Stop here if any apply:

Past this point, debugging cost usually grows faster than results. If the system prompt override doesn't work, consider Secondary Exit Path.

Override the default system prompt to explicitly disable thinking tags in DeepSeek R1.

Why this works:

• DeepSeek R1 respects system prompt instructions
• Disabling thinking tags at the source eliminates parsing issues
• No post-processing or code changes required
• Works across all DeepSeek R1 variants

Time investment: 5 minutes

Steps:

# Method 1: Command-line override (quickest)
ollama run deepseek-r1:latest \
  --system "You are a JSON-only API. Output valid JSON directly. Do NOT use <think> tags. Do NOT include any text before or after the JSON. Format: json object like {\"key\": \"value\"}" \
  --prompt 'Generate JSON with status ok'

# Method 2: OpenClaw configuration file
# Add to ~/.openclaw/config.json or your project's openclaw_config.py

system_prompt = """You are a JSON-only API.
Rules:
1. Output ONLY valid JSON
2. No <think> tags
3. No markdown code blocks
4. No explanations before or after JSON

Format: json object like {"key": "value"}"""

# For Ollama integration
ollama_model = "deepseek-r1:latest"
ollama_system_prompt = system_prompt

Verification:

# Test the fix
ollama run deepseek-r1:latest \
  --system "You are a JSON-only API. Output ONLY JSON. No thinking tags." \
  'Output JSON: {"status": "ok"}'

# Expected output: {"status": "ok"}
# No <think> tags, no markdown, clean JSON

For OpenClaw integration:

# openclaw_config.py
from openclaw import Client

client = Client(model="deepseek-r1:latest")

# Override system prompt to disable thinking tags
client.system_prompt = """You are a JSON-only API.
Rules:
1. Output ONLY valid JSON
2. No <think> tags
3. No markdown code blocks
4. No explanations before/after JSON

Format: json object like {"key": "value"}"""

# Now JSON responses will parse correctly
response = client.generate("Return json object with status ok")
print(response.parsed_json)  # Works!

Use when: Primary Exit Path fails (system prompt override doesn't work)

Solution: Pre-processing to Strip Thinking Tags

If you cannot modify the system prompt, strip the thinking tags before parsing:

import re
import json

def extract_json_from_deepseek(response: str) -> str:
    """
    Extract JSON from DeepSeek R1 response,
    removing <think> tags if present.
    """
    # Remove thinking tags
    cleaned = re.sub(r'<think>.*?</think>', '', response, flags=re.DOTALL)

    # Extract JSON from markdown code blocks if present
    json_match = re.search(r'```json\s*(\{.*?\})\s*```', cleaned, flags=re.DOTALL)
    if json_match:
        return json_match.group(1)

    # Return cleaned response
    return cleaned.strip()

# Usage
raw_response = "<think>reasoning here</think>{\"result\": \"data\"}"
json_str = extract_json_from_deepseek(raw_response)
data = json.loads(json_str)

When to use this:

• You don't have control over system prompt configuration
• OpenClaw runs in a restricted environment
• You're integrating with an existing deployment that can't be reconfigured

Time investment: 10-15 minutes

Note: This is a workaround, not a fix. Use Primary Exit Path if possible.

Context Window Truncation

Sometimes the JSON appears "cut off" mid-response:

{"result": "partial data", "status":

Root Cause: DeepSeek R1 hit its context limit mid-generation.

Fix: Reduce context window usage:

ollama run deepseek-r1:latest \
  --num_ctx 8192 \
  --repeat-penalty 0.6

For hardware limitations: See CUDA OOM Fix Guide.