gemini 3 support

[bitromortac]

When naively trying to use the gemini-3-pro-preview model it requires a thought_signature, see https://ai.google.dev/gemini-api/docs/thought-signatures. It would be great if that could be added to the plugin, I'd be happy to test out any changes.

[ERROR] 2025-11-19 09:22:23
Error: [{
  "error": {
    "code": 400,
    "message": "Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call `default
_api:read_file` , position 8. please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",
    "status": "INVALID_ARGUMENT"
  }
}
]

[olimorris]

Sigh...thanks for sharing this. The sooner we can move to a standardized protocol for LLM comms, the better.

Can you give me more background to this? What your prompt was etc? I can see that gemini-3 works with Copilot for conversation but not function calling

[bitromortac]

Right, the protocol divergence is annoying 😢.

An example prompt that leads to the error is: Can you see files @{file_search}?

[olimorris]

Thanks for the doc links too. I'll prioritize getting support added in over the next day or two

[bitromortac]

Thank you, feel free to ping if you need a tester

[JPFrancoia]

Very interested in this too, happy to test.

[olimorris]

This has now been added in #2411 👍🏼

[bitromortac]

Awesome, thanks a lot! 🚀 It seems to work, will test it out more 🙏

[JPFrancoia]

I know it's a slight tangent, but in the past I was able to basically copy-paste the gemini adapter to make a vertex adapter(here). Sure there are a couple of small changes to make for the endpoint and the env vars, but nothing too tricky.

I did the same after your latest changes, I sent a request the usual way, but codecompanion doesn't display the response from the model. There is no error in the UI (if I use an endpoint that doesn't exist, I do get a bit red error, unmissable). Here I'm not getting any error, see screenshot.

I checked the logs, and this is what I got:

[DEBUG] 2025-11-21 10:08:00
Output data:
data: {"choices":[{"delta":{"refusal":"Malformed function call: call:memory:search{query:\"user name preferences\"}","role":"assistant"},"finish_reason":"malformed_function_call","index":0,"logprobs":null}],"created":1763719667,"id":"8zkgafWKBpvotfAP5oT4OA","model":"google/gemini-3-pro-preview","object":"chat.completion.chunk","system_fingerprint":"","usage":{"extra_properties":{"google":{"traffic_type":"ON_DEMAND"}},"prompt_tokens":812,"total_tokens":812}}
[DEBUG] 2025-11-21 10:08:00
Output data:
data: [DONE]
[INFO] 2025-11-21 10:08:00
Chat request finished

I have no tool enabled for this session. I even disabled the tools in the vertex adapter (in the opts map). If I change the model to gemini-2.5-pro, it works and I get a response.

Would you have any idea about what's happening? If you could point me in the right direction I could try and fix the problem myself. Right now I'm not sure the problem is caused by codecompanion, I'd be tempted to say the error is from the API.

[olimorris]

Can you share the JSON body that was sent as part of the request? CodeCompanion logs the filepath and it will be called something like 0.json.

Your streamed logs imply a tool was sent but that seems strange as it's not something you can do accidentally.

[JPFrancoia]

Ah 🤦 this is so obvious looking at the request:

{
    "model": "google/gemini-3-pro-preview",
    "stream": true,
    "messages": [
        {
            "role": "system",
            "content": "You are an AI programming assistant named \"CodeCompanion\". You are currently plugged in to the Neovim text editor on a user's machine.\nYou may have access to a memory tool, which provides persistent storage for user-specific information such as preferences, past interactions, and relevant code snippets.  \nTreat this memory as a long-term context to improve the usefulness and relevance of your responses.\n\n## Your role\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 in 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- Helping the user increase their knowledge of programming.\n- Build a rich understanding of the user's programming environment, goals, and preferences over time.\n\n## Memory Usage\n\nYou must:\n- ALWAYS use the `@memory` tools for all memory and context retrieval.\n- Treat memory search as a DEFAUL STEP in your reasoning \u2014 never wait for the user to ask.\n- Use the `@memory` tools to SEARCH, RETRIEVE and INCORPORATE relevant context from prior interactions.\n\n### When to Retrieve Memory\nRetrieve and apply relevant memory context whenever:\n- The user asks about their codebase, environment, or workflow.\n- The user encounters bugs, errors, or recurring issues.\n- You are asked to extend or modify prior solutions.\n- Proposing new ideas, tools, or approaches where alignment with past context matters.\n\n### How to Use Memory\n- Perform a TARGETED SEARCH with `@memory` whenever new queries could benefit from prior knowledge.\n- INCORPORATE retrieved context directly into your reasoning and solutions.\n- Do not repeat irrelevant details \u2014 summarize or reframe memory as needed.\n- If no relevant memory is found, STATE THAT EXPLICITLY and proceed with fresh reasoning.\n\nONLY use the @memory tools for all memory operations. DO NOT confuse it with the @basic_memory tool. DO NOT use any other memory tool.\n\n## General Behavior\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- 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\u2019t 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 %s language indicated.\n- Multiple, different tools can be called as part of the same response.\n\nWhen given a task:\n1. Proactively use memory to build a contextual understanding of the user's goals, environment, and preferences.\n2. Think step-by-step and, unless the user requests otherwise or the task is very simple, describe your plan in detailed pseudocode.\n3. Output the final code in a single code block, ensuring that only relevant code is included.\n4. End your response with a short suggestion for the next user turn that directly supports continuing the conversation.\n5. Provide exactly one complete reply per conversation turn."
        },
        {
            "role": "user",
            "content": "Hello"
        }
    ],
    "stream_options": {
        "include_usage": true
    },
    "reasoning_effort": "medium"
}

I do mention a memory tool in my custom system prompt. I had disabled it at work (where I'm forced to use the gemini models) because gemini-2.5-pro wasn't able to use it properly. So in fact, disabling the memory tool made gemini-3 fail.

I re-enabled the memory tool, and now it works.

So it was really my fault, sorry about that.

On a positive note, I can confirm that the vertex adapter still works: I copy-pasted the gemini adapter, changed the endpoint/env vars (same changes as in the previous post about the vertex adapter), and it works perfectly.

Thank you very much for your work and help!