claude agent prompt updated

[shivammittal274]

No description provided.

[greptile-apps]

Greptile Overview

Greptile Summary

This PR updates the system prompt for the Claude SDK agent to reflect a fundamental architectural shift from Chrome DevTools Protocol (CDP) page-based operations to BrowserOS controller tab-based operations. The key changes include: replacing the implicit page selection workflow (list_pages/select_page) with explicit tab ID management using browser_list_tabs/browser_switch_tab, renaming all CDP-style tools (take_snapshot, click, fill) to controller-style tools (browser_get_interactive_elements, browser_click_element, browser_type_text), switching from uid-based element targeting to nodeId-based targeting, and making tabId a required parameter for most browser operations. The updated prompt now documents 15+ tools across categories including Tab Management, Navigation & Content, Interaction, Scrolling, and Advanced operations. This change aligns the agent's understanding with the new controller-extension-based architecture where a browser extension communicates via WebSocket with the agent, replacing direct CDP control with a more explicit, tab-centric model.

Important Files Changed

Changed Files

Filename	Score	Overview
packages/agent/src/agent/ClaudeSDKAgent.prompt.ts	4/5	Updated Claude agent system prompt from CDP page-based workflow to controller tab-based workflow with explicit tab ID management and nodeId-based element targeting

Confidence score: 4/5

• This PR is generally safe to merge but requires verification that the documented tool signatures match actual implementations
• Score reflects the critical dependency on alignment between this prompt and the actual tool implementations in packages/tools/src/controller-based/ - any mismatch will cause agent failures when Claude attempts to call tools with incorrect parameters or non-existent tool names
• Pay close attention to packages/agent/src/agent/ClaudeSDKAgent.prompt.ts to ensure all documented tool signatures (especially parameter requirements like tabId being required vs optional) match the actual tool definitions in the controller-based tools package

Sequence Diagram

sequenceDiagram
    participant User
    participant Agent as ClaudeSDKAgent
    participant Browser as BrowserOS
    
    User->>Agent: "Request browser automation task"
    
    rect rgb(240, 248, 255)
        Note over Agent,Browser: Tab Identification
        Agent->>Browser: "browser_list_tabs() or browser_get_active_tab()"
        Browser-->>Agent: "Return tab information with IDs"
        
        opt If different tab needed
            Agent->>Browser: "browser_switch_tab(tabId)"
            Browser-->>Agent: "Tab switched confirmation"
        end
    end
    
    rect rgb(255, 250, 240)
        Note over Agent,Browser: Navigation (if needed)
        Agent->>Browser: "browser_navigate(url, tabId)"
        Browser-->>Agent: "Navigation complete"
    end
    
    rect rgb(240, 255, 240)
        Note over Agent,Browser: Content Analysis
        Agent->>Browser: "browser_get_interactive_elements(tabId)"
        Browser-->>Agent: "Return elements with nodeIds"
        
        opt If visual context needed
            Agent->>Browser: "browser_get_screenshot(tabId)"
            Browser-->>Agent: "Return screenshot with bounding boxes"
        end
        
        opt If text extraction needed
            Agent->>Browser: "browser_get_page_content(tabId, type)"
            Browser-->>Agent: "Return page content"
        end
    end
    
    rect rgb(255, 240, 245)
        Note over Agent,Browser: Interaction
        alt Click action
            Agent->>Browser: "browser_click_element(tabId, nodeId)"
            Browser-->>Agent: "Click executed"
        else Type action
            Agent->>Browser: "browser_type_text(tabId, nodeId, text)"
            Browser-->>Agent: "Text typed"
        else Clear input
            Agent->>Browser: "browser_clear_input(tabId, nodeId)"
            Browser-->>Agent: "Input cleared"
        else Scroll to element
            Agent->>Browser: "browser_scroll_to_element(tabId, nodeId)"
            Browser-->>Agent: "Scrolled to element"
        end
    end
    
    rect rgb(245, 245, 245)
        Note over Agent,Browser: Advanced Operations
        opt Execute JavaScript
            Agent->>Browser: "browser_execute_javascript(tabId, code)"
            Browser-->>Agent: "Execution result"
        end
        
        opt Send keyboard keys
            Agent->>Browser: "browser_send_keys(tabId, key)"
            Browser-->>Agent: "Keys sent"
        end
    end
    
    Agent-->>User: "Return task results"

Loading

Context used:

• Context from dashboard - CLAUDE.md (source)

_{1 file reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}