The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted project data such as logs, HTML, JSON, and CSS to identify files for cleanup. A malicious actor could attempt to trigger accidental deletion of legitimate files by crafting fake references in these documents.
Ingestion points: Scans .workflow-control/work-items/, execution state JSON files, and project resource directories for path references.
Boundary markers: Operations are explicitly restricted to the current Work Item scope and project root; the skill refuses to perform deletions outside these defined boundaries.
Capability inventory: The skill has the capability to delete files via PowerShell's Remove-Item command.
Sanitization: The skill mitigates risks by requiring recursive dependency checks, SHA-256 verification for deletions, and strict path normalization that rejects parent directory traversal (..) or symbolic link escapes.
[COMMAND_EXECUTION]: The skill instructs the agent to perform file system deletions using PowerShell 7. To prevent filenames with special characters from being interpreted as command flags or wildcards, it mandates the use of the -LiteralPath parameter and requires that all paths be resolved to absolute paths before execution.

phaser4-game-artifact-cleanup