PR #63783: fix(webchat): preserve image attachments for text-only models

• Repository: openclaw/openclaw
• Author: helloworldtang
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/63783

Description (problem / solution / changelog)

Summary

• Problem: When the primary model (e.g. GLM-4.7) does not support image input, image attachments from webchat/Control UI are silently dropped. Users receive no feedback and cannot analyze images even when agents.defaults.imageModel is configured with a vision-capable model.
• Why it matters: This makes webchat appear broken for image uploads on any non-vision primary model, regardless of imageModel configuration.
• What changed: Instead of discarding attachments when supportsImages=false, parseMessageWithAttachments now saves images to the media store and injects media://inbound/... references into the message text. This allows the image tool (which uses imageModel) to analyze them via the standard media:// resolution path.
• What did NOT change: The behavior for vision-capable primary models is unchanged. The supportsImages gate itself is unchanged. No changes to detectAndLoadPromptImages or the agent runner.

Change Type

• Bug fix

Scope

• Gateway / orchestration

Linked Issue

• Closes #60644
• Related #61103
• This PR fixes a bug

Root Cause

• Root cause: parseMessageWithAttachments in chat-attachments.ts treats supportsImages=false as a signal to silently discard all image data. The design assumes that if the primary model can't see images, there's no value in preserving them.
• Missing detection / guardrail: No consideration for the case where agents.defaults.imageModel provides a separate vision-capable model. The image tool can process media:// references regardless of the primary model's capabilities.
• Contributing context: The original code comment explicitly warns against leaking media:// markers to text-only models, but these markers are opaque strings that don't cause issues — they're resolved by the image tool, not the primary model.

Regression Test Plan

• Unit test
• Target test: src/gateway/chat-attachments.test.ts — new describe("parseMessageWithAttachments with supportsImages=false") block
• Scenario: When supportsImages=false, images should be saved to disk with media:// refs injected into message text, not dropped. Non-images should be skipped. Empty attachments should return empty.
• Why this is the smallest reliable guardrail: The test directly exercises the changed code path (supportsImages=false branch) and verifies the contract (media refs injected, images array empty, offloadedRefs populated).

User-visible / Behavior Changes

• Users on text-only primary models (e.g. GLM-4.7) with a vision-capable imageModel configured can now successfully upload and analyze images in webchat/Control UI.
• Previously: images were silently dropped with a log warning attachment(s) dropped — model does not support images.
• Now: images are saved and a media:// reference appears in the message, allowing the image tool to process them.

Diagram

Before:
[webchat image] → [chat.send] → supportsImages=false → DROP all attachments → user sees nothing

After:
[webchat image] → [chat.send] → supportsImages=false → save to disk → inject media:// ref in message
  → [agent receives message with media:// ref] → [image tool reads file via media://] → [imageModel analyzes] → response

Security Impact

• New permissions/capabilities? No
• Secrets/tokens handling changed? No
• New/changed network calls? No
• Command/tool execution surface changed? No
• Data access scope changed? No — images are saved to the same media/inbound directory that was already used for large image offloading.

Repro + Verification

Environment

• OS: macOS (ARM)
• Runtime: Node.js 25.8.1
• Model/provider: zai/glm-4.7 (primary, text-only), zai/glm-4.6v (imageModel)
• Integration/channel: webchat / Control UI

Steps

1. Configure a text-only primary model (e.g. glm-4.7) and a vision imageModel (e.g. glm-4.6v)
2. Open OpenClaw Control UI (webchat)
3. Take a screenshot and paste into the chat input
4. Send the message with "analyze this image"

Expected

The agent receives the message with a [media attached: media://inbound/<id>] reference and can use the image tool to analyze it.

Actual (before fix)

Gateway log shows: parseMessageWithAttachments: 1 attachment(s) dropped — model does not support images. The agent receives the text message only, with no indication an image was attached.

Evidence

• Failing test/log before + passing after
  • Before: attachment(s) dropped — model does not support images in gateway.err.log
  • After: attachment(s) saved for text-only model in gateway.log, agent successfully analyzes image

Human Verification

• Verified scenarios:
  • Pasted screenshot in webchat → agent received media:// ref → image tool analyzed successfully
  • Multiple consecutive image uploads → all processed correctly
  • Non-image attachment → correctly skipped
• Edge cases checked: empty message with image, large image (>100KB)
• What I did not verify: ACP bridge clients (the persistChatSendImages function returns early for ACP clients, which is existing behavior)

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

Compatibility / Migration

• Backward compatible? Yes
• Config/env changes? No
• Migration needed? No

Risks and Mitigations

• Risk: Additional disk I/O for text-only models — images are now saved even when the primary model can't use them.
  • Mitigation: This matches the existing behavior for large images (>2MB offload threshold). The media store has existing cleanup mechanisms. For models that support images, the code path is unchanged.

Changed files

• PR_DESCRIPTION.md (added, +113/-0)
• src/gateway/chat-attachments.test.ts (modified, +72/-0)
• src/gateway/chat-attachments.ts (modified, +92/-14)

PR #67184: fix(chat): support non-image attachments in webchat file upload

• Repository: openclaw/openclaw
• Author: harinwu-netizen
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/67184

Description (problem / solution / changelog)

Summary

Fix Issue #60644: WebChat/Control-UI file upload only supports images.

Problem

• CHAT_ATTACHMENT_ACCEPT only accepts image/*
• isSupportedChatAttachmentMimeType() only validates image/* types
• Users cannot upload PDF, video, audio or text files

Solution

• Expand CHAT_ATTACHMENT_ACCEPT to include video/*, audio/*, application/pdf, text/*, application/json
• Update isSupportedChatAttachmentMimeType() to validate all supported types

Testing

• Upload PDF document via webchat
• Upload video file via webchat
• Upload audio file via webchat
• Existing image upload still works

Fixes #60644

Changed files

• ui/src/ui/chat/attachment-support.ts (modified, +12/-2)