The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to interact with a local CLI utility named orca (or its variants orca-dev and orca-ide). The agent is instructed to resolve the executable path using environment variables like ORCA_CLI_COMMAND and ORCA_DEV_REPO_ROOT, or fallback paths, and then execute management commands against a 'watcher fleet'.
[INDIRECT_PROMPT_INJECTION]: The instructions mandate that the agent must run ORCA skills get orca-heimdall to load a 'version-matched guide' before performing any actions. This creates a dependency where the agent processes and follows instructions provided by external tool output at runtime. This ingestion path lacks boundary markers or sanitization, allowing the tool output to potentially steer the agent's subsequent behavior.
Ingestion points: Output of the shell command ORCA skills get orca-heimdall.
Boundary markers: None; the agent is told to 'read the full guide before acting'.
Capability inventory: Management of watchers, budget adjustments, worker steering, and process management (pause/resume/remove) via the orca CLI.
Sanitization: None specified for the dynamic content loaded from the CLI.
[PRIVILEGE_ESCALATION]: The skill includes an explicit instruction for handling access errors: 'If it fails with runtime_access_denied, your sandbox blocked the connection: re-run it with escalated permissions'. This encourages the agent or user to bypass security boundaries and sandboxing mechanisms, which can lead to unauthorized system-level access.

orca-heimdall