The goal is to create an LLM validator that cleans the code, export it to other frameworks / languages and ideally run it in a small container to check bugs.
For code cleaning, sometimes multiple versions are required for the CodeAgent to generate the code so it is good to have a validator that makes sure we don't have multiple versions for getting the same data.

Summary by cubic

Adds an LLM-powered CodeValidator with sandbox execution and new session export options (Python, Playwright, TypeScript) to improve reliability and make sharing automation scripts easier.

• New Features

  • Added CodeValidator with structured results (issues, improved code, summary), a SandboxExecutor, and DOM-safe validation rules.
  • session_to_python_script is now async; validates the full script and can auto-apply improvements; preserves JS blocks and fixes escape sequences.
  • New async exports: session_to_playwright_script and session_to_typescript_script with LLM conversion and safe fallbacks.
  • CodeAgent captures browser state for validation before clearing history; init re-exports new helpers and validator classes; docs updated for async usage.

• Migration

  • session_to_python_script is async. Update calls to: script = await session_to_python_script(agent).
  • Optional validator_llm for all export helpers; set apply_improvements=True to also return an improved script.

^{Written for commit fd52684. Summary will update automatically on new commits.}