The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the gemini-cli binary using asyncio.create_subprocess_exec to process research tasks. The path to the executable is dynamically detected from common locations or configured in the skill settings.
[INDIRECT_PROMPT_INJECTION]: The skill ingests research queries from external sources (such as other agents or MCP clients) and processes them through an AI model via the CLI. This creates a surface for indirect prompt injection where malicious instructions embedded in research data could influence the agent's behavior.
Ingestion points: Research queries are ingested through the execute_research_task and stream_research_response methods in gemini_cli_skill.py.
Boundary markers: None identified; user input is passed directly to the CLI tool.
Capability inventory: The skill has the capability to execute local CLI commands and read/write stream data.
Sanitization: No explicit sanitization or filtering of the task description is performed before it is passed to the command-line interface.

gemini-cli-integration