Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Bug: All tool outputs display "(see attached image)" after uploading image to a non-multimodal model

Environment

• OpenClaw version: 2026.6.11 (e085fa1)
• OS: macOS (arm64)
• Mode: WebChat (direct)
• Model: deepseek/deepseek-v4-flash (only model used, no other provider)

Description

After uploading an image (.jpg/.png) to a non-multimodal model (deepseek-v4-flash), all tool outputs in the current session become corrupted. Specifically:

• Every tool (exec, read, write, web_fetch, sessions_spawn, cron, session_status) returns (see attached image) as its output
• Tools actually execute successfully — files are written, commands run, sub-agents are created — but the output cannot be rendered
• session_status and cron list occasionally return correct JSON, but most calls fail

Steps to Reproduce

1. Start a WebChat session using a non-multimodal model (e.g., deepseek-v4-flash)
2. Upload an image (.jpg/.png)
3. Execute any tool and observe the output
4. All tool responses are replaced with (see attached image)

Additional Observations

• Creating a brand new session: the first message works correctly, but the second message fails with:

  Error: reply session initialization conflicted for agent:main:dashboard:xxx
  

• This suggests the issue is not limited to a single corrupted session — it affects the initialization of subsequent sessions as well
• Restarting the Gateway and context compaction did not resolve the issue

Steps to reproduce

Steps to reproduce the behavior:

Start a WebChat session with a non-multimodal model (e.g., deepseek/deepseek-v4-flash)
Upload an image (.jpg/.png) to the session
Execute any tool (e.g., exec, read, write, web_fetch)
Observe that the tool response is (see attached image) instead of actual output
Create a new session and send a message — first reply works correctly
Send a second message in the new session — fails with Error: reply session initialization conflicted

Expected behavior

Non-multimodal models should reject image input gracefully or skip the image data
Tool outputs should render correctly regardless of previous input types
Creating a new session should fully reset the state

Actual behavior

All tool outputs are replaced with (see attached image)
Tools actually execute successfully in the background (files are written, commands run)
New session creation is also affected — initialization fails on the second message
Gateway restart and context compaction do not resolve the issue

OpenClaw version

2026.6.11

Operating system

macOS 26.4.1

Install method

No response

Model

deepseek-v4-flash

Provider / routing chain

DeepSeek

Additional provider/model setup details

No response

Logs

Screenshots, recordings, and evidence

No response

Impact and severity

No response

Additional information

No response