Local Ollama Models Not Responsive

[moosetunes]

Problem (one or two sentences)

I have several Ollama models downloaded to my PC and I'm using the latest Roo Code build 3.45. I was using Perplexity.ai to help me debug the VS Code Studio output and it was reported - "Roo Code 3.45 has a fundamental bug - it creates a new task/shadow Git repo for every single chat, regardless of tool settings. The rm -rf temporarily clears it, but Roo immediately recreates the corrupted task system."

Context (who is affected and when)

Happens when I attempt to use a local model. Ubuntu, VS Code Studio, Roo Code Plugin, 32GB Ram, Ryzen 7

Reproduction steps

1 - Select Provider - Ollama
2 - Select Model - llama3.1:latest

Here is my config.yaml contents:

name: Local Config
version: 1.0.0
schema: v1

models:

• name: Llama 3.1 (Ollama)
  provider: ollama
  apiBase: http://127.0.0.1:11434
  model: llama3.1:latest
  roles:

  • chat

• name: Mistral (Ollama)
  provider: ollama
  apiBase: http://127.0.0.1:11434
  model: mistral:latest
  roles:

  • chat

• name: Phi-3 Medium (Ollama)
  provider: ollama
  apiBase: http://127.0.0.1:11434
  model: phi3:medium
  roles:

  • chat

Expected result

"What is Python?" or "List the files in my workspace." Should get response.

Actual result

"API Request", then nothing.

Variations tried (optional)

Models will run from bash.

App Version

3.45.0

API Provider (optional)

Ollama

Model Used (optional)

llama3.1:latest

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[createTask] parent task 019c05ab-1af1-722e-9d7c-dc20b32daf5c.5be4536c instantiated
[Task#getCheckpointService] initializing shadow git
[t#create] git = 2.51.0
[t#initShadowGit] creating shadow git repo at /home/mooseshoes/.config/Code/User/globalStorage/rooveterinaryinc.roo-cline/tasks/019c05ab-1af1-722e-9d7c-dc20b32daf5c/checkpoints
[t#initShadowGit] initialized shadow repo with base commit 3308c9fbd634374a0d5c3ca54231ed1a60a08f89 in 563ms
[Task#getCheckpointService] service initialized

[roomote]

I'd like to help with this issue. Based on my analysis of the codebase, I've identified some recent changes that may be related:

1. PR fix: filter Ollama models without native tool support #10735 (Jan 15): Added filtering to hide Ollama models that don't have native tool support (models must have "tools" in their capabilities array)
2. PR chore: remove XML tool calling support #10841 (Jan 20): Removed XML tool calling support, making native tool support mandatory for all models

This means models like llama3.1:latest will only appear in the dropdown if Ollama reports them as having native tool support via the capabilities API field.

To help diagnose this, could you please provide more details:

1. Ollama version: What version of Ollama are you running? (ollama --version) - The capabilities field is a newer API feature.

2. Model visibility: Does llama3.1:latest appear in the Roo Code model dropdown, or did you type it manually?

3. Ollama model info: Can you run ollama show llama3.1:latest and share if it shows a capabilities section?

4. VS Code output: In VS Code, go to View > Output > select "Roo Code" from the dropdown - are there any errors shown when you try to send a message?

5. Ollama server logs: When you send a message, does the Ollama server show any activity or errors in its terminal output?

The shadow Git repo mentioned in your issue description is expected behavior for checkpoints and is not related to this problem.

Please tag @roomote in your response with these details, and I'll be happy to help further.

[moosetunes]

To help diagnose this, could you please provide more details:

Ollama version: What version of Ollama are you running? (ollama --version) - The capabilities field is a newer API feature.

ollama version is 0.13.5

Model visibility: Does llama3.1:latest appear in the Roo Code model dropdown, or did you type it manually?

in the dropdown

Ollama model info: Can you run ollama show llama3.1:latest and share if it shows a capabilities section?

I don't see it.

VS Code output: In VS Code, go to View > Output > select "Roo Code" from the dropdown - are there any errors shown when you try to send a message?

2 test messages sent, Roo Code responded but no Roo Code output was created

Ollama server logs: When you send a message, does the Ollama server show any activity or errors in its terminal output?

Terminal/Server had no activity after sending message

The shadow Git repo mentioned in your issue description is expected behavior for checkpoints and is not related to this problem.

Please tag @roomote in your response with these details, and I'll be happy to help further.

Done.

[roomote]

Thanks for the detailed information @moosetunes!

Based on your responses, I can see a few important clues:

1. Model appears in dropdown - This means Roo Code can connect to Ollama and read the model list
2. No Roo Code output - This is unusual, there should be output showing API activity
3. No Ollama server activity - This suggests the request is not reaching Ollama at all

The fact that the model appears in the dropdown but no request is being sent to Ollama when you try to chat suggests there may be an issue in the request flow after model selection.

I will investigate the Ollama handler code to understand why requests might be silently failing. I will open a PR to share progress and any potential fixes.

[roomote]

I have opened a draft PR with diagnostic improvements to help debug this issue: #11055

The PR adds logging to show:

1. Which Ollama models are being filtered out due to missing tool support
2. Why each model was filtered (missing capabilities vs capabilities without "tools")
3. Request details when Ollama API calls are made
4. A warning when a selected model is not in the list of tool-capable models

@moosetunes - With these changes, the VS Code Output panel (View > Output > Roo Code) should show more diagnostic information that can help identify the root cause of the unresponsive behavior.

Feedback welcome!

[moosetunes]

@roomote Both qwen2.5-coder AND llama3.1:latest work perfectly via curl /api/generate.

Roo Code issue:
✅ Ollama server running
✅ Models respond (45s first load)
❌ Roo creates shadow Git task every chat → never reaches Ollama (no server activity)

Even simple "print('hello')" → [createTask] → shadow git → hangs forever

All 15 tools EXCLUDED in settings. PR #11055 logging will show if model filtering despite working curl.

[roomote]

Thanks for the additional details @moosetunes!

The key observation that all 15 tools are EXCLUDED in settings is very important. I am going to investigate if there is an issue with how Roo Code handles requests when all tools are disabled - this could potentially cause the system to hang before reaching Ollama.

I will explore the request flow to see what happens when:

1. A chat message is sent
2. Tool capabilities are checked
3. The model request is constructed

I will open a PR to share my findings and any potential fixes. The PR #11055 logging improvements will also be helpful for diagnosing this.