The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill manages Solana keypairs at ~/.config/solana/id.json and ~/.proxygate/keypair.json. These files contain private keys providing full control over cryptocurrency funds and identity.
[EXTERNAL_DOWNLOADS]: The skill performs global installation of the @proxygate/cli package from the npm registry and downloads additional functional components from the ProxyGate gateway.
[REMOTE_CODE_EXECUTION]: The proxygate skills install command downloads and installs new code into the agent's skills directory. It also registers a SessionStart hook, establishing a persistence mechanism that executes code automatically at the beginning of each agent session.
[COMMAND_EXECUTION]: Multiple shell commands are executed to check system versions, generate cryptographic keys using solana-keygen, and configure CLI settings.
[DATA_EXFILTRATION]: The CLI communicates with an external gateway at https://gateway.proxygate.ai. This is functional for the service but involves transmitting transaction-related metadata and requires the agent to handle sensitive wallet configuration.
[PROMPT_INJECTION]: The skill processes untrusted data from API listings and documentation (ingestion points: proxygate apis and proxygate listings docs in references/commands.md). No boundary markers or sanitization processes are defined. Given the skill's capabilities (capability inventory: proxygate CLI execution, solana-keygen, and file system access), this creates an attack surface for indirect prompt injection.

pg-setup