The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the evaluation of other agents and models, meaning it must ingest and process their outputs, including code, test results, and artifacts. This creates a vulnerability surface where malicious instructions embedded within the data being evaluated could attempt to influence the agent's judging process or subsequent actions. The skill provides remediation guidance by stating that artifacts returned by an actor should be treated as candidates requiring inspection, but it lacks automated sanitization or boundary markers for these inputs.
Ingestion points: Artifacts, code, and session traces generated by agents/models under evaluation.
Boundary markers: Relies on manual inspection and session isolation provided by the host.
Capability inventory: File writes to .qp/adanwo/, command execution for profilers/probes, and interaction with other agent skills like irinse and alaga.
Sanitization: Not explicitly defined for prompt interpolation.
[DYNAMIC_EXECUTION]: The 'Exploration' branch of the skill involves building 'truthfully exercisable representations' and 'runnable interactions' to settle uncertainties. This involves the runtime generation and execution of code prototypes. Although the skill mandates that these be 'throwaway' and stored in a specific vendor-prefixed directory (.qp/adanwo/exploration/), generating executable artifacts based on potentially unvetted model logic carries inherent risks.
[COMMAND_EXECUTION]: The skill utilizes system-level capabilities to measure performance, including the use of native project tools, runtime probes, and profilers. These operations involve executing shell commands and interacting with the local environment to collect evidence for experimental claims.

adanwo