The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill explicitly instructs the agent to re-run commands with escalated permissions if it encounters a runtime_access_denied error, which constitutes a prompt to bypass sandbox-enforced security restrictions.
[COMMAND_EXECUTION]: The agent is directed to resolve the path to the ORCA executable via environment variables (ORCA_CLI_COMMAND, ORCA_DEV_REPO_ROOT) or specific naming conventions (orca-ide, orca) and execute shell commands using that binary.
[INDIRECT_PROMPT_INJECTION]: The skill dynamically loads instructions from the output of the ORCA skills get computer-use command, creating a surface where the agent's instructions can be influenced by the tool's output.
Ingestion points: Output of ORCA skills get computer-use command referenced in SKILL.md.
Boundary markers: No markers or safety delimiters are provided to isolate the dynamic guide from the core instructions.
Capability inventory: GUI manipulation (typing, clicks, menus, dialogs), screenshots, and shell command execution via the ORCA binary.
Sanitization: The instructions do not specify any validation or sanitization for the content returned by the dynamic lookup.

computer-use