Repository navigation
Bug: opencode-go/kimi-k2.6 fails in child sessions with "image.png" error #24454
Description
Description
Bug: opencode-go/kimi-k2.6 fails in child sessions with image.png error
Description
When using opencode-go/kimi-k2.6 as the default model in a custom agent configuration, it works correctly in the main conversation thread (including with screenshots/images). However, when OpenCode creates a child session (new thread, subagent call), it fails with:
ERROR: Cannot read "image.png" (this model does not support image input).
This error occurs even when no actual image file exists. It appears OpenCode is generating an internal reference to image.png in child sessions and then incorrectly rejecting it based on the model's reported capabilities.
Steps to Reproduce
- Configure
opencode-go/kimi-k2.6as the default model inopencode.jsonc - Open OpenCode desktop app (Windows) in a project directory
- In the main thread, send a message — it works correctly
- Open a new thread/tab in the same project
- The new thread fails with:
Provider returned error/Cannot read "image.png"
Expected Behavior
opencode-go/kimi-k2.6 should work identically in both main and child sessions, since Kimi K2.6 does support image input.
Actual Behavior
- Main thread: ✅ Works fine with images
- Child session / new thread: ❌ Fails with image.png error
- No
image.pngfile exists on disk
Workaround
Switching to opencode-go/glm-5.1 or opencode-go/qwen3.5-plus resolves the issue.
Environment
- OpenCode version: 1.14.25
- Platform: Windows desktop app
- Provider: OpenCode Go
- Model: opencode-go/kimi-k2.6
Additional Notes
The opencode-go/glm-5.1 model works correctly in the same child session context, suggesting this is specifically a model capabilities metadata issue with kimi-k2.6 in the OpenCode Go provider — it appears to be incorrectly flagged as not supporting image input, causing OpenCode to reject it before the request even reaches the model.
Confirmed working in child sessions:
opencode-go/glm-5.1✅opencode-go/qwen3.5-plus✅opencode-go/qwen3.6-plus✅openai/gpt-5.5✅openai/gpt-5.4-mini✅
Plugins
No response
OpenCode version
No response
Steps to reproduce
No response
Screenshot and/or share link
No response
Operating System
No response
Terminal
No response
Activity
github-actions commentedon Apr 26, 2026
This issue might be a duplicate of existing issues. Please check:
If your issue is distinct (e.g., specific to the opencode-go/kimi-k2.6 model or the child session context), please add that clarification so it can be tracked separately.
samuel-pro-ai commentedon Jun 12, 2026
just tested 7 opencode-go models on v1.17.3 (windows, cursor terminal).
3 work, 4 are broken, 1 inconclusive:
WORKING (both direct chat attach + read tool path):
BROKEN (same "Cannot read clipboard (this model does not support image input)" error on both paths):
INCONCLUSIVE (connection error, cant tell if vision is broken or its just a server issue):
so this is definitely a per-model MediaUnderstandingProvider registration gap. the underlying models can all see images (zen prefix works on the same names per the research notes) but the go provider just isnt registering them properly.
I also tried the user-level modalities config workaround (per issue #9897) - doesnt work. the provider registry ignores it.
a maintainer fix would be to either:
would love to see this consolidated. happy to retest when a fix lands.
github-actions commentedon Aug 12, 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.