Roo-Code: Roo Code does not know of the `browser_action` tool if necessary preconditions are unmet

App Version

3.19.5

API Provider

Google Gemini

Model Used

gemini-2.5-pro-preview-05-06

šŸ” Steps to Reproduce

On Mac OSX.

  1. Start a new Roo Code chat in Ask mode.
  2. Ask List all tools, even those not immediately available in Ask mode.
  3. Read result, looking for browser_action.

šŸ’„ Outcome Summary

Expected: While necessary conditions to use browser_action may not have ben met, Roo Code should at least be aware that the tool exists and could be made available. That way, Roo can guide the user into achieving the necessary prerequisites.

Without this knowledge, a user is left wondering if their Roo Code is broken. Use of browser tools is a prominent feature listed in the Roo Code project’s main README.md file.

Actual:

Roo Code has no knowledge of the browser_action tool if the necessary prerequisites have not been met. Example output:

Okay, I can list the tools available. Please note that not all tools are usable in every mode.

Here are the available tools:

*   **[`read_file`](command:read_file)**: Reads the contents of a specified file. It can read the entire file or specific line ranges.
*   **[`fetch_instructions`](command:fetch_instructions)**: Fetches detailed instructions for performing specific tasks, like creating an MCP server or a new mode.
*   **[`search_files`](command:search_files)**: Performs a regex search across files in a specified directory and provides context-rich results.
*   **[`list_files`](command:list_files)**: Lists files and directories within a specified path, either recursively or at the top level.
*   **[`list_code_definition_names`](command:list_code_definition_names)**: Lists definition names (classes, functions, etc.) from source code in a file or a directory.
*   **[`use_mcp_tool`](command:use_mcp_tool)**: Executes a tool provided by a connected MCP (Model Context Protocol) server.
*   **[`access_mcp_resource`](command:access_mcp_resource)**: Accesses a resource (like a file or API response) provided by a connected MCP server.
*   **[`ask_followup_question`](command:ask_followup_question)**: Asks the user a question to gather more information needed to complete a task.
*   **[`attempt_completion`](command:attempt_completion)**: Presents the final result of a task to the user, optionally with a command to demonstrate the result.
*   **[`switch_mode`](command:switch_mode)**: Requests to switch to a different operational mode (e.g., from "Ask" to "Code").
*   **[`new_task`](command:new_task)**: Creates a new task instance in a chosen mode with an initial message.

Additionally, there are tools for editing files, which are typically available in modes like "Code" mode:

*   **`apply_diff`**: Replaces lines in existing files.
*   **`write_to_file`**: Creates new files or completely rewrites existing files.
*   **`insert_content`**: Adds lines of text to existing files at a specific line number.
*   **`search_and_replace`**: Finds and replaces text or regex patterns in files.

There is also a tool for running commands:

*   **`execute_command`**: Runs CLI commands on the user's computer.

šŸ“„ Relevant Logs or Errors (Optional)


About this issue

  • Original URL
  • State: closed
  • Created a year ago
  • Reactions: 1
  • Comments: 18

Most upvoted comments

@jimbojw I don’t believe there’s a way to override this without editing the list of compatible models

I see. Thank you for that explanation! I would recommend adding all of that guidance to Browser Use - Using Browser Use documentation.

Gemini 2.5 Pro is my exclusive model for almost everything. I only switch to Claude rarely to solve niche programming issues. Gemini 2.5 has a 1M token limit, so it should be competent to run browser_action but of course I can’t say this with certainty as I haven’t tried it on Roo.

I was previously able to successfully use Cline’s integrated browser feature using Gemini 2.5 Pro, which is why I was perplexed to find it missing in Roo. This missing Roo feature isn’t enough to drive me back to Cline. Roo is best-in-class among directly-connectable models in my opinion (as of this writing).

To answer your question about Claude 3.5+ and browser_action:

There’s nothing mechanically special about Claude 3.5+ that makes it exclusively capable of browser actions. The restriction exists because only certain models have been officially trained and validated by Anthropic for computer use capabilities. This isn’t a Roo Code limitation - it’s based on what the model providers have actually released and tested.

The approved models (Claude 3.5 Sonnet, 3.7 Sonnet variants, Claude 4, etc.) are manually added to the OPEN_ROUTER_COMPUTER_USE_MODELS list when providers officially announce computer use support. It’s not an automated process or arbitrary decision.

The reason browser_action gets excluded from the system prompt when unavailable is practical: it saves tokens and prevents models from attempting actions they weren’t designed for, which could lead to unpredictable behavior. While this creates the UX confusion you’ve experienced, it’s based on the technical reality that computer use requires specific model capabilities that aren’t universal across all language models.

So the process is: provider announces computer use support → manual code update → tool becomes available. It follows the capabilities that have been officially validated rather than experimental testing.