The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill requires and manages highly sensitive environment variables, including WALLET_PRIVATE_KEY for Solana blockchain transactions, HELIUS_API_KEY for RPC access, and a GITHUB_TOKEN for issue tracking. While standard for this application type, these secrets are at risk if the environment is compromised.
[COMMAND_EXECUTION]: The skill requests the terminal capability and executes the prism CLI via shell scripts (scripts/prism-status.sh, scripts/prism-checkin.sh). It also spawns child processes for agent runtime integration.
[DYNAMIC_EXECUTION]: The skill uses the AGENT_ACP_COMMAND environment variable to spawn a child process (e.g., hermes acp) for communication via JSON-RPC. This allows the execution of arbitrary binaries specified in the local environment.
[REMOTE_CODE_EXECUTION]: The AUTO_UPDATE feature, which defaults to true, enables the skill to automatically check for and potentially install software updates. This represents a supply chain risk where malicious code could be introduced remotely.
[DATA_EXFILTRATION]: The skill provides an MCP server and an optional HTTP API (AGENT_HTTP_PORT) that exposes portfolio status, positions, and configuration data. If incorrectly configured or accessed by local malicious processes, this could lead to the exposure of sensitive trading information.
[INDIRECT_PROMPT_INJECTION]: The skill makes autonomous trading decisions (ENTER, EXIT, REBALANCE) based on external data sources and "second opinions" from agent runtimes.
Ingestion points: Fetches real-time pool metrics (TVL, volume, bin arrays) from the Solana blockchain via Helius RPC; receives decision reviews from Hermes or OpenClaw runtimes.
Boundary markers: None explicitly defined in the provided integration scripts or documentation to separate untrusted data from decision logic.
Capability inventory: Performs on-chain financial transactions, executes shell commands via the CLI, and maintains network connections via WebSockets and HTTP.
Sanitization: While the engine uses deterministic rules, the "Agent Runtime Overlay" allows external agent responses to influence or veto trading actions without explicit sanitization of those responses.

prism