The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where recipes can ingest data from external sources, creating a path for malicious instructions to enter the agent context.
Ingestion points: SKILL.md specifies that recipes may read health check configurations from a "paste" (e.g., pastebin sites).
Boundary markers: There are no instructions to use delimiters or to disregard instructions found within the data retrieved from external pastes.
Capability inventory: The skill enables the agent to execute shell scripts (live-check.sh), Python scripts (approval.py), database migrations (migrate --apply), and version rollbacks.
Sanitization: The instructions lack any requirement for sanitizing or validating the content fetched from external sources before processing it.
[COMMAND_EXECUTION]: The skill explicitly instructs the agent to run various shell and Python scripts located in other skills' directories.
It triggers scripts/live-check.sh with parameters like --address and --needs.
It utilizes scripts/approval.py from the records skill to log approvals.
It facilitates running infrastructure commands like migrate --apply and host-specific redeploy commands.
[EXTERNAL_DOWNLOADS]: The skill references recipes that fetch health check data from "paste" services. These services are frequently used to host untrusted or ephemeral content that bypasses standard source control.
[CREDENTIALS_UNSAFE]: The skill involves the management of sensitive secrets and database passwords.
While it recommends passing secrets by location to avoid direct value exposure, it accounts for scenarios where secrets are provided directly in the conversation, which can lead to sensitive data being captured in permanent conversation logs.

live-copy