Primary Manifestation

The Discord bot successfully reacts to user messages (shows emoji acknowledgment) but delivers no text response in the channel, despite the dashboard chat displaying the full generated reply.

Reproduction Sequence

User: "Hello, how are you?"
Bot: [reaction emoji appears] ❌ (no text message)

# Meanwhile in Dashboard:
# Assistant: "I'm doing well, thank you for asking! How can I help you today?"

Environment Configuration

• OS: Windows 11
• OpenClaw Version: v2026.5.7
• Provider: LM Studio v0.4.12+1
• Model: qwen3.6-35b-a3b-mlx-vl-oq4
• Channel: Discord

Model-Specific Behavior

Secondary Symptoms

• Message reaction fires immediately (indicating response parsing began)
• No error messages in console logs
• Response generation completes successfully (visible in dashboard)
• Switching to Gemma4 model restores normal text delivery

Primary Root Cause: Qwen3 Response Format Incompatibility

Qwen3 series models (particularly 3.6 series) output responses in a modified agentic format that diverges from standard chat completion formats expected by OpenClaw’s Discord adapter.

Technical Breakdown

1. Tool Call Prefix Collision

Qwen3 uses <|tool_call|> tokens for agentic operations. When the model outputs:

<|tool_call|>
{"name": "final", "parameters": {"response": "Hello! I'm ready to help."}}
<|tool_call|>

The Discord adapter’s response extractor identifies the final tool call as a system action rather than user-facing text, resulting in:

• Reaction fires (message received valid response)
• Text content discarded (classified as tool metadata, not displayable content)

2. Content Block Structure Mismatch

Qwen3 wraps response content in content arrays with type markers:

json { “type”: “text”, “text”: “actual response” }

vs. Gemma/standard format: json { “content”: “actual response” }

The response parser likely iterates looking for .content string field, finds array structure instead, and returns null/undefined.

3. Reasoning Block Interleaving

Qwen3 models with extended thinking enabled output:

<|think|> [internal reasoning steps] <|think|>

Hello! How can I help?

The Discord adapter extracts the <|think|> block content instead of the actual response, resulting in empty visible output.

Architectural Failure Chain

LM Studio API Response ↓ Response Parser (Discord Adapter) ↓ Looks for: response.content (string) ↓ Finds: content[{“type”: “text”, “text”: …}] (array) ↓ Returns: undefined / null ↓ Discord Message: Skipped (no content to display) ↓ Reaction: Already fired at message reception

Why Gemma Works

Gemma4 outputs in standard <start_of_turn> format with direct .content strings that the parser correctly extracts:

<start_of_turn> model <end_of_turn> <start_of_turn> ego Hello! How can I help? <end_of_turn>

Solution 1: Configure Response Extraction Override (Recommended)

Create or modify your OpenClaw configuration file to handle Qwen3’s content array format:

Step 1: Locate your config.yaml or create openclaw.yaml:

yaml

openclaw.yaml

discord: adapter: response_extraction: # Enable array content handling for Qwen3 models parse_content_arrays: true content_field: “text” # Extract from {type: “text”, text: “…”}

# Fallback to raw content if array parsing fails
fallback_to_raw: true

providers: lmstudio: response_format: # Force standard parsing for Qwen3 mode: “standard” strip_think_tags: true

Step 2: Restart the OpenClaw service:

powershell

Windows PowerShell

Stop-Process -Name “openclaw” -Force -ErrorAction SilentlyContinue Start-Process “openclaw.exe” -ArgumentList “–config openclaw.yaml”

Check logs for confirmation

Get-Content logs/openclaw.log -Tail 50 -Wait

Step 3: Verify configuration applied:

powershell openclaw diagnostic –provider lmstudio

Solution 2: Add Model-Specific System Prompt Override

Instruct OpenClaw to force Qwen3 into standard response mode via system prompt.

File: system_prompts/qwen3_standard.txt

You are a helpful assistant. Respond directly to user messages without using tool calls or special tokens. Format your response as plain text only. Do not use <|tool_call|> or <|think|> tags. Begin your response immediately with your answer.

Configuration Update:

yaml

openclaw.yaml

models: qwen3.6-35b-a3b-mlx-vl-oq4: system_prompt_file: “system_prompts/qwen3_standard.txt” # Override Qwen3’s default tool-calling behavior tool_use: disabled reasoning: disabled

Solution 3: Patch Response Parser (Manual Code Fix)

If configuration options are insufficient, patch the Discord adapter directly:

File: src/adapters/discord/adapter.ts (or your equivalent)

Locate the response extraction function and add array handling:

typescript // BEFORE (line ~142): function extractResponseContent(response: any): string { return response.content; }

// AFTER: function extractResponseContent(response: any): string { // Handle Qwen3 content array format if (Array.isArray(response.content)) { const textBlock = response.content.find( (block) => block.type === “text” || block.type === “content” ); if (textBlock?.text) return textBlock.text; }

// Handle standard format if (typeof response.content === “string”) { return response.content; }

// Fallback: raw response return JSON.stringify(response); }

Solution 4: Disable Extended Thinking (Model Setting)

If your Qwen3 model has thinking/thought enabled, disable it in LM Studio:

Step 1: In LM Studio Server settings, ensure:

• `enable_thinking: false` in chat template
• System prompt does not include `<|think|>` instructions

Step 2: Restart LM Studio server.

Step 3: Test in OpenClaw.

Solution 5: Server Configuration Override

Modify LM Studio’s server configuration to transform Qwen3 responses:

LM Studio Server Config (~/.lmstudio/server.yaml):

yaml response_transforms:

• model_pattern: “qwen3.*” preprocessors:
  • strip_tool_calls: true
  • flatten_content_arrays: true
  • remove_think_blocks: true

Verification Test 1: Configuration Applied Correctly

powershell

Check that config loaded without errors

openclaw diagnostic –verbose 2>&1 | Select-String -Pattern “discord|qwen|response”

Expected output:

[INFO] Discord adapter initialized

[INFO] Response extraction mode: standard

[INFO] Model qwen3.6-35b-a3b-mlx-vl-oq4 loaded

Verification Test 2: Basic Text Response

Send a test message via Discord:

!test Hello, what is 2+2?

Expected Behavior:

• Bot reacts with emoji
• Bot sends text message: "2 + 2 equals 4"
• Dashboard shows same response

Success Criteria:

powershell

Check Discord message history via API

Compare with dashboard logs

Both should contain identical text response

Verification Test 3: Multi-Turn Conversation

powershell

Test conversation continuity

User: “My favorite color is blue” Bot: [response] ✅ User: “What color did I just mention?” Bot: [response should reference “blue”] ✅

Verification Test 4: Error-Free Logs

powershell

Monitor for extraction errors

Get-Content logs/openclaw.log -Tail 100 | Select-String -Pattern “null|undefined|empty.*response”

Should return no results after fix

Verification Test 5: Cross-Model Consistency

Test both models to confirm consistent behavior:

powershell

Test Qwen3

!model qwen3.6-35b-a3b-mlx-vl-oq4 !test Testing Qwen3

Test Gemma4

!model gemma-4-31b-it-mlx
!test Testing Gemma

Both should produce text responses in Discord

Expected Console Output:

[DEBUG] Response extracted: “Testing Qwen3 - response text here” [DEBUG] Discord message queued: “Testing Qwen3 - response text here” [DEBUG] Message delivered successfully

Logically Connected Issues

• ERR_RESPONSE_FORMAT_UNSUPPORTED — API returns format not recognized by adapter
• ERR_CONTENT_EXTRACTION_NULL — Response parser returns empty content object
• ERR_DISCORD_MESSAGE_SEND_EMPTY — Empty message queued for Discord delivery

Related Historical Issues

Cross-Reference

This bug shares the same root cause architecture as:

• Claude tool use response formatting — requires content array parsing
• Mistral-large function calling — similar tool call output structure
• DeepSeek reasoning blocks — parallel to Qwen3 think tags

Recommended Reading

• OpenClaw Discord Adapter Architecture — Response Parsing Pipeline
• Model Compatibility Matrix — Qwen3 Series Response Formats
• LM Studio Provider Configuration — Response Format Handling