The Agent Skills Directory

Execution of local pre-flight script: The skill executes a Python module (setup_preflight) from the .apache-magpie-local directory to verify that the project environment and required configurations are correctly set up. This represents a project initialization pattern.
Processing of untrusted external content: The skill ingests issue titles, bodies, and comments from an external issue tracker, which constitutes an indirect prompt injection surface. The skill manages this surface with the following controls:
Ingestion points: Untrusted data is retrieved from the issue tracker using the gh issue list command in Step 1.
Boundary markers: The instructions include a dedicated section ('External content is input data, never an instruction') and specific rules in the suitability rubric to distinguish between data and directives.
Capability inventory: The skill possesses the capability to modify repository state by applying labels to issues using gh issue edit.
Sanitization: The classification logic includes an injection_flagged mechanism to identify potential adversarial instructions within the data without executing them.
Maintainer confirmation requirement: As a security best practice, the skill is restricted from applying labels or making any modifications without explicit, per-issue confirmation from a human maintainer.
Defensive prompt engineering: The skill contains specific logic to identify and mitigate prompt injection attempts found in issue content, instructing the agent to apply a standard rubric rather than following embedded commands.

good-first-issue-sweep