The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute npx convex run --prod debug:getEvalDebugInfo. This command interacts with the production database to retrieve debug information. While this is the intended functionality for the vendor's debugging tool, it grants the agent access to production-level data based on user-supplied IDs.
[EXTERNAL_DOWNLOADS]: The skill utilizes npx, which may download the convex package if it is not locally available. The convex CLI is a standard tool provided by the vendor (get-convex).
[INDIRECT_PROMPT_INJECTION]: The skill processes data that could be attacker-controlled, creating a vulnerability surface.
Ingestion points: The skill ingests outputFiles (generated by models) and evalSourceFiles (containing TASK.txt and grader code) via the output of a shell command.
Boundary markers: There are no explicit boundary markers or instructions to the agent to ignore embedded instructions within the ingested code or task files.
Capability inventory: The agent has the capability to execute shell commands (npx convex).
Sanitization: No sanitization or filtering of the retrieved file contents is specified before the agent analyzes them.

analyze-eval