Runs and tasks

A run is one task conversation connected to one repository context and one isolated workspace.

Run identity

Each run has its own identifier, transcript, workspace, tool activity, and review state. That separation is what lets multiple tasks run without sharing writable files.

You normally do not need the internal run identifier. Include it only when reporting a bug or debugging a failed run.

Run lifecycle

A typical run moves through these stages:

1. Context: resolve the user, repository, branch, provider, and model.
2. Workspace: prepare the isolated cloud workspace for the task.
3. Execution: stream the agent response and approved tool activity.
4. Review: collect changed files and diffs for inspection.
5. Outcome: complete, fail, or stop with a visible state.

If a run fails, keep the run open long enough to read the error and changed files. Starting another run may be useful, but it should be a deliberate retry with a clear scope.

Writing effective tasks

State the outcome, scope, and verification together:

1Fix the empty-state copy in the repository picker. Reuse the existing component styles, do not change authentication behavior, and run its focused test.

Avoid combining unrelated refactors, deployment changes, and product work in one task. Smaller run scopes produce clearer diffs and make recovery easier.