The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from $ARGUMENTS and external handoff files (e.g., probe-to-probe-llm.md) which can influence the generated probe plan.
Ingestion points: Instructions are loaded from the $ARGUMENTS variable and any markdown files referenced as "handoff context" in Phase 1.
Boundary markers: A mandatory human-in-the-loop checkpoint is implemented in Phase 1.5 ("Entry gate"), which requires user confirmation of the probe plan before execution.
Capability inventory: The agent possesses a broad toolset including Bash, Write, WebFetch, and Task, which are used to execute the probe's logic.
Sanitization: The skill relies on manual user validation of the presented plan at the entry gate; it does not perform automated validation or sanitization of the input data or handoff context.
[DYNAMIC_EXECUTION]: The skill's primary workflow involves the runtime assembly of "probe steps" in Phase 1, which are subsequently executed in Phase 2. This dynamic generation of executable logic is a core feature but presents an inherent risk if the planning process is compromised.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute the experiment steps defined during the qualification phase. This enables the agent to run arbitrary shell commands as part of its intended "probing" functionality.

probe