The Agent Skills Directory

[SAFE]: The skill operates entirely within the local environment. All scripts (goal_router.py, goal_state.py, loop_compiler.py) utilize the Python standard library only, with no external dependencies or network capabilities.
[DATA_EXPOSURE]: Analysis of the source code confirms that no sensitive files (e.g., SSH keys, AWS credentials, environment variables) are accessed. File operations are restricted to the project-specific directory (defaulting to ./my-agent) for state management of goal.json and execution plans.
[REMOTE_CODE_EXECUTION]: There are no patterns of remote code execution, dynamic code generation, or unsafe deserialization. The logic is based on keyword scoring and structured data manipulation.
[CREDENTIALS_UNSAFE]: The skill explicitly avoids handling API keys. It follows a BYOK (Bring Your Own Key) pattern, instructing the user to provide their own key when running generated commands, and does not store or print credentials.
[INDIRECT_PROMPT_INJECTION]: While the skill ingests untrusted user data via the goal string in goal.json, the vulnerability surface is minimal. The data is used for keyword matching and state tracking; it is not executed as code, passed to shell commands, or used in a way that could influence agent behavior beyond selecting the appropriate workflow lane.

agent-launcher-orchestrator