The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources to define its implementation scope and requirements, creating an attack surface where malicious instructions could influence agent behavior.
Ingestion points: Reads GitHub issue requirements, repository comments (SKILL.md), and performs web research for external evidence (references/approach-selection.md).
Boundary markers: The instructions do not specify the use of delimiters or warnings to ignore instructions embedded within the processed data.
Capability inventory: The skill possesses the ability to execute git commands, edit local files, and run arbitrary project commands (SKILL.md).
Sanitization: No sanitization or validation of the external content is performed before interpolation into the agent's workflow.
[COMMAND_EXECUTION]: The skill is instructed to execute commands derived from the local project environment.
Evidence: SKILL.md contains instructions to 'Run the named reproduction' and use 'declared project commands.'
[DYNAMIC_EXECUTION]: The verification process involves the dynamic execution of local scripts and test commands.
Evidence: SKILL.md directs the agent to run 'focused developer checks' and execute the project's 'completion contracts.'

implement