The Agent Skills Directory

Persistence Mechanisms (MEDIUM): The deployment guide (deployment.md) provides a systemd service unit definition for a 'Keeper Bot'. This establishes long-term persistence on the host machine. Although consistent with the skill's primary purpose of maintaining protocol health, automated execution should be reviewed.
Unverifiable Dependencies & Remote Code Execution (MEDIUM): The skill directs users to download and build code from several GitHub repositories (github.com/aeyakovenko/percolator*) that are not on the predefined trusted list. These repositories contain the core logic and scripts executed via cargo and pnpm.
Data Exposure & Exfiltration (MEDIUM): The configuration and setup guides (SETUP.md, PERCOLATOR_SKILL_README.md) reference access to sensitive local files, specifically the default Solana wallet path (~/.config/solana/id.json). This constitutes exposure of high-value credentials, though it is typical for the intended use case.
Indirect Prompt Injection (LOW): The skill describes an attack surface where the agent ingests data from external sources such as oracle accounts and market state (oracles.md).
Ingestion points: Oracle account data and market slab state.
Boundary markers: Not explicitly defined in the prompt templates.
Capability inventory: Execution of CLI tools and blockchain transaction submission.
Sanitization: Code-level validation (price bounds/staleness) is described in the documentation.

percolator