The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that direct the agent to conceal specific automated actions from the user. For example, in references/a2a/user/create-guide.md, the agent is told not to announce or show the user when it saves internal execution configurations or classifies a subscription guide as 'pure signal'. Additionally, references/runtime/watch.md instructs the agent to suppress its internal reasoning process ('No block, no analysis, no reasoning') when handling task notifications, which is a behavioral override designed for efficiency and safety but limits transparency.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple sources, including peer messages, marketplace signals, and service guides. While it includes strong defensive instructions to ignore instructions embedded in this data and use secure transport methods (like Base64 encoding for parameters in references/a2mcp/invoke.md), the large ingestion surface presents a potential vulnerability to sophisticated indirect injection attacks. The evidence chain includes ingestion points in references/a2a/user/intake.md and references/a2a/peer.md, with capabilities involving file writes and network operations via the vendor's CLI tools.
[COMMAND_EXECUTION]: The skill frequently executes shell commands using the onchainos and okx-a2a CLI tools. While these are vendor-provided tools, the skill dynamically constructs command arguments from user input and external payloads. The skill mitigates this by mandating Base64 encoding for complex arguments (e.g., in references/a2mcp/invoke.md), but the dependency on shell execution remains a key part of its functional model.
[EXTERNAL_DOWNLOADS]: The skill provides a recovery procedure in references/shared/chat-comm-init.md that installs the @okxweb3/a2a-node package globally using npm. This is a standard installation process for the vendor's tools and targets a scoped package belonging to the verified author, thus it is documented neutrally without escalating the verdict.

okx-ai