The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The protocol establishes a workflow where agents ingest data from external sources, including a 'TARGET-PIN' specification file, supporting documents, and an 'erratum channel', which may contain untrusted content.
Ingestion points: The protocol explicitly directs agents to read from shared specification files and supporting documentation.
Boundary markers: The skill does not define specific delimiters or security headers for separating external data from agent instructions.
Capability inventory: The agent has the ability to write and execute scripts, perform file system operations (journaling), and manage background processes.
Sanitization: No specific filtering or sanitization of the input text is described, though computational results are validated against expected values.
[DYNAMIC_EXECUTION]: The skill requires agents to generate and execute code at runtime to verify mathematical identities and perform arithmetic checks.
Evidence: Section 1.2 describes provers drafting chains where identities are 'tied to an executed exact-arithmetic script', and Section 4a calls for building 'exact evaluators'.
[COMMAND_EXECUTION]: The protocol instructs the agent to use shell commands for process management during long-running computations.
Evidence: Section 3 ('Infrastructure rules') advises the agent to 'Detach all compute expected to outlive the agent's turn' using tools like 'setsid/nohup'.

prove-protocol