Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Version: 2026.4.26 (be8c246) Regression introduced: Worked correctly before 2026.4.15

After updating to be8c246, the think parameter for Ollama models is always sent as false regardless of any configuration. Thinking cannot be enabled through any method — Discord slash commands, Web dashboard, or openclaw.json.

Steps to Reproduce

1. Use any Ollama model with thinking capability (e.g. qwen3.6:35b-a3b-mxfp8, gemma4)
2. Attempt to enable thinking via /think medium (also for low, minimal, etc.,) in Discord or Web dashboard
3. Observe that thinking level options are limited to off only

Expected behavior

• The think parameter should be configurable and passed correctly to Ollama (e.g. think: true when thinking is enabled), matching the behavior prior to 2026.4.15.
• Thinking levels (low, medium, high) should be selectable via Discord slash commands and the Web dashboard for Ollama models that support thinking.

Actual behavior

Discord /think command:

Choose level for /think. Options: off.

Discord /think medium:

Thinking level "medium" is not supported for ollama/qwen3.6:35b-a3b-mxfp8. Use one of: off.

Web dashboard /think command:

Current thinking level: medium.
Options: off.

Web dashboard /think medium:

Unsupported thinking level "medium" for this model. Valid levels: off.

Traffic Observation (mitmproxy)

Request to Ollama POST /api/chat:

{
  "options": {
    "temperature": 0.9,
    "top_p": 0.95,
    "num_ctx": 200000
  },
  "think": false
}

Note: num_ctx is set to 200000 here as it was modified during testing to verify that custom parameter values are being applied at all. temperature and top_p are correctly passed through — think is the only parameter that remains false regardless of configuration.

OpenClaw version

2026.4.26

Operating system

macOS 26.4.1

Install method

No response

Model

ollama/qwen3.6:35b-a3b-mxfp8 / ollama/qwen3.6:27b-nvfp4 / ollama/gemma4:e4b-nvfp4

Provider / routing chain

ollama (local ollama), local gateway

Additional provider/model setup details

• Ollama 0.21.1
• Mac16,11 (Mac mini M4 Pro, 64 GB RAM)

Logs, screenshots, and evidence

#### Ollama api/show: Model supports thinking

curl http://localhost:11434/api/show -d '{
  "model": "qwen3.6:35b-a3b-mxfp8"
}'


{ ... "parameters":"min_p                          0\npresence_penalty               1.5\nrepeat_penalty                 1\ntemperature                    1\ntop_k                          20\ntop_p                          0.95","template":"{{ .Prompt }}","details":{"parent_model":"","format":"safetensors","family":"qwen3_5_moe","families":null,"parameter_size":"35.1B","quantization_level":"mxfp8"},"model_info":{"general.architecture":"qwen3_5_moe","general.parameter_count":35106839040,"qwen3_5_moe.block_count":40,"qwen3_5_moe.context_length":262144,"qwen3_5_moe.embedding_length":2048},"capabilities":["completion","vision","thinking","tools"],"modified_at":"2026-04-24T17:04:13.123285384+09:00","requires":"0.19.0"}%   

---

#### Traffic Observation (mitmproxy)

**Request to Ollama `POST /api/chat`:**

{
  "options": {
    "temperature": 0.9,
    "top_p": 0.95,
    "num_ctx": 200000
  },
  "think": false
}


> **Note:** `num_ctx` is set to `200000` here as it was modified during testing to verify that custom parameter values are being applied at all. `temperature` and `top_p` are correctly passed through — `think` is the only parameter that remains `false` regardless of configuration.

---

Impact and severity

• Affected: All users running Ollama models with thinking capability (confirmed: qwen3.6:35b-a3b-mxfp8, gemma4)
• Severity: Blocks workflow — thinking is completely non-functional; no workaround available
• Frequency: Always reproducible after be8c246
• Consequence: Thinking capability is entirely unavailable for all Ollama models, despite the models supporting it

Additional information

Regression: Worked correctly before 2026.4.15

Attempted Workarounds (all failed)

1. agents.defaults.thinkingDefault in openclaw.json:

"agents": {
  "defaults": {
    "thinkingDefault": "medium"
  }
}

Tried values: low, medium, on — none applied.

2. Per-model params.think override:

{
            "id": "qwen3.6:35b-a3b-mxfp8",
            "name": "qwen3.6:35b-a3b-mxfp8",
            "api": "ollama",
            "reasoning": true,
            "input": [
              "text",
              "image"
            ],
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            },
            "contextWindow": 131072,
            "maxTokens": 8192,
            "params": {
              "thinking": "medium",
              "temperature": 0.9,
              "top_p": 0.95,
              "num_ctx": 131072,
              "think": true
            },
            "compat": {
              "supportsTools": true
            }
          }

Tried values: true, "medium", "low" — none applied. Traffic confirms think: false is still sent.

3. Reset provider/model list and re-ran onboarding — no change

4. Started a new chat session — no change

5. Removed per-agent model files; tried both models.mode: "merge" and "replace" — no change

Suspected Related Changes

• Ollama api/show handling: #72407, 7a45743
• Ollama params handling: 2dba9e6, aa071e0
• Related issue: #48010