Rebuilt how AI Studio knows what a model can do (#960)

app/MindWork AI Studio/Models/Google/GeminiFamily.cs

@@ -0,0 +1,91 @@
+using static AIStudio.Provider.Capability;
+
+namespace AIStudio.Models.Google;
+
+/// <summary>
+/// The Gemini chat models.
+/// </summary>
+/// <remarks>
+/// A Gemini reads everything -- text, images, audio, speech, video -- writes text, and calls tools.
+/// That is the first rule, and it is the family's own fallback for a Gemini nobody has written a
+/// rule for yet. What the generations add to it is how they think, and the older exceptions take
+/// something away instead.
+///
+/// Every generation gets a line of its own, including the dotted ones. The dot is a version
+/// boundary rather than a name part boundary, deliberately -- it is what keeps llama3 and llama3.1
+/// apart -- so a rule for "gemini-3" does not answer for "gemini-3.1", and each has to say so
+/// itself. The previous rules searched for "gemini-3" anywhere in the name and covered unreleased
+/// versions by accident; the price of not doing that is a line per generation, and the verification
+/// run names any model of the corpus which finds no rule.
+/// </remarks>
+public sealed class GeminiFamily : ModelFamily
+{
+    /// <inheritdoc />
+    public override ModelVendor Vendor => ModelVendor.GOOGLE;
+
+    /// <inheritdoc />
+    public override ModelSource Source => new("https://ai.google.dev/gemini-api/docs/gemini-3", new DateOnly(2026, 9, 12), "Capabilities ported unchanged from ProviderExtensions.Google.cs: one shape for all of Gemini, one sentence per generation about thinking. The Gemini 3 guide states a one million token input window; the 2.5 model pages state their input limit as 1,048,576, and both numbers are written here as their page gives them.");
+
+    /// <inheritdoc />
+    public override IReadOnlyList<ModelSource> FurtherSources =>
+    [
+        new("https://ai.google.dev/gemini-api/docs/image-understanding", new DateOnly(2026, 9, 12), "States one number for the whole family: \"Gemini models support a maximum of 3,600 image files per request.\" The 20 MB it also names is a limit on the request body rather than on the number of images."),
+        new("https://ai.google.dev/gemini-api/docs/tokens", new DateOnly(2026, 9, 12), "Google publishes no tokenizer file for Gemini. Counting happens through the countTokens method of the API, which returns the number of tokens of the input alone.")
+    ];
+
+    /// <inheritdoc />
+    protected override void Declare(ModelFamilyBuilder builder)
+    {
+        //
+        // Google states the image limit once, for all of Gemini, so it sits on the fallback and
+        // every generation inherits it. The two rules below which do not inherit from here say
+        // nothing about it: the live model looks at no still images at all, and for the 1.0 vision
+        // model Google's current pages state no number any more.
+        //
+        builder.Rule("gemini").AsSegment()
+            .Capabilities(TEXT_INPUT | MULTIPLE_IMAGE_INPUT | AUDIO_INPUT | SPEECH_INPUT | VIDEO_INPUT | TEXT_OUTPUT | FUNCTION_CALLING)
+            .Apis(CHAT_COMPLETION_API)
+            .Images(maxPerRequest: 3_600)
+            .Tokenizer(TokenizerKind.PROVIDER_API, "countTokens");
+
+        // The one Gemini which only ever read text and images:
+        builder.Rule("gemini-1.0-pro-vision").AsPrefix()
+            .Capabilities(TEXT_INPUT | MULTIPLE_IMAGE_INPUT | TEXT_OUTPUT)
+            .Apis(CHAT_COMPLETION_API);
+
+        //
+        // The live model, which belongs to a different API: it speaks back, and it is the one
+        // Gemini that does not look at still images.
+        //
+        builder.Rule("gemini-2.0-flash-live").AsPrefix()
+            .Capabilities(TEXT_INPUT | AUDIO_INPUT | SPEECH_INPUT | VIDEO_INPUT | TEXT_OUTPUT | SPEECH_OUTPUT | FUNCTION_CALLING)
+            .Apis(CHAT_COMPLETION_API);
+
+        //
+        // Google states an input limit and an output limit rather than one window. The input limit
+        // is the one a conversation is measured against, because that is where the conversation
+        // accumulates, so that is the number written here.
+        //
+        builder.Rule("gemini-2.5").AsPrefix().InheritsFrom("gemini")
+            .Reasoning(ReasoningSupport.ALWAYS)
+            .ContextWindow(1_048_576);
+
+        //
+        // The one exception of the 2.5 line: it can think, but only when asked. From the 3.x line
+        // on, even the Flash Lite models think at their lowest level.
+        //
+        builder.Rule("gemini-2.5-flash-lite").AsPrefix().InheritsFrom("gemini-2.5")
+            .Reasoning(ReasoningSupport.OPTIONAL);
+
+        builder.Rule("gemini-3").AsPrefix().InheritsFrom("gemini")
+            .Reasoning(ReasoningSupport.ALWAYS)
+            .ContextWindow(1_000_000);
+
+        builder.Rule("gemini-3.1").AsPrefix().InheritsFrom("gemini-3");
+        builder.Rule("gemini-3.7").AsPrefix().InheritsFrom("gemini-3");
+
+        // The two rolling aliases, which carry no version number and point at the current line:
+        builder.Rule("gemini-flash-latest").AsExact().InheritsFrom("gemini-3");
+        builder.Rule("gemini-pro-latest").AsExact().InheritsFrom("gemini-3");
+    }
+}

app/MindWork AI Studio/Models/Google/GeminiImageFamily.cs

@@ -0,0 +1,42 @@
+using static AIStudio.Provider.Capability;
+
+namespace AIStudio.Models.Google;
+
+/// <summary>
+/// The Gemini models which draw as well as write.
+/// </summary>
+/// <remarks>
+/// They are named like every other Gemini, with a version and a size, and the only thing setting
+/// them apart is the name part "image". So the rules here are the generation rules of the chat
+/// family with that one part required on top, and requiring it is exactly what makes them win: two
+/// rules reaching equally far into a name are separated by how many conditions they carry.
+///
+/// What they can do is nearly the opposite of what their generation can. They write images, which
+/// no chat Gemini does, and they call no tools, which every chat Gemini does. Reading them as chat
+/// models of their line -- which is what happens when nobody asks about the image part first --
+/// promises tool calling that is not there.
+/// </remarks>
+public sealed class GeminiImageFamily : ModelFamily
+{
+    /// <inheritdoc />
+    public override ModelVendor Vendor => ModelVendor.GOOGLE;
+
+    /// <inheritdoc />
+    public override ModelSource Source => new("https://ai.google.dev/gemini-api/docs/image-generation", new DateOnly(2026, 9, 11), "Ported unchanged from the rules in ProviderExtensions.Google.cs: images out, no tool calling, and thinking from the 3 line on.");
+
+    /// <inheritdoc />
+    protected override void Declare(ModelFamilyBuilder builder)
+    {
+        builder.Rule("gemini-2.5").AsPrefix().AlsoContains("image")
+            .Capabilities(TEXT_INPUT | MULTIPLE_IMAGE_INPUT | TEXT_OUTPUT | IMAGE_OUTPUT)
+            .Apis(CHAT_COMPLETION_API);
+
+        // From the 3 line on they think about a complicated prompt, and it cannot be switched off:
+        builder.Rule("gemini-3").AsPrefix().AlsoContains("image").Inherits()
+            .Reasoning(ReasoningSupport.ALWAYS);
+
+        // Only the 3.1 Flash image models watch video:
+        builder.Rule("gemini-3.1").AsPrefix().AlsoContains("image").Inherits()
+            .Capabilities(VIDEO_INPUT);
+    }
+}

app/MindWork AI Studio/Models/Google/GemmaFamily.cs

@@ -0,0 +1,78 @@
+using static AIStudio.Provider.Capability;
+
+namespace AIStudio.Models.Google;
+
+/// <summary>
+/// Gemma, the open weights Google publishes next to Gemini.
+/// </summary>
+/// <remarks>
+/// Two generations and one spelling problem. Ollama writes "gemma3:27b" and the hub writes
+/// "gemma-3-27b-it", and no normalization turns one into the other, so each statement stands twice.
+/// What it buys is that the rules never have to ask who served the model.
+///
+/// Tool calling is the line between the generations. What Google documents for Gemma 3 is writing
+/// the tool descriptions into the prompt by hand, which is a different thing from what the tools
+/// field of an OpenAI-compatible request does: the chat template has neither a tool role nor tool
+/// tokens, and Ollama refuses a request carrying tools for these models. Gemma 4 is the first with
+/// tokens of its own, and the first that thinks -- when the request opens the thinking channel.
+/// </remarks>
+public sealed class GemmaFamily : ModelFamily
+{
+    /// <inheritdoc />
+    public override ModelVendor Vendor => ModelVendor.GOOGLE;
+
+    /// <inheritdoc />
+    public override ModelSource Source => new("https://ai.google.dev/gemma/docs/core", new DateOnly(2026, 9, 11), "Ported unchanged from the Gemma block of ProviderExtensions.OpenSource.cs.");
+
+    /// <inheritdoc />
+    protected override void Declare(ModelFamilyBuilder builder)
+    {
+        // The early generations take text only and were not built for tools:
+        builder.Rule("gemma").AsSubstring()
+            .Capabilities(TEXT_INPUT | TEXT_OUTPUT)
+            .Apis(CHAT_COMPLETION_API);
+
+        // Gemma 3 reads pictures from the 4B checkpoint upwards:
+        builder.Rule("gemma3").AsSubstring()
+            .Capabilities(TEXT_INPUT | MULTIPLE_IMAGE_INPUT | TEXT_OUTPUT)
+            .Apis(CHAT_COMPLETION_API);
+
+        builder.Rule("gemma-3").AsSubstring().Inherits();
+
+        // The 1B checkpoint is the one that does not:
+        builder.Rule("gemma3").AsSubstring().AlsoContains("1b")
+            .Capabilities(TEXT_INPUT | TEXT_OUTPUT)
+            .Apis(CHAT_COMPLETION_API);
+
+        builder.Rule("gemma-3").AsSubstring().AlsoContains("1b").Inherits();
+
+        //
+        // The 3n checkpoints listen as well. Video is not a modality of any Gemma: the model cards
+        // list text, image, and audio, and mention video only as frames somebody else cut it into.
+        //
+        builder.Rule("gemma3n").AsSubstring()
+            .Capabilities(TEXT_INPUT | MULTIPLE_IMAGE_INPUT | AUDIO_INPUT | TEXT_OUTPUT)
+            .Apis(CHAT_COMPLETION_API);
+
+        builder.Rule("gemma-3n").AsSubstring().Inherits();
+
+        // Every checkpoint of Gemma 4 is multimodal; there is no text-only variant of it:
+        builder.Rule("gemma4").AsSubstring()
+            .Capabilities(TEXT_INPUT | MULTIPLE_IMAGE_INPUT | TEXT_OUTPUT | FUNCTION_CALLING)
+            .Apis(CHAT_COMPLETION_API)
+            .Reasoning(ReasoningSupport.OPTIONAL);
+
+        builder.Rule("gemma-4").AsSubstring().Inherits();
+
+        // Three of its checkpoints hear, and they are named one by one because that is all they
+        // have in common:
+        builder.Rule("gemma4").AsSubstring().AlsoContains("e2b").Inherits().Capabilities(AUDIO_INPUT);
+        builder.Rule("gemma-4").AsSubstring().AlsoContains("e2b").Inherits();
+
+        builder.Rule("gemma4").AsSubstring().AlsoContains("e4b").Inherits();
+        builder.Rule("gemma-4").AsSubstring().AlsoContains("e4b").Inherits();
+
+        builder.Rule("gemma4").AsSubstring().AlsoContains("12b").Inherits();
+        builder.Rule("gemma-4").AsSubstring().AlsoContains("12b").Inherits();
+    }
+}

app/MindWork AI Studio/Models/Google/GoogleEmbeddingFamily.cs

@@ -0,0 +1,35 @@
+using AIStudio.Provider;
+
+using static AIStudio.Provider.Capability;
+
+namespace AIStudio.Models.Google;
+
+/// <summary>
+/// Google's embedding models, which turn text into a vector and answer nothing.
+/// </summary>
+/// <remarks>
+/// The previous rules answered for these with the Google default: images in, text out, tool
+/// calling. None of it is true, and the app already knows better -- it asks every provider for its
+/// embedding models through a method of its own.
+///
+/// The Gemini one needs a rule of its own for another reason: its name begins with "gemini", so
+/// without one it would be read as a chat model of the family.
+/// </remarks>
+public sealed class GoogleEmbeddingFamily : ModelFamily
+{
+    /// <inheritdoc />
+    public override ModelVendor Vendor => ModelVendor.GOOGLE;
+
+    /// <inheritdoc />
+    public override ModelSource Source => new("https://ai.google.dev/gemini-api/docs/embeddings", new DateOnly(2026, 9, 11), "The app lists these under IProvider.GetEmbeddingModels, which is where the statement that they embed comes from.");
+
+    /// <inheritdoc />
+    protected override void Declare(ModelFamilyBuilder builder)
+    {
+        builder.Rule("text-embedding-004").AsExact()
+            .Capabilities(TEXT_INPUT | EMBEDDING)
+            .Kind(ModelKind.EMBEDDING);
+
+        builder.Rule("gemini-embedding").AsPrefix().Inherits();
+    }
+}

app/MindWork AI Studio/Models/Google/ImagenFamily.cs

@@ -0,0 +1,32 @@
+using AIStudio.Provider;
+
+using static AIStudio.Provider.Capability;
+
+namespace AIStudio.Models.Google;
+
+/// <summary>
+/// Imagen, which draws a picture from a description and does nothing else.
+/// </summary>
+/// <remarks>
+/// The previous rules had no branch for it. Its name does not contain "gemini", so it fell to the
+/// last line of the Google function and was answered as a chat model: reads images, writes text,
+/// calls functions. Not one of the three is true, and the one thing it does -- writing an image --
+/// was not said at all.
+///
+/// Whole name parts, not a substring: "imagen" also sits inside "imagenet" and "reimagined", and a
+/// chat model carrying such a word would be turned into an image generator by a careless match.
+/// </remarks>
+public sealed class ImagenFamily : ModelFamily
+{
+    /// <inheritdoc />
+    public override ModelVendor Vendor => ModelVendor.GOOGLE;
+
+    /// <inheritdoc />
+    public override ModelSource Source => new("https://ai.google.dev/gemini-api/docs/imagen", new DateOnly(2026, 9, 11), "A description goes in and an image comes out; there is no conversation and no tool calling.");
+
+    /// <inheritdoc />
+    protected override void Declare(ModelFamilyBuilder builder) =>
+        builder.Rule("imagen").AsSegment()
+            .Capabilities(TEXT_INPUT | IMAGE_OUTPUT)
+            .Kind(ModelKind.IMAGE_GENERATION);
+}