[Tharun-Varshan-S]

Description

Fixes #406

WebArenaInstance.full_reset() previously returned None in both scenarios:

1. When the environment was successfully reset via the automated endpoint.
2. When the reset was skipped because WA_FULL_RESET (or VWA_FULL_RESET) was unset and skip_if_not_set=True.

Because prepare_backend() called full_reset() without arguments, a missing reset endpoint only logged a single warning while returning None — exactly matching a successful reset. As a result, downstream benchmark runs could silently proceed against un-reset / dirty instances without callers having any programmatic signal to verify reset execution.

This PR implements the "fail legibly" approach:

• WebArenaInstance.full_reset() now returns bool:
  • True if the automated reset was successfully executed.
  • False if the reset was skipped because the reset URL environment variable was unset or empty.
  • Continues raising exceptions when a configured reset endpoint fails or when skip_if_not_set=False.
• Updated prepare_backend() at the call sites for webarena, webarena_verified, and visualwebarena to consume this return value and prominently log:
  • Environment reset: DONE (when True)
  • Environment reset: SKIPPED (<RESET_URL_VAR> unset) (when False)
• Added comprehensive unit tests in tests/webarena/test_instance_reset.py covering all reset scenarios and log outputs.

What Changed

File	Change
browsergym/webarena/src/browsergym/webarena/instance.py	Updated full_reset() to track was_reset state, return bool, and updated docstrings.
browsergym/experiments/src/browsergym/experiments/benchmark/utils.py	Updated prepare_backend() for webarena, webarena_verified, and visualwebarena to check full_reset() result and log reset status.
tests/webarena/test_instance_reset.py	Added 12 unit tests verifying return values (True/False), exception handling on failure, VisualWebArena inheritance of VWA_FULL_RESET, and prepare_backend logging.

How It Was Tested

1. Unit Tests:
   Ran uv run pytest -v tests/webarena/test_instance_reset.py — all 12 tests passed:

   • test_full_reset_skipped_when_env_unset
   • test_full_reset_skipped_when_env_empty
   • test_full_reset_raises_when_env_unset_and_skip_false
   • test_full_reset_raises_when_env_empty_and_skip_false
   • test_full_reset_success
   • test_full_reset_success_already_running
   • test_full_reset_raises_on_reset_endpoint_error
   • test_visualwebarena_uses_vwa_env_var
   • test_prepare_backend_webarena_logs_done
   • test_prepare_backend_webarena_logs_skipped
   • test_prepare_backend_visualwebarena_logs_skipped
   • test_prepare_backend_webarena_verified_logs_skipped

2. Code Formatting & Linting:
   Ran uv run black . --check — passes cleanly with all files compliant.

3. Core BrowserGym Tests:
   Verified core Playwright action tests pass without regressions.

Security & Edge Cases Considered

• Empty vs Unset Environment Variables: Handled both None and "" values.
• Backward Compatibility: Non-breaking change; existing callers that do not inspect the return value continue unaffected.
• Fail-closed opt-in: Callers can still explicitly pass skip_if_not_set=False to abort with RuntimeError if the reset URL is not configured.
• Inheritance: Verified VisualWebArenaInstance inherits full_reset() while honoring its VWA_FULL_RESET variable.

Mentions

cc @Surya-5555 (maintainer) and @nikolas-sapa (issue author) — thank you both for the helpful discussion and feedback on the issue!