The Agent Skills Directory

[COMMAND_EXECUTION]: The skill configures automated shell command execution through Claude settings hooks. The PreToolUse hook is configured to execute scripts/icpg-pre-edit.sh whenever an 'Edit' or 'Write' tool is matched. The Stop hook is configured to execute scripts/tdd-loop-check.sh and scripts/icpg-stop-record.sh upon session completion. This creates a persistent execution surface triggered by routine agent actions.
[COMMAND_EXECUTION]: The skill instructions direct the installation of a local package via pip install ./scripts/icpg. This involves executing the installation logic (e.g., setup.py) of local scripts, which could be modified to perform arbitrary actions.
[PROMPT_INJECTION]: The icpg bootstrap command ingests untrusted data in the form of git commit history and processes it through LLMs (OpenAI or Claude) to infer developer intent. This is a vulnerability surface for indirect prompt injection, where an attacker could craft commit messages designed to manipulate the inferred ReasonNodes or the agent's understanding of the codebase.
[EXTERNAL_DOWNLOADS]: The skill references the use of external LLM services (OpenAI, Claude) and vector databases (ChromaDB) for its core functionality, requiring network access and external API interactions.

icpg