The Roadmap

Fulltrace used to build only itself. This track opened on 1 October, and by 7 October the runner, its board and stats pages, and Studio's Build mode all took a project. One command starts a new project with a starter roadmap, kept in that project's own repo. On 5 October the first slice outside Fulltrace, and so far the only one, ran end to end in a run started by hand. Still open: Studio's Start button runs Fulltrace only, so another project's queue waits for a run started by hand.

A workflow that answers a question from chosen files, and from the web when a box is ticked. Every claim quotes its source, and plain code checks the quote before the claim is kept. For a web source the check is against the text the search returned. Since 9 October a claim whose quote fails is dropped and named, and the rest of the answer stands.

A maintenance track that became the dark factory. Queued slices up to risk class R3 are built and shipped by a session with nobody in it, around the clock. Each slice has a time limit, a slice that stops short gets a second try, and a slice that needs a person's decision is set aside. Under it, every slice that adds a check has to prove the check works: a tool breaks the code on purpose and confirms the new check fails.

Build mode in Studio, a board that shows the chain at work, a stats page and a record of every run. After each deploy a click pass lists what to click to prove the new build, and it can be finished inside Studio. Five open rows would let a model take a click pass by itself, inside a sandbox with no network.

The written process no longer assumes Claude. A file at the top of the repo tells any assistant where the rules live and what only a person decides. Since 28 September Codex can build a slice with a person present, inside a sandbox and behind the same guard as Claude. It cannot run the main test suites in that sandbox yet, so it hands those checks back. The unattended chain still runs on Claude only, and Cursor cannot be handed a slice yet.

A Chat tab shipped in Studio on 25 August. Pick a model and send, and every reply says which model answered and what it cost, under a limit that stops the tab. Since September it can set how hard a model thinks and take attachments to read. The part that would pick the model automatically is built and has no button yet. That button is the next row.

Shipped 11 to 23 August. A workflow is described in a file. It is checked against a fixed shape, and nothing can run it until a person accepts that exact file. Steps can carry conditions. A workflow that writes to files is refused, and what it could reach is shown before acceptance. Tying the acceptance to that reach as well is the open row. No workflow has been written this way yet.

Each of the nineteen standing rules has a record in code that says how it is enforced, checked against the real code, and a screen under Config > Governance. Eighteen were ruled permanent and one temporary. A project can be exempted from a rule only through a list held in code, and that list is empty.

The outbound rule is standing rule G19, in place since August. Fourteen places data can leave are listed in code and locked by tests, and nine of them are this machine talking to itself. The push tool refuses a remote owned by anyone else, the assistants' shells refuse the common download commands, and the extension keeps a checked list of the outside services it may reach.

Task runs can be put on an overnight schedule, and nothing launches until the limits are set by hand. Jobs are keyed to calendar dates, so a daylight saving change cannot run one twice. The Schedule tab switches jobs on and off and creates and edits them. It is built and tested. No job has been scheduled on it yet.

A spend page for what the runs cost, filterable by model, with two runs side by side. Scheduled runs can be capped in dollars. A soft daily ceiling shows at the moment Run is pressed. It is a warning and does not block a launch.

A saved run can be replayed with no model called and compared with the original. In August, 89 of 101 recorded runs replayed exactly, and the other 12 were set aside as not replayable. Replay tests the machinery around a model and says nothing about the model itself. Comparing models against each other on a fixed set of tasks was added to the roadmap on 8 October, starting with the test harness.

Push notifications and a read-only phone page were built in July, over a private network. With notifications on, the phone hears which slice a run or decision is for, when a closeout step is confirmed, when the chain is nearly out of work, and why it stopped. Since 16 September the phone can read click pass pages. A finished pass shows its counts, and the answers stay on this machine. Approving from the phone is not built, and waits until there is a real need.

A set of models runs on the local machine. Config > Models shows which are loaded and can load one. Their first job was a search index over the project's own docs, searched at $0 a query, and the Chat tab can use one too. They stay out of task runs. On 28 September Hugging Face joined the roadmap as a source of small local models. Nothing is built yet, and downloading from it stays off until its own row ships.

A browser page on this machine that can act, with the same preview, rehearse, approve and apply steps as Studio. Nine actions shipped in July, and each needs a write session that is started at the machine itself. Even a caller on the same machine is treated as hostile.

The foundation for several machines reporting to one place that can be watched shipped in July. Each machine has an identity and a publisher, and the centre can only look. It never approves or applies anything, and it has no way to send a command back. The live check with two machines is still to do.

Opened on 18 September. The next row is the design for what a stranger would need to install Fulltrace on their own machine, and which of today's limits that would open. The licence decision comes after it.