The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests potentially untrusted data from the proxy's operational environment, which could be manipulated to influence agent behavior.
Ingestion points: Data enters the agent context through command outputs from ocx logs, ocx inspect catalog, and ocx usage, which read request history and model metadata from the running proxy.
Boundary markers: The instructions do not specify the use of delimiters (e.g., XML tags or triple backticks) to separate tool output from instructions, nor do they provide 'ignore embedded instructions' warnings for log processing.
Capability inventory: The agent can execute a variety of state-changing shell commands, including account pausing, service restarts (ocx system codex-restart), and storage cleanup.
Sanitization: No explicit sanitization, validation, or filtering of the ingested log content is described, allowing raw external data to be interpolated into the prompt context.
[COMMAND_EXECUTION]: The skill's primary function is to drive the ocx CLI tool. It includes capabilities for state-altering operations such as modifying routing policies, pausing accounts, and performing disk cleanup. The skill incorporates safety boundaries by instructing the agent to perform 'previews' and requiring the explicit --yes flag and human approval before committing destructive changes.

opencodex-guide