SuperQode now supports OpenAI's Decisions API with gpt-6-luna and Google's Decisions API with gemini-decision-1. These join TypeSafe Jev, which we support with jev-1.13.0 and jev-1.14.0. You choose your decision model separately from your coding model, in the Mac app, the terminal UI or the CLI.
What SystemOne does during a coding run
A coding harness connects a model to tools and applies the rules that govern how those tools are used. SystemOne adds a decision model to that loop. Your coding model reads files, explains bugs and proposes changes. When it requests a tool, SystemOne can ask focused questions about the proposed action: whether it is relevant to the task, whether its arguments are plausible, and whether it looks destructive.
SystemOne can group several questions into a single request and get back structured answers, such as yes/no probabilities, choices from a fixed set or scores against a rubric. SuperQode uses those answers directly to form an allow, ask or deny recommendation, which avoids a separate chat turn for each question. A hosted check still takes a network round trip, so the effect on total task time depends on your workflow.
New configurations start in shadow mode. SystemOne records its recommendations while your existing permission policy stays in control of execution. On macOS, Agent activity → SystemOne decisions shows the tool, decision provider, model, HTTP result and recommendation for each recorded check. For setup details, see the SystemOne provider guide.
Demo: Why does a $50 order still pay for shipping?
The demo pairs a local coding model running in Ollama with Google Gemini Decision. The Python project contains a boundary bug: orders of $50 or more should ship free, but shipping.py uses > where it should use >=. The prompt is a single line:
Read ./shipping.py once and tell me why exactly $50 still gets charged.The coding model reads the file and explains the bug. In Agent activity, you can find the SystemOne check for that same read_file request. The lines "Google · gemini-decision-1" and "HTTP 200" show that Gemini answered during the coding turn, and "Would allow" is an example of how a shadow recommendation appears in the app.
Try it
Download SuperQode for macOS, open a project and choose New thread → SystemOne. Pick a decision provider and model, enter its API key, then click Connect and choose coding model. Saving the decision key to Keychain is optional. Test decision… sends a small synthetic request to check API access and answer formats.
Hosted decision providers receive relevant task and tool context, even when your coding model runs locally, and provider API access and usage charges apply. If you prefer the terminal, install SuperQode 2.31.17 and run the following command to see the supported models:
superqode harness decision-providers