The Agent Skills Directory

[COMMAND_EXECUTION]: Orchestrates a complex workflow through extensive use of the evo CLI. It manages background processes and parallel sub-agent lifecycles across various host environments using tools like evo dispatch and native host primitives (e.g., Bash(run_in_background=true) on Claude Code).
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing external data from benchmark outcomes and execution traces.
Ingestion points: Reads outcome.json and traces/task_*.json in steps 1, 2, 3, and 6 to identify failure patterns.
Boundary markers: Employs dedicated "scan sub-agents" explicitly restricted to read-only operations and forbidden from running experiments or editing code to ensure analysis remains isolated from execution.
Capability inventory: Employs evo new and evo run which facilitate code modification and execution, along with background bash processes for orchestration.
Sanitization: Relies on orchestrator verification steps and structured brief generation to ensure findings are supported by verbatim evidence rather than speculative or injected content.
[DYNAMIC_EXECUTION]: The skill is designed to perform runtime code generation and execution as part of its primary optimization purpose. It directs sub-agents to form edits, modify source code within experiment worktrees, and evaluate those changes using the evo run command within an isolated environment.

optimize