The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes unvalidated user-supplied feedback and stores it in defect records. This data is subsequently consumed by implementation skills like implement-task, which performs code modifications based on the content. While the skill uses metadata markers to delimit data sections, it lacks semantic sanitization of the input to prevent malicious instructions from being passed to downstream tools. Ingestion points: SKILL.md describes taking feedback as an argument to the $task-cycle command. Boundary markers: The skill utilizes specific HTML-style comment tags (e.g., DEFECT-METADATA:START) in references/defect-record-template.md to mark managed data blocks. Capability inventory: The skill is designed to trigger implement-task (code generation and file modification) and unity-cli (test execution and build management). Sanitization: The protocol only specifies basic escaping for markdown pipe characters used in tables.
[COMMAND_EXECUTION]: The skill defines a workflow that requires executing various system and development CLI tools. Evidence: It executes git rev-parse to locate the git common directory for managing its custom locking protocol. Evidence: It invokes the unity-cli toolset and unity_test_batch.py for monitoring test outcomes. Evidence: It implements a file-based locking mechanism within the repository's .git/ccgs/ directory, involving direct file system writes and process ID (PID) checks to ensure exclusive ownership.

task-cycle