Repository navigation
GLM-5.2 session broken when model foolishly tries to view a screenshot #34113
Description
Description
For context, the GLM-5.2 model does not support image input.
Testing OpenCode desktop with GLM-5.2 via LiteLLM.
The agent found a skill that involved looking at screenshots and seems to have triggered image input somehow.
The error message itself is fine, since GLM-5.2 does not in fact support image inputs. The problem is that this borked my session. Now every message I send gives me that same error message.
Need to prevent image inputs from contaminating the context for GLM-5.2.
The best case would be an image viewer tool that models like GLM-5.2 could use to get descriptions of images vai a subagent with image capabilities ... but that's a feature request.
Plugins
No response
OpenCode version
OpenCode Desktop v1.17.11
Steps to reproduce
- Setup LiteLLM with a fireworks.ai API key
- Start a thread in OpenCode desktop running GLM-5.2
- Tell it to take / look at a screenshot
- Observe "invalid_request_error: This model does not support image inputs"
- Same error for subsequent user messages
Operating System
Ubuntu 24.04.3 LTS
Terminal
No response
Metadata
Metadata
Assignees
Labels
No labels
Activity
[-]GLM-5.2 session broken when model foolishly tries to read a screenshot[/-][+]GLM-5.2 session broken when model foolishly tries to view a screenshot[/+]github-actions commentedon Jun 26, 2026
This issue might be a duplicate of existing issues. Please check:
andrewlucas-bg commentedon Jun 26, 2026
It is the same issue as 29571. But the bug is deeper than GitHub Copilot or GLM. Harness issue.
Olbrasoft commentedon Jun 27, 2026
Same use case here — running GLM-5.2 as primary via OpenCode Go subscription and needing a multimodal subagent for occasional screenshot analysis.
After checking all OpenCode Zen models, the only free multimodal option is MiMo-V2.5 Free (text + image + video + audio). All other free Zen models (Big Pickle, DeepSeek V4 Flash Free, North Mini Code Free, Nemotron 3 Ultra Free) are text-only. MiniMax-M3 and GLM-5V-Turbo are multimodal but pay-per-use.
Strongly +1 for merging PR #32302 — the "forward parent attachments to subagents" fix is exactly what's needed to make this image-via-subagent workflow work with GLM-5.2 as primary. With proper AGENTS.md instructions directing GLM-5.2 to delegate image inputs to a MiMo-V2.5 Free subagent via Task tool, this would solve the broken-session problem (image error contaminates parent context) without forcing users to manually switch models back and forth.
Would love to see #32302 merged so we can build an
image-analyzersubagent config on top of it.github-actions commentedon Aug 27, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.