Do not infer vision from the model name.Model discovery supplies identity and capacities. The operator owns the modality claim, and the endpoint remains the final capability test.

Declare the narrowest owner.

One vision modelmodels[].input

Changes only the exact route entry that has been verified.

Every undeclared modeldefaultInput

Broadens the provider fallback. Use only when the provider contract covers all models.

Existing provider rowStop the writer and preserve the original settings file first.
llm-pi-ai:
  providers:
    aiping:
      models:
        - id: GLM-5.3-Flash
          input: [text, image]

Four visible fields. Hidden data still matters.

ID

visible

Name

visible

Context

visible

Max output

visible

The rc.2 editor keeps model drafts structurally open, so an unrelated visible edit is intended to preserve hidden fields such as input. Removing a row, resetting the model array, or replacing the provider still removes its data deliberately.

Six proof gates.

A YAML write is not a capability test.

If the provider rejects the image, remove the incorrect claim. Do not disable attachment validation or broaden every model through defaultInput.

Primary evidence.

Keep the complete provider workflow.

The canonical guide covers custom providers, Ollama, image metadata, round-trip safety, route verification, provider recovery, and troubleshooting.

Read the complete guide