The Agent Skills Directory

[COMMAND_EXECUTION]: The script claude-delegate.mjs sets the environment variable CLAUDE_CODE_SKIP_PERMISSIONS to 1. This configuration instructs the delegated Claude agent to execute all tool calls—including arbitrary shell commands and file system modifications—without requiring manual user confirmation.
[EXTERNAL_DOWNLOADS]: The skill uses pnpm dlx to download and execute the @zed-industries/claude-code-acp package from the npm registry at runtime.
[PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection by accepting arbitrary input text and passing it to an agent session with elevated, unattended tool access.
Ingestion points: Command-line arguments passed to claude-delegate.mjs are joined into the final prompt text.
Boundary markers: None; the prompt text is interpolated directly into the message without delimiters or instructions to ignore embedded commands.
Capability inventory: The delegated agent has access to Claude Code's native capabilities, including full filesystem access and shell command execution, with auto-approval enabled.
Sanitization: No validation or sanitization is performed on the input prompt text.

gemini-use-claude-acp