[Ollama] Tool Calls don't work reliably anymore

[xpomul]

Bug Description:

I have noticed that in most cases, tool calls from Ollama fail.
It looks like the tool is executed correctly, but when expanding the tool call response, the following JSON is shown:

{
  "kind": "toolCall",
  "_id": "ollama_2026-02-11T09:04:08.338165Z_0",
  "_name": "getFileContent",
  "_arguments": "{\"file\":\"frontend/src\"}",
  "_finished": true,
  "_result": "{\"content\":[{\"type\":\"error\",\"data\":\"Error executing tool 'getFileContent': Tool call content for tool getFileContent not found in the response\"}]}",
  "_confirmed": {},
  "_whenFinished": {}
}

and the model then says the tool call failed and will try again (but will fail again).

Steps to Reproduce:

1. Configure Ollama with a tool-capable model (e.g. glm-4.7-flash)
2. Ask the Architect Agent to describe the workspace contents
3. Inspect the tool calls in the chat by expanding them

Additional Information

• Operating System: macOS
• Theia Version: 1.68.200 (maybe earlier, have not tried)

[xpomul]

First analysis

The error comes from this line:

theia/packages/ai-chat/src/browser/chat-tool-request-service.ts

Line 131 in bec8ac2

throw new Error(`Tool call content for tool ${toolRequest.id} not found in the response`);

I have placed a breakpoint and detected that the contentArray only contains the messages before the tool call message, but not the tool call message itself.

Most likely there has been a change in how Theia AI expects tool call message parts to be returned from the Language Model implementation.

The relevant code in ollama-language-model.ts is:

theia/packages/ai-ollama/src/node/ollama-language-model.ts

Lines 157 to 167 in bec8ac2

	if (toolCalls && toolCalls.length > 0) {
	chatRequest.messages.push({
	role: 'assistant',
	content: currentContent,
	thinking: currentThought,
	tool_calls: toolCalls
	});
	const toolCallsForResponse = await that.processToolCalls(toolCalls, chatRequest, lastUpdated);
	yield { tool_calls: toolCallsForResponse };

As can be seen we call processToolCalls() in line 165. This method calls tool.handle() and then prepares a ToolCallResponsePart including both the tool call and the result and returns that. In line 166, yield the response part for the tool call.

This has worked fine so far, but now it does not work anymore, because calling the tool handler calls the logic above to find the message associated with the tool call.

Before I dig in the code to find out what to do, can someone with knowledge about the tool call handling and/or the recent changes here point me to what is expected? Should I just send the ToolCallResponsePart without the result part before calling tool.handle()?
And if so, will the result be automatically added or do I have to send another ToolCallResponsePart with the result (and the same ID?).

After that, I'll try to open a new PR with a fix.

[xpomul]

I have had a look into the code and it seems that, e.g., google-language-model.ts actually creates the ToolCallResponsePart and later updates it by yielding it once again with the same ID and the result set.

I have created a PR that implements the same logic for the Ollama model as well.

[rajj28]

@ndoschek can i work on this issue

[ndoschek]

Hi @rajj28, there is already a PR open for this issue (see GH-16976)