osworld-reset_skill

Overview

This skill resets an OSWorld environment either softly or hard. Soft mode clears only the internal step counter; hard mode reverts the desktop environment to a saved snapshot by calling DesktopEnv.reset(). Use it to control experiment state and support reflection or replay workflows.

How this skill works

The skill accepts a required string parameter: "soft" or "hard". In soft mode it updates the environment metadata to set the step counter back to zero without changing any running state. In hard mode it invokes DesktopEnv.reset() on the OSWorld instance (using OSWORLD_URL or supplied config) to restore the snapshot state and reset counters. The skill returns a note containing formatted text and metadata about success, mode, and step_counter.

When to use it

• Clear the step counter before a new evaluation without altering the environment (soft).
• Restore the entire desktop environment to a known snapshot for reproducible replay or counterfactual tests (hard).
• Prepare a clean baseline before running a different policy or experiment variant.
• Recover from a corrupted or undesired environment state by reverting to snapshot.
• Run controlled reflection cycles where you want to compare outcomes from the same initial state.

Best practices

• Choose soft resets for quick reflection or logging when environment continuity matters.
• Use hard resets when you need deterministic replay or must undo environment-side changes.
• Ensure OSWORLD_URL is set correctly or provide osworld_url in the character config to target the right instance.
• Verify success metadata after reset before proceeding with sensitive actions.
• Limit hard resets in parallel runs to avoid snapshot conflicts; serialize heavy operations.

Example use cases

• Soft reset to zero the step counter between evaluation episodes while preserving user apps and windows.
• Hard reset to return to a clean snapshot before launching an alternative task policy for A/B testing.
• Automate a test loop: hard reset, run scripted interactions, observe results, repeat for reproducible metrics.
• Use soft reset after a brief reflection step to continue exploration without returning to snapshot state.
• Recover from accidental GUI changes by issuing a hard reset to restore the original desktop layout.