The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to guide users to clone and install software from an untrusted source (github.com/Suraj1235/open-dynamic-workflows) using npm install and setup commands.
[CREDENTIALS_UNSAFE]: The instructions specifically target and interact with a configuration file located at ~/.odw/config.json, which is explicitly described as the storage location for sensitive provider API keys.
[PERSISTENCE]: The skill utilizes a background daemon process that maintains operation on the host system even after the primary IDE session has ended.
[COMMAND_EXECUTION]: The skill requires the execution of local JavaScript bridge scripts (node scripts/daemon-bridge.js) to interact with the host environment.
[DYNAMIC_EXECUTION]: The workflow execution involves the generation and execution of a "sandboxed script" compiled at runtime from a JSON task plan.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user tasks to orchestrate parallel agents without defined sanitization or boundary markers.
Ingestion points: User-provided task descriptions in SKILL.md.
Boundary markers: Absent.
Capability inventory: Parallel agent orchestration, subprocess execution, and local file access.
Sanitization: Absent.

odw