The Agent Skills Directory

[COMMAND_EXECUTION]: The skill leverages the Elixir mix build tool to execute various evaluation commands such as mix skill.eval SUITE train, tune, and verify. These commands are used to perform the core functions of the Dojo runner for testing skill candidates.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to ingest and process "candidate skills" and "evaluation suites," which are external data sources. This establishes an attack surface where a malicious skill package could attempt to influence the agent's actions.
Ingestion points: External skill packages (via --candidate PATH) and evaluation suites (SUITE).
Boundary markers: The instructions mandate that the user must explicitly request the "Dojo" skill to activate it, preventing accidental or inferred activation. It also specifies a bounded-execution profile.
Capability inventory: The skill is authorized for local-write to produce evidence artifacts and uses subprocess execution for mix commands.
Sanitization: The skill contains strict directives against inventing runner commands and prohibits installing new dependencies or repairing the runner during evaluation.
[DYNAMIC_EXECUTION]: The evaluation process involves running logic from suites and candidates at runtime. The instructions mitigate risk by requiring fixed suites and representative tasks rather than arbitrary code execution.

kata-dojo