Step 3 of 6

Keep heavy files off the system drive

Move the folders that grow to a second drive before the first long job. A full system drive does not slow one task down, it stops everything at once.

What happened

What it costs when you skip it

A long job writes intermediate files, model weights or exports. The system drive reaches zero free space in the middle of it. The job fails, the app cannot save, and the cleanup that follows is an afternoon of deciding what is safe to delete — on a machine that is already misbehaving.

The cost is not the gigabytes. It is that recovery happens while the machine is degraded, which is the worst possible moment to make decisions about your own files.

Do this

1. Check the numbers first. Windows: Get-PSDrive C. macOS: df -h /. Write down free space and total size.
2. Create the destinations on the second drive. D:\agent-work for outputs, D:\cache\temp for scratch files. Keep the two apart so clearing scratch never touches deliverables.
3. Move the scratch folder first. Windows: set TEMP and TMP to D:\cache\temp. This is the cheapest win and it affects every app, not just the agent.
4. Move tool caches one at a time. Each tool has its own switch — npm (npm config set cache), pip (pip config set global.cache-dir), Hugging Face (HF_HOME), Playwright browser binaries, and model weights. Check the tool's own documentation for the current variable name before setting it.
5. Point the agent's workspace at the second drive. The prompt from step 1 already does this if the workspace path is right.
6. Verify with a real job. Run the heaviest thing you normally run and watch free space before and after. If it did not move, the variable is not being read.
7. Keep a floor. Treat 20 percent free as the minimum. Below that, stop starting long jobs until you have cleared space.

Checklist

• Free space and total size recorded before any change.
• TEMP and TMP point at the second drive.
• At least one tool cache has been moved and verified.
• The agent workspace is on the second drive.
• A real job ran, and the system drive did not shrink.

Common mistakes

• Moving one cache and calling it done. The drive fills from whichever folder grows fastest, and that is rarely the one you moved. Measure after a real job.
• Moving app data while the app is running. Close the app first. A half-moved cache is worse than a full drive, because it breaks the tool instead of warning you.
• Expecting the agent to clean up after itself. Cleanup is a decision about your files. It is yours to make, with a rule, not a polite request.
• Deleting to make room during a failure. When the machine is already unstable, move files to the second drive. Delete nothing you cannot identify.

What you need for this step

The Drive mover guide is still being written. The delivery rules pack covers the workspace half of this step today, including where the workspace should live.