The Agent Skills Directory

[COMMAND_EXECUTION]: The skill operates primarily by executing shell commands via the evo CLI to manage code experiments, benchmarks, and sub-agent workflows. It also uses standard system utilities like nvidia-smi and ps for resource monitoring within the evo-optimize.js workflow driver.
[INDIRECT_PROMPT_INJECTION]: The skill processes output from external benchmark runs (outcome.json and trace files) to inform its optimization strategy. This pattern analysis is a core function but constitutes an ingestion point for data generated by potentially untrusted experiment code.
Ingestion points: Evaluation outcomes and task traces are read by scan sub-agents in Step 3 of the orchestrator loop to identify shared failure causes.
Boundary markers: Prompt instructions for sub-agents mandate the use of verbatim evidence and focus strictly on optimization goals to mitigate the impact of speculative or malicious content in logs.
Capability inventory: The orchestrator possesses capabilities to spawn sub-agents, execute bash commands, and perform file system operations.
Sanitization: The skill relies on structural constraints and specific sub-agent briefs rather than automated sanitization of log data.
[DATA_EXFILTRATION]: The skill includes a telemetry reporting mechanism for providing feedback to the developers.
Evidence: The skill body for the evo telemetry feedback command contains explicit safety instructions: "Do not include repo names, company names, file paths, commands, prompt text, raw logs, URLs, secrets, dataset names, or exact task examples" to protect user privacy.

optimize