[BUG] Gemini 3 on OpenRouter API tool result error - All Tools Broken

[mrkmg]

Problem (one or two sentences)

When using OpenRouter and Gemini 3, anytime roo tries to use a tool, it fails with a 400 Provider Error.

This leaves Gemini completely broken.

Context (who is affected and when)

Using Gemini 3 pro preview on openrouter, no tool calls work

Reproduction steps

1. Use Open Router and Gemini 3 Pro Preview
2. Try to do anything

Expected result

No error

Actual result

Errors

Variations tried (optional)

I don't have access to other providers

App Version

3.37.0

API Provider (optional)

OpenRouter

Model Used (optional)

Gemini 3 Pro Preview

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

400
OpenRouter completion error: Gemini models require OpenRouter reasoning details to be preserved in each request. Please refer to our docs: https://openrouter.ai/docs/guides/best-practices/reasoning-tokens#preserving-reasoning-blocks. Upstream 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:new_task` , position 2. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",
    "status": "INVALID_ARGUMENT"
  }
}



Date/time: 2025-12-23T13:22:49.746Z
Extension version: 3.37.0
Provider: openrouter
Model: google/gemini-3-pro-preview

400
OpenRouter completion error: Gemini models require OpenRouter reasoning details to be preserved in each request. Please refer to our docs: https://openrouter.ai/docs/guides/best-practices/reasoning-tokens#preserving-reasoning-blocks. Upstream 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:update_todo_list` , position 2. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",
    "status": "INVALID_ARGUMENT"
  }
}



// Please share this file with Roo Code Support (support@roocode.com) to diagnose the issue faster
// Just make sure you're OK sharing the contents of the conversation below.

{
  "error": {
    "timestamp": "2025-12-23T13:16:07.592Z",
    "version": "3.37.0",
    "provider": "openrouter",
    "model": "google/gemini-3-pro-preview",
    "details": "400\nOpenRouter completion error: Gemini models require OpenRouter reasoning details to be preserved in each request. Please refer to our docs: https://openrouter.ai/docs/guides/best-practices/reasoning-tokens#preserving-reasoning-blocks. Upstream error: {\n  \"error\": {\n    \"code\": 400,\n    \"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:new_task` , position 2. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.\",\n    \"status\": \"INVALID_ARGUMENT\"\n  }\n}"
  },
  "history": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "REMOVED"
        },
        {
          "type": "text",
          "text": "REMOVED"
        }
      ],
      "ts": 1766495649479
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "tool_use",
          "id": "tool_new_task_k7LktBScQZtG5uIJYm6g",
          "name": "new_task",
          "input": {
            "mode": "code",
            "message": "Update overview-channel.ts to support multiple filters",
            "todos": "[ ] Update `checkAndUpdateOverviews` to include filters in the query\n[ ] Update `checkIfUpdateNeeded` to use multiple filters logic\n[ ] Update `fetchFilteredTickets` to use multiple filters logic\n[ ] Update `buildOverviewEmbeds` to handle titles for multiple filters"
          }
        }
      ],
      "ts": 1766495688590,
      "reasoning_details": [
        {
          "type": "reasoning.text",
          "text": "REMOVED",
          "format": "google-gemini-v1",
          "index": 0
        },
        {
          "type": "reasoning.encrypted",
          "data": "REMOVED",
          "id": "tool_new_task_k7LktBScQZtG5uIJYm6g",
          "format": "google-gemini-v1",
          "index": 0
        }
      ]
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tool_new_task_k7LktBScQZtG5uIJYm6g",
          "content": "{\"status\":\"denied\",\"message\":\"The user denied this operation and provided the following feedback\",\"feedback\":\"No need to to use subtasks\"}"
        },
        {
          "type": "text",
          "text": "<environment_details>\n# VSCode Visible Files\nbot\\src\\services\\overview-channel.ts\n\n# VSCode Open Tabs\nbot/src/services/overview-channel.ts,bot/src/commands/overview.ts,shared/prisma/schema.prisma\n\n# Current Time\nCurrent time in ISO 8601 UTC format: 2025-12-23T13:15:04.300Z\nUser time zone: America/New_York, UTC-5:00\n\n# Current Cost\n$0.12\n\n# Current Mode\n<slug>code</slug>\n<name>💻 Code</name>\n<model>google/gemini-3-pro-preview</model>\n<tool_format>native</tool_format>\nYou have not created a todo list yet. Create one with `update_todo_list` if your task is complicated or involves multiple steps.\n</environment_details>"
        }
      ],
      "ts": 1766495704303
    }
  ]
}



// Please share this file with Roo Code Support (support@roocode.com) to diagnose the issue faster
// Just make sure you're OK sharing the contents of the conversation below.

{
  "error": {
    "timestamp": "2025-12-23T13:22:58.105Z",
    "version": "3.37.0",
    "provider": "openrouter",
    "model": "google/gemini-3-pro-preview",
    "details": "400\nOpenRouter completion error: Gemini models require OpenRouter reasoning details to be preserved in each request. Please refer to our docs: https://openrouter.ai/docs/guides/best-practices/reasoning-tokens#preserving-reasoning-blocks. Upstream error: {\n  \"error\": {\n    \"code\": 400,\n    \"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:update_todo_list` , position 2. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.\",\n    \"status\": \"INVALID_ARGUMENT\"\n  }\n}"
  },
  "history": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "REMOVED"
        },
        {
          "type": "text",
          "text": "REMOVED"
        }
      ],
      "ts": 1766496074639
    },
    {
      "role": "assistant",
      "content": [
        {
          "type": "tool_use",
          "id": "tool_update_todo_list_mL716LwhunURIXx4ZUZf",
          "name": "update_todo_list",
          "input": {
            "todos": "[ ] Update `fetchFilteredTickets` in `bot/src/services/overview-channel.ts` to support multiple filters\n[ ] Update `checkIfUpdateNeeded` in `bot/src/services/overview-channel.ts` to support multiple filters\n[ ] Update `buildOverviewEmbeds` in `bot/src/services/overview-channel.ts` to handle titles for multiple filters"
          }
        }
      ],
      "ts": 1766496112549,
      "reasoning_details": [
        {
          "type": "reasoning.text",
          "text": "REMOVED",
          "format": "google-gemini-v1",
          "index": 0
        },
        {
          "type": "reasoning.encrypted",
          "data": "REMOVED",
          "id": "tool_update_todo_list_mL716LwhunURIXx4ZUZf",
          "format": "google-gemini-v1",
          "index": 0
        }
      ]
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "tool_update_todo_list_mL716LwhunURIXx4ZUZf",
          "content": "Todo list updated successfully."
        },
        {
          "type": "text",
          "text": "<environment_details>\n# VSCode Visible Files\n..\\..\\..\\AppData\\Local\\Temp\\roo-diagnostics-a7c08862-1766495767594.json\n\n# VSCode Open Tabs\nbot/src/services/overview-channel.ts,../../../AppData/Local/Temp/roo-diagnostics-a7c08862-1766495767594.json,bot/src/commands/overview.ts,shared/prisma/schema.prisma\n\n# Current Time\nCurrent time in ISO 8601 UTC format: 2025-12-23T13:21:52.685Z\nUser time zone: America/New_York, UTC-5:00\n\n# Current Cost\n$0.11\n\n# Current Mode\n<slug>code</slug>\n<name>💻 Code</name>\n<model>google/gemini-3-pro-preview</model>\n<tool_format>native</tool_format>\n====\n\nREMINDERS\n\nBelow is your current list of reminders for this task. Keep them updated as you progress.\n\n| # | Content | Status |\n|---|---------|--------|\n| 1 | Update `fetchFilteredTickets` in `bot/src/services/overview-channel.ts` to support multiple filters | Pending |\n| 2 | Update `checkIfUpdateNeeded` in `bot/src/services/overview-channel.ts` to support multiple filters | Pending |\n| 3 | Update `buildOverviewEmbeds` in `bot/src/services/overview-channel.ts` to handle titles for multiple filters | Pending |\n\n\nIMPORTANT: When task status changes, remember to call the `update_todo_list` tool to update your progress.\n\n</environment_details>"
        }
      ],
      "ts": 1766496112695
    }
  ]
}

[roomote]

I've analyzed this issue and identified the root cause.

Problem Analysis

The error occurs because Gemini 3 models require thought signatures (reasoning.encrypted blocks) to be preserved and linked to their respective tool calls via the id field. When a tool result is sent back to the model, the original reasoning_details must include the encrypted thought signature tied to the tool call ID.

In src/api/transform/openai-format.ts (lines 194-198), ALL id fields are currently being stripped from reasoning_details:

baseMessage.reasoning_details = messageWithDetails.reasoning_details.map((detail: any) => {
    const { id, ...rest } = detail
    return rest
})

However:

• For reasoning.text and reasoning.summary types, the id is only used internally for accumulation and should be stripped
• For reasoning.encrypted types, the id is the tool call ID that links the thought signature to the function call and must be preserved

Without this id, Gemini's API cannot match the thought signature to the function call, resulting in the error:

"Function call is missing a thought_signature in functionCall parts"

Proposed Implementation Plan

1. Modify src/api/transform/openai-format.ts: Update the reasoning_details mapping logic to only strip the id field for non-encrypted types:

   baseMessage.reasoning_details = messageWithDetails.reasoning_details.map((detail: any) => {
       // Keep id for reasoning.encrypted (tool call thought signatures)
       // Strip id for other types (used internally for accumulation)
       if (detail.type === 'reasoning.encrypted') {
           return detail
       }
       const { id, ...rest } = detail
       return rest
   })

2. Add test coverage in src/api/transform/__tests__/openai-format.spec.ts to verify:

   • reasoning.encrypted blocks preserve their id field
   • reasoning.text and reasoning.summary blocks have id stripped

This is a minimal, targeted fix that addresses the specific issue without affecting other providers.

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

[mrkmg]

I reverted to 3.36.16 and XML tool calls for now. No problems with that.

[leonardas103]

Tested locally the @roomote suggestion and it does solve the issue

   if (detail.type === 'reasoning.encrypted') {
       return detail
   }

[roomote]

Thanks @leonardas103 for testing and confirming the fix works. I'll proceed with implementing the suggested change to preserve the id field for reasoning.encrypted blocks. I'll open a PR shortly.

[roomote]

Opened PR #10310 with my attempt at addressing this issue. The fix preserves the id field for reasoning.encrypted blocks while stripping it for other reasoning types. Tests have been added to cover the new behavior. Feedback welcome!