The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directs the agent to manage workspaces and directories using system commands.
Instructions include the use of git worktree add to establish isolated environments for distinct task candidates.
The workflow involve creating and writing to temporary directories (e.g., /tmp/arena-<slug>/).
[INDIRECT_PROMPT_INJECTION]: The skill architecture, which involves one agent processing the output of several others, presents a surface for indirect prompt injection.
Ingestion points: Data is ingested from local grounding documents (e.g., local://arena-grounding.md) and the textual outputs/rationales produced by parallel runner agents in Phase B.
Boundary markers (absent): The skill does not mandate the use of XML tags, delimiters, or specific 'ignore' instructions when runner outputs are passed to the 'reviewer' or 'picker' agents.
Capability inventory: The skill allows for spawning multiple agent seats, writing to agent-specific URIs (agent://), terminating processes (proc://), and accessing local configuration files such as ~/.omp/agent/config.yml (SKILL.md).
Sanitization (absent): There are no prescribed sanitization or validation steps before 'grafting' logic from losing candidates into the final base artifact.

arena