Ollama reasoning models hang or enter degenerate generation loops through OpenCode, while direct /v1/chat/completions is fast

Description

Description

When using local Ollama reasoning models through OpenCode with the
@ai-sdk/openai-compatible provider, very simple requests can take several
minutes, more than one hour, or enter a degenerate generation loop.

The same models queried directly through Ollama's OpenAI-compatible
/v1/chat/completions endpoint respond in seconds.

Disabling reasoning directly at the Ollama API level with:

"reasoning_effort": "none"

makes the response almost instantaneous and removes the reasoning field.

However, I have not found a working way to make OpenCode send the equivalent
setting. Using --variant none did not solve the problem.

Environment

OpenCode:

1.18.18

Ollama:

0.32.13

OS:

Arch Linux

Ollama model execution:

NAME           SIZE     PROCESSOR    CONTEXT
qwen3.8:27b    17 GB    100% GPU     65536

System RAM:

90 GiB total
55 GiB available during testing

OpenCode provider configuration:

"provider": {
  "ollama": {
    "npm": "@ai-sdk/openai-compatible",
    "name": "Ollama (local)",
    "options": {
      "baseURL": "http://localhost:11434/v1"
    },
    "models": {
      "qwen3.8:27b": {
        "name": "Qwen3.8 27B"
      },
      "qwen3-coder:30b": {
        "name": "Qwen3 Coder 30B"
      },
      "glm-4.7-flash:latest": {
        "name": "GLM 4.7 Flash"
      }
    }
  }
}

Reproduction

1. Direct Ollama test

Running:

ollama run glm-4.7-flash:latest

Prompt:

Reply with exactly: READY

returns successfully in less than approximately 12 seconds.

The model generates a thinking/reasoning block and then:

READY

2. Direct OpenAI-compatible API test

Request:

time curl -s http://localhost:11434/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "glm-4.7-flash:latest",
    "messages": [
      {"role": "user", "content": "Reply with exactly: READY"}
    ],
    "stream": false
  }'

Result:

{
  ...
  "message": {
    "role": "assistant",
    "content":"READY",
    "reasoning":"..."
  },
  ...
}

Measured time:

real    0m4.271s

So the Ollama OpenAI-compatible endpoint itself works correctly and quickly.

3. Direct API with think: false

I also tested:

"think": false

The request still returned a reasoning field.

Measured time:

real    0m11.846s

So think:false did not disable reasoning for this model through this endpoint.

4. Direct API with reasoning_effort: "none"

Request:

time curl -s http://localhost:11434/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "glm-4.7-flash:latest",
    "messages": [
      {"role": "user", "content": "Reply with exactly: READY"}
    ],
    "stream": false,
    "reasoning_effort": "none"
  }'

Result:

{
  ...
  "message": {
    "role": "assistant",
    "content":"READY"
  },
  ...
}

There is no reasoning field.

Measured time:

real    0m0.381s

This is a very large difference compared with reasoning enabled.

5. Same trivial request through OpenCode

Using GLM 4.7 Flash in OpenCode Build mode:

Reply with exactly: READY

OpenCode reported:

Thought: 1h 5m
Build · GLM 4.7 Flash · 1h 6m

This is despite the same Ollama model responding directly in seconds.

On another simple request with Qwen3.8 27B, OpenCode reported:

Thought: 50.2s
Plan · Qwen3.8 27B · 3m 13s

Another earlier simple OpenCode request with Qwen3.8 27B took:

Build · Qwen3.8 27B · 3m 37s

while direct Ollama requests were much faster.

6. Degenerate generation loop

During another OpenCode test with GLM 4.7 Flash, generation entered a loop
producing text similar to:

However, However, Other
...
Finally,
However,
Therefore,
Thus,
...
Therefore, However, However, Therefore...
...

The generation did not converge to a useful answer and had to be stopped.

7. OpenCode --variant none test

OpenCode 1.18.18 exposes:

--variant  model variant (provider-specific reasoning effort,
           e.g., high, max, minimal)

I tested:

time timeout 30s opencode run \
  -m ollama/glm-4.7-flash:latest \
  --variant none \
  "Reply with exactly: READY"

Output:

> build · glm-4.7-flash:latest

Then it timed out:

real    0m30.051s

No READY response was produced.

Expected behavior

OpenCode should either:

1. correctly handle the reasoning field returned by Ollama's
   OpenAI-compatible endpoint without hanging or entering a generation loop,

or

2. provide a reliable way to pass:

   "reasoning_effort": "none"
   

   to Ollama for models where reasoning should be disabled.

Ideally both should work.

Important observation

The performance difference is reproducible at the API level:

Ollama /v1 with reasoning:
    ~4-12 seconds depending on test

Ollama /v1 with reasoning_effort:none:
    0.381 seconds

OpenCode:
    30+ second timeout
    several minutes
    up to 1h06
    or degenerate generation loop

The Ollama model is running 100% on GPU, so this does not appear to be caused
by CPU offloading.

The direct /v1/chat/completions tests also suggest that the Ollama server and
OpenAI-compatible endpoint themselves are responsive.

Question

Is OpenCode correctly handling Ollama's reasoning response field when using
@ai-sdk/openai-compatible?

Also, what is the supported way in OpenCode 1.18.18 to send the exact
OpenAI-compatible request parameter:

"reasoning_effort": "none"

to Ollama?

I found previous reports concerning OpenAI-compatible reasoning handling, but
the behavior described above is still reproducible with OpenCode 1.18.18 and
Ollama 0.32.13.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response