[CodeCompanion] MCP call not working with Gemini models

[Eckii24]

Hi @ravitemer,

thanks again for your great work.
Its works like expected when using the Copilot models.

When changing to an gemini adapter, the model is not able to call every tool.

## Me

> Context:
> - <group>mcp</group>

get docs about dapr using context7 mcp server 
@{mcp}

##   Gemini (gemini-2.5-flash)

I can help you with that. I can use the `access_mcp_resource` tool to get resources from the `context7` MCP server. However, I need to know the specific URI for the Dapr documentation on that server. Could you please provide the URI?

## Me

> Context:
> - <group>mcp</group>

context7 has a tool to search for the url

##   Gemini (gemini-2.5-flash)

To search for the Dapr documentation URL on `context7`, I need to know the name of the tool and what input it requires. Could you please provide the name of the tool on `context7` that searches for URLs, and what parameters it needs to search for Dapr documentation?

## Me

> Context:
> - <group>mcp</group>

In debug view, I can see that all information about the tools are provided:

-- Adapter: "Gemini"
-- Buffer: 26
-- Context: "" (1)

local settings = {
  model = "gemini-2.5-flash",  -- "gemini-1.5-flash", "gemini-1.5-pro", "gemini-2.0-flash-lite", "gemini-2.0-flash", "gemini-2.5-flash-preview-05-20", "gemini-2.5-pro", 
  max_tokens = nil,
  temperature = nil,
  top_p = nil,
  reasoning_effort = "medium",
}

local messages = 
{ {
    content = "You are an AI programming assistant named \"CodeCompanion\". You are currently plugged into the Neovim text editor on a user's machine.\n\nYour core tasks include:\n- Answering general programming questions.\n- Explaining how the code in a Neovim buffer works.\n- Reviewing the selected code from a Neovim buffer.\n- Generating unit tests for the selected code.\n- Proposing fixes for problems in the selected code.\n- Scaffolding code for a new workspace.\n- Finding relevant code to the user's query.\n- Proposing fixes for test failures.\n- Answering questions about Neovim.\n- Running tools.\n\nYou must:\n- Follow the user's requirements carefully and to the letter.\n- Use the context and attachments the user provides.\n- Keep your answers short and impersonal, especially if the user's context is outside your core tasks.\n- Minimize additional prose unless clarification is needed.\n- Use Markdown formatting in your answers.\n- Include the programming language name at the start of each Markdown code block.\n- Do not include line numbers in code blocks.\n- Avoid wrapping the whole response in triple backticks.\n- Only return code that's directly relevant to the task at hand. You may omit code that isn’t necessary for the solution.\n- Avoid using H1, H2 or H3 headers in your responses as these are reserved for the user.\n- Use actual line breaks in your responses; only use \"\\n\" when you want a literal backslash followed by 'n'.\n- All non-code text responses must be written in the English language indicated.\n- Multiple, different tools can be called as part of the same response.\n\nWhen given a task:\n1. Think step-by-step and, unless the user requests otherwise or the task is very simple, describe your plan in detailed pseudocode.\n2. Output the final code in a single code block, ensuring that only relevant code is included.\n3. End your response with a short suggestion for the next user turn that directly supports continuing the conversation.\n4. Provide exactly one complete reply per conversation turn.\n5. If necessary, execute multiple tools in a single turn.",
    cycle = 1,
    id = 126646444,
    opts = {
      tag = "from_config",
      visible = false
    },
    role = "system"
  }, {
    content = "get docs about dapr using context7 mcp server \naccess_mcp_resource, use_mcp_tool",
    cycle = 1,
    id = 1388759219,
    opts = {
      visible = true
    },
    role = "user"
  }, {
    content = "# MCP SERVERS\n\nThe Model Context Protocol (MCP) enables communication between the system and locally running MCP servers that provide additional tools and resources to extend your capabilities.\n\n# Connected MCP Servers\n\nWhen a server is connected, you can use the server's tools via the `use_mcp_tool` tool, and access the server's resources via the `access_mcp_resource` tool.\nNote: Server names are case sensitive and you should always use the exact full name like `Firecrawl MCP` or `src/user/main/time-mcp` etc \n\n## context7\n\n### Available Tools\n\n- resolve-library-id: Resolves a package/product name to a Context7-compatible library ID and returns a list of matching libraries.\n\nYou MUST call this function before 'get-library-docs' to obtain a valid Context7-compatible library ID UNLESS the user explicitly provides a library ID in the format '/org/project' or '/org/project/version' in their query.\n\nSelection Process:\n1. Analyze the query to understand what library/package the user is looking for\n2. Return the most relevant match based on:\n- Name similarity to the query (exact matches prioritized)\n- Description relevance to the query's intent\n- Documentation coverage (prioritize libraries with higher Code Snippet counts)\n- Trust score (consider libraries with scores of 7-10 more authoritative)\n\nResponse Format:\n- Return the selected library ID in a clearly marked section\n- Provide a brief explanation for why this library was chosen\n- If multiple good matches exist, acknowledge this but proceed with the most relevant one\n- If no good matches exist, clearly state this and suggest query refinements\n\nFor ambiguous queries, request clarification before proceeding with a best-guess match.\n    Input Schema:\n    {\n      [\"$schema\"] = \"http://json-schema.org/draft-07/schema#\",\n      additionalProperties = false,\n      properties = {\n        libraryName = {\n          description = \"Library name to search for and retrieve a Context7-compatible library ID.\",\n          type = \"string\"\n        }\n      },\n      required = { \"libraryName\" },\n      type = \"object\"\n    }\n\n- get-library-docs: Fetches up-to-date documentation for a library. You must call 'resolve-library-id' first to obtain the exact Context7-compatible library ID required to use this tool, UNLESS the user explicitly provides a library ID in the format '/org/project' or '/org/project/version' in their query.\n    Input Schema:\n    {\n      [\"$schema\"] = \"http://json-schema.org/draft-07/schema#\",\n      additionalProperties = false,\n      properties = {\n        context7CompatibleLibraryID = {\n          description = \"Exact Context7-compatible library ID (e.g., '/mongodb/docs', '/vercel/next.js', '/supabase/supabase', '/vercel/next.js/v14.3.0-canary.87') retrieved from 'resolve-library-id' or directly from user query in the format '/org/project' or '/org/project/version'.\",\n          type = \"string\"\n        },\n        tokens = {\n          description = \"Maximum number of tokens of documentation to retrieve (default: 10000). Higher values provide more context but consume more tokens.\",\n          type = \"number\"\n        },\n        topic = {\n          description = \"Topic to focus documentation on (e.g., 'hooks', 'routing').\",\n          type = \"string\"\n        }\n      },\n      required = { \"context7CompatibleLibraryID\" },\n      type = \"object\"\n    }\n\n## mcphub\nMCPHub server provides tools and resources to manage the mcphub.nvim neovim plugin. It has tools to toggle any MCP Server along with resources like docs, guides.\n\n### Available Tools\n\n- get_current_servers: Get the current state of all MCP servers (connected and disabled). This is useful when you need to know what servers are currently available, especially when restoring chat from history or when server state may have changed.\n    Input Schema:\n    {\n      properties = {\n        format = {\n          default = \"detailed\",\n          description = \"Response format: 'detailed' for full server info or 'summary' for compact list (default: detailed)\",\n          enum = { \"detailed\", \"summary\" },\n          type = \"string\"\n        },\n        include_disabled = {\n          default = true,\n          description = \"Whether to include disabled servers in the response (default: true)\",\n          type = \"boolean\"\n        }\n      },\n      type = \"object\"\n    }\n\n- toggle_mcp_server: Start or stop an MCP server. You can only start a server from one of the disabled servers.\n    Input Schema:\n    {\n      properties = {\n        action = {\n          description = \"Action to perform. One of 'start' or 'stop'\",\n          enum = { \"start\", \"stop\" },\n          type = \"string\"\n        },\n        server_name = {\n          description = \"Name of the MCP server to toggle\",\n          type = \"string\"\n        }\n      },\n      required = { \"server_name\", \"action\" },\n      type = \"object\"\n    }\n\n### Available Resources\n\n- mcphub://docs (text/plain)\n  MCPHub Plugin Docs\n  Documentation for the mcphub.nvim plugin for Neovim.\n\n- mcphub://native_server_guide (text/plain)\n  MCPHub Native Server Guide\n  Documentation on how to create Lua Native MCP servers for mcphub.nvim plugin.\nThis guide is intended for Large language models to help users create their own native servers for mcphub.nvim plugin.\nAccess this guide whenever you need information on how to create a native server for mcphub.nvim plugin.\n\n- mcphub://changelog (text/plain)\n  MCPHub Changelog\n  Changelog for the mcphub.nvim plugin for Neovim.\n\n# Disabled MCP Servers\n\nWhen a server is disabled, it will not be able to provide tools or resources. You can start one of the following disabled servers by using the `toggle_mcp_server` tool on `mcphub` MCP Server if it is connected using `use_mcp_tool`\n\n## repomix (Disabled)\n\n## ado (Disabled)\n\n## vectorcode (Disabled)\n\n## sequential-thinking (Disabled)\n\n## mcp-promptregistry (Disabled)\n\n## git (Disabled)\n\n## deepwiki (Disabled)\n\n## neovim (Disabled)\nNeovim MCP server provides a set of tools and resources that integrate with neovim.\n\n\n\n# Examples: \n\n## `use_mcp_tool`\n\nWhen you need to call a tool on an MCP Server, use the `use_mcp_tool` tool:\n\nPseudocode:\n\nuse_mcp_tool\n  server_name: string (One of the available server names)\n  tool_name: string (name of the tool in the server to call)\n  tool_input: object (Arguments for the tool call)\n\n## `access_mcp_resource`\n\nWhen you need to access a resource from a MCP Server, use the `access_mcp_resource` tool:\n\nPseudocode:\n\naccess_mcp_resource\n  server_name: string (One of the available server names)\n  uri: string (uri for the resource)\n\n## Toggling a MCP Server\n\nWhen you need to start a disabled MCP Server or vice-versa, use the `toggle_mcp_server` tool on `mcphub` MCP Server using `use_mcp_tool`:\n\nCRITICAL: You need to use the `use_mcp_tool` tool to call the `toggle_mcp_server` tool on `mcphub` MCP Server when `mcphub` server is \"Connected\" else ask the user to enable `mcphub` server.\n\nPseudocode:\n\nuse_mcp_tool\n  server_name: \"mcphub\"\n  tool_name: \"toggle_mcp_server\"\n  tool_input: \n    server_name: string (One of the available server names to start or stop)\n    action: string (one of `start` or `stop`)\n",
    cycle = 1,
    id = 550695658,
    opts = {
      reference = "<group>mcp</group>",
      tag = "tool",
      visible = false
    },
    role = "system"
  }, {
    content = "You can use the access_mcp_resource tool to get resources on MCP servers.\n",
    cycle = 1,
    id = 2044923708,
    opts = {
      reference = "<tool>access_mcp_resource</tool>",
      tag = "tool",
      visible = false
    },
    role = "system"
  }, {
    content = "You can use the use_mcp_tool tool to calls tools on MCP servers.\n",
    cycle = 1,
    id = 356754939,
    opts = {
      reference = "<tool>use_mcp_tool</tool>",
      tag = "tool",
      visible = false
    },
    role = "system"
  }, {
    content = "I can help you with that. I can use the `access_mcp_resource` tool to get resources from the `context7` MCP server. However, I need to know the specific URI for the Dapr documentation on that server. Could you please provide the URI?",
    cycle = 1,
    id = 547156487,
    opts = {
      visible = true
    },
    role = "llm"
  }, {
    content = "context7 has a tool to search for the url",
    cycle = 2,
    id = 361229934,
    opts = {
      visible = true
    },
    role = "user"
  }, {
    content = "To search for the Dapr documentation URL on `context7`, I need to know the name of the tool and what input it requires. Could you please provide the name of the tool on `context7` that searches for URLs, and what parameters it needs to search for Dapr documentation?",
    cycle = 2,
    id = 860933072,
    opts = {
      visible = true
    },
    role = "llm"
  } }

It looks like Gemini has no access to the system prompt, which explains the tools in detail.
When I paste the content of this system prompt as additional user prompt, it is working again.

Any idea, why this is the case and how to fix it?

[Eckii24]

Maybe the reason is the triple system message, also mentioned here: olimorris/codecompanion.nvim#1801

[ravitemer]

Thank you for raising this @Eckii24 . I can replicate this.

I'm not sure but it might be due to the order of system messages in the messages table. I will try to debug this more in some free time. Safe to say nothing to do from mcphub.nvim end but I'm also facing some issues with function tools with gemini adapter at codecompanion.nvim while sending tool results back to gemini.