The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for the agent to use curl for session probing. These commands target twin session REST APIs to verify environment state and are designed to utilize a session-specific agent_token held in environment variables to prevent accidental credential leakage.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes untrusted markdown content representing a task's source code and configuration. Malicious instructions within this content could attempt to influence the agent's verification logic.
Ingestion points: User-provided task_source is processed by verification tools and analyzed by the agent (SKILL.md, references/seed-fidelity-checklist.md).
Boundary markers: No specific markers or isolation techniques are established in the instructions to separate untrusted task content from the agent's reasoning process.
Capability inventory: The agent has access to run_task for session creation, curl for network interaction, and session management tools (stop_session) in SKILL.md and references/seed-fidelity-checklist.md.
Sanitization: The instructions emphasize behavioral safeguards, such as "read-only probes" and discarding sessions if state mutation is detected, rather than automated input validation.

pome-verify-seed