The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes JSON data representing the state of the workspace (via the parallax status command) to suggest the next command for the agent to execute. The helper script scripts/parallax_next.py interpolates identifiers like proposal references directly into command strings without sanitization. If a workspace is configured with adversarial content that influences the tool's output, it could lead to the generation of malicious command suggestions. The ingestion point is the parallax status --json output parsed by the Python script, which currently lacks explicit boundary markers or sanitization for interpolated values.
[COMMAND_EXECUTION]: The integration and unit tests for the skill (tests/test_integration_live_cli.py and tests/test_parallax_next.py) utilize the subprocess module to execute external binaries and the Python interpreter. Specifically, the px function in the integration tests drives the parallax CLI, and the _run function in the unit tests executes the script under test. While these are appropriate for testing environments, they represent active command execution capabilities within the skill's file structure.

parallax