The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The prediction demo included in the project (apps/dashboard/src/worker/predict.ts) interpolates external market data from Kalshi into LLM prompts. While this is part of the intended functionality, it represents a standard risk surface where malicious external descriptions could attempt to influence the model's output.
Ingestion points: Market titles and subtitles are fetched and interpolated into the prompt in predict.ts.
Boundary markers: The prompt uses text-based labels like MARKET INFORMATION: but lacks strictly unique delimiters.
Capability inventory: The service is capable of performing network requests and submitting on-chain attestations.
Sanitization: The implementation uses Zod to validate structured output from the LLM, though the input text from the external API is used directly.
[DATA_EXPOSURE]: The CLI documentation and example scripts (e.g., examples/shared/_env.ts) reference the default Solana CLI wallet path (~/.config/solana/id.json). This is standard behavior for Solana developer tools and does not constitute an unauthorized exposure, but users should be aware that the tool assumes the presence of a local filesystem wallet for signing transactions.
[EXTERNAL_DOWNLOADS]: The project references and fetches data from well-known and trusted services including Kalshi, Pinata, and official Solana RPC endpoints. All external dependencies and package references are standard for the Solana and AI ecosystems.

sati-sdk