The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses wrapper scripts (run_gemini.sh, run_gemini.ps1) to invoke the Gemini CLI tool. It explicitly implements the --approval-mode yolo flag, which bypasses all user confirmation for tool calls (such as writing files) initiated by the Gemini model during execution.
[EXTERNAL_DOWNLOADS]: The skill documentation instructs users to install external dependencies via NPM (@anthropic-ai/gemini-cli or @anthropic-ai/gemini). While these are associated with a known organization, they are external packages required for the skill's core functionality.
[DATA_EXFILTRATION]: Repository contents and task-specific descriptions are transmitted to Google's Gemini API via the CLI tool. This is intended for delegation but represents the movement of potentially sensitive local data to an external service.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8). It processes files from the local repository and passes them as context to an external model that has file-writing capabilities.
Ingestion points: Any repository files specified in the .ai/gemini_task_<name>.md context file.
Boundary markers: Absent; the skill does not use specific delimiters or instructions to prevent the model from obeying instructions embedded within the ingested data.
Capability inventory: The model has the ability to write and modify files (referenced as the write_file tool in the documentation).
Sanitization: No sanitization or validation of repository content is performed before it is interpolated into the prompt for the external model.

gemini-delegate