Custom OpenAI-compatible provider via LiteLLM is incorrectly classified as text-only / no-reasoning and blocks image input #22673

Activity

1. Promental

   added

   bugSomething isn't workingSomething isn't working

   on Apr 15, 2026
2. github-actions

   assigned

   Brendonovich

   on Apr 15, 2026
3. github-actions

   added

   coreAnything pertaining to core functionality of the application (opencode server stuff)Anything pertaining to core functionality of the application (opencode server stuff)

   webRelates to opencode on web / desktopRelates to opencode on web / desktop

   windows

   on Apr 15, 2026
4. freelanceagent1 commented on Apr 17, 2026

   

   freelanceagent1

   on Apr 17, 2026

   More actions

   you should define those in opencode.json
   no?
5. Promental commented on Apr 18, 2026

   

   Promental

   on Apr 18, 2026

   Author

   More actions

   Thanks. I considered that, but this seems not to be only an opencode.json issue.

   The same OpenAI-compatible endpoint and models accept image input in another client, while OpenCode classifies them as text-only / no-reasoning immediately after setup.

   That suggests the limitation is in capability detection / provider mapping for custom LiteLLM providers, not just in user configuration.

   If OpenCode expects explicit capability flags in opencode.json, could you share the exact supported fields and an example for enabling vision + reasoning for a custom provider?
6. XIYBHK commented on Apr 21, 2026

   

   XIYBHK

   on Apr 21, 2026

   More actions

   could you share the exact supported fields

   The field is modalities on each model entry. Capability detection in packages/opencode/src/provider/provider.ts reads it as:

   capabilities.input.image = model.modalities?.input?.includes("image") ?? false

   If absent (the default when users register a custom model), capabilities.input.image evaluates to false and opencode strips image attachments client-side before the request is sent — which matches your "classified as text-only / blocks image input" symptom exactly.

   Minimal config to unblock vision for your LiteLLM-backed models:

   "provider": {
     "litellm": {
       "npm": "@ai-sdk/openai-compatible",
       "options": { "baseURL": "https://your-litellm-gateway/v1" },
       "models": {
         "claude-sonnet-4-6": {
           "modalities": { "input": ["text", "image"], "output": ["text"] }
         },
         "gpt-5.3-codex": {
           "modalities": { "input": ["text", "image"], "output": ["text"] },
           "reasoning": true
         },
         "deepseek-v3.2": {
           "modalities": { "input": ["text"], "output": ["text"] }
         }
       }
     }
   }

   Restart opencode (capability cache is built once at startup, not hot-reloaded).

   See #20802 for the full root-cause writeup — this is the same gating check manifesting across every custom @ai-sdk/openai-compatible provider.

   👍React with 👍2Reacted by Kevin Wang and Nik
7. thdxr

   removed

   bugSomething isn't workingSomething isn't working

   windows

   webRelates to opencode on web / desktopRelates to opencode on web / desktop

   coreAnything pertaining to core functionality of the application (opencode server stuff)Anything pertaining to core functionality of the application (opencode server stuff)

   on May 3, 2026
8. github-actions commented on Jul 5, 2026

   

   github-actionsBot

   on Jul 5, 2026 – with GitHub Actions

   Contributor

   More actions

   To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.

   👎React with 👎1Reacted by Vlad L.
9. github-actions

   closed this as not plannedon Jul 5, 2026