The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script (run_opusplan.py) by interpolating user-supplied input directly into a shell command string (--task "TASK_HERE"). If the input is not properly escaped by the platform before substitution, a malicious user could potentially execute arbitrary shell commands on the host system.
[INDIRECT_PROMPT_INJECTION]: The skill implements a workflow where instructions for implementation are read from a generated file and external context sources, which could be influenced by malicious content.
Ingestion points: Untrusted data enters the agent context via the /tmp/opusplan_result.json file (output of the planning script) and any files provided via the --context argument.
Boundary markers: The instructions lack explicit boundary markers or warnings to disregard instructions embedded within the planning output or context files; instead, the agent is directed to "Follow the plan faithfully."
Capability inventory: The skill has access to powerful tools, including Bash, Write, and Edit, which are used to execute the steps defined in the plan.
Sanitization: There is no evidence of sanitization, filtering, or validation of the plan's content before the implementation phase begins.

opusplan