Manual · 9

The AI assistant

The assistant is a chat panel that answers about your code, the TIA, the 6502 and the cartridge formats. It is optional and off until you turn it on. With a model that runs on your computer, nothing leaves your machine.

Turn it on at the first-start question or in Settings → Assistant. Until you set up a model, nothing is sent anywhere and nothing is installed.

Choosing a model

VCSIDE ships a helper, llama.cpp's llama-server, that runs a model on this computer. The Built-in model list recommends one that fits your memory. Confirm the download once; the file is checked against a fixed checksum. You can also pick a .gguf file you already have.

• Detect local servers looks for llama.cpp (ports 8080 and 9931), Ollama and LM Studio already running.
• With Add a service... you can use any OpenAI-compatible address.
• Remote services (Gemini, Anthropic, OpenAI, OpenRouter or a compatible server) work on macOS only for now.
• The combo in the chat's header switches between the services you have set up.

The helper is llama.cpp, by Georgi Gerganov and contributors.

Asking

Ask about your code or the 2600. A chip above the box shows what goes along with your question: the open file or the selected lines, and the build errors. Take a chip out with its cross. @ adds files and folders.

• The names VCSIDE knows in your question (registers, instructions, cartridge formats) are looked up in its reference and given to the model, with their ids shown after the answer.
• The parts of the project that match your question go along with a map of its files, so it can answer about code that is not open. View the request shows exactly what was sent.
• Registers and instructions in an answer link to the Documentation panel. A name that is neither in the reference nor in your files is flagged: the model may have made it up.

Search by meaning is optional: turn on Search the project by meaning, then download a small embedding model. Without it the project is searched by keywords.

Ask and Agent

A switch in the panel chooses the mode.

• Ask answers and proposes changes in the chat. You review each one and apply it.
• Agent reads your project, looks things up and describes graphics and music files, and proposes changes itself.

Small local models work in Ask only. If you know a model is good enough, its settings can allow Agent. The agent only sees the project's folder, and files that may hold secrets are never read.

Reviewing changes

A change shows its place and a diff first. Nothing is written until you click Apply or Discard, and a file that changed in the meantime is never overwritten. An applied change is one undo step in the editor. Undo everything this conversation changed restores every file the assistant touched.

In Agent mode, building and running the game ask first: Allow once or Allow for this session. Changes to a Makefile or a .vcsproj are asked for each time.

Explain errors

With Explain errors on, the panel shows Explain the errors whenever the last build failed. Each row of Problems has Explain this error in its right-click menu.

Chats, remote models and keys

• Chats are tabs. They open again with their service and model. They and the project's search index live in VCSIDE's data folder, not in your project. New chat starts another one.
• A VCSIDE.md in the project holds notes the assistant reads first. Every template has one.
• A remote service is marked remote. VCSIDE asks before the first message is sent to it, and a token budget stops it when reached.
• API keys are kept in the system's secret store (the Keychain on macOS). If you prefer, a setting keeps them in a file only you can read, which the system does not guard.