The Agent Skills Directory

[SAFE]: The skill implements strong safety principles, explicitly stating that outward writes (Redmine updates, time logs) and modifications to the core priority file are never automatic and always require explicit user approval.
[COMMAND_EXECUTION]: A local Bash script (scripts/dave.sh) acts as the state layer for the skill. It manages local configuration and logs in the user's home directory. The script employs safe practices, such as using jq for JSON manipulation and a restrictive slugify function for filename generation to prevent directory traversal.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from Redmine tickets and pasted Kanban boards.
Ingestion points: Untrusted data enters the system through scripts/dave.sh intake (handling board text) and Redmine queries (fetching ticket subjects and comments).
Boundary markers: The skill delegates parsing to specialized sub-agents and uses a reconciliation loop where the user must approve a diff before changes are committed to the state.
Capability inventory: Filesystem writes via the local state script and external Redmine API calls (status changes, comments, time logging).
Sanitization: Filenames are sanitized via a restrictive regex-based slugification. External writes are protected by a manual approval gate where the exact payload is shown to the user before transmission.

dave