LLM request anatomy · public-safe edition

What the model actually receives

A simple question enters the IDE. The agent harness wraps it in operating instructions, capability catalogs, memories, and tool contracts before the model sees it. Sensitive captured values are omitted or masked.

userRequest who are you?
107,527serialized characters
26,504prompt tokens
25tool schemas
71skill descriptors
14agent descriptors
SERIALIZED INPUT · PUBLIC-SAFE MINIMAP TOP → BOTTOM

Area preserves the original serialized character count. Inspector values are public-safe: persistent memory and registries are omitted, and personal, machine, network, and session identifiers are masked.

The harness is part of the prompt

The model does not receive the sentence in isolation. Its apparent identity and behavior are shaped by the surrounding system instructions, injected context, available skills, and callable tools.

0.011%of serialized input

The 12 visible characters in “who are you?” are almost invisible at the scale of the complete request.