The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill instructions explicitly direct the agent to "re-run it with escalated permissions" if a sandbox blocks connection (noted as a runtime_access_denied error). This encourages the agent to attempt to bypass security controls and operate with higher privileges than intended.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using the orca CLI to manage orchestration, load guides, and start the runtime (e.g., ORCA skills get orchestration, ORCA open --json).
[DYNAMIC_EXECUTION]: The skill uses conditional logic to determine which executable to run (orca, orca-ide, or orca-dev) based on the environment context, such as ORCA_CLI_COMMAND or ORCA_DEV_REPO_ROOT environment variables. This allows the agent to target different binaries dynamically.
[INDIRECT_PROMPT_INJECTION]: The skill is a "discovery stub" that fetches its actual instructions and orchestration reference at runtime using ORCA skills get orchestration. This creates a vulnerability where the agent ingests untrusted instructions from external command output.
Ingestion points: Command output from ORCA skills get orchestration and ORCA skills get orchestration --reference. (SKILL.md)
Boundary markers: None identified; the skill instructions direct the agent to read and follow the returned guide verbatim.
Capability inventory: The skill possesses the capability to execute shell commands and manage multi-agent task DAGs through the orca binary.
Sanitization: None identified; the agent is instructed to use the guide to handle its next commands directly.

orchestration