session: image content in tool results permanently breaks sessions on text-only models #55044

Description

When a tool result adds image content (e.g. a screenshot returned by an MCP tool, or reading an image file) to a session whose model only accepts text input, every subsequent request fails with HTTP 400 from the provider. OpenCode surfaces the upstream error verbatim with no way to locate or remove the offending message, leaving the session permanently unusable with that model.

Example error, replayed on every turn:

upstream returned HTTP 400: {"object":"error","message":"Model only supports text input; received unsupported content type 'image_url'.","type":"BadRequestError","param":null,"code":400}

The provider config used (redacted):

{
  "model": "nio_ds/glm-5.3",
  "providers": {
    "nio_ds": {
      "package": "aisdk:@ai-sdk/openai-compatible",
      "settings": { "baseURL": "https://<internal-gateway>/openai/v1", "apiKey": "<redacted>" },
      "models": {
        "glm-5.3": {
          "name": "DS-GLM-5.3",
          "limit": { "context": 1048576, "output": 65536 }
        }
      }
    }
  }
}

Expected behavior:

• When capabilities.input: ["text"] is declared for a model (or once the provider has rejected image content), image parts should be stripped or replaced with a text placeholder when assembling requests, so the session stays usable.
• • At minimum, when a provider rejects a request because of image content, the error should identify the offending message and offer a recovery path (remove the message, revert to an earlier turn, or retry without image content).
    Additional context: the root trigger is the uncatalogued-model default of assuming image input support ("These are fallback assumptions, not detected capabilities" per the V2 models docs). With private OpenAI-compatible gateways this assumption is frequently wrong, and a single tool result then breaks the session.

Workarounds:

• Declaring "capabilities": { "tools": true, "input": ["text"], "output": ["text"] } on the model prevents new sessions from hitting this (whether it also repairs an already-affected session by filtering the stored history on send has not been verified).
• • Switching the affected session to a multimodal model (<leader>m) restores it; a compact under that model, then switching back, also unblocks the original model.
• • • <leader>u undo back to before the image-producing turn also works, at the cost of losing later turns.

Plugins

superpowers (git+https://github.com/obra/superpowers.git); MCP servers: playwright, chrome-devtools (local)

OpenCode version

2.0.27

Steps to reproduce

1. Configure a custom OpenAI-compatible provider whose model only accepts text input and is not in models.dev (e.g. a private gateway). Because the model is uncatalogued, OpenCode falls back to assuming text and image input.
2. 2. Start a session with that model and an MCP server that returns screenshots (e.g. playwright get-screenshot).
3. 3. Invoke the screenshot tool. The image enters session history as tool-result content.
4. 4. The next model request fails: upstream returned HTTP 400: ... "Model only supports text input; received unsupported content type 'image_url'".
5. 5. Every subsequent turn replays the same history, including the image, and fails identically. The session is permanently unusable with this model.

Screenshot and/or share link

No response

Operating System

macOS 15 (Darwin 24.6.0, darwin arm64)

Terminal

iTerm.app (TERM=xterm-256color, COLORTERM=truecolor), shell: /bin/zsh