[Bug] WeChat inbound images are downloaded but inaccessible to Claw analyze_image

Start by reading the WeChat inbound image conversion in src/crates/assembly/core/src/service/remote_connect/bot/weixin.rs around lines 220-255, then trace how text-only multimodal rendering handles the attachment in src/crates/assembly/core/src/agentic/execution/execution_engine.rs around lines 1908-1944. Reproduce the text-only Claw flow with a WeChat JPEG and verify that the image is inspectable without a missing workspace-relative path.

Summary

When using BitFun's built-in WeChat gateway with Claw, an inbound image is downloaded successfully but cannot be inspected by a text-only primary model through the analyze_image tool.

The image is present as an inline data URL in the turn payload, but no local attachment path is created. Claw receives only the remote_img_<uuid> identifier and the tool treats it as a workspace-relative path.

Environment

• BitFun: 0.2.17 (macOS Apple Silicon)
• Channel: built-in WeChat gateway / Claw
• Primary model: GLM-5.3 (text-only; image inputs disabled)
• Reproduced: 2026-08-18

Steps to reproduce

1. Configure a Claw session with a text-only primary model.
2. Send a JPEG image to the Claw bot through WeChat.
3. Let Claw try to inspect or describe the image.

Actual behavior

The image is received and displayed in the BitFun transcript, but Claw calls:

analyze_image(path="remote_img_<uuid>")

and fails with:

unable to locate image at <workspace>/remote_img_<uuid>: No such file or directory

The gateway logs show a successful CDN request immediately before the turn starts. The turn payload contains a valid data:image/jpeg;base64,... value, so the media download itself is not the failing step.

Expected behavior

One of these should happen:

• Save the downloaded/decoded image in a session-scoped temporary directory and expose its real image_path; or
• Make analyze_image accept the attachment ID / inline data URL directly; or
• Route the attachment directly to a configured vision-capable primary model when available.

The current prompt says to use analyze_image when a path is available, but the WeChat adapter only supplies a data URL.

Upstream references

The current implementation still follows this path in 0.2.18:

• WeChat inbound image conversion creates only a data URL.
• Text-only multimodal rendering exposes only an image ID when no path exists.

I have not tested outbound file delivery in this report; this issue is specifically about inbound image inspection from WeChat.