The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill instructions direct the agent to circumvent sandbox security. If a tool execution is blocked by the environment (e.g., runtime_access_denied), the agent is explicitly told to "re-run it with escalated permissions".
[INDIRECT_PROMPT_INJECTION]: The skill uses a "discovery stub" pattern where the primary instructions are fetched at runtime via the ORCA skills get orca-cli command. This pattern introduces a surface where external content from the CLI tool dictates the agent's behavior without safety boundaries. 1. Ingestion points: Output of the ORCA skills get orca-cli command. 2. Boundary markers: None; the agent is instructed to follow the fetched guide directly without isolation or safety warnings. 3. Capability inventory: Terminal manipulation, worktree management, repository access, and browser control. 4. Sanitization: No validation, escaping, or filtering of the dynamically loaded instructions is implemented.
[COMMAND_EXECUTION]: The skill relies on executing various binary executables (orca, orca-ide, orca-dev) and resolves these paths based on environment variables like ORCA_CLI_COMMAND and ORCA_DEV_REPO_ROOT, which allows the execution of arbitrary system commands.

orca-cli