openclaw - 💡(How to fix) Fix image_generate returns MEDIA path but Telegram auto-reply can finish text-only without sending image [2 comments, 2 participants]

Summary

In a Telegram auto-reply session, image_generate successfully generated and saved an image, returning a MEDIA:<local path> line, but the assistant then replied only with text (Klaar ...) and no image was delivered to Telegram.

This is confusing for users because the bot says the image is ready, but nothing appears in chat.

Environment

• OpenClaw: 2026.5.3-1
• OS: Windows 11 / win32
• Channel: Telegram direct chat
• Agent model/runtime after workaround: openai-codex/gpt-5.5 via PI runner
• Image generation model: openai/gpt-image-1.5

Observed behavior

Session transcript showed:

Generated 1 image with openai/gpt-image-1.5.
Warning: Ignored unsupported overrides for openai/gpt-image-1.5: aspectRatio=9:16.
MEDIA:C:\Users\...\.openclaw\media\tool-image-generation\bad-iphone-phone-screen-trial-in-bio---<uuid>.png

The generated file existed locally and was a valid PNG:

1024x1536, about 2.9 MB

But the final assistant reply was only:

Klaar — 9:16 verticaal, met jouw screenshot als phone screen.

No Telegram media was sent.

Manual delivery of the same file worked:

openclaw message send --channel telegram --account colorbestie --target <chat_id> --message "..." --media "C:\...\bad-iphone...png" --json

That returned ok: true.

Why this matters

The tool description says generated images are delivered automatically from the tool result as MEDIA paths, but in this real Telegram turn the MEDIA path stayed inside the transcript and did not become an outbound media payload.

Suggested fix

Consider one of these behaviors:

1. Convert MEDIA:<path> returned by image_generate into an outbound media payload automatically when the final response is built.
2. Make the tool result structured, e.g. details.media.mediaUrls, drive final payload construction from that, and prevent final text-only success replies when generated media is unsent.
3. Update the tool contract and agent instructions so that agents must call the message tool or emit a markdown/local media attachment before claiming completion.

Expected behavior

If image_generate succeeds during a Telegram/chat turn, the generated image should be delivered to the user automatically or the agent should be forced/guided to return a media payload instead of text-only completion.