v3.11.0: Claude models from the Models API, benchmark log buttons, per-drawing chat keeps choices

ClaudeRevit.PaneTests/ClaudeRevit.PaneTests.csproj

@@ -31,6 +31,7 @@
     <Compile Include="..\ClaudeRevit\UI\ChatMessage.cs" Link="ChatMessage.cs" />
     <Compile Include="..\ClaudeRevit\Services\McpAgentSelection.cs" Link="McpAgentSelection.cs" />
     <Compile Include="..\ClaudeRevit\Services\CodexModels.cs" Link="CodexModels.cs" />
+    <Compile Include="..\ClaudeRevit\Services\ClaudeModels.cs" Link="ClaudeModels.cs" />
     <Compile Include="..\ClaudeRevit\Services\DocumentSessions.cs" Link="DocumentSessions.cs" />
   </ItemGroup>
 </Project>

ClaudeRevit.PaneTests/Program.cs

@@ -46,7 +46,7 @@ private static void Main(string[] args)
                 effort.SelectedValue = "high";
                 Check(SettingsStore.GetAgentSelection("codex").Effort == "high", "Codex effort was not saved");
                 agent.SelectedIndex = 1;
-                model.SelectedValue = "opus";
+                model.SelectedValue = "claude-opus-5-5";
                 effort.SelectedValue = "medium";
                 agent.SelectedIndex = 2;
                 Check((string)model.SelectedValue == "gpt-test" && (string)effort.SelectedValue == "high", "Codex choices were not restored");
@@ -58,7 +58,7 @@ private static void Main(string[] args)
                 await (Task)send.Invoke(pane, null)!;
                 Check(ChatService.SentSelection?.Model == "custom-model-id", "Typed model was ignored on send");
                 Check(ChatService.SentSelection?.Agent == "codex", "Wrong agent on send");
-                Check(SettingsStore.GetAgentSelection("claudecode").Model == "opus", "Claude model was overwritten by Codex");
+                Check(SettingsStore.GetAgentSelection("claudecode").Model == "claude-opus-5-5", "Claude model was overwritten by Codex");
                 model.SelectedValue = "gpt-test";
                 effort.SelectedValue = "high";
                 ChatService.Pending = new();
@@ -117,15 +117,15 @@ private static void Main(string[] args)
                 Check(((ItemsControl)pane.FindName("AttachmentList")).Items.Count == 0, "A's attachments leaked into B");
                 Check(pane.Messages.Count == 0, "Another project's transcript was shown in B");
                 agent.SelectedIndex=1;
-                model.SelectedValue="sonnet";
+                model.SelectedValue="claude-sonnet-5-5";
                 effort.SelectedValue="low";
                 pane.Messages.Add(new ChatMessage { Role = "user", Text = "Project B" });
                 DocumentHarness.Change("a", false);
                 Check(((ItemsControl)pane.FindName("AttachmentList")).Items.Count == 1, "Returning to A lost its pending attachment");
                 Check(pane.Messages.Single().Text == "Project A", "Returning to A lost its transcript");
                 Check((string)model.SelectedValue=="gpt-test" && (string)effort.SelectedValue=="high","A's selected model/effort leaked from B");
                 agent.SelectedIndex=1;
-                Check((string)model.SelectedValue=="opus" && (string)effort.SelectedValue=="medium","A's other provider choices leaked from B");
+                Check((string)model.SelectedValue=="claude-opus-5-5" && (string)effort.SelectedValue=="medium","A's other provider choices leaked from B");
                 agent.SelectedIndex=2;
                 ChatService.Pending = new();
                 ((TextBox)pane.FindName("InputBox")).Text = "Pending A turn";

ClaudeRevit.PaneTests/Stubs.cs

@@ -46,6 +46,12 @@ public void ClearHistory() { }
         public void RecreateClient() { }
         public void Activate() { }
     }
+    public static class ClaudeModelCatalog
+    {
+        public static bool LastWasLive { get; private set; }
+        public static Task<IReadOnlyList<ClaudeModelInfo>> ReadAsync(CancellationToken ct, bool forceRefresh = false) =>
+            Task.FromResult(ClaudeModels.Current);
+    }
     public static class CodexModelCatalog
     {
         public static Task<IReadOnlyList<CodexModel>> ReadAsync(string exe, string workDir, CancellationToken ct, bool forceRefresh = false) =>

ClaudeRevit.Tests/ClaudeModelsTests.cs

@@ -10,7 +10,9 @@ public class ClaudeModelsTests
     [InlineData("sonnet-5-5", "claude-sonnet-5-5")]
     [InlineData("haiku-5-5", "claude-haiku-5-5")]
     [InlineData("fable-5-1", "claude-fable-5-1")]
-    [InlineData("auto", "claude-sonnet-5")]
+    [InlineData("fable-5-2", "claude-fable-5-2")]
+    [InlineData("claude-opus-5-5", "claude-opus-5-5")]
+    [InlineData("auto", "claude-sonnet-5-5")]
     public void TagsMapToApiIds(string tag, string id) => Assert.Equal(id, ClaudeModels.ApiId(tag));
 
     // Served ids come back dashed or in the SDK's enum spelling; the newer ids must not fall
@@ -27,12 +29,40 @@ public class ClaudeModelsTests
     [InlineData("claude-fable-5", "fable-5")]
     [InlineData("claude-opus-4-8", "opus-4-8")]
     [InlineData("claude-sonnet-5", "sonnet-5")]
+    [InlineData("claude-fable-5-2", "fable-5-2")]
+    [InlineData("claude-opus-4-5-20251101", "opus-4-5")]
+    [InlineData("claude-sonnet-5-20260101", "sonnet-5")]
     public void ServedModelsMapToRateTags(string id, string tag) => Assert.Equal(tag, ClaudeModels.TagFromModel(id));
 
+    // The pickers follow the Models API: Claude ids only, in the API's order, without repeats,
+    // and a missing display name is derived from the id.
+    [Fact]
+    public void ServedCatalogBecomesPickerEntries()
+    {
+        var list = ClaudeModels.FromServed(new (string, string?)[]
+        {
+            ("claude-fable-5-2", "Claude Fable 5.2"), ("claude-opus-5-5", null), ("gpt-5", "GPT"),
+            ("claude-opus-5-5", "dup"), ("claude-haiku-4-5", "Claude Haiku 4.5"),
+        });
+        Assert.Equal(new[] { "fable-5-2", "opus-5-5", "haiku-4-5" }, list.Select(m => m.Tag));
+        Assert.Equal("Claude Opus 5.5", list[1].Name);
+        Assert.Equal(new[] { "fable-5-2", "opus-5-5", "sonnet-5-5", "haiku-5-5" }, ClaudeModels.Current.Select(m => m.Tag));
+    }
+
+    [Theory]
+    [InlineData("opus-5-5", true)]
+    [InlineData("claude-sonnet-4-6", true)]
+    [InlineData("fable-5-2", true)]
+    [InlineData("haiku-4-5", false)]
+    [InlineData("opus-4-5-20251101", false)]
+    [InlineData("sonnet-4-5", false)]
+    public void AdaptiveThinkingFromThe46GenerationOn(string tag, bool expected) =>
+        Assert.Equal(expected, ClaudeModels.SupportsAdaptiveThinking(tag));
+
     [Fact]
     public void PreservedThinkingModelsKeepAppendOnlyHistory()
     {
-        foreach (var tag in new[] { "fable-5-1", "opus-5-5", "sonnet-5-5", "haiku-5-5" })
+        foreach (var tag in new[] { "fable-5-2", "fable-5-1", "opus-5-5", "sonnet-5-5", "haiku-5-5", "claude-opus-5-5" })
             Assert.True(ClaudeModels.AppendOnlyHistory(tag), tag);
         foreach (var tag in new[] { "sonnet-5", "opus-5", "opus-4-8", "fable-5", "haiku-4-5", "auto" })
             Assert.False(ClaudeModels.AppendOnlyHistory(tag), tag);
@@ -57,6 +87,9 @@ public void PricesMatchTheCurrentListAndUnknownTagsOverestimate()
         // Fable 5.1 reads its cache at $0.25 per million, Opus 5.5 at $0.20.
         Assert.True(ClaudeModels.TryPrice("fable-5-1", 1000, out var fable51));
         Assert.Equal(0.25m, ClaudeModels.Cost(fable51, 0, 0, 0, 1_000_000));
+        Assert.True(ClaudeModels.TryPrice("fable-5-2", 1000, out var fable52));
+        Assert.Equal(40m, ClaudeModels.Cost(fable52, 0, 1_000_000, 0, 0));
+        Assert.Equal(0.20m, ClaudeModels.Cost(fable52, 0, 0, 0, 1_000_000));
         Assert.True(ClaudeModels.TryPrice("opus-5-5", 1000, out var opus55));
         Assert.Equal(0.20m, ClaudeModels.Cost(opus55, 0, 0, 0, 1_000_000));
         // 1-hour cache writes are twice the base input rate.

ClaudeRevit/Services/ChatService.cs

@@ -863,8 +863,8 @@
         // invalidate the 1h prompt-cache prefix for the rest of the turn, and rebuilding
         // the tool list would let a Settings change yank tools out from under the model
         // mid-plan (plus ~130 tool schemas re-serialized per round for nothing).
-        // "Auto" model mode. DEFAULT (advisorMode): Sonnet 5 runs every round and consults Opus
-        // 4.8 via the advisor tool only when it needs a plan — Sonnet's prompt cache stays warm
+        // "Auto" model mode. DEFAULT (advisorMode): Sonnet 5.5 runs every round and consults Opus
+        // 5.5 via the advisor tool only when it needs a plan — Sonnet's prompt cache stays warm
         // all session, Opus is billed only for the short advice sub-inference. The legacy
         // alternative (Settings → AutoUseAdvisor off) switches the whole turn to Opus on a hard
         // task/error/deep-loop; see the escalation vars below.
@@ -926,7 +926,7 @@
         // to a stronger one when the task looks hard up front, or once the fast model starts
         // erroring / the task turns into a deep multi-step loop. Two paths share this machinery:
         //   • Alt path — fast alt model → the configured "reasoning" alt model.
-        //   • Claude "Auto" — Sonnet 5 (cheap) → Opus 4.8 (strong). This is the big cost lever:
+        //   • Claude "Auto" — Sonnet 5.5 (cheap) → Opus 5.5 (strong). This is the big cost lever:
         //     Sonnet is $3/$15 vs Opus $5/$25 per 1M, so routine turns run at ~60% of the price
         //     and Opus is reserved for the turns that actually need it.
         // Legacy model-switch path (alt reasoning model, or Auto with the advisor turned off).
@@ -1005,11 +1005,11 @@
                 // the reasoning model still takes precedence, since that is a deliberate switch.
                 var altModelForTurn = escalate ? reasoningModel : AltModelOverride(model);
                 // Claude Auto: in advisor mode the executor is the configured cheap/fast model
-                // (Sonnet 5 default, or Haiku 4.5 for the cheapest tier); the advisor is reached
+                // (Sonnet 5.5 default, or Haiku 5.5 for the cheapest tier); the advisor is reached
                 // via the tool, not by switching. In legacy mode: Sonnet until escalated to Opus.
                 var claudeModelForTurn = advisorMode
                     ? SettingsStore.AutoExecutorModel
-                    : (claudeAuto ? (escalate ? "opus-4-8" : "sonnet-5") : model);
+                    : (claudeAuto ? (escalate ? ClaudeModels.DefaultAdvisor : ClaudeModels.DefaultExecutor) : model);
                 BackendTurn turn;
                 try
                 {
@@ -1570,7 +1570,7 @@
             ContextManagement = ClaudeModels.AppendOnlyHistory(model)
                 ? new BetaContextManagementConfig { Edits = [new BetaCompact20260112Edit()] }
                 : null,
             Fallbacks = fallback ? new Default() : null,
             Betas = BetasFor(model, useAdvisorBeta, fallback, binding)
         };
 
@@ -1674,7 +1674,7 @@
             {
                 turn.Blocks.Add(new ChatFallbackBlock(JsonSerializer.Serialize(fb)));
                 string? category = null;
                 try { category = fb.Trigger?.Category.Raw(); } catch { }
                 turn.FallbackNote = $"[Answered by {fb.To.Model.Raw()}: {fb.From.Model.Raw()} declined this request" +
                                     (string.IsNullOrEmpty(category) ? "" : $" (classifier category: {category})") +
                                     ". Server-side fallback can be switched off in Settings.]";
@@ -1726,7 +1726,7 @@
                             adv.CacheCreationInputTokens, adv.CacheReadInputTokens));
                     else if (fellBack && iter.TryPickBetaMessageIterationUsage(out var declined))
                         turn.AdvisorUsages.Add(new AdvisorUsage(
                             ClaudeModels.TagFromModel(declined.Model.Raw() ?? ""),
                             declined.InputTokens, declined.OutputTokens,
                             declined.CacheCreationInputTokens, declined.CacheReadInputTokens));
                 }
@@ -1741,7 +1741,7 @@
     // The advisor tool definition for Auto mode: the executor consults the advisor model at most
     // once per request (a per-turn total cap is enforced by the caller), with the advisor's own
     // transcript cached for an hour so repeated consults in a session don't re-bill the whole
-    // prefix. Advisor model is user-configurable (Opus 4.8 default, or Fable 5 for hardest tasks).
+    // prefix. Advisor model is user-configurable (Opus 5.5 default, or Fable 5.2 for hardest tasks).
     private static BetaToolUnion AdvisorTool(string advisorModelId) => new BetaAdvisorTool20260301
     {
         Model = advisorModelId,

ClaudeRevit/Services/ClaudeCodeBackend.cs

@@ -42,26 +42,17 @@ public sealed class Result
         public string? Subtype;
     }
 
-    // Map our internal model tag to what Claude Code accepts for `--model`.
-    //
-    // The older tags use the coarse family aliases ("opus"/"sonnet"/"haiku"), which resolve to
-    // whatever the CLI currently considers that family's model. That is fine for them, but it can't
-    // express "Opus 5 specifically" — the family alias would just pick the CLI's current Opus and
-    // make two dropdown entries behave identically. So the newer models pass their FULL model id,
-    // which Claude Code also accepts, and the pane's choice actually decides.
+    // Map our internal model tag to what Claude Code accepts for `--model`: the full model id,
+    // so the pane's choice decides exactly which model runs (a family alias such as "opus"
+    // would silently pick whatever the CLI currently considers that family's model). Tags come
+    // from the Models API catalog, so any model the account can use maps without a table here.
     //
     // Returns null for "auto" — there we let the CLI pick, because the advisor-escalation that
-    // "auto" means on the API has no equivalent inside Claude Code's own loop. Fable 5 used to
-    // map to null too, which silently ran the CLI's default model instead of the one chosen.
-    // A tag the CLI doesn't recognise surfaces as a CLI error rather than a silent fallback.
-    public static string? ModelAlias(string? tag) => tag switch
-    {
-        "opus-5-5" or "sonnet-5-5" or "haiku-5-5" or "opus-5" or "fable-5" or "fable-5-1" => ClaudeModels.ApiId(tag),
-        "opus-4-8" or "opus-4-7" or "opus-4-6" => "opus",
-        "sonnet-5" or "sonnet-4-6" => "sonnet",
-        "haiku-4-5" => "haiku",
-        _ => null
-    };
+    // "auto" means on the API has no equivalent inside Claude Code's own loop. A tag the CLI
+    // doesn't recognise surfaces as a CLI error rather than a silent fallback.
+    public static string? ModelAlias(string? tag) =>
+        string.IsNullOrWhiteSpace(tag) || tag == "auto" || tag!.StartsWith("alt", StringComparison.Ordinal)
+            ? null : ClaudeModels.ApiId(tag);
 
     public static async Task<Result> RunAsync(
         string exe, string prompt, string workDir, string mcpConfigPath, string? resumeSessionId,

ClaudeRevit/Services/ClaudeModelCatalog.cs

@@ -0,0 +1,65 @@
+using System;
+using System.Collections.Generic;
+using System.Threading;
+using System.Threading.Tasks;
+using Anthropic;
+using Anthropic.Core;
+using Anthropic.Models.Models;
+
+namespace ClaudeRevit.Services;
+
+// The Claude models this account can use, read from the Models API (GET /v1/models) — the
+// counterpart of CodexModelCatalog, so the pickers follow what Anthropic serves instead of a
+// hand-kept list that goes stale with each release. Listing models is free: no inference runs.
+// Without an API key (Claude Code on a subscription) the built-in current generation is used.
+public static class ClaudeModelCatalog
+{
+    private static readonly SemaphoreSlim CacheGate = new(1, 1);
+    private static string? _cacheKey;
+    private static IReadOnlyList<ClaudeModelInfo>? _cached;
+    private static DateTime _expires;
+
+    // True when the last answer came from the Models API rather than the built-in list.
+    public static bool LastWasLive { get; private set; }
+
+    public static async Task<IReadOnlyList<ClaudeModelInfo>> ReadAsync(CancellationToken ct, bool forceRefresh = false)
+    {
+        var apiKey = ApiKeyStore.Load();
+        if (string.IsNullOrWhiteSpace(apiKey)) { LastWasLive = false; return ClaudeModels.Current; }
+        // The key itself is never kept; its hash only tells a changed key from the cached one.
+        var key = Convert.ToBase64String(System.Security.Cryptography.SHA256.Create()
+            .ComputeHash(System.Text.Encoding.UTF8.GetBytes(apiKey)));
+        await CacheGate.WaitAsync(ct);
+        try
+        {
+            ct.ThrowIfCancellationRequested();
+            if (!forceRefresh && _cached != null && _cacheKey == key && DateTime.UtcNow < _expires)
+            { LastWasLive = true; return _cached; }
+            // A failed explicit refresh must not leave stale entries available to the next run.
+            _cached = null;
+            var models = await DiscoverAsync(apiKey!, ct);
+            if (models.Count == 0) throw new InvalidOperationException("The Models API returned no Claude models.");
+            _cacheKey = key; _cached = models; _expires = DateTime.UtcNow.AddMinutes(5);
+            LastWasLive = true;
+            return models;
+        }
+        finally { CacheGate.Release(); }
+    }
+
+    private static async Task<IReadOnlyList<ClaudeModelInfo>> DiscoverAsync(string apiKey, CancellationToken ct)
+    {
+        using var timeout = CancellationTokenSource.CreateLinkedTokenSource(ct);
+        timeout.CancelAfter(TimeSpan.FromSeconds(30));
+        var client = new AnthropicClient(new ClientOptions { ApiKey = apiKey });
+        var served = new List<(string, string?)>();
+        // Newest first, as the API orders them; bounded so a misbehaving cursor cannot loop.
+        IPage<ModelInfo> page = await client.Models.List(new ModelListParams { Limit = 100 }, timeout.Token);
+        for (var pages = 0; ; pages++)
+        {
+            foreach (var model in page.Items) served.Add((model.ID, model.DisplayName));
+            if (pages >= 20 || !page.HasNext()) break;
+            page = await page.Next(timeout.Token);
+        }
+        return ClaudeModels.FromServed(served);
+    }
+}