chatgpt provider strips all image input (attachments and tool-result images); capabilities: {image: true} override has no effect #4521

Versions: docker-agent v1.142.0 and v1.148.0 (linux/amd64, inside a Docker Sandbox), config version 16, provider chatgpt (ChatGPT subscription login), model gpt-6.1-sol.

What happens

Images never reach the model when the agent uses the chatgpt provider:

• docker-agent run agent.yaml --exec --attach screenshot.png "Describe the image" → the model answers that no image is attached.
• Images returned by MCP tools (e.g. a screenshot tool whose result carries images: [{data: <base64 png>}]) are also dropped: the next request's new input is only ~230 tokens.

The model itself accepts images: models.dev lists openai/gpt-6.1-sol with input modalities text, image, pdf, and the same account and model see images through Codex CLI (codex exec -i).

Minimal config

version: '16'
agents:
  root:
    model: sol
    instruction: Describe the attached image.
models:
  sol:
    provider: chatgpt
    model: gpt-6.1-sol
    capabilities:
      image: true

Debug log (--debug), same with and without the capabilities block:

level=DEBUG msg="Failed to resolve model capabilities for message transforms" model=chatgpt/gpt-6.1-sol error="provider \"chatgpt\" not found"
level=DEBUG msg="strip_unsupported_modalities: stripped media part" kind=image role=user reason="model does not support image input"
level=DEBUG msg="Stripped media content from message" role=user original_parts=2 remaining_parts=1

Analysis (from reading the source)

• LocalRuntime.prepareMessagesForModel (pkg/runtime/transforms.go) looks the model up in models.dev as chatgpt/gpt-6.1-sol. models.dev has no chatgpt provider, so catalogModel is nil.
• It then uses cfg.CapsOverride() (pkg/model/provider/contracts), which reads ModelConfig.Capabilities. Because the image is still stripped with capabilities: {image: true} set, the override appears to be nil for the chatgpt provider by the time the runtime sees it. We did not find where it is lost.
• With no catalogue entry and no override, providerFallbackCaps returns empty capabilities, and strip_unsupported_modalities removes every image part before convertDocumentToResponseInput (which would have produced a correct input_image) is reached.

Expected

• The chatgpt provider resolves capabilities from the openai models.dev entry (it is the same model family on the Responses API), and/or
• an explicit capabilities: override is honoured for the chatgpt provider, as the warning text in modelinfo.warnCapsLookupMiss suggests.