The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a specialized command-line interface (jev ladder, jev supervise, jev lane step, jev ask) to manage task delegation and monitor execution states. These tools are central to the skill's documented purpose of managing 'frontier' model workloads.
[INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface by ingesting and monitoring output from external models. It addresses this risk by providing a mandatory evidence chain for the agent:
Ingestion points: External data is read via jev supervise --tail-file and citations from delegated reports.
Boundary markers: The instructions explicitly mention an injection_seen flag and advise that instructions like 'ignore previous instructions' found in data must be treated as data only.
Capability inventory: The agent uses the jev toolset and grep to verify claims.
Sanitization: The skill mandates independent verification of all 'done' claims through tests or manual state checks rather than trusting model output.
[PROMPT_INJECTION]: A static analysis flag was triggered by the phrase 'ignore previous instructions'. However, context confirms this is a defensive security instruction teaching the agent to recognize and disregard injection attempts in third-party data, rather than an attempt to compromise the agent itself.

jev-frontier-work