Add Browser Env Integration

[filip-michalsky]

Description

Adds BrowserEnv - a unified browser automation integration for the verifiers library supporting two operational modes:

DOM Mode (mode="dom")

• Uses the Stagehand Python SDK for natural language browser control
• Tools: navigate, observe, act, extract - Stagehand's AI-driven primitives
• Ideal for tasks that benefit from semantic understanding of page elements

CUA Mode (mode="cua")

• Vision-based primitives for Computer Use Agent workflows
• Tools: click, double_click, type_text, keypress, scroll, goto, back, forward, wait, screenshot
• Requires companion TypeScript server (included) for CDP connection via Stagehand internals
• Automatic screenshot management with context trimming for VLM input

Both modes support local browser execution or Browserbase cloud infrastructure.

What's included:

• verifiers/envs/integrations/browser_env/ - Core integration (BrowserEnv, DOMMode, CUAMode)
• verifiers/envs/integrations/browser_env/cua-server/ - TypeScript server for CUA mode
• environments/browser_dom_example/ - Minimal DOM mode example
• environments/browser_cua_example/ - Minimal CUA mode example
• New [browser] extra: uv add 'verifiers[browser]'

Benchmarks (GAIA, WebVoyager, Mind2Web) have been pushed to Prime Hub under the browserbase/ namespace.

Type of Change

• New feature (non-breaking change which adds functionality)

Testing

# DOM mode
prime eval run browserbase/browser-dom-example -m openai/gpt-4.1-mini

# CUA mode (start server first: cd verifiers/envs/integrations/browser_env/cua-server && ./start.sh)
prime eval run browserbase/browser-cua-example -m qwen/qwen3-vl-30b-a3b-instruct

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Additional Notes

Future work:

• Compile CUA TypeScript server to binary to remove Node.js dependency
• Additional benchmark environments available on Prime Hub under browserbase/ org
  ~

---

Note

Adds a unified browser automation integration with two modes and supporting assets.

• New BrowserEnv in verifiers/envs/integrations/browser_env with mode="dom" (Stagehand tools) and mode="cua" (vision primitives + screenshots); default system prompts; env var validation; custom tool call handling for multipart CUA responses; screenshot filtering
• Exports BrowserEnv via verifiers/__init__.py and integration package __init__.py (lazy imports)
• Examples: environments/browser_dom_example and environments/browser_cua_example with minimal datasets, judge rubric, README, and pyproject.toml
• CUA server: TypeScript Fastify service under browser_env/cua-server/ (actions API, session management, README, scripts, env templates)
• Docs: add BrowserEnv to docs/environments.md and integrations/README.md, including install extras and mode descriptions
• Deps: new [project.optional-dependencies].browser extra (stagehand, aiohttp, python-dotenv)
• Tests: tests/test_browser_env.py covering env var checks, prompts, CUA formatting/filtering, DOM LLM config, example datasets; update tests/test_envs.py to skip new browser examples and mcp_env

^{Written by Cursor Bugbot for commit 906a836. This will update automatically on new commits. Configure here.}

[CLAassistant]

All committers have signed the CLA.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Screenshot filtering ineffective due to wrong placement

High Severity

The filter_screenshots_in_messages call in env_response doesn't actually filter screenshots from the context sent to the model. In MultiTurnEnv.get_prompt_messages, the original unfiltered messages variable is used in the final concat_messages([messages, env_response]) call. The filtered version only affects what's passed to super().env_response() for tool call extraction, which only examines messages[-1] anyway. Screenshots will accumulate unbounded in the conversation, potentially causing context length issues despite the filtering code existing.

 

[cursor]

Missing synchronization causes resource leak in concurrent rollouts

Medium Severity

DOMMode._create_session lacks synchronization when creating the shared stagehand_client. If multiple rollouts call _create_session concurrently when stagehand_client is None, each creates a new AsyncStagehand instance. Only the last one is stored; the others are orphaned with unclosed connections. CUAMode correctly uses _thread_lock and _client_lock to protect its shared client creation, but DOMMode has no such protection.