openclaw - 💡(How to fix) Fix image_generate outputs should be accessible from sandbox workspace for follow-up workflows [1 comments, 2 participants]

Summary

image_generate returns a MEDIA: path under /home/<user>/.openclaw/media/tool-image-generation/..., but sandboxed agents cannot read that file or use it as input for follow-up tools such as exec, basecamp attach, local image processing, OCR/text verification, or archiving.

For agent workflows, generated media should ideally be materialized in a workspace-readable location, or exposed through a first-class copy/export handle.

Observed

A successful image generation returned:

MEDIA:/home/aaajiao/.openclaw/media/tool-image-generation/image-1---c8371631-4715-4cd8-bebb-b8ac962cf45c.png

The user can see the image in chat, but the agent cannot access it from the sandbox workspace:

read /home/aaajiao/.openclaw/media/... -> Path escapes sandbox root (~/.openclaw/workspace)
exec ls /home/aaajiao/.openclaw/media/... -> No such file or directory

So the generated image can be delivered to the user, but cannot be composed into the next step of the workflow from inside the sandbox.

Why this matters

A common agent media workflow is:

1. Generate image with image_generate
2. Inspect / verify it
3. Optionally modify it locally
4. Upload it elsewhere, e.g. Basecamp, website, archive, repo, etc.

Currently step 1 succeeds, but step 2+ fail unless the agent regenerates the image manually into /workspace via a separate script/API call. That defeats the purpose of having image_generate as a first-class tool.

Expected behavior

One of these would solve it:

• Save/copy generated media into a workspace-readable path such as /workspace/media/generated/...
• Or return both paths:
  • delivery path for MEDIA: rendering
  • sandbox/workspace path for agent follow-up work
• Or provide a first-class media export/copy tool that can copy MEDIA: outputs into the current workspace
• Or allow tools that accept file inputs to consume prior MEDIA: artifact references directly

Environment

• OpenClaw: v2026.4.29
• Agent: main session, sandboxed runtime
• Workspace: /workspace
• Host repo: /home/aaajiao/.openclaw/workspace
• Tool: image_generate

Concrete use case

We generated poster mockups with image_generate. The output was visible in Telegram, but could not be uploaded to Basecamp from the agent because basecamp attach runs in the sandbox and only sees /workspace files. The workaround was to bypass image_generate and call the OpenAI Images API from a Python script to save the image directly under /workspace/media/generated/, then upload that file.

That workaround is unnecessary friction if generated tool media can be workspace-addressable.