[feat]: add support for safety confirmation callback for OpenAI + Google CUA #1473

shrey150 · 2025-12-24T04:20:56Z

why

Previously, we auto-approved all safety checks for CUA models
There are cases in which the caller wants to have the human approve/deny the request
Therefore, we should allow the handler to pass through any safety checks

what changed

Added support for safety checks and safety confirmations, which have a handler to intercept requests

test plan

Performed a risky request on a banking site and observed a safety check pop up

Summary by cubic

Adds safety confirmation support for OpenAI and Google CUA so risky actions can be reviewed and approved by a user. Defaults to auto-acknowledge when no handler is provided.

New Features
- Added SafetyCheck, SafetyConfirmationResponse, and SafetyConfirmationHandler types.
- Added onSafetyConfirmation to AgentExecuteOptions.callbacks for non-streaming CUA agents; streaming not supported and V3 prevents using it on non-CUA agents.
- OpenAICUAClient now requests confirmation for pending_safety_checks and sets acknowledged_safety_checks when approved.
- GoogleCUAClient now requests confirmation for safety_decision and sets safety_acknowledgement when approved.
- v3CuaAgentHandler wires the handler to CUA clients via setSafetyConfirmationHandler.

^{Written for commit 657fafc. Summary will update automatically on new commits.}

changeset-bot · 2025-12-24T04:20:59Z

🦋 Changeset detected

Latest commit: 657fafc

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 3 packages

Name	Type
@browserbasehq/stagehand	Patch
@browserbasehq/stagehand-evals	Patch
@browserbasehq/stagehand-server	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

cubic-dev-ai

No issues found across 5 files

greptile-apps · 2025-12-24T04:24:13Z

Greptile Summary

This PR adds safety confirmation support for OpenAI and Google CUA models, allowing callers to intercept and approve/deny risky actions instead of auto-approving all safety checks.

Key Changes

Type Definitions: Added SafetyCheck, SafetyConfirmationResponse, and SafetyConfirmationHandler types with comprehensive documentation
OpenAI CUA: Implemented handleSafetyConfirmation to process pending_safety_checks and set acknowledged_safety_checks based on user response
Google CUA: Implemented handleSafetyConfirmation to process safety_decision and set safety_acknowledgement based on user response
Handler Integration: V3CuaAgentHandler passes callbacks.onSafetyConfirmation through to CUA clients via setSafetyConfirmationHandler
Validation: Added check in prepareAgentExecution to prevent non-CUA agents from using onSafetyConfirmation callback
Auto-acknowledge Fallback: Both implementations auto-acknowledge when no handler is provided, maintaining backward compatibility

Implementation Quality

The implementation is well-structured and follows existing patterns in the codebase. Safety confirmation logic is consistently handled across both providers, with proper logging and error handling. Unit tests cover the core logic paths including acknowledgement, rejection, and auto-acknowledgement scenarios.

Confidence Score: 5/5

This PR is safe to merge with no identified issues
The implementation is clean, well-tested, and maintains backward compatibility through auto-acknowledgement. The safety confirmation feature is properly scoped to CUA agents only with appropriate validation. Code follows existing patterns, includes comprehensive type definitions, and has unit test coverage for the core logic.
No files require special attention

Important Files Changed

Filename	Overview
packages/core/lib/v3/types/public/agent.ts	Added SafetyCheck, SafetyConfirmationResponse, and SafetyConfirmationHandler types with clear documentation and examples. Added onSafetyConfirmation to AgentExecuteCallbacks and appropriate unavailability markers for streaming mode.
packages/core/lib/v3/agent/OpenAICUAClient.ts	Added safety confirmation handler support for OpenAI CUA. Handles pending_safety_checks by requesting user confirmation or auto-acknowledging. Implementation correctly handles both success and error paths.
packages/core/lib/v3/agent/GoogleCUAClient.ts	Added safety confirmation handler support for Google CUA. Handles safety_decision by requesting user confirmation or auto-acknowledging. Correctly converts safety decisions to SafetyCheck format.
packages/core/lib/v3/handlers/v3CuaAgentHandler.ts	Added setSafetyConfirmationHandler method that delegates to CUA clients. Handler is set at the start of execute() from callbacks.onSafetyConfirmation.
packages/core/lib/v3/v3.ts	Added validation to prevent non-CUA agents from using onSafetyConfirmation callback. The check correctly throws an error for DOM mode agents while allowing CUA agents to use the callback.

Sequence Diagram

sequenceDiagram
    participant User
    participant V3CuaAgentHandler
    participant CUAClient as OpenAICUAClient/GoogleCUAClient
    participant Provider as OpenAI/Google API
    participant Handler as SafetyConfirmationHandler

    User->>V3CuaAgentHandler: execute(options with callbacks.onSafetyConfirmation)
    V3CuaAgentHandler->>V3CuaAgentHandler: setSafetyConfirmationHandler(callback)
    V3CuaAgentHandler->>CUAClient: setSafetyConfirmationHandler(callback)
    
    V3CuaAgentHandler->>CUAClient: execute task
    CUAClient->>Provider: send request
    Provider-->>CUAClient: response with pending_safety_checks/safety_decision
    
    alt Handler is set
        CUAClient->>Handler: call handler(safetyChecks)
        Handler-->>User: prompt for confirmation
        User-->>Handler: approve/deny
        Handler-->>CUAClient: {acknowledged: true/false}
        alt User approved
            CUAClient->>Provider: send acknowledged_safety_checks/safety_acknowledgement
        else User denied
            CUAClient->>Provider: continue without acknowledgement
        end
    else No handler set
        CUAClient->>CUAClient: auto-acknowledge
        CUAClient->>Provider: send acknowledged_safety_checks/safety_acknowledgement
    end
    
    Provider-->>CUAClient: final response
    CUAClient-->>V3CuaAgentHandler: agent result
    V3CuaAgentHandler-->>User: return result

greptile-apps

_{5 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

packages/core/lib/v3/agent/OpenAICUAClient.ts

seanmcguire12 · 2025-12-24T22:54:41Z

@greptileai

@tkattkat

Add safety confirmation support

29ab411

cubic-dev-ai bot reviewed Dec 24, 2025

View reviewed changes

greptile-apps bot reviewed Dec 24, 2025

View reviewed changes

packages/core/lib/v3/agent/OpenAICUAClient.ts Show resolved Hide resolved

packages/core/lib/v3/agent/OpenAICUAClient.ts Show resolved Hide resolved

shrey150 added 5 commits December 23, 2025 20:24

Fix lint

7179658

Add changeset

cbe577a

Add tests

14e0b8b

Fix flaky test

e0b6bed

Fix lint

b45ece8

shrey150 mentioned this pull request Dec 24, 2025

Add safety confirmation support browserbase/gemini-browser#35

Open

pass safety check callback through callbacks param in .execute()

7f579ce

use actual cua clients in test instead of duplicating logic

657fafc

seanmcguire12 changed the title ~~Add safety confirmation support for OpenAI + Google CUA~~ [feat]: add support for safety confirmation callback for OpenAI + Google CUA Dec 24, 2025

seanmcguire12 approved these changes Dec 24, 2025

View reviewed changes

pkiv approved these changes Dec 24, 2025

View reviewed changes

seanmcguire12 merged commit e021674 into main Dec 24, 2025
20 checks passed

This was referenced Dec 24, 2025

Version Packages #1414

Merged

Version Packages CloudEngineHub/stagehand#1

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[feat]: add support for safety confirmation callback for OpenAI + Google CUA #1473

[feat]: add support for safety confirmation callback for OpenAI + Google CUA #1473

Uh oh!

shrey150 commented Dec 24, 2025 •

edited by cubic-dev-ai bot

Loading

Uh oh!

changeset-bot bot commented Dec 24, 2025 •

edited

Loading

Uh oh!

cubic-dev-ai bot left a comment

Uh oh!

greptile-apps bot commented Dec 24, 2025 •

edited

Loading

Uh oh!

greptile-apps bot left a comment

Uh oh!

Uh oh!

Uh oh!

seanmcguire12 commented Dec 24, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

[feat]: add support for safety confirmation callback for OpenAI + Google CUA #1473

[feat]: add support for safety confirmation callback for OpenAI + Google CUA #1473

Uh oh!

Conversation

shrey150 commented Dec 24, 2025 • edited by cubic-dev-ai bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

why

what changed

test plan

Summary by cubic

Uh oh!

changeset-bot bot commented Dec 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🦋 Changeset detected

Uh oh!

cubic-dev-ai bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot commented Dec 24, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Key Changes

Implementation Quality

Confidence Score: 5/5

Important Files Changed

Sequence Diagram

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

seanmcguire12 commented Dec 24, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

shrey150 commented Dec 24, 2025 •

edited by cubic-dev-ai bot

Loading

changeset-bot bot commented Dec 24, 2025 •

edited

Loading

greptile-apps bot commented Dec 24, 2025 •

edited

Loading