[BUG] Invalid schema in CodeCompanion

[blankenshipz]

I sporadically get this bug when adding vectorcode to the context, not exactly sure if this is a CodeCompanion issue or a vectorcode issue. I'll add more details if I can nail down a way to reproduce

[ERROR] 2025-05-13 16:12:50
Error: {
  "error": {
    "message": "Invalid schema for function 'vectorcode': In context=('properties', 'options'), 'required' is required to be supplied and to be an array including every key in properties. Missing 'count'.",
    "type": "invalid_request_error",
    "param": "tools[2].function.parameters",
    "code": "invalid_function_parameters"
  }
}

[Davidyz]

Looks like this is caused by the model failing to generate the correct JSON object for the tool call arguments. Next time you encounter this, could you press gd in the codecompanion chat buffer to bring up the debug window and drop the content here? This'll include the actual tool arguments generated by the model and will help me locate the error.

[Davidyz]

Also, when I was testing the function calling tools with various LLM providers and models I noticed that some models simply fails to produce a correct JSON (some produced broken JSON with missing parentheses, some didn't follow the schema). I suspect that this is due to the providers not using the correct chat template (some LLMs don't use JSON for openai-compatible tool calling under the hood, and the providers have to do the conversion for the API).

[blankenshipz]

@Davidyz It seems like when this happens it kind of breaks codecompanion, the debug buffer won't open. I'll see if I can get some more debug info from neovim

[Davidyz]

Ok I see. Could you lmk what adapter/model you're using? It's possible that it's not happy with the schema I wrote (I've only tested with a handful of openai-compatible providers and Gemini, and other providers may have their own tool calling implementation).

[blankenshipz]

This particular example was openai with gpt-4.1

Thanks for looking at this!

[ahakanbaba]

I am getting this error as well.

Here is the output of gd

-- Adapter: "OpenAI"
-- Buffer: 9
-- Context: "install_tensorrt.sh" (2)

local settings = {
  model = "o3",  -- "gpt-4.1", "gpt-4o", "gpt-4o-mini", "gpt-4-turbo-preview", "gpt-4", "gpt-3.5-turbo", "o3-mini-2025-01-31", "o4-mini-2025-04-16", "o1-2024-12-17", "o3-2025-04-16", 
  temperature = 1,
  top_p = 1,
  stop = nil,
  max_tokens = nil,
  presence_penalty = 0,
  frequency_penalty = 0,
  logit_bias = nil,
  user = nil,
}

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- 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- Avoid including 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 = 1191187114,
    opts = {
      tag = "from_config",
      visible = false
    },
    role = "system"
  }, {
    content = "vectorcode \n\nHow to test tensrorrt ?",
    cycle = 1,
    id = 1217350477,
    opts = {
      visible = true
    },
    role = "user"
  }, {
    content = "### VectorCode, a repository indexing and query tool.\n\n1. **Purpose**: This gives you the ability to access the repository to find information that you may need to assist the user.\n\n2. **Key Points**:\n  - The path of a retrieved file will be wrapped in `<path>` and `</path>` tags. Its content will be right after the `</path>` tag, wrapped by `<content>` and `</content>` tags. Do not include the `<path>``</path>` tags in your answers when you mention the paths.\n  - If you used the tool, tell users that they may need to wait for the results and there will be a virtual text indicator showing the tool is still running\n  - Include one single command call for VectorCode each time. You may include multiple keywords in the command\n  - VectorCode is the name of this tool. Do not include it in the query unless the user explicitly asks\n  - Use the `ls` command to retrieve a list of indexed project and pick one that may be relevant, unless the user explicitly mentioned 'this project' (or in other equivalent expressions)\n  - **The project root option MUST be a valid path on the filesystem. It can only be one of the results from the `ls` command or from user input**\n\n  - If the user did not specify how many documents to retrieve, **start with 10 documents**\n  - If you decide to call VectorCode tool, do not start answering the question until you have the results. Provide answers based on the results and let the user decide whether to run the tool again\n  - **Use at your discretion** when you feel you don't have enough information about the repository or project\n  - **Don't escape** special characters\n  - separate phrases into distinct keywords when appropriate\n  - If a class, type or function has been imported from another file, this tool may be able to find its source. Add the name of the imported symbol to the query\n  - Avoid retrieving one single file because the retrieval mechanism may not be very accurate\n  - When providing answers based on VectorCode results, try to give references such as paths to files and line ranges, unless you're told otherwise (but do not include the full source code context)\n  - VectorCode is the name of this tool. Do not include it in the query unless the user explicitly asks\n  - If the retrieval results do not contain the needed context, increase the file count so that the result will more likely contain the desired files\n  - If the returned paths are relative, they are relative to the root of the project directory\n  - Do not suggest edits to retrieved files that are outside of the current working directory, unless the user instructed otherwise\n  - When specifying the `project_root` parameter when making a query, make sure you run the `ls` tool first to retrieve a list of valid, indexed projects\n  - If a query failed to retrieve desired results, a new attempt should use different keywords that are orthogonal to the previous ones but with similar meanings\n  - Do not use exact query keywords that you have used in a previous tool call in the conversation, unless the user instructed otherwise, or with different count/project_root\n  - The current working directory is /home/coder/src/corp/pi-gpucompute. Assume the user query is about this project, unless the user asked otherwise or queries from the current project fails to return useful results. \n",
    cycle = 1,
    id = 194284111,
    opts = {
      reference = "<tool>vectorcode</tool>",
      tag = "tool",
      visible = false
    },
    role = "system"
  } }

[ahakanbaba]

I was actually trying to see how the code-companion calls the OpenAI LLM.

Looking at the gd output, I do not see other parameters in the call body.
Ideally the tool's schema should be send via the tools field in the request body

tools
array

Optional
A list of tools the model may call. Currently, only functions are supported as a tool. Use this to provide a list of functions the model may generate JSON inputs for. A max of 128 functions are supported.


Hide properties
function
object

Required

Show properties
type
string

Required
The type of the tool. Currently, only function is supported.

Looking at the gd output, I see a detailed description of the tool in the system prompt only.

[ahakanbaba]

Sometimes I also get this error

[ERROR] 2025-05-14 15:52:27
Error: {
  "error": {
    "message": "Invalid schema for function 'vectorcode': In context=('properties', 'options'), 'required' is required to be supplied and to be an array including every key in properties. Missing 'project_root'.",
    "type": "invalid_request_error",
    "param": "tools[0].function.parameters",
    "code": "invalid_function_parameters"
  }
}

With project root.
I should have the version that has the project_root related fix.

cat nvim/lazy-lock.json | grep -i vectorcode
  "VectorCode": { "branch": "main", "commit": "0be48d75aff47c9d29d448766eacb946844bc718" },

[ahakanbaba]

Enabling debug mode in code-companion lets me see the full request body to openai.
It looks like this

{
  "stream_options": {
    "include_usage": true
  },
  "messages": [
    {
      "role": "system",
      "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- 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- Avoid including 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."
    },
    {
      "role": "user",
      "content": "vectorcode \nHow to test tensorrt ?"
    },
    {
      "role": "system",
      "content": "### VectorCode, a repository indexing and query tool.\n\n1. **Purpose**: This gives you the ability to access the repository to find information that you may need to assist the user.\n\n2. **Key Points**:\n  - The path of a retrieved file will be wrapped in `<path>` and `</path>` tags. Its content will be right after the `</path>` tag, wrapped by `<content>` and `</content>` tags. Do not include the `<path>``</path>` tags in your answers when you mention the paths.\n  - If you used the tool, tell users that they may need to wait for the results and there will be a virtual text indicator showing the tool is still running\n  - Include one single command call for VectorCode each time. You may include multiple keywords in the command\n  - VectorCode is the name of this tool. Do not include it in the query unless the user explicitly asks\n  - Use the `ls` command to retrieve a list of indexed project and pick one that may be relevant, unless the user explicitly mentioned 'this project' (or in other equivalent expressions)\n  - **The project root option MUST be a valid path on the filesystem. It can only be one of the results from the `ls` command or from user input**\n\n  - If the user did not specify how many documents to retrieve, **start with 10 documents**\n  - If you decide to call VectorCode tool, do not start answering the question until you have the results. Provide answers based on the results and let the user decide whether to run the tool again\n  - **Use at your discretion** when you feel you don't have enough information about the repository or project\n  - **Don't escape** special characters\n  - separate phrases into distinct keywords when appropriate\n  - If a class, type or function has been imported from another file, this tool may be able to find its source. Add the name of the imported symbol to the query\n  - Avoid retrieving one single file because the retrieval mechanism may not be very accurate\n  - When providing answers based on VectorCode results, try to give references such as paths to files and line ranges, unless you're told otherwise (but do not include the full source code context)\n  - VectorCode is the name of this tool. Do not include it in the query unless the user explicitly asks\n  - If the retrieval results do not contain the needed context, increase the file count so that the result will more likely contain the desired files\n  - If the returned paths are relative, they are relative to the root of the project directory\n  - Do not suggest edits to retrieved files that are outside of the current working directory, unless the user instructed otherwise\n  - When specifying the `project_root` parameter when making a query, make sure you run the `ls` tool first to retrieve a list of valid, indexed projects\n  - If a query failed to retrieve desired results, a new attempt should use different keywords that are orthogonal to the previous ones but with similar meanings\n  - Do not use exact query keywords that you have used in a previous tool call in the conversation, unless the user instructed otherwise, or with different count/project_root\n  - The current working directory is /home/coder/src/corp/pi-gpucompute. Assume the user query is about this project, unless the user asked otherwise or queries from the current project fails to return useful results. \n"
    }
  ],
  "model": "o3",
  "frequency_penalty": 0,
  "presence_penalty": 0,
  "tools": [
    {
      "function": {
        "name": "vectorcode",
        "description": "Retrieves code documents using semantic search or lists indexed projects",
        "strict": true,
        "parameters": {
          "type": "object",
          "additionalProperties": false,
          "required": [
            "command"
          ],
          "properties": {
            "command": {
              "enum": [
                "query",
                "ls"
              ],
              "type": "string",
              "description": "Action to perform: 'query' for semantic search or 'ls' to list projects"
            },
            "options": {
              "type": "object",
              "additionalProperties": false,
              "required": [
                "query"
              ],
              "properties": {
                "project_root": {
                  "type": "string",
                  "description": "Project path to search within (must be from 'ls' results)"
                },
                "count": {
                  "type": "integer",
                  "description": "Number of documents to retrieve, must be positive"
                },
                "query": {
                  "items": {
                    "type": "string"
                  },
                  "description": "Query messages used for the search.",
                  "type": "array"
                }
              }
            }
          }
        }
      },
      "type": "function"
    }
  ],
  "top_p": 1,
  "temperature": 1,
  "stream": true
}

[ahakanbaba]

After enabling trace logging in code-companion, I see this response which looks very revelant.

Output data:
    "message": "Invalid schema for function 'vectorcode': In context=('properties', 'options'), 'required' is required to be supplied and to be an array including every key in properties. Missing 'project_root'.",

[ahakanbaba]

Hah !
It turns out all fieds should be required
https://community.openai.com/t/api-rejects-valid-json-schema/906163/2

Please let me know if there is anything else I can help with.

[Davidyz]

It turns out all fieds should be required

Thanks for spotting this. I'll open a PR for the fix so that you can give it a try.

[Davidyz]

See #141