The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill implements a proactive security control in persist_goal_contract.py that scans for and rejects the persistence of hardcoded secrets, such as private keys and API tokens, during the file-writing process.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute local Python scripts and a narrow set of read-only Git commands (status, branch, rev-parse). These operations are restricted by the allowed-tools configuration and are used exclusively for read-only project reconnaissance.
[DYNAMIC_EXECUTION]: The persist_goal_contract.py script invokes subprocess.run to query Git metadata. These calls use hardcoded arguments and do not incorporate unsanitized user input into the shell command.
[PRIVILEGE_ESCALATION]: The skill performs authorized file writes to the project root to save handoff contracts. It mitigates escalation risks by strictly validating that the target path is a direct child of the project root and by refusing to follow symbolic links or reparse points.
[INDIRECT_PROMPT_INJECTION]: As a tool that transforms user tasks into agent instructions, the skill handles untrusted input. It mitigates this risk by enforcing a governed state machine (DRAFT → APPROVED) that prevents the automatic execution of generated payloads and uses fenced blocks to maintain clear instruction boundaries.

goal-meta-skill