Chatting
How GlirOS checks an answer
After an answer with code, GlirOS checks the code without running it (brackets, tags, colons, JSON and JavaScript syntax) and gives the errors it finds back to the model, which writes those files again, at most twice.
Under the answer it says "Code checked · errors fixed", or how many are left, with Fix with AI. When the page in the preview shows errors, Fix with AI in its console sends them too.
Keeping the model on track
If the model makes the same tool call again and again, GlirOS stops it. A long answer is asked to finish after its time limit, and a cloud model's answer after its token limit (Settings > Models; Off is possible). A plan step marked done is checked against what really happened.
On this PC (desktop app), once a model starts a tool call it can only name tools that are switched on and write valid JSON (Settings > Models > Advanced > Keep tool calls well-formed).
Seeing what happened
Under each answer, "How this answer was made" lists its steps, the model, the time and tokens, and what left this device. For a model on this PC or in the browser it also shows what the model read, part by part (who it is and how GlirOS works, the tools, the chat so far, your message, web results, documents) with about how many tokens each took of its working window. A chat's ⋯ menu has "What left this device": everything the chat sent to a cloud company, a search service, a connector or a tool's sites (to whom and what kind, never the words).
After an interruption
If GlirOS was closed in the middle of an answer or a plan, a reminder next time takes you back to that chat; ask it to go on from there.
This guide is the same text GlirOS itself uses to answer questions about GlirOS.