The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from architectural handoffs and dormant metadata seeds to influence the design and planning phases.
Ingestion points: The references/design.md file reads JSON data from the adr/handoffs/ directory and trigger conditions from .seeds/index.json.
Boundary markers: The skill utilizes an external validator (scripts/handoff.py) to check for absolute paths, traversal, shell metacharacters, and symlink escapes before adoption.
Capability inventory: The skill has access to the Bash tool for command execution, the Agent tool for sub-task dispatch, and can write to the repository using the repository_artifact.py utility.
Sanitization: Path and metadata validation is performed at the boundary, though the validated content is directly interpolated into subsequent prompt contexts to drive the design dialogue.
[DYNAMIC_EXECUTION]: The skill uses shell redirection and heredocs to pass dynamically generated content to local scripts for processing and writing.
Evidence: references/design.md pipes LLM-generated Architectural Decision Record (ADR) content into python3 scripts/repository_artifact.py write using a heredoc (<<'ADR').
Context: This is used to persist design decisions to the repository. The skill relies on local scripts to handle the security of the write operations (atomic replaces, fsyncs, and path containment).

feature-lifecycle