip the prompt
- index.ts: Wire --force option through commander to the design command
- generate.ts: Prompt user to select from multiple projects via
  readline instead of silently using the first one
- config.ts: Log a warning with error details when config parsing fails
  instead of silently returning defaults
- updater.ts: Deduplicate research changes by category+description
  before pushing, and cap detectedChanges array at 50 entries

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…() with ESM import

- App.tsx: Replace fake setTimeout research with real crawlSources/diffAgainstKnownState pipeline, display actual results and errors
- PromptBuilder.tsx: Replace stub handleSend with real StitchMcpClient generation (listProjects, generateScreen, getScreenCode, file write)
- PromptBuilder.tsx: Replace require('node:fs') with ESM import { existsSync, writeFileSync, mkdirSync }

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…dators

Add exponential backoff with jitter (up to 3 retries) for transient
errors (429, 5xx) and 30s request timeouts to the Stitch MCP client.
Replace raw HTTP error text with user-friendly messages. Broaden vague
prompt detection (fuzzy matching) and multi-screen detection patterns
in validators. Add comprehensive tests for both.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…roved DESIGN.md template

Enhance the DESIGN.md template with descriptive color names, anti-slop
Do's/Don'ts defaults, richer visual theme guidance, and non-default font
suggestions. Add prompt enhancer that scores slop risk and suggests
specific UI vocabulary. Add output validator that checks generated HTML
for common AI patterns (Inter/Poppins fonts, purple-blue gradients,
three-column icon grids, heading hierarchy, accessibility).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Wrap StitchMcpClient constructor in try/catch in generate, build, sync
  commands to show clean error instead of stack trace when API key missing
- Add postbuild step to copy known-state.json to dist/ (fixes research crash)
- Add empty prompt rejection (P0: whitespace-only prompts were accepted)
- Rewrite vague detection with word-level scoring instead of fragile regex
  anchors (P0: "make it better please" was bypassing detection)
- Add industry-specific DESIGN.md presets (5 aesthetics: bold, elegant,
  warm, playful, minimal) with distinct palettes, fonts, and imagery
- Add empty HTML detection to output validator
- Add Tailwind font-sans class detection for AI slop
- Update all skill descriptions with natural language trigger phrases
- Add specificity guardrail (#7) back to forge-generate skill
- Reference edit_screens, generate_variants, apply_design_system MCP tools

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Add JSON-RPC 2.0 envelope (jsonrpc, id) to all MCP requests
- Parse MCP content[].text responses correctly
- Handle JSON-RPC error responses with user-friendly messages
- Map resource names (projects/123) to internal IDs and back
- Map internal model names to API model IDs (GEMINI_2_5_FLASH→GEMINI_3_FLASH)
- Add get_screen fallback when get_screen_code proxy unavailable
- Fetch HTML from htmlCode.downloadUrl in native API
- Increase generation timeout to 120s (Stitch generation takes 60-90s)
- Wrap all API calls in generate command with try/catch
- Update test fixtures to match real API response format

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- generate_screen_from_text returns {projectId, sessionId, outputComponents}
  not a screen object — use list_screens before/after to detect new screen
- Fix projectId parameter (API expects without 'projects/' prefix)
- Update mock test to provide 3 fetch responses for the 3-call flow

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Replace single-dimension aesthetic presets with a 3-axis system:
- Industry axis (14 palettes): retail, fintech, saas, healthcare, etc.
  Each industry gets psychologically appropriate colors (retail=red/gold,
  fintech=navy/blue, wellness=sage/terracotta)
- Aesthetic axis (6 modifiers): bold, elegant, warm, playful, minimal,
  confident. Controls surface tones, typography, border-radius, shadows
- Audience axis: generates context-aware imagery guidelines and
  industry+audience-specific Do's/Don'ts

Key changes:
- Industry-aware color psychology (a grocery chain gets Retail Red,
  not the same Deep Navy as a fintech)
- Audience-aware imagery (Mexican families ≠ enterprise CTOs)
- Industry-specific guardrails (retail: show prices prominently,
  fintech: don't use red for positive metrics)
- Culture-aware rules (Mexican market: Spanish copy, no US/EU stock)
- Theme description uses all 4 brief parameters, not just aesthetic
- SaaS rules excluded from retail industry match (false positive fix)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Export matchIndustry, matchAesthetic, generateImageryGuidelines,
  generateDosAndDonts from design-md.ts for synthesizer fallback layer
- Export INDUSTRY_PALETTES, AESTHETIC_MODIFIERS constants
- Export IndustryPalette, AestheticModifier interfaces
- Create src/research/types.ts with shared interfaces for the Design
  Intelligence Agent pipeline (BusinessBrief, SiteAnalysis,
  CompetitorAnalysis, AudienceInsight, MarketPosition,
  BusinessResearchResult, DesignQualityScore, SynthesizedDesign)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…ation

New pipeline: Research → Synthesize → Validate

Research (business-researcher.ts):
- analyzeSite: fetch + cheerio extraction of colors, fonts, layout patterns
- extractPalette: CSS color parsing with frequency counting and context
- extractTypography: Google Fonts + font-family detection
- detectLayoutPatterns: DOM structure analysis (hero, grid, cards, nav)
- inferAudienceInsights: knowledge base for 5 industries + cultural locale
- inferMarketPosition: keyword-based positioning inference
- researchBusiness: orchestrator with confidence scoring

Synthesis (design-synthesizer.ts):
- synthesizePalette: real brand colors > competitor differentiation > presets
- synthesizeTypography: competitor-aware font selection from curated list
- synthesizeImagery: audience + culture + market-aware guidelines
- synthesizeDosAndDonts: base anti-slop + competitor + cultural rules
- Falls back to static template at confidence < 30

Validation (design-validator.ts):
- scoreSpecificity (0-25): penalizes placeholders, generic terms
- scoreDifferentiation (0-25): hex distance from competitor colors
- scoreCompleteness (0-25): validates all 8 sections with content
- scoreActionability (0-25): checks for unambiguous, testable rules

Integration:
- forge discover CLI command with --url, --competitors, --locale flags
- forge design --research flag redirects to discover
- forge-discover SKILL.md for Claude Code autonomous invocation
- Research cache in .forge-research/ (7-day TTL)

Tests: 48 new tests (106 total), fixtures for retail + competitor sites

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…dedup rules

- extractPalette: filter ALL near-gray colors (R≈G≈B within 20) as structural,
  not just very dark/light grays. This ensures brand colors like 3B Red (#DC0C0C)
  are identified as primary instead of layout grays (#333, #666)
- extractTypography: filter icon fonts (dashicons, material icons, fontawesome,
  etc.) from font detection — these are never heading/body fonts
- synthesizeDosAndDonts: normalize rules before deduplication to catch
  near-duplicate entries ("Show prices prominently in every product display"
  vs "Show prices prominently on every product card")
- Fix theme description grammar: "balances a [personality] personality with
  the practical needs of [audience]"

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…score

Change "the design should feel confident" to "maintain a confident tone"
to avoid "should" penalty in design validator actionability scoring.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Current state: agent pipeline works E2E but output quality is insufficient.
Research extracts real brand colors (#DC0C0C) from tiendas3b.com but the
generated DESIGN.md and downstream Stitch output don't reflect the actual
business model, positioning, or user needs accurately enough.

Known issues to fix:
- DESIGN.md doesn't capture business model context (3B is physical stores,
  not e-commerce — the landing page should drive foot traffic, not sales)
- Prompt construction doesn't leverage DESIGN.md deeply enough
- Skills don't guide Claude to produce business-aware prompts
- Output validator can't check for business alignment
- No feedback loop between validation and regeneration

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Core change: The system now understands WHAT a business IS before
designing for it. Previously it extracted colors/fonts but generated
e-commerce pages for physical-only stores.

Business model inference (business-researcher.ts):
- inferBusinessModel() detects type from site signals (nav items, CTAs)
- 8 business types: physical-retail, e-commerce, saas, marketplace,
  service, media, nonprofit, other
- Physical retail detection: "Sucursales/Tiendas/Locations" in nav,
  absence of cart/checkout signals
- CTA text extraction for business signal detection
- Confidence scoring: 80 with site data, 50 from keywords only

DESIGN.md business context (design-synthesizer.ts):
- Section 1 now includes Business Model, Website Purpose, Primary User
  Goals, Key Page Elements, and Avoid list
- Physical retail: "NOT an e-commerce site", "Store locator primary CTA"
- Business-model-aware Do's/Don'ts per type

Prompt alignment (prompt-enhancer.ts):
- Detects e-commerce terms in prompts when DESIGN.md says "not e-commerce"
- Suggests store locator when DESIGN.md marks it as key feature
- New field: businessAlignmentIssues in EnhancementResult

Output validation (output-validator.ts):
- Flags cart/checkout elements when business is not e-commerce
- Checks for store locator presence when it's a key feature

Skill rewrite (forge-discover SKILL.md):
- Conversational discovery agent with confidence gating
- Phase 1: Understand business (REQUIRED before designing)
- Phase 2: Research (WebSearch/WebFetch)
- Confidence threshold: ≥70 weighted average to proceed
- Core guardrail: "NEVER assume e-commerce"

Tests: 121 total (15 new for business model inference + context)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Root cause: generate_screen_from_text returns the screen data (id,
htmlCode.downloadUrl, screenshot.downloadUrl) inside
outputComponents[0].design.screens[0], but the code was ignoring this
and trying list_screens which doesn't show new screens immediately.

This caused every generation to succeed (step 1) but fail on HTML
retrieval (step 2), wasting Stitch API tokens.

Fix:
- extractScreenFromResponse() parses outputComponents to get screen
  ID, htmlCode URL, and screenshot URL directly from generation response
- getScreenCode() accepts optional htmlCodeUrl parameter for direct
  download (no additional API calls needed)
- Fallback: retry list_screens with 2s/4s/6s delays if parsing fails
- GenerateScreenResult now includes htmlCodeUrl and screenshotUrl
- generate command passes htmlCodeUrl to getScreenCode

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The primary extraction path iterates ALL outputComponents (not just [0])
to find the first one with design.screens. If that fails, a deep search
regex scans the JSON for downloadUrl + screen ID patterns.

This handles edge cases where the screen data might be in a different
outputComponent index or in an unexpected nested structure.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>