Prompt, context, and extension assembly

This stage is the prompt-building workshop. Just before the system sends a turn to the model, it gathers everything the model is allowed to see and shapes it into one clear input package. Some parts provide the raw shelves and labels: shared prompt templates, context fragment formats, and extension hooks so built-in code and add-ons can add text in predictable places.

Other parts create the actual note cards that go into the prompt. They turn settings and events into model-readable messages: project instructions, permissions, network rules, available skills, memories, goals, IDE details, plugin and app connectors, review requests, and warnings about space limits. They also keep older saved warning formats understandable.

The turn assembly parts then combine these pieces with the conversation history. They trim old messages when the model’s memory space is limited, remove items the chosen model cannot use, and send only changed context when possible. Realtime startup and developer prompt-debugging use the same machinery to build their own opening briefings.