Repl Scratchpad

What it does

The REPL Scratchpad provides a persistent Python REPL environment for coding agents, allowing variables and data to survive across multiple turns of conversation. This prevents context window overflow by ensuring only the output from print() statements are sent back to the main conversation, while raw data and intermediate results remain within the scratchpad's memory. The system executes code written by the agent in a tmux-backed Python REPL session.

When to use it

• Analyzing large datasets: When an agent needs to process many files (e.g., 142) or API responses, preventing context clutter.
• Complex calculations: For tasks involving multiple steps and intermediate variables that don't need to be visible in the main conversation.
• Iterative development: When refining code logic across several turns without repeatedly re-querying data sources.
• Summarization of file contents: To generate concise summaries from a large number of files, as demonstrated by the example provided.

Key capabilities

• Persistent Python REPL environment
• Variables survive across conversation turns
• Only print() output enters context
• Tmux-backed session management
• Supports Claude Code, OpenAI Codex, Gemini CLI, Cursor, GitHub Copilot, OpenCode, and other Agent Skills compatible tools.

Example prompts

• "Analyze the TypeScript files in this directory and print a summary of file counts per directory."
• "Process these API responses and calculate the average response time; then print the result."
• "Write Python code to glob all .ts files, group them by directory, and print a sorted list of directories with their file counts."

Tips & gotchas

• Requires tmux and python3 (version 3.8 or higher) to be installed. Most Linux/macOS systems and WSL on Windows have these pre-installed.
• The scratchpad prevents raw data from entering the main context, which can make debugging more challenging if you need to inspect intermediate values.