When a run stops
Most runs end at factory:pr-open. The rest end in one of two labels, each with a comment on the issue that says what happened. Either way the factory archives the session's T3 thread and removes its worktree, and keeps the branch.
needs-info
The session decided the issue can't be built as written: the outcome is unclear, two readings would give different code, or the issue holds two separate pieces of work. It ends with its questions, and the factory posts them as a comment, sets factory:needs-info and unassigns the issue.
Answer in a comment or edit the issue so it settles every question, then an instructor sets factory:ready. The factory doesn't read the answer by itself: the label starts a new run, which reads the issue from the top.
stuck
The factory gave up. The comment gives the reason and, when the session had made commits, the branch they are on. Common reasons:
| Reason | What to do |
|---|---|
| The gate failed after the allowed attempts | Read the gate output on the run's page, fix the cause or the issue, then set factory:ready. |
| The run passed its time limit | Split the issue, or raise limits.minutes in .dw/config.yml. |
| The session gave up, or ended without a result line | Read the session in its T3 thread. |
| The issue text was changed by someone who isn't an instructor | Review the text, then set factory:ready. |
| The session ran as the wrong GitHub account | Restart the server so it has the right GH_CONFIG_DIR (see the machine config). |
| The required checks failed, or didn't report in time | Check the repo's CI and required_checks. |
| No model to launch with | Set a default model on the T3 project or in T3's Settings. |
| Stopped by hand | Nothing: someone stopped it. |
Fix the cause, then an instructor sets factory:ready to start a new run. A new run starts from the base branch, not from the old branch.
The time limit
Each run may spend limits.minutes (180 by default) claimed and building. Time waiting at factory:pr-open doesn't count. Past the limit, the factory stops the session and sets factory:stuck with the reason "the run passed its time limit".
Other limits in the repo's factory block cap the gate retries, the gate reruns when the base branch moves, and the wait for required checks. Each ends the run with factory:stuck when it runs out.
If T3 itself had trouble (the server restarted, or several sessions failed at once), the factory resumes the session instead of giving up, up to limits.interruption_resumes times. A session someone stopped in T3 waits rather than counting as a failure.
Stopping a run by hand
A run in factory:building can be stopped from its run page, or on the server's Mac with:
factory/scripts/stop-run.sh <run id>The run id is on the Factory page, for example original-solutions__t3code-factory__20__1. On its next pass the factory archives the session, removes the worktree and sets factory:stuck with the reason "stopped by hand". If a gate is running for the run, the stop waits for it.
A run in any other state can't be stopped this way. The script and the page say why and what to do instead.
Abandoning a run
Close the issue. A run that hasn't reached factory:pr-open is abandoned on the next pass: the factory archives the session, removes the worktree and writes nothing on GitHub. The server's Mac gets a notification that the run was abandoned.
A run at factory:pr-open ends when its pull request is merged or closed. Close the pull request to end it early.