What a run looks like

The feed, the controls, a usage limit crossed, and what you read in the morning. All of it from what the panel and the CLI show.

The feed #

The panel opens on the Feed. The thread is the app, and the other destinations are lenses on it. The rail has six: Feed, CLIs, Report, Memory, Needs you and Config.

Participants take turns on one shared thread, each on its own provider, model and effort, and The shared thread describes a turn. A message you type is a post: the participant decides what it means. The composer says when a message will not be read at once: Drained at the next turn boundary — never mid-turn.

Controls #

There is no start button. A message starts the run, and it keeps going across usage limits and restarts. The controls are for stepping in: Pause, Stop, Restore, New run and Kill, in the same places every time, and Approve and Deny while the run waits on a gate. A button that does not apply is disabled, and its title says why. Restore and New run work only while the run is stopped. Kill sits apart from the others and takes a second press, because it is the last resort.

Each button has a command: autodev pause, autodev stop, autodev restore, autodev new-run, autodev kill, autodev approve and autodev deny. Every CLI command describes each one. Approve and Deny answer an irreversible-action gate, and Checkpoints and restore covers Restore.

A limit crossed #

When a provider hits a usage limit, AutoDev looks for somewhere else to run the turn instead of sleeping until the limit clears. It tries the next entry in the participant's own failover chain first, then a teammate whose provider is free. It sleeps only when every participant and every chain is limited. A team on one provider with no chain has nothing to hand the turn to. Providers and failover describes each step, and what AutoDev cannot detect.

The Needs you destination holds what is waiting on a person: a gate, a notice a participant left you, a question from a participant, or a run that is paused or in error with no daemon alive, which will not resume on its own. You answer a notice in the panel or with autodev answer, and a message of yours resumes a stopped run. The status bar shows the same count. It shows (stopped) when a run claims to be active and no daemon is alive, so a dead run is never reported as running.

What you read in the morning #

These three commands read a run back:

autodev status
autodev report --hours 12
autodev quota --days 7

autodev report prints what ran overnight: work done, what blocked, quota burned and what needs you. autodev quota prints subscription capacity used, left idle and converted into commits. A run that stopped records its own reason, and autodev status and autodev report read it.

The report window defaults to 12 hours and the quota period to 7 days. The panel has a Report destination as well, and Reading the quota panel says what each word in it means.

You do not have to be at the machine to hear how the night ended. notify.url sends a message to your phone when a run finishes, stops on its own or needs you, and suspendOnCompletion lets the machine sleep once the run is over. Every config key covers both.