ghostwriter

name	ghostwriter
description	Targeted code reducer for minimizing token usage. Essential for analyzing bugs in large files without creating reproduction scripts or debug files. Replaces unrelated code with comments in its output.

Core Rules

1. No Reproduction Files: Never create reproduction scripts, debug files, or unit tests to "find" the bug. Fix the bug directly in the existing source files.
2. Targeted Reading: Use ghostwriter prune to read only the necessary logic.
3. In-Memory Output: The reduced output exists only in your context (via stdout) or the clipboard. It must not be saved to a file.
4. Direct Modification: Apply your code changes directly to the original source file using the agent's normal editing tools.

Execution Pattern

1. Identify Target: Determine the file and line number (from error trace or user prompt).
2. Execute Prune: Run ghostwriter prune <file> --line <number> --stdout --silent.
3. Analyze: Use the stdout output to understand the local logic.
4. Modify Source: Edit the original file directly after analyzing the reduced output.

Advantages

• No Temporary Files: Avoids creating and cleaning up reproduction scripts.
• Reduced Token Usage: Excludes unneeded code blocks from the context window.
• High-Signal Context: Provides only the relevant imports, class structures, and the targeted code section.

Codex Notes

• In Codex, run ghostwriter prune <file> --line <number> --stdout --silent from the terminal.
• Read the reduced output in context, then patch the original source file normally.
• Use clipboard-based auto-detection only when the command has access to the relevant terminal trace context.