The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes standard system utilities including mkdir, ls, and cat for local file management, and the GitHub CLI (gh) for repository operations. These operations are scoped to a local staging path and are triggered by explicit user commands.
[DATA_EXFILTRATION]: No unauthorized data transmission detected. Interaction with external services is limited to the official GitHub API and repository hosting via the gh tool. These operations, such as creating a public repository and pushing code, are only performed after clear user confirmation steps.
[PROMPT_INJECTION]: The skill is exposed to indirect prompt injection as it processes external project files. The assessment identified the following evidence chain:
- Ingestion points: Project source files are read from user-defined paths into the agent context via the opensource-forker and opensource-sanitizer agents.
- Boundary markers: Variables are interpolated into sub-agent prompts (e.g., {SOURCE_PATH}, {STAGING_PATH}) without specialized delimiters to segregate data from instructions.
- Capability inventory: The skill maintains capabilities to execute shell commands and perform network operations via the GitHub CLI.
- Sanitization: A dedicated opensource-sanitizer sub-agent is implemented as a security gate, specifically instructed to scan for secrets, PII, and dangerous files before the pipeline proceeds.

opensource-pipeline