[Bug]: opencode-go vision-capable models cannot use image tool — missing MediaUnderstandingProvider #70482

🐛 Bug Report (한국어 요약 아래)

Bug type

Behavior bug (incorrect output/state without crash)

Summary

Users subscribed to the OpenCode Go provider (via opencode-go) cannot use the image tool with vision-capable models (kimi-k2.6, minimax-m2.7, glm-5, mimo-v2-omni, qwen3.6-plus, etc.) even though these models support multimodal/vision capabilities through the OpenCode Go API.

한국어 요약: OpenCode Go 프로바이더를 구독하여 OpenClaw에 연결한 사용자들이, 해당 프로바이더의 멀티모달/비전 모델들(kimi, minimax, glm, mimo, qwen 등)로 이미지 분석 기능을 사용할 수 없는 버그입니다.

Steps to reproduce

1. Subscribe to OpenCode Go and configure agents.defaults.model.primary to a vision-capable model (e.g., opencode-go/kimi-k2.6)
2. Do NOT set agents.defaults.imageModel
3. Send an image to the agent
4. Observe that the image tool is either missing or fails with No media-understanding provider registered for opencode-go

재현 방법: OpenCode Go를 구독하여 agents.defaults.model.primary를 opencode-go/kimi-k2.6 등으로 설정하고, imageModel은 설정하지 않은 채 이미지를 전송하면 이미지 툴이 노출되지 않거나 "No media-understanding provider registered for opencode-go" 에러가 발생합니다.

Expected behavior

The image tool should work with opencode-go vision-capable models, analyzing the uploaded image using the same model.

기대 동작: OpenCode Go 프로바이더의 비전 모델로 이미지를 분석할 수 있어야 합니다.

Actual behavior

• If no explicit imageModel is configured: createImageTool() returns null because resolveImageModelConfigForTool() cannot find a config-driven fallback for opencode-go
• Even with PR fix(image): keep tool available for vision-capable runtime models #62266's runtime fallback: the actual invocation may still fail because opencode-go has no registered MediaUnderstandingProvider

실제 동작: imageModel 설정이 없으면 createImageTool()이 null을 반환하고, PR #62266의 런타임 fallback이 적용되어도 실제 실행 시 opencode-go에 등록된 MediaUnderstandingProvider가 없어서 실패합니다.

Verified Affected Models

The following opencode-go models are confirmed to show input: ["text"] only in the catalog, despite being vision-capable:

• opencode-go/kimi-k2.6 (my current session model)
• opencode-go/minimax-m2.7
• opencode-go/glm-5
• opencode-go/mimo-v2-omni
• opencode-go/qwen3.6-plus

Note: opencode-go/glm-5.1 is not vision-capable and is not affected by this issue.

확인된 영향 모델: 위 모델들은 실제로 비전 기능을 지원하지만, OpenClaw 카탈로그에서는 input: ["text"]로만 표시됩니다. glm-5.1은 비전 기능이 없어 해당되지 않습니다.

Root Causes

1. Missing MediaUnderstandingProvider registration: extensions/opencode-go/index.ts does not call api.registerMediaUnderstandingProvider(). Other providers like minimax and openai do register one.
2. Missing image input capability in model catalog: Vision-capable models in the opencode-go catalog are listed with only input: ["text"] instead of input: ["text", "image"].

근본 원인:

1. extensions/opencode-go/index.ts에 api.registerMediaUnderstandingProvider() 호출이 없음. minimax, openai 등 다른 프로바이더는 등록되어 있음.
2. 모델 카탈로그에서 비전 모델들이 input: ["text"]로만 표시됨. 실제로는 input: ["text", "image"]여야 함.

Related Issues/PRs

• [Bug]: image tool not available when using provider aliases (e.g. opencode/xxx models) #61789 — image tool not available when using provider aliases (similar runtime issue)
• fix(image): keep tool available for vision-capable runtime models #62266 — fix(image): keep tool available for vision-capable runtime models (Greptile review flagged the exact same missing provider gap for opencode-go)
• [Bug] Image tool broken in 2026.3.22 - No media-understanding provider registered #53093 — Image tool broken in 2026.3.22 (historical regression)

관련 이슈: #61789(유사한 런타임 이슈), #62266(Greptile이 opencode-go의 동일한 provider 누락 지적), #53093(과거 회귀)

OpenClaw version

2026.4.21

Operating system

macOS (but affects all platforms)

Model

opencode-go/kimi-k2.6, opencode-go/minimax-m2.7, opencode-go/glm-5, opencode-go/mimo-v2-omni, opencode-go/qwen3.6-plus

Provider / routing chain

OpenClaw → opencode-go provider → OpenCode Go API

Suggested Fix

1. Add media-understanding-provider.ts to extensions/opencode-go/ following the OpenAI-compatible pattern (OpenCode Go API is OpenAI-compatible)
2. Register it in extensions/opencode-go/index.ts
3. Update opencode-go model catalog to mark vision-capable models with input: ["text", "image"]

제안하는 수정:

1. extensions/opencode-go/에 media-understanding-provider.ts 추가 (OpenAI-compatible 패턴 사용, OpenCode Go API는 OpenAI-compatible)
2. extensions/opencode-go/index.ts에서 등록
3. 비전 모델들의 카탈로그를 input: ["text", "image"]로 업데이트