The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including user trade intents and live account state from the exchange, creating a theoretical attack surface for indirect prompt injection.
Ingestion points: Workflow steps in SKILL.md ingest network identity, account state (balances, positions), and user intent.
Boundary markers: The skill contains multiple explicit safety instructions, such as 'Do not construct a signature', 'Planning only; not for market selection... signing, or submission', and 'Route explicit execution to hyperliquid-operate only after the ticket is complete'.
Capability inventory: The skill executes a bundled local script scripts/risk.mjs via Node.js to perform arithmetic (referenced in references/risk-sizing.md).
Sanitization: The scripts/risk.mjs utility implements strict input validation, checking that arguments are finite numbers, enforcing ranges for risk percentages and size decimals, and using a Map for duplicate-safe flag parsing.
[COMMAND_EXECUTION]: The skill uses a local script for calculations, which is a standard and safe practice for deterministic arithmetic in financial planning.
Evidence: SKILL.md and references/risk-sizing.md reference node scripts/risk.mjs for risk budget and size calculations.
Security check: The script is bundled with the skill, does not perform network operations, and relies solely on command-line arguments passed by the agent based on validated state.

hyperliquid-plan