Prompt Construction

Prompt generation is deterministic and scoped. Juno does not query external sources or chain data during prompt time everything is pulled from in-memory session state.

Inputs to the prompt include:

• Active wallet address

• Filtered signal set

• Current filter state

• Token-level risk flags

• Compressed user interaction history (if available)

Signals below confidence threshold are excluded. Token metadata is reduced to flags only (mintable, paused, unverified). No token pricing, no tx logs, no unnecessary fields.

The system constructs a system message payload formatted for LLM use. Assistant responses are driven purely by this structure. There’s no fallback prompt, no retry, no rehydration logic.

If memory is empty, the assistant returns with no context. If the user changes wallets, a full reset occurs. If nothing has changed since the last prompt, Juno reuses the last compiled message.

Prompt construction is fast, reproducible, and minimal by design.

Just current state → assistant input.