The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs users to perform command-line operations such as installing the Claude Code CLI and building the extension from source using npm commands. It also manages terminal instances for agent execution.\n- [DATA_EXFILTRATION]: The extension logic (demonstrated in the watchAgentTranscript function) reads from the sensitive file path .claude/projects/*/transcript.jsonl. This directory contains the internal state and interaction logs of the Claude Code agent, posing a high risk of data exposure as it includes source code modifications and terminal outputs.\n- [PRIVILEGE_ESCALATION]: The skill provides a specific user interface option to launch agents with the --dangerously-skip-permissions flag. This action intentionally disables the human-in-the-loop safety mechanism, allowing the AI agent to execute commands and modify files without user confirmation.\n- [EXTERNAL_DOWNLOADS]: The skill guides users to download and install a third-party VS Code extension (pablodelucca.pixel-agents) and clone source code from an external GitHub repository.\n- [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its processing of agent transcripts.\n
Ingestion points: Reads activity logs from transcript.jsonl in the .claude directory.\n
Boundary markers: None identified in the provided logic.\n
Capability inventory: Can spawn agents with elevated autonomy via the --dangerously-skip-permissions flag.\n
Sanitization: Uses JSON.parse to identify tool names and message types, which then directly trigger UI states and animations without validation of the content's safety.

pixel-agents-vscode