The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on local CLI commands (icpg) and shell scripts (scripts/icpg-pre-edit.sh, scripts/icpg-stop-record.sh, scripts/tdd-loop-check.sh) to perform its functions. These scripts are integrated via hooks in the agent configuration to execute automatically during edit and stop events.
[PROMPT_INJECTION]: The skill processes untrusted data to generate metadata that influences the agent's context.
Ingestion points: Git history and commit messages are parsed via the icpg bootstrap command to infer intent.
Boundary markers: The skill does not define specific delimiters or instructions to prevent the agent from obeying instructions embedded within inferred commit data.
Capability inventory: The skill possesses capabilities for local script execution and file modifications in the .icpg/ directory.
Sanitization: No evidence of sanitization for inferred intent text was found before it is injected into the agent's prompt context via the PreToolUse hook.
[EXTERNAL_DOWNLOADS]: The skill requires standard Python packages, including chromadb, scikit-learn, and openai, which are fetched from official registries.
[DATA_EXFILTRATION]: Codebase metadata, including commit history and symbol structures, is transmitted to well-known LLM services (OpenAI and Claude) for intent inference during the bootstrap and creation phases.

icpg