Development

The Work group (Advanced: Development).

Trains

(Advanced name: Streams · trains.) A stream is one unit of work with its own worktree and branch; a train is a goal — a set of streams, its wagons, that go one after another under one order.

The screen lists the trains with their wagons in order. Each wagon carries a state chip — merged, closed, needs you, in review, working, stuck, not started — and hovering a chip says what it means. Click a chip to show only the wagons in that state; click a wagon to read its story: its dates, its review verdict, its quote, cap and cost.

Find searches every train, with a date range, the order's type (TRAIN ORDER, MASTER ORDER) and the lane (Active, Backlog, Done, Archived, Any). View order opens the train's order document. A train can be renamed; its slug cannot change.

Editing a wagon. A wagon's title, quote, cap, review mode, wall clock and model can be changed only while it has not been dispatched; once it is in flight, in review, at your gate or merged, the fields are locked and say why. Save opens your passkey prompt: the server plans the edit, your passkey signs exactly that edit, and only then is it written.

Runs

(Advanced name: Runners.) Every run on the record, newest first, with how many are running now and how many stopped without finishing. Pick Running, Stopped, Finished, Refused or Never closed; filter by date, machine, seat, vendor, model and work; sort by newest, oldest, longest or costliest. The engine's own housekeeping runs are counted apart, under System runs.

Source map

(Advanced name: Source · map.) The branches and commits of the workspace's git repository against a base: which areas — directories — have changed paths on branches ahead of the base, which run is behind each branch, and the commits. Search takes a path, a commit or a branch; a filter narrows to one agent. A changed path opens its diff.

What it cannot read, it says: an unreadable repository, a missing base, a shallow history that cannot prove ancestry, or a fleet record that did not answer each draw a banner, and no figure on the screen is shown as current when it is not.

Pull requests

The pull requests of the forge this workspace is connected to — GitHub or Gitea — and, for each, its merge steps.

Connecting a forge is done here: the forge, its API address and the repository, and the access tokens of two different bot accounts — one that opens and merges pull requests, one that approves them, since the account that opens a pull request cannot approve it. The tokens go to the secret store, Save with passkey records the change, and the gateway reads the forge after it is restarted.

Merging a pull request is signed with your passkey.

Event Log

The committed events of the record, newest first — gate decisions, orders registered, runs started and closed, returns, refusals by a guardrail or the egress door — with a detail pane for the event you select. Tiles count what is on the page: events, gate decisions, blocked actions, agents, machines, runs. A date range narrows it; Live view adds events as they are committed; Export writes out the range.

Bugs

(Advanced name: Bug Harvester.) Defects the fleet filed about itself, by class, with the count of each class's instances. A defect enters with a measurement; nothing here is a human's opinion.

Reports

Decision-ready views over the record — every number is read, never sampled. A picker chooses the plate:

Dashboard · Delivery · Gates · Cost · Trains · Flow · Capacity · Tests · Defects · Laws · Compliance audit · Metrics.

Each plate leads with its key figure and what it counts, then its trend and its evidence, then the one screen to go to next (Open Costs →, Open Bugs → …). The range is Today, This week, Month or All time; each plate says when it was read, and Export writes it out. The console remembers the last plate you picked.

⚠ A part of a plate this build does not answer says "Not answered on this build"; it is not a zero.

Prompts

(Advanced name: Prompt Library.) The prompts the roles use, as documents. Built-in prompts ship with the product; yours, workspace and customer prompts are written or copied from them. The library says where it was read from.

Filter by role, language, scope, category and domain, or search titles and purposes. A prompt shows its template, the inputs a caller fills — words in {{ }} — and a filled preview that updates as you type. Copy puts the filled text on the clipboard; Use in a conversation opens a role's conversation with the text in the composer, unsent — no run starts. Picking a prompt never activates a skill.

New prompt starts blank or from a copy of a built-in: a name, what it is for, who can use it, and the body. Save to my prompts signs it on save; Save and try with the Advisor also opens it in the Advisor's composer. A built-in is never edited in place — Edit a copy.