Description

The image tool returns "Image model returned no text" for all OpenRouter vision models, both free and paid. Local models (LM Studio) work fine.

Environment

• OpenClaw: 2026.4.9
• Node: v22.22.0
• OS: macOS arm64 (Darwin 25.3.0)

Models Tested (all fail)

• openrouter/openai/gpt-4.1-mini (paid, non-reasoning)
• openrouter/google/gemini-2.5-flash (paid)
• openrouter/google/gemma-4-31b-it:free (free)
• openrouter/qwen/qwen3.5-flash-02-23 (paid)

Working

• lm-studio/qwen/qwen3.5-9b (local) — works in 14-21 seconds

Analysis

1. Models resolve correctly — no "Unknown model" error
2. API auth works — OpenRouter accepts the request (no 401/403)
3. Response: finish_reason: "stop" but content is empty string
4. Format: pi-ai correctly converts {type: "image", data: base64} to OpenAI {type: "image_url", image_url: {url: "data:..."}}

Source Code Findings

In image-BekbXrJh.js, buildImageContext puts the prompt in systemPrompt but the user message content array contains only images, no text:

function buildImageContext(prompt, images) {
    return {
        systemPrompt: prompt,
        messages: [{
            role: "user",
            content: images.map((image) => ({
                type: "image",
                data: image.buffer.toString("base64"),
                mimeType: image.mime ?? "image/jpeg"
            })),
            // ^ No text content in user message!
        }]
    };
}

OpenRouter docs recommend: "we recommend sending the text prompt first, then the images" — the current implementation only sends images in the user message.

I tried patching this to include {type: "text", text: prompt} in the user message content array, but it still returned empty text after a full gateway restart. The issue may be deeper — possibly in how extractAssistantText processes the response, or in pi-ai's response handling for OpenRouter.

Suggested Fix

1. Add {type: "text", text: prompt} to user message content alongside images (per OpenRouter docs)
2. Add debug logging in coerceImageAssistantText to log the raw stopReason, errorMessage, and message content when text is empty
3. Investigate if pi-ai openai-completions.js handles OpenRouter's response format correctly for vision requests

Error Log

[tools] image failed: All image models failed (2): openrouter/openai/gpt-4.1-mini: Image model returned no text (openrouter/openai/gpt-4.1-mini). | openrouter/google/gemini-2.5-flash: Image model returned no text (openrouter/google/gemini-2.5-flash).