double-shot-latte

Double Shot Latte

Stops "Would you like me to continue?" interruptions.

Claude Code plugin that uses Claude to judge whether Claude should continue working.

The Problem

Claude Code interrupts multi-step work to ask permission even when next steps are obvious:

• Work is incomplete
• Clear next steps exist
• You requested multi-step implementation
• Claude stated what it will do next

This breaks flow.

The Solution

When Claude tries to stop, a separate Claude instance evaluates the context and decides whether continuation makes sense.

Key principle: If you can type "continue" and Claude knows what to do, the plugin continues automatically.

Installation

/plugin install double-shot-latte@superpowers-marketplace

Prerequisites: Claude Code and jq command-line tool.

How It Works

1. Stop Hook Intercepts - Catches stop attempts
2. Context Analysis - Extracts recent conversation
3. Judge Evaluation - Separate Claude instance evaluates continuation
4. Smart Decision - Blocks inappropriate stops, allows legitimate ones

When It Continues

• Work is incomplete
• Obvious next steps exist (more files, functions, tests)
• Claude mentioned follow-up work
• Implementation has TODOs or placeholders
• Multi-step process with remaining steps

When It Stops

• Claude asks for user decisions
• Claude requests clarification on requirements
• Work is complete and documented
• Claude explicitly needs user input

Use Cases

Multi-File Projects

"Refactor codebase to TypeScript with strict types and comprehensive tests"

API Development

"Create REST API with authentication, CRUD operations, and test coverage"

Component Libraries

"Build React components: Button, Input, Modal, Table with TypeScript and Storybook"

Development Environment

"Set up complete dev environment: Docker, CI/CD, linting, testing, deployment"

Features

• Intelligent decisions using Claude's reasoning
• Aggressive continuation - continues unless clear stop signal
• Smart throttling - max 3 continuations per 5 minutes
• Recursion prevention - judge Claude can't trigger own hooks
• Graceful fallback - allows stopping if evaluation fails
• Zero configuration - works after installation

Configuration

Model Selection

By default, the plugin uses Claude Haiku for fast, cost-effective judgments. You can configure a different model via environment variable:

Configuration (~/.claude/settings.json):

{
  "env": {
    "DOUBLE_SHOT_LATTE_MODEL": "sonnet"
  }
}

Available models:

• haiku (default) - Fast and cost-effective
• sonnet - More capable reasoning
• opus - Highest capability

Note: The judge model runs on every stop attempt. Using more expensive models will increase costs proportionally.

Technical Details

• Default Model: Claude Haiku (fast, cost-effective)
• Context: Last 10 transcript entries
• Throttling: 3 continuations per 5-minute window
• Performance: Minimal latency on stop decisions only

Troubleshooting

Hook not triggering:

• Check installation: /plugin list
• Restart Claude Code
• Verify hook appears in Claude Code hooks UI

Unexpected behavior:

• Review hook reasoning in Claude Code logs
• Check throttle files: /tmp/.claude-continue-throttle-*
• Verify jq is installed

Contributing

1. Fork repository
2. Create feature branch
3. Test changes
4. Submit pull request

License

MIT License - see LICENSE