The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external agent CLIs (Claude, Codex, Agy, and Grok) using high-risk flags that explicitly bypass security and permission prompts, such as --dangerously-skip-permissions, --dangerously-bypass-approvals-and-sandbox, and --always-approve. Additionally, the skill intentionally avoids inlining task text in commands to prevent security guard hooks from scanning the input, which is a technique for evading certain types of monitoring.
[DYNAMIC_EXECUTION]: The skill utilizes parallel subagents to run multiple CLI tools simultaneously. It dynamically constructs execution commands based on task shape (Prompt vs. Review) and manages execution state through background processes and scratch files.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data, such as user-provided prompts and repository diffs, which are then passed to other agent CLIs. This creates an attack surface where instructions embedded in the analyzed code or prompt could influence the behavior of the panelist agents.
Ingestion points: Processes user prompt text and the output of git diff for code reviews.
Boundary markers: Prepends an instruction to subagents: "You are one panelist. Answer the task below yourself, directly. Do not run a panel, launch another agent CLI, or start subagents."
Capability inventory: The subagents run with elevated privileges (auto-approval) and can access the repository's file system and execute bash commands via the spawned agents.
Sanitization: Employs safe practices such as writing tasks to files and using quoted shell variables to prevent local shell injection, though it does not filter the content of the ingested data.

panel