The Agent Skills Directory

Command Execution: The skill workflow utilizes a shell command to execute a local script (find_latest_release_tag.sh) for identifying release versions. This execution allows the agent to interact with the environment to manage compatibility baselines.
Indirect Prompt Injection: The skill is designed to ingest and process external 'review feedback' and 'design changes' to guide the implementation process. (1) Ingestion points: External feedback and scope change requests are processed in the Workflow and Review-feedback gate sections of SKILL.md. (2) Boundary markers: The implementation utilizes a 'Review checkpoint' structure to scope changes, though it lacks explicit delimiters for raw feedback content. (3) Capability inventory: The skill has the capability to execute shell scripts, run git commands, and propose code changes. (4) Sanitization: The skill includes robust guidelines in the SDK-specific decision rules section, specifically instructing the agent to prevent untrusted state (such as RunState or serialized session data) from overriding security-critical settings like credentials, secretRefs, and environment.

implementation-strategy