The Agent Skills Directory

[SAFE]: The skill architecture prioritizes security through 'refusal lines' that are explicitly stated as absolute. It defines a clear separation between local development and the remote staging environment, ensuring that code changes only reach the server through a verified CI/CD pipeline.
[COMMAND_EXECUTION]: The skill utilizes legitimate developer tools (git and gh) to perform its primary tasks. These commands are executed within a strictly defined workflow that requires the agent to present commit messages and PR drafts to the user for approval before any remote operations are performed.
[DATA_EXFILTRATION]: While the skill allows for the inspection of staging data (logs, database state), this is restricted to read-only access and is gated by a requirement for the user to explicitly provide SSH access and approval within the current session. No hardcoded credentials or unauthorized data transmission paths were found.
[PROMPT_INJECTION]: The skill uses instructional language to define its own safety properties (e.g., 'Production is never accessed') rather than attempting to override the base model's safety filters. There are no patterns indicative of jailbreaking or instruction bypass techniques.

staging-fix