Which AI model should write and review your tests?
CycleMantle does not ship with one fixed brain. You choose a cloud AI provider, a model running on your PC, or the built-in model, and you can change that choice per chat. The choice decides where your prompts go and how good the proposed tests are.
Three ways to connect a model
A cloud AI provider
You enter the address of a standard API endpoint and your key in the settings, and CycleMantle sends the chat to that provider. Large cloud models usually cope better with long classes and tricky logic. The key stays on your PC. With a cloud model, your prompt and the file text the agent reads go to that provider; with a local model nothing leaves your PC. Any usage charges come from the provider, so read its terms. Keep passwords and private keys out of the files you ask the agent to read.
A model on your PC
You point CycleMantle at an endpoint that runs on your own machine. The prompts then go to that local address, and with a local model nothing leaves your PC. This suits code you may not send to a third party. Smaller local models can be slower and weaker at long files, so choose tasks that fit: one class, one fixture, one failing test. You start and update the local model yourself. The chat cards, the diffs and the approval modes look exactly the same as with a cloud model.
The built-in model
The built-in model is ready to use without your own endpoint, so it is the quickest way to try the app after installing it. Use it for a first run: ask for tests of one small function, read the diff, and see how the cards and the approval modes behave. If a task needs more depth, switch the chat to a cloud or local model. Your project folder, the chat history and the approval mode stay as they are, because they belong to the app and not to the model.
Check the connection with the Test button
After you enter an endpoint, press the Test button before you give the agent real work. It sends a small request, then shows the response time and the context window of the model. The response time tells you how long each step of a long run will wait. The context window tells you how much code and output fit into one conversation. A small window is fine for one function but struggles with a whole folder of tests. If the test fails, fix the address, the key or the model name and press the button again.
Match the model to the task
Use a bigger model for refactors and complex logic, and a smaller one for small edits such as renaming a fixture or adding one assertion. You do not have to switch the whole chat. The run_on_model tool sends one hard step to a bigger model, while routine steps stay on the model you chose. For example, let a small model run your test command and read the exit code, and hand the design of a tricky test class to the larger one. Read the result in the card either way.
Quality varies between models, so treat every proposal as a draft. Use Ask first so that each change waits for your OK, read the diff line by line, and run your test command again after you accept it. If a test looks too clever, reject it and ask for a plainer one. Keep your own version control history so you can step back. A passing run is useful evidence, but it is not a guarantee.