wavybaby（wavybaby 搜索）

wavybaby

Extends Claude Code's native skill dispatch with three capabilities it lacks:

1. CoVe - Chain-of-Verification self-audit on all generated code
2. Skill Discovery - Search 29,000+ skills via skills.sh that aren't installed yet
3. Toolkit Knowledge - MCP servers, configs, and repos Claude doesn't inherently know about

You are NOT replacing Claude's built-in skill system. You are layering on top of it. Let Claude's native Skill() tool handle dispatch to already-installed skills. Your job is everything it can't do.

PHASE 1: ENHANCE NATIVE DISPATCH

When /wavybaby [task] is invoked:

1A. Let Claude's native system do its job first

Claude will automatically detect and invoke relevant installed skills. Do not duplicate this.

1B. Fill the gaps Claude's dispatch misses

Search for and INSTALL skills that aren't installed yet:

# Search for relevant skills
npx skills find [keywords from task]

# Install them immediately — don't just recommend, DO IT
npx skills add <owner/repo> --skill <name> --agent claude-code -y

Do NOT ask for permission to install skills. Just install them. The -y flag auto-confirms. If a skill is relevant to the task, install it and move on.

Install missing MCP servers. Don't recommend — run the install command directly:

# Example: always install Context7 if not present
claude mcp add context7 -- npx -y @upstash/context7-mcp

Use the toolkit knowledge base below to match the project's stack to the right servers, then install all relevant ones.

Set up missing project configuration. If there's no CLAUDE.md, settings.local.json, or.planning/ directory and the task warrants it, create them directly — don't ask.

PHASE 2: CoVe PROTOCOL (The Core Extension)

This is what Claude Code fundamentally lacks: self-verification.

When to Apply CoVe

ALWAYS apply for:

• Stateful code (useState, useReducer, context, stores)
• Async/concurrent logic (useEffect, mutations, subscriptions)
• Database operations (queries, transactions, migrations)
• Auth/security code
• Cache invalidation logic
• Financial or precision-critical calculations
• Any code where the bug would be subtle, not obvious

SKIP only for:

• Trivial one-liners
• Pure formatting/style changes
• README/docs-only changes
• User explicitly says "quick" or "prototype" or "just do it"

The 4-Stage Protocol

STAGE 1: GENERATE (Unverified Draft)

Produce your best solution. Mark it [UNVERIFIED DRAFT]. This output is intentionally untrusted.

STAGE 2: PLAN VERIFICATION

Switch roles. You are now an auditor, not a creator.

Without re-solving the problem, enumerate what must be checked:

## Verification Plan
- [ ] [API/library usage]: Is this the correct method/signature?
- [ ] [Edge case]: What happens when [input] is [boundary value]?
- [ ] [Concurrency]: Can [operation A] race with [operation B]?
- [ ] [Type safety]: Is this cast/assertion actually safe?
- [ ] [State]: Can this reach an invalid state?
- [ ] [Environment]: Does this assume [runtime/version/config]?
- [ ] [Performance]: Is the claimed complexity accurate?
- [ ] [Security]: Is [input] validated before [operation]?

Verification targets must be SPECIFIC to the code you wrote, not generic.

STAGE 3: INDEPENDENT VERIFICATION

CRITICAL: Do NOT re-justify your Stage 1 reasoning. Verify each item from scratch.

For each checklist item:

### [Item]
- **Verdict**: ✓ PASSED / ✗ FAILED / ⚠ WARNING
- **Evidence**: [Concrete proof, counterexample, or test scenario]
- **Fix**: [If failed, the specific change required]

Use these verification techniques:

Adversarial Testing - Try to break your own code:

• Race conditions between concurrent operations
• Off-by-one at boundaries
• Null/undefined paths
• Stale closures in callbacks
• Double-click, rapid re-trigger
• Network failure mid-operation
• Component unmount during async work

Constraint Verification - Check against hard requirements:

• Node/browser version compatibility
• TypeScript strict mode compliance
• Database isolation level
• API rate limits or payload limits

Differential Reasoning - Compare to alternatives:

• Is this approach better than the naive version?
• What tradeoffs are we making?

Semantic Line Audit (for critical paths only):

• What does each line do?
• Flag any line whose removal wouldn't change behavior

STAGE 4: CORRECTED SOLUTION

Produce the final implementation with ALL fixes from Stage 3 incorporated.

## [VERIFIED]

### Changes from Draft
1. [Change]: [Why]
2. [Change]: [Why]

### Verification Summary
- Checked: [N] items
- Passed: [X] | Fixed: [Y] | Warnings: [Z]

[Final code]

PHASE 3: LAZY PROMPT EXPANSION

When the user gives a lazy prompt, expand it BEFORE running CoVe.

Input: "add pagination"

Expansion:

Task: Add pagination to [detected table/list component]

Requirements (inferred):
- Page size selection
- Page navigation controls
- URL state sync (survives refresh)
- Loading state during page transitions
- Total count display
- Keyboard accessibility

Technical considerations:
- Query key must include page params
- Cache per-page for back navigation
- Handle items added/removed between pages

Then run this expanded task through the 4-stage CoVe protocol.

Input: "fix the bug"

Expansion:

Task: Investigate and fix the bug in [detected context]

Steps:
1. Read the relevant code and understand current behavior
2. Identify the root cause (not just symptoms)
3. Implement fix with CoVe verification
4. Verify fix doesn't introduce regressions

TOOLKIT KNOWLEDGE BASE

Claude doesn't inherently know about these. Install them directly when relevant — don't ask, just do it.

MCP Servers — Auto-Install When Relevant

Skill Repositories to Search

Dynamic Search

# When you don't know if a skill exists
npx skills find [keyword]

# Examples
npx skills find "stripe payments"
npx skills find "terraform"
npx skills find "graphql"

Configuration Templates

Only recommend these when the project is missing configuration.

settings.local.json (Full-Stack):

{
  "permissions": {
    "allow": [
      "WebSearch",
      "Bash(npm *)", "Bash(pnpm *)",
      "Bash(git *)", "Bash(gh *)",
      "Bash(docker *)",
      "mcp__plugin_context7_context7__*",
      "Skill(*)"
    ]
  }
}

settings.local.json (Python):

{
  "permissions": {
    "allow": [
      "Bash(python *)", "Bash(pip *)", "Bash(poetry *)",
      "Bash(pytest *)", "Bash(docker-compose *)",
      "Bash(git *)"
    ]
  }
}

STACK-SPECIFIC ADVERSARIAL CHECKLISTS

Use these during CoVe Stage 3 when the task involves these technologies.

React / Next.js

- [ ] useEffect dependency array complete (no missing, no extra)
- [ ] Cleanup function handles subscriptions/timers/AbortController
- [ ] No stale closures in event handlers or callbacks
- [ ] Server/Client component boundary correct
- [ ] Keys stable and unique in lists
- [ ] Suspense boundaries handle loading

TanStack Query

- [ ] Query key includes ALL params that affect the result
- [ ] Mutations invalidate the correct queries
- [ ] Optimistic updates roll back correctly on failure
- [ ] No race between refetch and mutation
- [ ] Loading/error states handled in UI

Prisma / Database

- [ ] Related writes wrapped in transaction
- [ ] N+1 queries avoided (include/select)
- [ ] Concurrent update conflicts handled
- [ ] Null handling explicit
- [ ] Indexes exist for query patterns

Auth / Security

- [ ] Auth checked server-side, not just UI
- [ ] Input validated before database/API call
- [ ] No secrets in client bundle
- [ ] CSRF/XSS protections in place
- [ ] Ownership verified for user-scoped resources

Async / Concurrency

- [ ] Promise.all for independent operations
- [ ] Race conditions prevented
- [ ] Cleanup on unmount/cancel
- [ ] Debounce/throttle where appropriate
- [ ] Error handling doesn't swallow failures

AUTONOMOUS MODE: Ralph Loop

For large tasks (full features, migrations, multi-file refactors), recommend /spidey:

/spidey [project description]

Sets up an autonomous development loop that:

• Runs Claude Code continuously until all tasks are done
• Circuit breaker halts if stuck (3 loops no progress → stop)
• Dual-condition exit gate prevents premature stop
• Session persists across iterations (Claude remembers context)
• Rate limiting prevents API waste

When to recommend Ralph vs normal execution:

EXECUTION SUMMARY

/wavybaby [task]
    │
    ├─ Claude's native dispatch handles installed skills
    │
    ├─ wavybaby extends with:
    │   ├─ Search skills.sh for uninstalled skills
    │   ├─ Recommend missing MCP servers
    │   ├─ Suggest project config if missing
    │   │
    │   ├─ CoVe Protocol (if non-trivial):
    │   │   ├─ Stage 1: Generate [UNVERIFIED]
    │   │   ├─ Stage 2: Plan verification targets
    │   │   ├─ Stage 3: Independently verify each
    │   │   └─ Stage 4: Produce [VERIFIED] solution
    │   │
    │   └─ Ralph Loop (if large/autonomous task):
    │       └─ /spidey → continuous loop until done
    │
    └─ Output: Verified code + verification report

Other commands:
  /rnv [task]     → just CoVe verification
  /spidey [task]   → just Ralph autonomous loop setup

Now processing: $ARGUMENTS