AI

Every AI feature in the app goes through one thin layer. That layer builds a prompt, sends it, and hands back the answer. So you can swap in any OpenAI-compatible service and everything keeps working.

Supported providers

A local model for fully offline use

If nothing should leave your machine, this is the way.

Ollama

ollama pull llama3.1
ollama pull qwen2.5

Then in the app:

1. Set the provider to Ollama
2. Use the model name exactly as you pulled it, for example llama3.1
3. Leave the API key empty
4. Click Test connection

LM Studio

1. Download a model in LM Studio and load it
2. Start the local server
3. Set the provider to LM Studio and click Test connection

A local model still knows plenty of general things, but being scoped to GitHub usually makes it weaker than the large hosted ones. For structured work like writing a README it is entirely adequate.

Suggested settings

For structured output such as JSON, drop the temperature to 0.2–0.4 so results stay consistent.

The limits on the assistant

This is a feature, not a restriction the app could not overcome. The assistant is deliberately limited to GitHub.

What it will do

• Write and improve READMEs, profile READMEs and documentation
• Repository settings, topics, visibility, releases and tags
• Issues and pull requests: drafting, triage, review replies
• Commits, branch names, merge and rebase questions
• GitHub Actions workflows, .gitignore, licences, security and tokens
• Read code and suggest a project structure

What it will not do

• Anything outside GitHub: recipes, medical advice, legal or financial questions, general tutoring
• Writing code for targets unrelated to a GitHub repository
• Working around GitHub's own limits

How the limit is enforced

Three independent layers, one after another:

1. Classified before sending. The request is checked. If it is off-topic it never reaches the model, and you get an out-of-scope reply.
2. System prompt. An explicit allow and deny list travels with every request.
3. Output validated. The answer is checked too. If it strayed, it is replaced with the refusal message.

The first layer also blunts prompt injection: if a README or an issue contains "ignore all previous instructions", that text arrives as content rather than as a command, and the repository context is always placed after the scope reminder.

These layers are not an absolute guarantee; no text filter is. What they do is make the failure unlikely and the consequences small.

What is sent to the model

Only when you ask for it:

• The repository name and description
• The file tree
• The languages used and the topics
• The current README
• A few key files, capped in size
• The issue or comment you are working on

Never sent: your GitHub token, your API key, local file paths, or anything in the data folder.

The app/core/ package deliberately imports nothing from Qt, which is what makes the logic testable on its own — and what lets the integration suite drive the real client over real HTTP.

When it does not work