The Agent Skills Directory

[SAFE]: The skill implements secure credential handling by advising the agent to store API keys in a dedicated local configuration file (~/.config/treasures/credentials.json) using restricted file permissions (chmod 600). It explicitly instructs against logging or inlining secrets.
[SAFE]: Network activity is restricted to the official vendor domain api.treasures.io. The instructions emphasize that the agent only submits intents and never holds private keys, with the vendor handling on-chain signing and execution.
[SAFE]: The skill includes robust transaction safeguards, including mandatory idempotency keys to prevent duplicate trades, polling to terminal job states with defined timeouts, and the use of the big.js library to ensure mathematical precision for asset amounts.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from the Treasures API which acts as an attack surface for indirect instructions.
Ingestion points: Untrusted data enters the agent context through API responses from endpoints such as /quotes, /balances, and trade status polling documented in SKILL.md and references/endpoints.md.
Boundary markers: The instructions guide the agent to perform strict JSON parsing and validate state machine outcomes (e.g., confirmed, failed, rejected) rather than relying on loose string matching.
Capability inventory: The skill has the capability to perform network operations via fetch and write a local credentials file to the filesystem.
Sanitization: The implementation uses structured JSON handling and dedicated numeric parsing libraries to handle external data safely.

treasures-wallet