start-vibing 2.0.42 → 2.0.44

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "start-vibing",
-	"version": "2.0.42",
+	"version": "2.0.44",
 	"description": "Setup Claude Code agents, skills, and hooks in your project. Smart copy that preserves your custom domains and configurations.",
 	"type": "module",
 	"bin": {

package/template/.claude/CLAUDE.md

@@ -46,17 +46,16 @@ All agents MUST use these MCP servers when applicable:
 | `context7`            | Up-to-date library documentation   | research, analyzer, tester              |
 | `sequential-thinking` | Complex problem-solving            | orchestrator, analyzer, final-validator |
 | `memory`              | Persistent knowledge graph         | domain-updater, commit-manager          |
-| `playwright`          | Browser automation and E2E testing | tester, ui-ux-reviewer                  |
+| `playwright`          | Browser automation (via MCP tools) | ui-ux-reviewer                          |
 | `nextjs-devtools`     | Next.js specific development tools | analyzer (Next.js projects)             |
 | `mongodb`             | MongoDB database operations        | tester, security-auditor                |
 
 ### Agent MCP Usage Rules
 
 - **research agent**: MUST use `context7` for library docs before recommending patterns
-- **tester agent**: MUST use `playwright` for E2E tests
 - **analyzer agent**: SHOULD use `context7` to verify current API patterns
 - **domain-updater**: SHOULD use `memory` to persist patterns across sessions
-- **ui-ux-reviewer**: MUST use `playwright` to verify UI implementations
+- **ui-ux-reviewer**: MUST use `playwright` MCP tools to verify UI implementations
 
 ---
 
@@ -205,9 +204,8 @@ All implementations MUST:
 - [ ] Pass typecheck (command from config)
 - [ ] Pass lint (command from config)
 - [ ] Pass unit tests (command from config)
-- [ ] Pass E2E tests (command from config)
 - [ ] Pass build (command from config)
-- [ ] Have E2E tests with real auth
+- [ ] Use Playwright MCP for browser testing when needed
 - [ ] Have documentation in `docs/`
 - [ ] Be security audited
 - [ ] Be committed with conventional commits
@@ -215,6 +213,48 @@ All implementations MUST:
 
 ---
 
+## Opus 4.6 Best Practices
+
+### Effort Levels
+
+| Level | Use For |
+|-------|---------|
+| `max` | Orchestrators, complex architecture (Opus only) |
+| `high` | Default - complex agentic tasks |
+| `medium` | Most workloads, balance speed/cost |
+| `low` | Subagents, high-volume tasks |
+
+### Prompting Rules
+
+- **AVOID**: Words like "CRITICAL", "MUST", "be thorough" - causes overtriggering
+- **USE**: "Use this tool when it would enhance understanding"
+- Keep prompts minimal - Opus 4.6 overthinks with verbose instructions
+- Explicit instructions work better than implied expectations
+
+### Model Selection
+
+| Model | Use Case |
+|-------|----------|
+| **Opus 4.6** | Team leads, orchestrators, deep reasoning |
+| **Sonnet 4.6** | Teammates, subagents (5x cheaper, near-identical performance) |
+| **Haiku 4.5** | Simple classification, no thinking needed |
+
+### Interleaved Thinking
+
+- Enabled automatically with adaptive thinking
+- Claude reasons between every tool call
+- Pass thinking blocks back unmodified in tool loops
+- Provides transparency into Claude's reasoning process
+
+### Cost Control
+
+- Set `max_tokens` to 64K+ at high effort
+- Use `low` effort for subagents
+- Billed for FULL thinking tokens (not summarized output)
+- Thinking tokens count toward context window
+
+---
+
 ## Domain Updater Agent
 
 The **domain-updater** runs BEFORE commit-manager to ensure git stays clean.

package/template/.claude/hooks/user-prompt-submit.ts

@@ -55,17 +55,38 @@ async function main(): Promise<void> {
 
 	const systemMessage = `TASK WORKFLOW (English only):
 
-0. READ both CLAUDE.md (project root) and .claude/CLAUDE.md before making any changes. These contain project rules, architecture, and context you need.
-1. CREATE a detailed todo-list (TaskCreate) breaking down the user's request into subtopics/steps.
+0. READ both CLAUDE.md (project root) and .claude/CLAUDE.md before making any changes.
+
+1. CREATE a detailed todo-list (TaskCreate) breaking down the request into steps.
+   - This is MANDATORY for ALL tasks, not optional.
+   - Include "Update CLAUDE.md with changes" as final task.
+
 2. WORK through each item sequentially — mark in_progress when starting, completed when done.
-3. COMMIT using conventional commits (feat/fix/refactor/docs/chore). Use the commit-manager agent for proper git workflow: branch → implement → commit → merge to main.
-4. UPDATE CLAUDE.md before finishing — this is the project's fast-recall memory across sessions:
-   a. "## Last Change" section (date: ${today}, branch, summary). Keep only the latest, never stack.
-   b. If you changed how the app works (new flows, changed workflows, new patterns): update the relevant CLAUDE.md sections (Architecture, Workflow, Rules, UI patterns, etc.) so the next session knows.
-   c. If you added new rules, gotchas, or forbidden patterns: add them to the appropriate section.
-   d. If you changed config, stack, or tooling: update the Stack/Configuration sections.
-   CLAUDE.md is the single source of truth for quick context. Anything not recorded here will be forgotten next session.
-5. RUN stop-validator before finishing: npx tsx .claude/hooks/stop-validator.ts`;
+
+3. UI/UX IMPLEMENTATION:
+   - BEFORE implementing any UI: invoke ui-mobile, ui-tablet, ui-desktop agents in PARALLEL.
+   - RESEARCH competitors in the same niche using competitor-analyzer or research-web agent.
+   - UI changes without market research will be REJECTED by final-validator.
+
+4. TESTING (via Playwright MCP):
+   - Use the Playwright MCP server (mcp__playwright__*) for browser testing.
+   - Do NOT run bun run test:e2e - use MCP tools directly.
+   - Test on 3 viewports: mobile (375px), tablet (768px), desktop (1280px).
+
+5. COMMIT using conventional commits via commit-manager agent.
+
+6. UPDATE CLAUDE.md BEFORE finishing (MANDATORY):
+   a. "## Last Change" section (date: ${today}, branch, summary). Keep only latest.
+   b. Architecture changes: Update Architecture section with new patterns, file locations.
+   c. Business rules: Add new rules to relevant sections (not just summary).
+   d. UI patterns: Document new components, design decisions, aesthetic rules.
+   e. Workflows: Update Workflow section if process changed.
+   f. Gotchas: Add to FORBIDDEN or NRY sections if discovered.
+   g. Config/Stack: Update if tooling changed.
+
+   CLAUDE.md is the SINGLE SOURCE OF TRUTH. Document everything that affects future sessions.
+
+7. RUN stop-validator before finishing: npx tsx .claude/hooks/stop-validator.ts`;
 
 	console.log(JSON.stringify({ continue: true, systemMessage }));
 	process.exit(0);

package/template/.claude/settings.json

@@ -1,5 +1,5 @@
 {
-	"model": "claude-opus-4-5-20251101",
+	"model": "opus",
 	"max_tokens": 8192,
 	"max_turns": 100,
 
@@ -239,7 +239,7 @@
 		},
 		"testing": {
 			"unit_tests_required": true,
-			"e2e_for_ui_features": true,
+			"playwright_mcp_for_browser_tests": true,
 			"data_testid_required": true,
 			"edge_cases_research_required": true
 		},
@@ -266,11 +266,6 @@
 			"required": true,
 			"blocking": true
 		},
-		"test_e2e": {
-			"command": "bun run test:e2e",
-			"required": "when_ui_changes",
-			"blocking": true
-		},
 		"build": {
 			"command": "bun run build",
 			"required": true,