Tasks and runs
Separate the requested work from each execution attempt and its evidence.
Tasks
A task owns the requested outcome, numbered requirements, repository, workflow and accountable person. Create a task from the organization's Work list or a repository, or import a GitHub issue. A task keeps one durable conversation; clarify assumptions there before a run writes code.
Runs
A run is one execution attempt through the six workflow stages. It records workspace identity, stage attempts and the provider connection each stage used, tool decisions, validation, review findings, cost and delivery status. The run console shows live progress and links the draft pull request once it exists.
Before a run starts, the task's run preflight shows the readiness checks and the review requirement: every 1.0.0 run needs a reviewer agent from a different provider and model family than every Implement step, and a human technical review before publication. See review requirements for every run.
The run console's controls depend on the stage:
- Redirect a Codex subscription turn. While a Codex subscription turn runs, you can steer it with a Steering instruction. Other stages accept no input while they run, and no turn can be interrupted midway.
- Output. Codex subscription and API-key stages show live output. A Claude subscription stage's result appears when the stage settles.
- Pause takes effect at the next stage boundary, after the running turn settles. Resume continues the run.
- Cancel run stops the whole run. A single stage cannot be cancelled on its own.
Retry without erasing history
Start a new run after a failed or rejected attempt. The prior run stays available, so reviewers can compare attempts and see why the next one exists.