Fulltrace puts AI to work in three ways. A software factory builds its roadmap day and night, task runs check and fix projects, and a chat in the editor answers quick questions. Plain code checks the work, and every step leaves a record that can be replayed.
On a phone, turn it sideways to watch.
From the most hands-off to the most hands-on. Each one has its own page if you want the detail.
The factory builds Fulltrace from its own roadmap, one small planned piece of work at a time. Each piece is called a slice. Every slice starts with a written plan that names its checks, and ends with a written closeout and a commit. Since October it can also take a slice from another project's roadmap.
Low and medium risk slices run with nobody in the session, one after another, day and night. A slice that needs a person's decision is set aside so the queue keeps moving. The chain runs on a fixed copy of the runner's code, so a slice cannot change the runner it is running under.
Riskier slices change how the system writes, approves or is set up. They run with a person in the session, and so does design work. They need a written design before any code.
A person chooses the order of the queue, and which Claude model the runner uses, with a cheaper one for low-risk slices.
Audits and fixes across projects, and research from chosen sources, all run from Studio in the editor. Before a run starts, Studio estimates its cost from past runs, and every step shows live. Plain code checks the results, and a change to a file waits for that check and for a person's approval of that exact change. A scheduler for overnight runs is built too, and nothing is scheduled on it yet.
A person chooses the job, the projects, and a model for each stage: planning, doing, checking and fixing.
A chat inside the editor that can use any model on OpenRouter, or one running on the same computer at no cost. It has its own API key, and every reply shows which model answered and what it cost, under a spending limit. It reads only what is attached to a message, such as the files that describe the person using it or the open file, and it cannot change anything.
A person chooses the model, and how hard it thinks when the model supports it.
Each way has its own models, its own checks and its own record. The factory and task runs share one local server, which also briefs the coding assistants. Chat stands apart from both, and reads only what is attached.
Each column, top to bottom: how the work starts, the models it uses, how it is checked, and what it leaves behind.
Every new AI session used to start from zero, with no memory of the person, the projects or the way of working. The first ten minutes went on re-explaining yesterday.
Now Claude Code, Codex and Cursor all read the same ten plain-text files about the person using them: who they are, what they are working on and how they like things done. The files come through one small server on the same machine. Change a file once and every assistant picks it up. They also share one set of memory notes and 17 commands, kept in git and copied to each assistant by a sync script.
It all started as a final project for the AI Daily Brief AgentOS program.
Live and in daily use. The factory ships slices most days, and task runs have made real fixes to project code once a person approved the exact change.