What happens

Assign a model from an openai-compatible provider (llama.cpp, vLLM, llama-server, and friends) with netclaw model set, and the saved config comes back with "InputModalities": "Text" and "OutputModalities": "Text" — even when the model handles images or video. Once that value is on disk the daemon never corrects it. A config-level modality is treated as authoritative and short-circuits capability resolution at startup, so the vLLM/llama.cpp backend strategy, the OpenRouter oracle, and the HuggingFace resolver never get a vote.

The visible result is a multimodal model that quietly behaves as text-only. Image attachments get dropped before they ever reach the model, and netclaw daemon status reports input: Text.

Why

The OpenAI-style /v1/models listing has no modality field, so the parser never sets one. It builds each DiscoveredModel with an id and a context window and nothing else:

• netclaw/src/Netclaw.Providers/ProbeHelpers.cs

  Lines 36 to 43 in 60601c6

  	if (model.TryGetProperty("id", out var id))
  	{
  	models.Add(new DiscoveredModel
  	{
  	ModelId = new(id.GetString()!),
  	ContextWindowTokens = readContextWindow(model)
  	});
  	}

• (openai-compatible routes straight through that helper)

  netclaw/src/Netclaw.Providers/SelfHosted/OpenAiCompatibleDescriptor.cs

  Lines 50 to 51 in 60601c6

  	internal static ProviderProbeResult ParseModels(string json)
  	=> ProbeHelpers.ParseOpenAiStyleModels(json, TryReadContextWindow);

DiscoveredModel then falls back to its defaults, which are Text:

• netclaw/src/Netclaw.Configuration/DiscoveredModel.cs

  Lines 31 to 34 in 60601c6

  	public ModelModality InputModalities { get; init; } = ModelModality.Text;
  	/// <summary>Content types the model can produce as output.</summary>
  	public ModelModality OutputModalities { get; init; } = ModelModality.Text;

So discovery hands back Text not because it detected a text-only model, but because it never looked. The persistence step then takes that default and writes it as if an operator had deliberately chosen it:

• netclaw/src/Netclaw.Cli/Model/ModelCommand.cs

  Lines 183 to 186 in 60601c6

  	if (discoveredModel is not null)
  	{
  	modelEntry["InputModalities"] = discoveredModel.InputModalities.ToString();
  	modelEntry["OutputModalities"] = discoveredModel.OutputModalities.ToString();

That is the trap. An unknown gets promoted to a hard assertion, and on the next daemon boot the override beats real detection.

Same bug, second code path

The init wizard does the identical thing through its own code, so a fix in one place won't cover the other:

• netclaw/src/Netclaw.Cli/Tui/Wizard/Steps/ProviderStepViewModel.cs

  Lines 319 to 329 in 60601c6

  	builder.Model = new ModelConfigSection
  	{
  	Provider = providerName,
  	ModelId = SelectedModelId,
  	ContextWindow = selectedModel?.ContextWindowTokens,
  	Provenance = selectedModel is null ? ModelDiscoverySource.Manual : ModelDiscoverySource.Live,
  	InputModalities = selectedModel?.InputModalities,
  	OutputModalities = selectedModel?.OutputModalities,
  	};
  	}

Worth fixing both together, and ideally collapsing them onto one shared write path.

Suggested direction

• When discovery can't actually determine modalities, leave them unset rather than defaulting to Text, and don't persist a value the provider never reported. Unknown should mean "let the daemon resolve this," not "Text, forever."
• Only write modalities to config when they come from a source that genuinely knows. The OpenAI Codex OAuth catalog, for example, already resolves real input_modalities/output_modalities; a self-hosted /v1/models listing does not.

Repro

1. Configure an openai-compatible provider pointed at a server hosting a vision-capable model.
2. netclaw model set Main <provider> <model-id>
3. Look at the saved model entry: InputModalities is Text.
4. netclaw daemon status reports text-only input, and image attachments are dropped, regardless of the model's real capability.

Related

#1267 is about surfacing modalities in model discover/list/TUI. This is the upstream cause of the bad data it would surface: discovery defaults modalities to Text and then bakes them into config as an override.