[BUG] Url mentions don't seem to be functional

[eliannelavoie]

Problem (one or two sentences)

Pasting a URL to fetch it's content in a @ mention doesn't work. The model is not given the contents of the page in the API request.

Context (who is affected and when)

Happens to users trying to give a web page as context to their agent

Reproduction steps

1. VS Code Extension, Windows 11
2. Start a new task
3. Write a prompt with a URL mention

My exact test prompt is:

@https://docs.roocode.com/basic-usage/context-mentions 

What's in the above page?

Expected result

The model receives the markdownified contents of the page and gives a brief summary

Actual result

The model only sees the url and attempts to read the page with an external tool (ie. curl)

Variations tried (optional)

No response

App Version

3.52.0

API Provider (optional)

None

Model Used (optional)

No response

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

I used the "Enable debug mode" toggle to get the API History of my test conversation.

We can see the the first "user" message doesn't send the contents of the page to the API. I would expect there to be an additional text content object showing the contents of the mention, just like what happens when we mention a file.


[
  {
    "role": "user",
    "content": [
      {
        "type": "text",
        "text": "<user_message>\n'https://docs.roocode.com/basic-usage/context-mentions' \n\nWhat's in the above page?\n</user_message>"
      },
      {
        "type": "text",
        "text": "<environment_details>Cleared for privacy...</environment_details>"
      }
    ],
    "ts": 1775726555085
  },
  {
    "role": "assistant",
    "content": [
      {
        "type": "tool_use",
        "id": "tool_execute_command_3861XvsLetyeUFr9efsh",
        "name": "execute_command",
        "input": {
          "command": "curl -L https://docs.roocode.com/basic-usage/context-mentions"
        }
      }
    ],
    "ts": 1775726561154,
    "reasoning_details": [
      {
        "type": "reasoning.encrypted",
        "text": "The user wants to know the content of the page `https://docs.roocode.com/basic-usage/context-mentions`.\nI don't have a tool to directly browse the web, but I can use `execute_command` to run a command that fetches the content of the URL, such as `curl` or `powershell`'s `Invoke-WebRequest`.\n\nSince I am on Windows 11, I can use `curl` (which is available in cmd.exe) or PowerShell.\n\nPlan:\n1. Use `execute_command` to fetch the content of the URL.\n2. Analyze the fetched content and summarize it for the user.",
        "id": "tool_execute_command_3861XvsLetyeUFr9efsh",
        "format": "google-gemini-v1",
        "index": 0
      },
      {
        "type": "reasoning.encrypted",
        "data": "EiYKJGUyNDgzMGE3LTVjZDYtNDJmZS05OThiLWVlNTM5ZTcyYjljMw==",
        "id": "tool_execute_command_3861XvsLetyeUFr9efsh",
        "format": "google-gemini-v1",
        "index": 0
      }
    ]
  },
  {
    "role": "user",
    "content": [
      {
        "type": "tool_result",
        "tool_use_id": "tool_execute_command_3861XvsLetyeUFr9efsh",
        "content": "{\"status\":\"error\",\"message\":\"The tool execution failed\",\"error\":\"Tool \\\"execute_command\\\" is not allowed in architect mode.\"}",
        "is_error": true
      },
      {
        "type": "text",
        "text": "<environment_details>Cleared for privacy...</environment_details>"
      }
    ],
    "ts": 1775726561505
  },
  {
    "role": "assistant",
    "content": [
      {
        "type": "tool_use",
        "id": "tool_switch_mode_LGRJECozuNxG5sGrdv1F",
        "name": "switch_mode",
        "input": {
          "mode_slug": "code",
          "reason": "I need to use `execute_command` to fetch the content of the provided URL using `curl`, and this tool is not available in Architect mode."
        }
      }
    ],
    "ts": 1775726584661,
    "reasoning_details": [
      {
        "type": "reasoning.encrypted",
        "text": "The user wants to know the content of the page `https://docs.roocode.com/basic-usage/context-mentions`.\nI tried to use `execute_command` with `curl`, but it failed because `execute_command` is not allowed in `architect` mode.\n\nI need to find a way to get the content of that URL.\nSince I am in `architect` mode and cannot use `execute_command`, I should check if there are other tools available.\nWait, the system prompt says: \"Some modes have restrictions on which files they can edit.\" but it doesn't explicitly say `execute_command` is restricted in `architect` mode in the *general* rules, but the tool execution error clearly stated: `Tool \"execute_command\" is not allowed in architect mode.`\n\nI need to switch to a mode that allows `execute_command` or has a tool to fetch web content.\nActually, `code` mode or `ask` mode might allow it, but usually, `execute_command` is for system operations.\nWait, if I can't use `execute_command` in `architect` mode, I should switch to a mode that can.\nHowever, the user just asked \"What's in the above page?\". This is an informational request.\n`ask` mode is for explanations, documentation, or answers to technical questions.\nLet's try switching to `ask` mode or `code` mode.\nActually, `code` mode is usually more permissive with tools.\n\nBut wait, I can just use `switch_mode` to `ask` or `code`.\nLet's try `ask` mode first as it's more aligned with \"What's in the above page?\".\nWait, if `execute_command` is restricted in `architect`, is it restricted in `ask`?\nThe prompt says: \"You have access to tools that let you execute CLI commands on the user's computer...\".\nIf I switch to `code` mode, I can probably use `curl`.\n\nActually, I should check if I can use `execute_command` in `ask` mode.\nBut the most direct way to get the content of a URL if I don't have a `fetch` tool is `curl` via `execute_command`.\n\nLet's try switching to `code` mode to use `curl`.\n\nWait, I can also try to see if there's any other way.\nThe user is asking about Roo Code documentation.\nMaybe I can just switch to `ask` mode and see if it has better capabilities or if I can use `execute_command` there.\nActually, `code` mode is the safest bet for tool access.\n\nWait, I'll try to switch to `ask` mode first because it's a question. If that fails, I'll go to `code`.\nActually, let's just go to `code` mode to be sure I can run the command.\n\nWait, I should explain why I'm switching.\n\"I need to fetch the content of the URL, and `execute_command` is not available in Architect mode. I will switch to Code mode to use `curl` and retrieve the page content.\"\n\nActually, let's try `ask` mode first.\nNo, `code` mode is definitely allowed to run commands.\n\nWait, I can just use `switch_mode` to `code`.",
        "id": "tool_switch_mode_LGRJECozuNxG5sGrdv1F",
        "format": "google-gemini-v1",
        "index": 0
      },
      {
        "type": "reasoning.encrypted",
        "data": "EiYKJGUyNDgzMGE3LTVjZDYtNDJmZS05OThiLWVlNTM5ZTcyYjljMw==",
        "id": "tool_switch_mode_LGRJECozuNxG5sGrdv1F",
        "format": "google-gemini-v1",
        "index": 0
      }
    ]
  }
]

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

Root cause: When the browser-use feature was removed (PR #11392), URL content fetching for @url mentions was also lost. The parseMentions() function in src/core/mentions/index.ts currently replaces URL mentions with just a quoted reference (e.g. 'https://example.com') but never actually fetches and includes the page content. File mentions, git mentions, terminal mentions, etc. all populate content blocks or append content, but URLs are skipped entirely.

Proposed fix (3 files changed, plus tests):

1. Create src/core/mentions/fetchUrlContent.ts - A new utility that fetches URL content using axios (already a dependency) and extracts readable text from HTML using cheerio (also already a dependency). It will:

   • Use axios with a 15s timeout, 5 redirect limit, and 5MB max response size
   • For HTML responses: use cheerio to strip scripts/styles/nav elements and extract main content text
   • For non-HTML responses: use raw text directly
   • Truncate to 50k characters max to avoid context window overflow

2. Modify src/core/mentions/index.ts - Two changes in parseMentions():

   • Update the URL text replacement (line ~165) from '${mention}' to '${mention}' (see below for fetched content) so the model knows content follows
   • Add a new else if (mention.startsWith("http")) block (after git commit handling, ~line 222) that calls fetchUrlContent() and pushes a MentionContentBlock with type: "url" containing the fetched content

3. Add tests:

   • Unit tests for fetchUrlContent (mocking axios)
   • Integration tests in the existing parseMentions test file for URL mention handling

This approach reuses existing dependencies and follows the same pattern used by file/folder mentions (separate content blocks). It's essentially the same fix that was developed in commit e29c073 for a related issue (#11856) but never made it to main.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.