smart-compact

Claude Code Toolkit

A Claude Code marketplace that cuts token costs, preserves context across sessions, and extends Claude with intelligent project exploration, workflow automation, and development standards.

🚀 Quick Start

Add this marketplace to your Claude Code configuration:

# Add the marketplace (replace with actual installation command)
/marketplace add https://github.com/thoeltig/claude-code-toolkit

Then install any plugin:

/plugin install <plugin-name>@claude-code-toolkit

📦 Available Plugins

Plugin	Description	Version
smart-compact	Remove duplicate file reads and script outputs from transcripts to reduce token waste, lower hallucination risk, and preserve context priority when resuming sessions	2.2.3.0
project-intel	Lightweight reconnaissance system that provides semantic direction before code exploration with persistent knowledge across sessions	1.6.0.0
read-efficient	Machine-efficient file reading tool that converts files to compact JSON and removes redundant whitespace, saving 30-70% tokens across 10+ formats	1.1.1.0
fetch-full-content	Download full page content from URLs to markdown for complete information retrieval without summarization (⚠️ trusted sources only)	1.2.1.0
session-protocol	Save your active tasks between sessions and never loss context again	1.2.0.0
changelog	Create, update, and maintain CHANGELOG.md files following Keep a Changelog and Common Changelog standards	1.1.0.0
documentation	Create and maintain high-quality project documentation with quality validation, style guides, and inclusive language standards	1.0.0.0
claude-code-capabilities	Comprehensive management of Claude Code features including skills, commands, hooks, prompts, subagents, and MCPs	1.8.0.0
cross-platform-notification	Send native system notifications for Claude Code hook events across Windows, macOS, and Linux	1.0.0.0

🎯 Plugin Highlights

Smart Compact

During development you naturally read the same file to keep important context "fresh" (higher priority) or execute the same bash script with the same output multiple times. These fill the conversation with outdated content which the LLM needs to process everytime a new message is added. Also when resuming a session Claude Code reconstructs the entire conversation from the transcript including all file reads which waste tokens without adding new information.

This plugin automatically deduplicates as your session ends, removing redundant reads using intelligent forward-chaining:

• Forward-chaining algorithm: Processes reads chronologically per file, efficiently catching identical and cascading duplicates
• Keep-last strategy: Keeps the latest occurrence of each file and output (represents current state), dedups earlier versions
• Partial dedup: When content differs, keeps only changed lines with ±N line context margin (configurable)
• Bash script dedup: Detects and deduplicates identical script execution outputs (python, npm, node, dotnet, ruby, java, go)
• Grep dedup: Safely deduplicates grep search results when later reads exist without intervening edits
• Smart edit overlap detection: Only skips dedup if edits touch the exact lines matched by grep (not conservative "any edit = skip")
• Write awareness: Detects reads immediately following Writes and marks them as redundant
• Cache validator hook: Blocks input if transcript is stale and duplicates exist, showing token savings to encourage resuming
• Duplicate tokens notification: Shows notification when user input awaited, with configurable threshold to reduce notification fatigue
• Self-documenting markers: Context-aware markers that clearly distinguish file reads from script outputs
• Configurable thresholds: Customize via environment variables: cache duration, context window size, min bytes to replace, notification triggers

The cleaned transcript benefits every future resume—lower cost, reduced hallucination risk from context bloat.

Install: /plugin install smart-compact@claude-code-toolkit

Project Intel

Lightweight reconnaissance system that provides semantic direction before expensive code exploration. Query first to get a ranked list of relevant files - saves exploration tokens and reduces context pollution.