The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill explicitly instructs agents to use flags that bypass security sandboxes and human-in-the-loop approvals. In references/cli-workers.md, it advocates for --dangerously-bypass-approvals-and-sandbox when executing Codex. In references/orca-workers.md, it notes that Orca starts Claude Code with --dangerously-skip-permissions, removing the requirement for user consent for sensitive actions.
[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface where untrusted 'briefs' or 'specs' from task steps are passed directly to sub-agents. It acknowledges that security constraints (such as read-only mode) in these environments are not enforced by the platform but are instead reliant on instructions within the injectable brief itself.
[COMMAND_EXECUTION]: The instruction set enables powerful CLI operations across multiple environments, including git force-pushing, gh pull request manipulation, and orca orchestration commands. This includes destructive operations like deleting simulators and resetting orchestration state.
[DYNAMIC_EXECUTION]: The skill's primary function is the dynamic construction and execution of shell commands based on 'capabilities' requested at runtime. This involves interpolating variables into command strings that are executed via agent shell tools.
[REMOTE_CODE_EXECUTION]: Through the spawn-worker and run-remote capabilities, the skill facilitates the execution of arbitrary instruction sets on secondary agents or remote environments, effectively creating a distributed execution network for potentially untrusted task data.

resolve-capability