The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill encourages the agent to download and execute "minimal reproducible repos" or code examples provided by GitHub issue reporters. Running arbitrary code from untrusted external sources poses a significant risk of system compromise through malicious build scripts or test configurations.\n- [COMMAND_EXECUTION]: The agent is authorized to use a wide range of local CLI tools such as flutter, patrol, git, adb, and gh. These tools are used to perform complex operations like running tests, viewing logs, and cleaning build artifacts, providing a large attack surface for potential exploitation.\n- [PROMPT_INJECTION]: The skill ingests untrusted data from GitHub issues and comments using gh issue view. This creates a vulnerability to indirect prompt injection where an attacker could embed malicious instructions in an issue. Evidence chain: 1. Ingestion: gh issue view command. 2. Boundaries: Strict output schema and read-only mandate. 3. Capabilities: Local CLI execution of flutter, patrol, and git. 4. Sanitization: None mentioned for processed external content.\n- [EXTERNAL_DOWNLOADS]: The skill references configuration and compatibility data from patrol.leancode.co. These resources belong to the vendor and are used appropriately within the skill's stated purpose to validate version compatibility.

issue-triage