The Agent Skills Directory

[PROMPT_INJECTION]: The skill acknowledges an attack surface for indirect prompt injection via untrusted data sources.
Ingestion points: Pull request bodies, commit messages, and file contents as noted in the 'External Side Effects' section.
Boundary markers: No specific delimiters are defined for these inputs.
Capability inventory: The skill delegates tasks to several external skills including 'deploy' and 'ec2-backend-deployer' which likely perform file and network operations.
Sanitization: The skill contains explicit instructions to treat all observed content as data and specifically warns never to obey instructions embedded within that data.
[COMMAND_EXECUTION]: This skill functions strictly as a router and does not execute any shell commands or subprocesses directly. All execution logic is delegated to other skills.
[SAFE]: The skill follows security best practices by using the disable-model-invocation: true flag, ensuring it only runs when explicitly called by the user. It also maintains a clear separation of concerns, preventing the injection of deployment logic into the routing layer.

deploy-dispatch