Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

image tool cannot use Ollama vision models despite correct configuration and confirmed vision capability

Steps to reproduce

1. Configure Ollama provider with a vision-capable model (e.g., gemma4:e4b) in openclaw.json
2. Set agents.defaults.imageModel.primary to "ollama/gemma4:e4b"
3. Enable the Ollama plugin
4. Attempt to analyze an image using the image tool
5. Observe error: "No API provider registered for api: ollama"

Expected behavior

image tool should route to ollama/gemma4:e4b and successfully analyze images, since Ollama's native /api/chat supports vision models

Actual behavior

Error: "All image models failed: ollama/gemma4:e4b: No API provider registered for api: ollama". Ollama chat (text) works fine with the same configuration.

OpenClaw version

2026.4.9

Operating system

Windows 11

Install method

npm global

Model

ollama/gemma4:e4b

Provider / routing chain

openclaw -> ollama (local, http://localhost:11434)

Additional provider/model setup details

Ollama plugin is enabled and working for text chat. The model gemma4:e4b has confirmed vision capability via /api/show (capabilities: ["completion", "vision", "audio", "tools"]). The Ollama plugin contracts only register memoryEmbeddingProviders and webSearchProviders, NOT image or mediaUnderstanding contracts.

Logs, screenshots, and evidence

Impact and severity

Affected: Users who want to use local Ollama vision models for image analysis Severity: Medium (blocks use of local vision models, forces use of external API) Frequency: Always (100% reproducible) Consequence: Cannot use gemma4:e4b or other Ollama vision models via image tool despite correct setup

Additional information

Root cause: The Ollama plugin (openclaw.plugin.json) registers these contracts:

• memoryEmbeddingProviders: ["ollama"]
• webSearchProviders: ["ollama"] It does NOT register any image or mediaUnderstanding contract. The image tool uses a separate provider lookup path that doesn't find Ollama. Workaround: using doubao-vision-ocr skill as a temporary alternative.