Skip to Content

Run autonomous, self-running Claude Code sessions

Tares runs a software factory for Claude Code: a crew of Claude Code sessions that run in the background on your machine. A watch brings back any session that stops, the orchestrator checks for work on a schedule, and builders merge reviewed pull requests, so the work continues while you are away. The crew asks you before the decisions that are yours.

Start the crew

tares factory crew up

The crew lives in a folder of its own, ~/factory. The first run offers to open Claude Code there so you can trust the folder once, then asks how far the crew may go (below). The answers are kept on Tares. crew up also starts the watch in the background that brings back any station that stops.

When a project’s repo has a language Claude Code can search by symbol, and no language server is set up for it yet, crew up lists what is missing and offers to install it. It installs only on your yes. Troubleshooting

When more than one person is on the Tares, the first crew up also asks for a crew name, so your stations are told apart from your teammates’. A crew per person

How far the crew may go

SettingChoicesWhat it decides
AutonomyL4-ship (recommended), L3-reviewAt L4-ship a builder merges once the reviewer passes the pull request and CI is green. At L3-review every merge waits for your yes.
Releaseroff, onWhether a releaser tags and deploys. Without one, work ends merged on main.
Second checkoff, onWhether a second model (Codex) checks every commit.
Builders at oncea numberHow many builders run side by side, across every project.
What you allowedyour wordsThings the crew may always do without asking.

Change them under Settings, Crew in the console, or tell the orchestrator. A single project can be stricter than the crew: say so in its spec, and every merge of that project waits for your yes. On a team, each person can set their own values, as strict as the team’s or stricter.

How the sessions keep running

Claude Code stops an idle session after about an hour. Tares plans for that:

  • tares factory crew up starts the crew and a watch in the background. The watch brings back any station that stops, under the same name. Its log is ~/factory/watch.log.
  • The orchestrator wakes on a schedule every 20 minutes and checks for work, so an idle crew still notices a new project, a stalled ticket or a merged pull request that unblocks the next one. With wake on events on, the watch does the waking instead.
  • A revived session reads its notes and the project on Tares, so it carries on from where the last one stopped. Tares keeps the plan, the tickets, the reviews and your answers.

tares factory crew down stops every station. Tares keeps the projects, tickets and answers, and tares factory crew up starts fresh sessions that carry on from them.

Crew settings

Three settings change how the crew spends your Claude plan and keeps up. Each is shared by every project, and the Crew page shows all three under How far the crew may go.

SettingDefaultWhat it does
idle_wakeoffWakes stations on events instead of fixed ticks.
paceoffSlows the crew before your plan’s limit.
claude_updateonLets the watch move stations onto a new Claude Code.

To change one, tell the orchestrator (tares factory crew attach orchestrator), or your own session if it has the Tares plugin: either saves it with the set_crew_settings tool. Pace also has a Turn on and Turn off button on the Crew page. The watch carries a change to the stations by itself.

Wake on events

By default every station holds a schedule of its own, and each tick costs a model turn even when there is nothing to do. With idle_wake on, the stations hold no schedule. The watch asks Tares what each station has to do and restarts an idle station, in the same conversation, with one message only when there is work:

  • a pull request without a verdict, for the reviewer
  • a change on one of its tickets, such as a verdict, changes asked for or failed CI on its pull request, for its builder
  • a release request, for the releaser
  • an answer on the desk, for the orchestrator
  • a new project, for the orchestrator

When nothing is known to be pending, a station gets a short liveness wake after 15, 20 or 30 minutes of quiet, then after 30, then every 45 minutes at most, until it does something. It checks its queue once and ends the turn, so an empty crew costs a few model turns an hour. Turn idle_wake off and each station is told to arm its own schedule again.

tares factory doctor warns when idle_wake is on and the watch is not running, because nothing would wake the stations then.

Pace the crew before the plan limit

Without pace, the crew runs until your Claude plan’s limit stops it, and every station restarts at once when the limit resets. With pace on, the watch reads the plan usage Tares already keeps (see the Crew page) and slows the crew first:

Plan usageWhat the crew does
85% of the 5-hour window, or 90% of the weekNo new builder starts. Each station is told once to tick every 45 minutes.
95% of the 5-hour windowA builder also finishes its step, pushes, writes down where it is and waits.

A station that stopped while a limit was full is not revived into it. After the reset the watch revives it in its own conversation, and starts a new one only after that fails twice. The Crew page and tares factory crew status say when the crew is pacing and until when.

While a plan limit is full, tares factory crew update refuses too, because a restart is a model turn.

Move onto a new Claude Code

Claude Code installs a new version by swapping a link, so a station keeps the version it started on until someone restarts it. With claude_update on (the default), tares factory crew watch sees a newer claude --version that has held for a whole round and restarts one idle station a round, in the same conversation, the orchestrator last. A busy station, or one you have open in a terminal, waits. The crew’s helpers move too, after the other stations and before the orchestrator, and a busy helper does not hold the orchestrator back.

To move every station right away:

tares factory crew update

It waits for each station to be idle, restarts it, and checks that it runs the new version and that there is one session per station name. A station still busy after --wait <seconds> (30 minutes by default, 0 for no deadline) is left to the watch. crew update skips a station that already runs the installed version. tares factory crew restart restarts every station, those too.

Claude Code’s background service keeps pre-warmed spares of the old version, and a restart can land on one. The factory stops such a spare and tries again. If a station keeps coming back on the old version, it is started again in a new conversation on the new one after three tries (a helper is not, since its task lives only in its conversation). If that fails too, the watch tries again after a pause that doubles from 10 minutes up to 4 hours, and crew update tries at once. The Crew page and tares factory roster show each station’s Claude Code version, flag an older one and mark a retrying one. tares factory doctor says so for each station, and tares factory crew status counts them.

Turn claude_update off to pin a version: the watch then moves no station, and one that restarts for another reason starts on the installed version. A crew you froze with tares factory crew freeze, which keeps the crew on today’s plugin kit, moves no station on its own either. The watch says once for each new version that one is there, and crew update still works.

How work gets picked up

You plan a project with Claude, and Tares keeps the plan as tickets, each with the repo it changes. After that you do not assign anything.

  • The orchestrator hands out work and starts builders itself, one per milestone and repo of a project.
  • A builder takes its tickets, opens pull requests and merges its own once they pass.
  • The reviewer reviews every pull request to a fixed bar and posts the verdict.

What still needs you

The crew does not guess at decisions that are yours. The orchestrator is the session you talk to, and it brings you what needs you, one question at a time and with a recommendation.

  • Questions wait on the desk until you answer them. When the plan does not answer a smaller question, the crew records an assumption and carries on, and the orchestrator confirms the assumptions with you once per milestone.
  • Merging is your setting. At L4-ship, a builder merges once the reviewer approves and CI is green. At L3-review, every merge waits for your yes.
  • Releasing is off unless you ask for it. Without a releaser, the crew stops when work is merged to main.

To answer, attach to the orchestrator:

tares factory crew attach orchestrator

Commands

Every crew command, from crew up to crew down, is on Commands.

Last updated on