Spend fewer tokens without changing your API.

Appends a system instruction asking for a direct, concise answer without preamble or repetition.

On conversations with more than 24 non-system messages, keeps system messages, the first user message and recent turns, preserving marked cache prefixes and complete tool-call groups.

Drops consecutive messages with an identical role and content, and removes exact-duplicate system messages that appear more than once in the conversation.

Serves eligible identical, non-streaming requests with temperature ≤ 0 from a per-user response cache, avoiding another upstream generation.

Detects fenced ```json blocks and inline JSON objects over 120 characters inside message text, validates them with a parser, then removes only insignificant whitespace from the original bytes.

When a request has no positive max_tokens or max_completion_tokens value, adds a default output cap of 2048 tokens.

Routes eligible short, simple prompts to a cheaper configured model using input-length and request-shape checks.

Collapses redundant whitespace in prompt text while preserving leading indentation.