[GH-ISSUE #17757] issue: Native tool-calling overwrites tool-emitted chat message updates #18387

New Issue

GiteaMirror · 2026-04-20T00:36:23-05:00

GiteaMirror commented

2026-04-20 00:36:23 -05:00

Originally created by @brandco on GitHub (Sep 25, 2025).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/17757

Check Existing Issues

I have searched for any existing and/or related issues.
I have searched for any existing and/or related discussions.
I am using the latest version of Open WebUI.

Installation Method

Docker

Open WebUI Version

v0.6.30

Ollama Version (if applicable)

No response

Operating System

ghcr.io/open-webui/open-webui:main

Browser (if applicable)

No response

Confirmation

I have read and followed all instructions in README.md.
I am using the latest version of both Open WebUI and Ollama.
I have included the browser console logs.
I have included the Docker container logs.
I have provided every relevant configuration, setting, and environment variable used in my setup.
I have clearly listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc).
I have documented step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation. My steps:
Start with the initial platform/version/OS and dependencies used,
Specify exact install/launch/configure commands,
List URLs visited, user input (incl. example values/emails/passwords if needed),
Describe all options and toggles enabled or changed,
Include any files or environmental changes,
Identify the expected and actual result at each stage,
Ensure any reasonably skilled user can follow and hit the same issue.

Expected Behavior

Tool-emitted message updates persist in the visible assistant message during generation, regardless of function-calling mode.

Actual Behavior

In native function-calling, subsequent "chat:completion" events with a full content snapshot replace the message content, effectively overwriting prior tool updates, causing flicker/disappearance.

Steps to Reproduce

In ghcr.io/open-webui/open-webui:main with a model that can support native tool calling:

Enable tools for the model and set params.function_calling = "native" and stream = true.
In a tool function, accept and use the injected event_emitter to emit mid-stream updates:
- Example: type = "message" with data.content to append, or type = "replace" to set content, or type = "chat:message".
Observe the UI: the tool’s text appears briefly, then disappears/gets overwritten during the model’s streamed response.
Repeat with params.function_calling = "default". In this path, the appended/replaced text persists.

"""
title: Native Tool Emit Bug
author: brandco
description: Minimal reproduction of native tool emit bug.
version: 0.0.1
"""

from __future__ import annotations

import asyncio
from typing import Awaitable, Callable, Optional


class Tools:
    """Expose a single tool that reproduces the message overwrite sequence."""

    def __init__(self):
        """Initialize the Tool."""
        self.citation = True

    async def message_event_reproducer(
        self,
        prompt: str = "Test message",
        __event_emitter__: Optional[Callable[[dict[str, object]], Awaitable[None]]] = None,
    ) -> str:
        """Emit a message to show overwriting."""

        chunk_one = "\n\nStreaming chunk 1.\n\n"
        chunk_two = "\n\nStreaming chunk 2.\n\n"
        final_text = "\n\nFinal text.\n\n"

        if not __event_emitter__:
            return final_text

        await __event_emitter__(
            {
                "type": "message",
                "data": {"content": chunk_one},
            }
        )
        await asyncio.sleep(1)

        await __event_emitter__(
            {
                "type": "message",
                "data": {"content": chunk_two},
            }
        )
        await asyncio.sleep(1)

        await __event_emitter__(
            {
                "type": "message",
                "data": {
                    "content": final_text,
                },
            }
        )
        await asyncio.sleep(1)

        return final_text

Logs & Screenshots

With native tool calling:

With default tool calling:

Additional Information

This issue is related to issue 12214 which was closed prematurely.

In native function-calling, the server constructs content_blocks from the streaming model output and repeatedly emits "chat:completion" events with a full serialized content snapshot (mid-stream and at completion). The UI treats these snapshots as authoritative and sets message.content = content.

File: backend/open_webui/utils/middleware.py:2441
- Adds tools for native path:
  - metadata.params.function_calling == "native" ⇒ metadata["tools"], form_data["tools"].
- Streaming handler (process_chat_response) emits repeated snapshots:
  - mid-stream and after tool calls, via event_emitter({ type: "chat:completion", data: { content: serialize_content_blocks(...) } }).
  - emits a final done=true snapshot at the end.

Client – Snapshots overwrite prior tool updates

File: src/lib/components/chat/Chat.svelte:350
- chatEventHandler:
  - type === 'message' or 'chat:message:delta' ⇒ message.content += data.content
  - type === 'chat:message' or 'replace' ⇒ message.content = data.content
  - type === 'chat:completion' ⇒ chatCompletionEventHandler(data, ...)
- chatCompletionEventHandler:
  - if (choices) deltas ⇒ appends to message.content
  - if (content) snapshot ⇒ message.content = content (REPLACE)

Originally created by @brandco on GitHub (Sep 25, 2025). Original GitHub issue: https://github.com/open-webui/open-webui/issues/17757 ### Check Existing Issues - [x] I have searched for any existing and/or related issues. - [x] I have searched for any existing and/or related discussions. - [x] I am using the latest version of Open WebUI. ### Installation Method Docker ### Open WebUI Version v0.6.30 ### Ollama Version (if applicable) _No response_ ### Operating System ghcr.io/open-webui/open-webui:main ### Browser (if applicable) _No response_ ### Confirmation - [x] I have read and followed all instructions in `README.md`. - [x] I am using the latest version of **both** Open WebUI and Ollama. - [x] I have included the browser console logs. - [x] I have included the Docker container logs. - [x] I have **provided every relevant configuration, setting, and environment variable used in my setup.** - [x] I have clearly **listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup** (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc). - [x] I have documented **step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation**. My steps: - Start with the initial platform/version/OS and dependencies used, - Specify exact install/launch/configure commands, - List URLs visited, user input (incl. example values/emails/passwords if needed), - Describe all options and toggles enabled or changed, - Include any files or environmental changes, - Identify the expected and actual result at each stage, - Ensure any reasonably skilled user can follow and hit the same issue. ### Expected Behavior Tool-emitted message updates persist in the visible assistant message during generation, regardless of function-calling mode. ### Actual Behavior In native function-calling, subsequent "chat:completion" events with a full content snapshot replace the message content, effectively overwriting prior tool updates, causing flicker/disappearance. ### Steps to Reproduce In ghcr.io/open-webui/open-webui:main with a model that can support native tool calling: 1. Enable tools for the model and set params.function_calling = "native" and stream = true. 2. In a tool function, accept and use the injected __event_emitter__ to emit mid-stream updates: - Example: type = "message" with data.content to append, or type = "replace" to set content, or type = "chat:message". 3. Observe the UI: the tool’s text appears briefly, then disappears/gets overwritten during the model’s streamed response. 4. Repeat with params.function_calling = "default". In this path, the appended/replaced text persists. ```python """ title: Native Tool Emit Bug author: brandco description: Minimal reproduction of native tool emit bug. version: 0.0.1 """ from __future__ import annotations import asyncio from typing import Awaitable, Callable, Optional class Tools: """Expose a single tool that reproduces the message overwrite sequence.""" def __init__(self): """Initialize the Tool.""" self.citation = True async def message_event_reproducer( self, prompt: str = "Test message", __event_emitter__: Optional[Callable[[dict[str, object]], Awaitable[None]]] = None, ) -> str: """Emit a message to show overwriting.""" chunk_one = "\n\nStreaming chunk 1.\n\n" chunk_two = "\n\nStreaming chunk 2.\n\n" final_text = "\n\nFinal text.\n\n" if not __event_emitter__: return final_text await __event_emitter__( { "type": "message", "data": {"content": chunk_one}, } ) await asyncio.sleep(1) await __event_emitter__( { "type": "message", "data": {"content": chunk_two}, } ) await asyncio.sleep(1) await __event_emitter__( { "type": "message", "data": { "content": final_text, }, } ) await asyncio.sleep(1) return final_text ``` ### Logs & Screenshots With native tool calling: <img width="1068" height="603" alt="Image" src="https://github.com/user-attachments/assets/59b5ba2a-6ef3-4e14-8b93-5ed14f115716" /> With default tool calling: <img width="1086" height="459" alt="Image" src="https://github.com/user-attachments/assets/227c58c6-a4bc-4548-8f2d-b2e86db24e95" /> ### Additional Information This issue is related to issue [12214](https://github.com/open-webui/open-webui/issues/12214) which was closed prematurely. In native function-calling, the server constructs content_blocks from the streaming model output and repeatedly emits "chat:completion" events with a full serialized content snapshot (mid-stream and at completion). The UI treats these snapshots as authoritative and sets message.content = content. - File: backend/open_webui/utils/middleware.py:2441 - Adds tools for native path: - metadata.params.function_calling == "native" ⇒ metadata["tools"], form_data["tools"]. - Streaming handler (process_chat_response) emits repeated snapshots: - mid-stream and after tool calls, via event_emitter({ type: "chat:completion", data: { content: serialize_content_blocks(...) } }). - emits a final done=true snapshot at the end. Client – Snapshots overwrite prior tool updates - File: src/lib/components/chat/Chat.svelte:350 - chatEventHandler: - type === 'message' or 'chat:message:delta' ⇒ message.content += data.content - type === 'chat:message' or 'replace' ⇒ message.content = data.content - type === 'chat:completion' ⇒ chatCompletionEventHandler(data, ...) - chatCompletionEventHandler: - if (choices) deltas ⇒ appends to message.content - if (content) snapshot ⇒ message.content = content (REPLACE)

GiteaMirror added the bug label 2026-04-20 00:36:23 -05:00

GiteaMirror closed this issue

2026-04-20 00:36:24 -05:00

GiteaMirror commented

2026-04-20 00:36:25 -05:00

@rgaricano commented on GitHub (Sep 25, 2025):

It not a bug.
In Native mode, the event emitter approach doesn't work because Native function calling bypasses Open WebUI's custom tool processing pipeline. When function_calling is set to "native", tools are added directly to the form data for the model to handle natively, rather than going through Open WebUI's chat_completion_tools_handler.

@rgaricano commented on GitHub (Sep 25, 2025): It not a bug. In Native mode, the event emitter approach doesn't work because Native function calling bypasses Open WebUI's custom tool processing pipeline. When function_calling is set to "native", tools are added directly to the form data for the model to handle natively, rather than going through Open WebUI's chat_completion_tools_handler.

GiteaMirror commented

2026-04-20 00:36:26 -05:00

@tjbck commented on GitHub (Sep 25, 2025):

@rgaricano is correct here, you generally should not use message event types in your tool in conjunction with a native function calling.

@tjbck commented on GitHub (Sep 25, 2025): @rgaricano is correct here, you generally should not use `message` event types in your tool in conjunction with a native function calling.

GiteaMirror referenced this issue

2026-04-20 00:47:58 -05:00

[GH-ISSUE #18387] issue: Stored XSS in Artifact Rendering #18581

GiteaMirror referenced this issue

2026-04-25 08:02:21 -05:00

[GH-ISSUE #18387] issue: Stored XSS in Artifact Rendering #34110

GiteaMirror referenced this issue

2026-05-05 20:46:50 -05:00

[GH-ISSUE #18387] issue: Stored XSS in Artifact Rendering #57247

Sign in to join this conversation.