The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data originating from on-chain smart contracts and third-party DeFi APIs.
Ingestion points: Blockchain data fetched via RPC in rpc.rs and CLI output from the onchainos tool in onchainos.rs.
Boundary markers: The skill contains an explicit 'Data Trust Boundary' section in SKILL.md that directs the agent to treat all returned data as untrusted and render only human-relevant fields.
Capability inventory: The skill can execute wallet transactions and contract calls via the onchainos CLI and make network requests to public RPC nodes.
Sanitization: Instructions mandate field filtering and forbid passing raw RPC responses or full calldata objects into the agent's context.
[EXTERNAL_DOWNLOADS]: Setup instructions in SKILL.md download configuration files, installation scripts, and plugin binaries from the official OKX GitHub organization.
Evidence: The 'Pre-flight Dependencies' block fetches assets from github.com/okx/onchainos-skills and github.com/okx/plugin-store.
Mitigation: The skill implements mandatory SHA256 checksum verification for all downloaded installers and binaries before they are permitted to execute.
[COMMAND_EXECUTION]: The skill facilitates blockchain operations by spawning subprocesses for the onchainos CLI.
Evidence: Intent routing in SKILL.md and logic in onchainos.rs involve executing shell commands like wallet contract-call and defi position-detail.
[SAFE]: The skill defines and enforces a mandatory 'Live Trading Confirmation Protocol' to safeguard user assets.
Safety Gates: This protocol requires paper/preview mode by default, typed user confirmation for live mode activation, mandatory previews before every write, and enforcement of risk-control limits defined in the configuration.

aave-v3-plugin