The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill workflow involves handling highly sensitive information such as raw account seeds and passphrases. Specifically, it instructs the agent to use tools like mcp__convex-testnet__transact with a raw seed if a signing service is unavailable. Passing such credentials as tool arguments can result in sensitive data being logged in cleartext or stored in the agent's conversation history.
[COMMAND_EXECUTION]: The skill facilitates the execution of arbitrary CVM (Convex Virtual Machine) code provided via the $ARGUMENTS parameter. While this is the intended functionality for blockchain interaction, it constitutes dynamic execution of user-supplied logic.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input from $ARGUMENTS as executable source code without technical sanitization or boundary markers.
Ingestion points: User-supplied input via the $ARGUMENTS variable in SKILL.md is interpolated directly into transaction tools.
Boundary markers: No technical boundary markers (e.g., XML tags or delimiters) are defined to separate the transaction code from other instructions, though the skill mandates manual user confirmation.
Capability inventory: The skill uses mcp__convex-testnet__signingTransact, mcp__convex-testnet__transact, and mcp__convex-testnet__prepare to execute code on the Convex network.
Sanitization: No automatic validation or escaping of the CVM source code is performed before execution.

transact