Tasks and jobs
A task is the conversation you see in the app. A job (also called a run) is one attempt by the agent inside it. Your first prompt starts the first job. A follow-up, an answer to a question, or a retry starts a new job on the same task. The new job continues from the same checkout and session, so the task keeps one history and one status. The status you see on a task is the state of its newest job, plus what you still owe it.Statuses
Queued Running Stopping Needs answer Waiting for review Done FailedA job whose runner vanished (for example a lost node) shows as Failed. Infrastructure losses are retried for you and listed on the task as interruptions. They do not count as a failed attempt.
What moves a task
- Start task creates the first job. It is Queued, then Running when a runner picks it up.
- When the agent asks a question, the task shows Needs answer until you answer. Your answer starts a new job.
- Stop moves a running task to Stopping, then Stopped · Needs review.
- A finished run waits for you. Follow up, Retry or Edit start a new job. Mark done closes the task. Reopen brings it back.
- When review on the pull request is pending, the task shows Waiting for review. Once that wait ends, the reason is added to the label.
When a run fails
A failed run shows a red pill with the reason. Open the run to read the output, the gates and the services.timed out
timed out
The run used its whole time budget and was killed. Gates still run on whatever the agent left behind. Split the task into smaller steps, or follow up and ask the agent to continue.
gate failed
gate failed
The agent finished, but a gate in
.bellows.yaml failed. Open the run and read the gate output. Follow up with the failing command’s error, or fix a flaky gate in the repo. See Configure a repo.config error
config error
Bellows could not start the run because of its configuration, for example an executor or
.bellows.yaml problem. Fix the setting, then retry. A retry without a fix fails the same way.service failed
service failed
A service declared in
.bellows.yaml (a database, for example) died during the run. Check the service image and settings, then retry.cache lost
cache lost
Bellows ended the run because the agent’s prompt cache was lost. Retry. The new job runs on the same task.
blocked
blocked
The agent reported itself blocked, for example a missing credential or an unclear requirement. Read its last message, give it what it needs, and follow up.
publish failed
publish failed
The change was ready but Bellows could not publish the branch or pull request. Typical causes are a conflict with the base branch or a refused push. Read the run output, resolve it, then follow up or retry.
helper failed
helper failed
A helper step of the workflow failed. Read the run output for the step’s error, then retry.
evidence required
evidence required
The agent said it succeeded, but your organization requires passing gates or a review as evidence, and the run did not have it. Bellows turned the result into a failure. Follow up and ask the agent to run the gates, or get the review.
checkpoint unrestorable
checkpoint unrestorable
The task’s workspace was lost and Bellows could not rebuild it from its saved checkpoint. Start a new task for the follow-up work.
runner error
runner error
The runner itself failed, or the run was cancelled mid-way. Retry. If it repeats, ask your admin to check the executor.
Retry is for failures that are not your prompt’s fault: runner error, cache lost, a flaky service. For gate, evidence and blocked failures, a Follow up with the details works better. See Writing good tasks.