The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of AI agents that ingest untrusted user input through the RunAgentInput structure and process it using tools with high-privilege capabilities such as bash execution and filesystem access. This architecture creates a surface for indirect prompt injection. * Ingestion points: Untrusted data enters the agent context through the messages and state fields defined in adapter-coze.md, adapter-langgraph.md, and server-quickstart.md. * Boundary markers: The provided documentation does not specify the use of delimiters or specific instructions to ignore embedded commands in user data. * Capability inventory: The framework supports powerful operations including shell command execution via create_bash_tool and full filesystem access through create_read_tool and create_write_tool, as described in skill.md and references/tools.md. * Sanitization: There is no evidence of mandatory sanitization or filtering of external content before it is processed by the agent.
[COMMAND_EXECUTION]: The deployment instructions in agent-deployment.md include shell scripts that perform local environment modifications, such as rm -rf ./env and directory creation, to prepare artifacts for cloud deployment.
[EXTERNAL_DOWNLOADS]: The setup process requires downloading and installing several vendor-specific packages from public registries, including cloudbase-agent-server, cloudbase-agent-langgraph, and cloudbase-agent-coze, which is standard for the framework's operation.

cloudbase-agent-python