(• ◡ -) Wink
Discipline & Insights for Claude Code
Enforces quality • Tracks context • Suggests improvements
What is Wink?
Wink is a discipline plugin for Claude Code that:
- Enforces verification - blocks stopping until checks pass
- Tracks context - monitors reads, edits, and patterns
- Reports insights - shows session metrics for you to analyze
- Suggests agents - based on hot folders and error patterns
Wink collects data and Claude reasons about it. You approve changes.
Key Features
| Feature | What It Does |
|---|
| Intent Guardian | Silently captures task intent, verifies completion before stopping |
| Smart Verification | Skips irrelevant checks (e.g., no tests for docs-only changes) |
| Stop Discipline | Blocks stopping until verification passes |
| Baseline Awareness | Distinguishes pre-existing failures from regressions |
| Context Hygiene | Monitors wasted reads, loops, and session efficiency |
| Agent Suggestions | Suggests specialized agents based on metrics |
Quick Start
1. Install
Option A: Claude Code Plugin (Recommended)
/plugin marketplace add skishore23/wink
/plugin install wink@wink
Option B: Manual
git clone https://github.com/skishore23/wink.git
cd wink && bun install && bun run build
Add to ~/.claude/settings.local.json:
{
"hooks": "/path/to/wink/hooks/hooks.json"
}
2. Run Setup
/setup
This will:
- Create
.wink/ directory and database
- Auto-detect your project type (Node, Go, Rust, Python)
- Auto-detect verification commands from
package.json scripts
- Show what was configured
3. Test Verification
/verify
This runs your configured checks. If verifiers weren't auto-detected, edit .wink/config.json:
{
"verifiers": {
"typecheck": "npm run typecheck",
"lint": "npm run lint",
"test": "npm test"
}
}
4. Add to .gitignore
.wink/
5. Confirm It's Working
Look for this in your prompt responses:
○ wink · ✓ verified
The /wink Command
Run /wink to see session analysis:
✦ wink · session analysis
node · 45 edits · 120 reads · 32min
hot folders
████████████ core (28)
████████ hooks (18)
context loss (files read 5+ times)
12x storage.ts
8x config.ts
context hygiene
███████████████ efficiency: 73/100
focus: 0.38 (120 read → 45 edited)
• 3 files read 3+ times
• 8 files read but never edited
agents
suggested
+ core-expert
28 edits in core/, 15 re-reads
current thresholds
folder-expert: 20
error-fixer: 3
context-keeper: 5
Based on this data, what improvements would you suggest?
Run /wink --apply to have Claude generate the suggested agents.
How It Works
Architecture
┌─────────────────────────────────────────────────────────────┐
│ │
│ Hooks collect data (reads, edits, errors) │
│ ↓ │
│ /wink shows raw metrics │
│ ↓ │
│ Claude analyzes and suggests improvements │
│ ↓ │
│ User approves, Claude applies │
│ │
└─────────────────────────────────────────────────────────────┘
What Wink Tracks
| Metric | Purpose |
|---|
| Hot Folders | Directories with 20+ edits → agent candidates |
| Read Loops | Files read 3+ times → context-keeper suggestions |
| Error Patterns | Logged by category for analysis |
| Context Hygiene | Wasted reads, dead files, search efficiency |
| Verification | Pass/fail history |
Agent Types
Wink suggests six types of specialized agents:
| Agent | Trigger | Purpose |
|---|
| folder-expert | 20+ edits | Expert on specific folder patterns |
| context-keeper | 5+ reads | Cache file summaries to prevent re-reads |
| error-fixer | 3+ errors | Focus on recurring error patterns |
| lang-specialist | Project detected | Language expert (commands from config) |
| quality-guard | Failing checks | Fix quality hotspots |
| regression-fixer | Regressions | Restore checks that used to pass |
How Agents Are Generated
When you run /wink --apply, Claude automatically generates agents:
