INSIDE WORKBENCH

A real task.
A closer look.

Imagine reopening your project and finding every terminal back in its default directory. Here’s how you could approach session recovery in Workbench.

ILLUSTRATIVE WALKTHROUGH

Restore the workspace.
Keep the scope specific.

The task: restore the layout, pane directories and selected pane after restarting the app. The examples below illustrate the work involved; they are not a recorded run or screenshots of the interface.

01 · DEFINE THE BEHAVIOR

What should survive a restart?

Write down what “restore my session” means before asking the agent to implement it. Save the requirement and the decisions that bound the task.

  • Restore the pane layout and selected pane
  • Reopen each pane in its previous directory
  • Keep restarting running commands outside this task
EXAMPLE REQUIREMENT

Session recovery

After closing and reopening the app, I can return to the same arrangement of panes.

Decision: save workspace state on close.
Edge case: a saved directory may no longer exist.
Fallback: open that pane at the repository root.

02 · WORK IN THE REPOSITORY

Give the agent a bounded task.

Select the requirement, relevant architecture and session-storage files. Use Codex or Claude Code with your local tools to make the change and inspect the diff.

  • Include the agreed fallback for missing directories
  • Keep unrelated startup behavior out of the change
  • Run the repository’s session tests in the terminal
EXAMPLE AGENT INSTRUCTION

Implement session recovery

Use the saved requirement and existing session storage. Restore the layout, directories and selected pane. Add coverage for a directory that has been deleted, then report the changes and test results.

03 · CHECK THE RESULT

Exercise the awkward case too.

Inspect the changed files and test output, then restart the app. Check that a missing directory uses the fallback without preventing the other panes from restoring.

  • Compare the restored layout with the original
  • Check the directories and selected pane
  • Record failures and follow-up work alongside the result
EXAMPLE REVIEW CHECKLIST

What would convince you?

Storage tests cover saving and loading the layout.
A restart demonstrates directory and selection recovery.
A deleted-directory check exercises the fallback.
The diff contains only changes needed for this task.

PRACTICAL QUESTIONS

How it fits
your setup.

Does Workbench replace my coding agent?
Workbench brings Codex and Claude Code into the workspace alongside project context, your repository and terminal. You still work with the agent you choose.
Can I use an existing repository?
Workbench is built around local codebases and installed development tools. The example uses session recovery, but the same workspace can hold tasks from your own project.
Where does my code run, and what does the provider receive?
Development tools run in your local environment. Agent-provider processing is a separate boundary; see the security page for the current disclosures and details still being confirmed.
Does passing a test mean the task is finished?
A test supports the behavior it checks. Review the diff and exercise the relevant user behavior as well, especially where a requirement depends on interaction or recovery.

WORKBENCH

Have a workflow question?

Ask about how Workbench could fit the way you develop software.

Get in touch