The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes output from Scenario API jobs (errors, hints, prompts) to determine recovery steps. This data is untrusted as it comes from external model providers or is a reflection of user input. The agent then uses these details to drive further tool calls like model_run and recommend.
Ingestion points: SKILL.md (Quick reference section) specifies reading error, hint, and metadata.input from job_get and jobs_wait tool calls.
Boundary markers: The instructions do not define delimiters or specific safety boundaries when the agent processes the ingested job error or prompt content.
Capability inventory: The skill uses model_run, recommend, job_get, and jobs_wait tools to perform actions based on the analysis of the potentially untrusted job data.
Sanitization: No specific sanitization, validation, or filtering of the ingested error messages or prompts is described before they are used to influence the agent's next steps.

scenario-moderation