The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell commands including git, gh (GitHub CLI), and glab (GitLab CLI) to probe the repository state and create issues or merge requests.
[DATA_EXFILTRATION]: The skill facilitates the transmission of distilled requirement specifications to remote git forges (GitHub/GitLab). This risk is mitigated by a 'Privacy Discipline' section in references/publish.md which explicitly forbids the inclusion of sensitive internal data like file paths, code snippets, or internal codenames in the output.
[DATA_EXPOSURE]: During 'Phase 1: Facts self-check', the skill reads the local codebase, documentation, and configuration files. While this is necessary for its purpose, it provides the agent with access to sensitive internal project information.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted content from the local codebase to build its internal context.
Ingestion points: Phase 1 retrieves content from the repository's code, docs, and configurations (SKILL.md).
Boundary markers: No specific delimiters are defined to isolate untrusted file content from the agent's instructions during the analysis phase.
Capability inventory: The skill possesses the ability to read the file system and perform network requests via git CLI tools.
Sanitization: A 'Whitelist Template' is enforced in references/publish.md to ensure only specific sections (Problem, Solution, Acceptance Criteria, etc.) are included in the final output, and a manual user review is mandatory before any data is sent externally.

clarify