Summary

The OpenClaw image tool fails with "Image model returned no text" when using any OpenAI-compatible reasoning-capable vision model that returns its response in reasoning_content instead of content.

Environment

• OpenClaw version: latest (container)
• Provider: Custom OpenAI-compatible provider
• Model: Any vision model with reasoning: true configuration

Steps to Reproduce

1. Configure a vision model with reasoning: true
2. Set it as agents.defaults.imageModel.primary
3. Use the image tool with any image
4. Error: "Image model returned no text (<provider>/<model>)"

Root Cause

The streaming provider correctly parses reasoning_content from the API response and creates thinking blocks in the message content array.

However, coerceImageAssistantText() in minimax-vlm.ts only calls extractAssistantText() which only reads text type blocks:

function coerceImageAssistantText(params) {
  // ...
  const text = extractAssistantText(params.message);
  if (text.trim()) return text.trim();
  throw new Error(`Image model returned no text (${params.provider}/${params.model}).`);
}

When a reasoning-capable model returns:

• content: "" (empty string)
• reasoning_content: "actual response text"

The message content array becomes: [{ type: "thinking", thinking: "actual response text" }]

extractAssistantText() returns "" → throws error.

Affected Models

Any OpenAI-compatible vision model with reasoning enabled that returns empty content. Tested with multiple models from the same provider family:

• Vision model A (reasoning enabled) → ❌ fails
• Vision model B (reasoning enabled) → ❌ fails
• Vision model C (reasoning enabled) → ❌ fails
• Vision model A (reasoning disabled) → ✅ works
• Vision model A (reasoning enabled, patched code) → ✅ works

Suggested Fix

In src/agents/minimax-vlm.ts, coerceImageAssistantText() should fall back to extractAssistantThinking() when text is empty:

function coerceImageAssistantText(params) {
  const stop = params.message.stopReason;
  const errorMessage = params.message.errorMessage?.trim();
  if (stop === "error" || stop === "aborted") throw new Error(/* ... */);
  if (errorMessage) throw new Error(/* ... */);
  
  const text = extractAssistantText(params.message);
  if (text.trim()) return text.trim();
  
  // Fallback: reasoning models may return content in thinking blocks
  const thinking = extractAssistantThinking(params.message);
  if (thinking.trim()) return thinking.trim();
  
  throw new Error(`Image model returned no text (${params.provider}/${params.model}).`);
}

Workaround

Set reasoning: false on the vision model config in openclaw.json. This causes the model to return content in the content field directly, bypassing the issue. However, this disables the reasoning capability.

Additional Context

• media-understanding.ts already handles this correctly with coerceOpenAiCompatibleVideoText() which checks reasoning_content as fallback
• The complete() function from @mariozechner/pi-ai correctly parses reasoning_content into thinking blocks via the streaming handler
• Only the image tool path (coerceImageAssistantText) is missing the fallback

Files Involved

• src/agents/minimax-vlm.ts — coerceImageAssistantText() function
• src/utils/pi-embedded-utils.ts — extractAssistantThinking() (already exists, used elsewhere)