diff --git a/Directory.Packages.props b/Directory.Packages.props
index 718fe4c23..d3ca8a741 100644
--- a/Directory.Packages.props
+++ b/Directory.Packages.props
@@ -86,6 +86,11 @@
     <PackageVersion Include="Aspire.Hosting.AppHost" Version="$(AspireHostingVersion)" />
     <PackageVersion Include="Aspire.Hosting.Testing" Version="$(AspireHostingVersion)" />
     <PackageVersion Include="CommunityToolkit.Aspire.Hosting.Ollama" Version="$(CommunityToolkitAspireVersion)" />
+    <!-- Transitive security pin: Aspire 13.4.x pulls MessagePack 2.5.192, which
+         trips NU1903 (GHSA-hv8m-jj95-wg3x, LZ4 decompression AccessViolation).
+         Keep this on the patched 3.x line already used by dev until Aspire ships
+         a non-vulnerable transitive dependency on its own. -->
+    <PackageVersion Include="MessagePack" Version="3.1.7" />
   </ItemGroup>
   <!-- Image normalization (Netclaw.Media). SkiaSharp ships Windows/macOS native
        assets in the base package; Linux needs the explicit headless native asset
@@ -106,13 +111,7 @@
   <ItemGroup>
     <PackageVersion Include="Microsoft.SourceLink.GitHub" Version="10.0.300" />
   </ItemGroup>
-  <!-- Transitive security pin: Aspire.Hosting.AppHost → StreamJsonRpc pulls in
-       MessagePack 2.5.192, flagged by NuGetAudit (GHSA-hv8m-jj95-wg3x, LZ4
-       decompression DoS). Netclaw uses MessagePack nowhere — this is sample-only
-       Aspire tooling RPC — but pin to the patched v2 (2.5.301) to clear the audit. -->
-  <ItemGroup>
-    <PackageVersion Include="MessagePack" Version="3.1.7" />
-  </ItemGroup>
+
   <!-- Transitive audit suppress: Microsoft.Data.Sqlite → SQLitePCLRaw.lib.e_sqlite3 2.1.11
        is flagged by NuGetAudit (GHSA-2m69-gcr7-jv3q, CVE-2025-6965 — SQLite < 3.50.2
        memory corruption in aggregate-term handling). No patched version of
@@ -123,4 +122,4 @@
   <ItemGroup>
     <NuGetAuditSuppress Include="https://github.com/advisories/GHSA-2m69-gcr7-jv3q" />
   </ItemGroup>
-</Project>
\ No newline at end of file
+</Project>
diff --git a/README.md b/README.md
index 42e4c218c..fdff5673b 100644
--- a/README.md
+++ b/README.md
@@ -77,7 +77,8 @@ dotnet run --project samples/Netclaw.Demo.AppHost
 ### Prerequisites
 
 - An LLM provider — [Ollama](https://ollama.com/) (local, default),
-  [OpenRouter](https://openrouter.ai/), or any OpenAI-compatible endpoint.
+  [OpenRouter](https://openrouter.ai/), [DwarfStar/ds4](https://github.com/antirez/ds4)
+  (local DeepSeek V4 on Apple Silicon / CUDA), or any OpenAI-compatible endpoint.
   See the full [provider documentation](https://netclaw.dev/configuration/managed-providers/)
   for all supported options.
 
diff --git a/docs/spec/SPEC-008-model-provider-abstraction.md b/docs/spec/SPEC-008-model-provider-abstraction.md
index 0b6915c4f..23783a4e0 100644
--- a/docs/spec/SPEC-008-model-provider-abstraction.md
+++ b/docs/spec/SPEC-008-model-provider-abstraction.md
@@ -25,6 +25,9 @@ Session actors depend on provider-neutral chat client behavior.
 - Anthropic direct
 - OpenAI direct
 - Ollama via OpenAI-compatible local endpoint
+- DwarfStar (ds4) — local DeepSeek V4 engine via the `openai-compatible`
+  provider type; its OpenRouter-shaped `context_length` metadata is read by a
+  dedicated backend strategy in the capability resolver
 
 ## Configuration Model
 
diff --git a/docs/spec/configuration.md b/docs/spec/configuration.md
index 71e284271..7267ed971 100644
--- a/docs/spec/configuration.md
+++ b/docs/spec/configuration.md
@@ -96,7 +96,7 @@ keys used by model references.
 
 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
-| `Type` | string | `"ollama"` | Provider SDK to use. Currently supported: `ollama`. Future: `openrouter`, `openai`, `anthropic`. |
+| `Type` | string | `"ollama"` | Provider SDK to use. Supported: `ollama`, `openai-compatible`, `openrouter`, `openai`, `anthropic`, `github-copilot`, `veniceai`. |
 | `Endpoint` | string | `"http://localhost:11434"` | Base URL for the provider API. |
 | `ApiKey` | string? | `null` | API key. Should go in `secrets.json` or an environment variable. |
 
@@ -512,6 +512,10 @@ export NETCLAW_Session__MaxToolIterationsPerTurn="60"
     "openrouter": {
       "Type": "openrouter",
       "Endpoint": "https://openrouter.ai/api/v1"
+    },
+    "ds4": {
+      "Type": "openai-compatible",
+      "Endpoint": "http://127.0.0.1:8000"
     }
   },
   "Models": {
diff --git a/feeds/skills/.system/files/netclaw-operations/SKILL.md b/feeds/skills/.system/files/netclaw-operations/SKILL.md
index cf6673348..342566054 100644
--- a/feeds/skills/.system/files/netclaw-operations/SKILL.md
+++ b/feeds/skills/.system/files/netclaw-operations/SKILL.md
@@ -3,7 +3,7 @@ name: netclaw-operations
 description: "REQUIRED when the user asks about scheduling, reminders, cron jobs, timers, background jobs, diagnostics, troubleshooting, MCP tools, daemon health, identity updates, or Netclaw capabilities and self-maintenance."
 metadata:
   author: netclaw
-  version: "2.16.0"
+  version: "2.17.0"
 ---
 
 # Netclaw Operations
@@ -803,7 +803,7 @@ and a `type` (well-known identifier). Manage them with `netclaw provider`:
 |------|------|-------|
 | `ollama` | Endpoint only | `--endpoint http://host:11434` |
 | `openai` | API key **or** OAuth (ChatGPT sub) | Codex backend for OAuth path |
-| `openai-compatible` | API key + endpoint | Generic OpenAI-shape proxies |
+| `openai-compatible` | Endpoint; optional API key | Generic OpenAI-shape proxies, llama.cpp, vLLM. Also DwarfStar (ds4): `--endpoint http://127.0.0.1:8000`, run `ds4-server` separately, model ids `deepseek-v4-flash` / `deepseek-v4-pro`, context window auto-detected |
 | `anthropic` | API key | `sk-ant-...` |
 | `openrouter` | API key | `sk-or-...` |
 | `github-copilot` | OAuth device flow only | Requires active Copilot subscription on the GitHub account |
diff --git a/openspec/specs/netclaw-model-capabilities/spec.md b/openspec/specs/netclaw-model-capabilities/spec.md
index a99f73850..37cc80fa6 100644
--- a/openspec/specs/netclaw-model-capabilities/spec.md
+++ b/openspec/specs/netclaw-model-capabilities/spec.md
@@ -79,6 +79,26 @@ it, as the highest-priority detection source.
 - **AND** map array values (`"text"`, `"image"`, `"audio"`, `"video"`) to
   corresponding `ModelModality` flags
 
+#### Scenario: OpenAI-compatible backend strategy detection
+
+- **GIVEN** a model is configured on an `openai-compatible` provider
+- **WHEN** capabilities are resolved for that model
+- **THEN** the system SHALL probe `/v1/models` and the optional llama.cpp
+  `/props` endpoint
+- **AND** dispatch the probe to the first matching backend strategy in priority
+  order: DwarfStar/ds4, vLLM, llama.cpp, generic OpenAI-compatible
+
+#### Scenario: DwarfStar ds4 context detection
+
+- **GIVEN** an OpenAI-compatible `/v1/models` entry for the configured model
+  has `owned_by: "ds4.c"`
+- **WHEN** capabilities are resolved for that model
+- **THEN** the DwarfStar/ds4 strategy SHALL match before broader backend
+  heuristics
+- **AND** context window SHALL be read from top-level `context_length`, falling
+  back to `top_provider.context_length`
+- **AND** input and output modalities SHALL resolve to `Text`
+
 ### Requirement: OpenRouter oracle for cross-provider lookup
 
 The system SHALL use OpenRouter's public `GET /api/v1/models` endpoint as a
diff --git a/openspec/specs/netclaw-model-providers/spec.md b/openspec/specs/netclaw-model-providers/spec.md
index 863ce4aed..40a7b7249 100644
--- a/openspec/specs/netclaw-model-providers/spec.md
+++ b/openspec/specs/netclaw-model-providers/spec.md
@@ -16,6 +16,8 @@ The system SHALL default to OpenRouter during first-run setup.
 ### Requirement: Multi-provider support
 
 The system SHALL support selecting one provider profile from a supported set.
+Supported provider type keys SHALL include `ollama`, `openai-compatible`,
+`openrouter`, `openai`, `anthropic`, `github-copilot`, and `veniceai`.
 All provider interactions SHALL use the Microsoft.Extensions.AI `IChatClient`
 abstraction layer, ensuring provider-agnostic model access throughout the
 application.
@@ -27,7 +29,8 @@ and `OutputModalities` fields populated from provider responses.
 #### Scenario: Switch provider
 
 - **GIVEN** OpenRouter is configured
-- **WHEN** operator selects Anthropic, OpenAI, or Ollama profile
+- **WHEN** operator selects Anthropic, OpenAI, Ollama, OpenAI-compatible,
+  OpenRouter, GitHub Copilot, or Venice.ai profile
 - **THEN** runtime uses selected provider through the `IChatClient` interface
   after validation
 
@@ -54,6 +57,16 @@ and `OutputModalities` fields populated from provider responses.
   `InputModalities` and `OutputModalities` populated from
   `architecture.input_modalities` and `architecture.output_modalities`
 
+#### Scenario: OpenAI-compatible discovery includes backend context metadata
+
+- **GIVEN** an OpenAI-compatible provider is configured
+- **WHEN** model discovery runs via `ProviderProbe`
+- **THEN** the returned `DiscoveredModel` records SHALL include context-window
+  metadata when the backend exposes a known field shape, including vLLM
+  `max_model_len`, DwarfStar/ds4 `context_length` or
+  `top_provider.context_length`, and llama.cpp `meta.n_ctx` or
+  `meta.n_ctx_train`
+
 ### Requirement: Optional live smoke provider checks
 
 The system SHALL support optional provider smoke checks against a local
@@ -318,4 +331,3 @@ by Copilot with `HTTP 400 "Authorization header is badly formatted"`.
 - **THEN** the SDK's credential auth policy emits the exchanged Copilot token,
   not the placeholder, because the shared credential was updated before the auth
   policy ran
-
diff --git a/src/Netclaw.Cli.Tests/Tui/ModelManagerViewModelTests.cs b/src/Netclaw.Cli.Tests/Tui/ModelManagerViewModelTests.cs
index 3f50decad..249954fb5 100644
--- a/src/Netclaw.Cli.Tests/Tui/ModelManagerViewModelTests.cs
+++ b/src/Netclaw.Cli.Tests/Tui/ModelManagerViewModelTests.cs
@@ -471,7 +471,7 @@ public void Refresh_PopulatesDisplayNameFromRegistry()
 
         Assert.Single(vm.Providers);
         Assert.Equal("my-vllm", vm.Providers[0].Name);
-        Assert.Equal("llama.cpp / vLLM", vm.Providers[0].DisplayName);
+        Assert.Equal("OpenAI-compatible (llama.cpp / vLLM / DwarfStar ds4)", vm.Providers[0].DisplayName);
     }
 
     [Fact]
diff --git a/src/Netclaw.Daemon.Tests/Providers/OpenAiCompatibleCapabilityResolverTests.cs b/src/Netclaw.Daemon.Tests/Providers/OpenAiCompatibleCapabilityResolverTests.cs
index 9f3eb2243..5660f7f06 100644
--- a/src/Netclaw.Daemon.Tests/Providers/OpenAiCompatibleCapabilityResolverTests.cs
+++ b/src/Netclaw.Daemon.Tests/Providers/OpenAiCompatibleCapabilityResolverTests.cs
@@ -19,6 +19,8 @@ public sealed class OpenAiCompatibleCapabilityResolverTests
     [Theory]
     [InlineData("{\"id\":\"Qwen3.6-27B-MTP-UD-Q4_K_XL.gguf\",\"meta\":{\"n_ctx\":131072,\"n_ctx_train\":262144}}", 131_072)]
     [InlineData("{\"id\":\"Qwen/Qwen3.6-VL-30B-FP8\",\"max_model_len\":256000}", 256_000)]
+    [InlineData("{\"id\":\"deepseek-v4-flash\",\"owned_by\":\"ds4.c\",\"context_length\":262144}", 262_144)]
+    [InlineData("{\"id\":\"deepseek-v4-pro\",\"owned_by\":\"ds4.c\",\"top_provider\":{\"context_length\":196608}}", 196_608)]
     public void ParseModels_ReadsSelfHostedContextMetadata(string modelJson, int expectedContextWindow)
     {
         var json = $$"""
@@ -168,6 +170,65 @@ public void ResolveFromProbe_LlamaCppRouterPropsZero_ReturnsUnknownContext()
         Assert.Null(result.OutputModalities);
     }
 
+    [Fact]
+    public void ResolveFromProbe_Ds4Shape_DispatchesToDs4Strategy_EvenWithProps()
+    {
+        const string modelsJson = """
+        {
+          "object": "list",
+          "data": [
+            {
+              "id": "deepseek-v4-flash",
+              "object": "model",
+              "owned_by": "ds4.c",
+              "context_length": 262144,
+              "top_provider": { "context_length": 262144 }
+            }
+          ]
+        }
+        """;
+        // A proxy in front of ds4 may answer /props; the ds4 strategy is
+        // ordered before llama.cpp so the owned_by signal still wins.
+        const string propsJson = """
+        { "default_generation_settings": { "n_ctx": 4096 } }
+        """;
+
+        var result = OpenAiCompatibleCapabilityResolver.ResolveFromProbe(
+            "deepseek-v4-flash", modelsJson, propsJson);
+
+        Assert.NotNull(result);
+        Assert.Equal(262_144, result.ContextWindowTokens);
+        Assert.Equal(ModelModality.Text, result.InputModalities);
+        Assert.Equal(ModelModality.Text, result.OutputModalities);
+    }
+
+    [Fact]
+    public void ResolveFromProbe_Ds4Shape_DispatchesToDs4Strategy_EvenWithMaxModelLen()
+    {
+        const string modelsJson = """
+        {
+          "object": "list",
+          "data": [
+            {
+              "id": "deepseek-v4-flash",
+              "object": "model",
+              "owned_by": "ds4.c",
+              "max_model_len": 4096,
+              "context_length": 262144
+            }
+          ]
+        }
+        """;
+
+        var result = OpenAiCompatibleCapabilityResolver.ResolveFromProbe(
+            "deepseek-v4-flash", modelsJson, propsJson: null);
+
+        Assert.NotNull(result);
+        Assert.Equal(262_144, result.ContextWindowTokens);
+        Assert.Equal(ModelModality.Text, result.InputModalities);
+        Assert.Equal(ModelModality.Text, result.OutputModalities);
+    }
+
     [Fact]
     public void ResolveFromProbe_UnknownShape_FallsThroughToGenericStrategy()
     {
diff --git a/src/Netclaw.Daemon.Tests/Providers/Strategies/Ds4BackendStrategyTests.cs b/src/Netclaw.Daemon.Tests/Providers/Strategies/Ds4BackendStrategyTests.cs
new file mode 100644
index 000000000..fb2f98d64
--- /dev/null
+++ b/src/Netclaw.Daemon.Tests/Providers/Strategies/Ds4BackendStrategyTests.cs
@@ -0,0 +1,114 @@
+// -----------------------------------------------------------------------
+// <copyright file="Ds4BackendStrategyTests.cs" company="Petabridge, LLC">
+//      Copyright (C) 2026 - 2026 Petabridge, LLC <https://petabridge.com>
+// </copyright>
+// -----------------------------------------------------------------------
+using System.Text.Json;
+using Netclaw.Configuration;
+using Netclaw.Providers.SelfHosted;
+using Xunit;
+
+namespace Netclaw.Daemon.Tests.Providers.Strategies;
+
+public sealed class Ds4BackendStrategyTests
+{
+    // Real ds4-server response shape (see append_model_json_values in
+    // ds4_server.c): OpenRouter-shaped metadata with owned_by "ds4.c".
+    private const string Ds4ModelsJson = """
+    {
+      "object": "list",
+      "data": [
+        {
+          "id": "deepseek-v4-flash",
+          "object": "model",
+          "created": 1767225600,
+          "owned_by": "ds4.c",
+          "name": "DeepSeek V4 Flash",
+          "context_length": 262144,
+          "top_provider": {
+            "context_length": 262144,
+            "max_completion_tokens": 8192,
+            "is_moderated": false
+          },
+          "supported_parameters": ["tools", "tool_choice", "stream"]
+        },
+        {
+          "id": "deepseek-v4-pro",
+          "object": "model",
+          "created": 1767225600,
+          "owned_by": "ds4.c",
+          "name": "DeepSeek V4 PRO",
+          "top_provider": { "context_length": 196608 }
+        }
+      ]
+    }
+    """;
+
+    [Fact]
+    public void Matches_OwnedByDs4()
+    {
+        using var doc = JsonDocument.Parse(Ds4ModelsJson);
+        var probe = new BackendProbe("deepseek-v4-flash", doc.RootElement, PropsRoot: null);
+        Assert.True(new Ds4BackendStrategy().Matches(probe));
+    }
+
+    [Fact]
+    public void Matches_False_WhenOwnedByOther()
+    {
+        const string json = """
+        { "object": "list", "data": [ { "id": "model-x", "owned_by": "vllm", "context_length": 131072 } ] }
+        """;
+        using var doc = JsonDocument.Parse(json);
+        var probe = new BackendProbe("model-x", doc.RootElement, PropsRoot: null);
+        Assert.False(new Ds4BackendStrategy().Matches(probe));
+    }
+
+    [Fact]
+    public void Matches_False_WhenModelMissing()
+    {
+        using var doc = JsonDocument.Parse(Ds4ModelsJson);
+        var probe = new BackendProbe("gpt-4", doc.RootElement, PropsRoot: null);
+        Assert.False(new Ds4BackendStrategy().Matches(probe));
+    }
+
+    [Fact]
+    public void Parse_ReadsTopLevelContextLength_AndTextModalities()
+    {
+        using var doc = JsonDocument.Parse(Ds4ModelsJson);
+        var probe = new BackendProbe("deepseek-v4-flash", doc.RootElement, PropsRoot: null);
+
+        var result = new Ds4BackendStrategy().Parse(probe);
+
+        Assert.NotNull(result);
+        Assert.Equal(262_144, result.ContextWindowTokens);
+        Assert.Equal(ModelModality.Text, result.InputModalities);
+        Assert.Equal(ModelModality.Text, result.OutputModalities);
+    }
+
+    [Fact]
+    public void Parse_FallsBackToTopProviderContextLength()
+    {
+        using var doc = JsonDocument.Parse(Ds4ModelsJson);
+        var probe = new BackendProbe("deepseek-v4-pro", doc.RootElement, PropsRoot: null);
+
+        var result = new Ds4BackendStrategy().Parse(probe);
+
+        Assert.NotNull(result);
+        Assert.Equal(196_608, result.ContextWindowTokens);
+    }
+
+    [Fact]
+    public void Parse_TreatsMissingContextAsUnknown()
+    {
+        const string json = """
+        { "object": "list", "data": [ { "id": "deepseek-v4-flash", "owned_by": "ds4.c" } ] }
+        """;
+        using var doc = JsonDocument.Parse(json);
+        var probe = new BackendProbe("deepseek-v4-flash", doc.RootElement, PropsRoot: null);
+
+        var result = new Ds4BackendStrategy().Parse(probe);
+
+        Assert.NotNull(result);
+        Assert.Null(result.ContextWindowTokens);
+    }
+}
diff --git a/src/Netclaw.Providers/SelfHosted/OpenAiBackendStrategy.cs b/src/Netclaw.Providers/SelfHosted/OpenAiBackendStrategy.cs
index e28400e28..d6ad15dbb 100644
--- a/src/Netclaw.Providers/SelfHosted/OpenAiBackendStrategy.cs
+++ b/src/Netclaw.Providers/SelfHosted/OpenAiBackendStrategy.cs
@@ -198,10 +198,65 @@ private static bool IsRouterProps(JsonElement props)
 }
 
 /// <summary>
-/// Last-resort fallback when neither vLLM nor llama.cpp signals are
-/// present. Returns the model id with null fields so downstream
-/// resolvers (OpenRouter oracle, HuggingFace) have the chance to fill
-/// in. Always matches.
+/// DwarfStar (ds4) strategy — antirez's native DeepSeek V4 inference engine.
+/// Recognized via <c>owned_by: "ds4.c"</c> on the <c>/v1/models</c> entry.
+/// ds4 emits OpenRouter-shaped metadata: the context window lives in
+/// <c>context_length</c> / <c>top_provider.context_length</c>, fields neither
+/// the vLLM nor llama.cpp strategies read. Without this strategy the engine's
+/// large context window (its defining feature, backed by compressed KV cache
+/// on SSD) would never be auto-detected and compaction would size against a
+/// default. Ordered before llama.cpp so a proxy that happens to serve
+/// <c>/props</c> in front of ds4 cannot steal the match.
+/// </summary>
+internal sealed class Ds4BackendStrategy : IOpenAiBackendStrategy
+{
+    public string Name => "ds4";
+
+    public bool Matches(BackendProbe probe)
+    {
+        if (!probe.TryFindModelEntry(out var model))
+            return false;
+
+        return model.TryGetProperty("owned_by", out var ownedBy) &&
+               ownedBy.ValueKind == JsonValueKind.String &&
+               string.Equals(ownedBy.GetString(), "ds4.c", StringComparison.OrdinalIgnoreCase);
+    }
+
+    public ResolvedModelCapabilities? Parse(BackendProbe probe)
+    {
+        if (!probe.TryFindModelEntry(out var model))
+            return null;
+
+        // ds4 serves text-only DeepSeek V4 models; report modalities explicitly
+        // (its model ids are not HuggingFace ids, so downstream resolvers would
+        // have nothing useful to fill in).
+        return new ResolvedModelCapabilities(
+            probe.ModelId, ModelModality.Text, ModelModality.Text, ReadContextLength(model));
+    }
+
+    /// <summary>
+    /// Reads the OpenRouter-shaped context window: a top-level
+    /// <c>context_length</c>, falling back to the nested
+    /// <c>top_provider.context_length</c> when the top-level field is absent.
+    /// Shared with <see cref="OpenAiCompatibleDescriptor"/> probe parsing.
+    /// </summary>
+    internal static int? ReadContextLength(JsonElement model)
+    {
+        var top = ProbeHelpers.TryReadPositiveInt32(model, "context_length");
+        if (top is not null)
+            return top;
+
+        return model.TryGetProperty("top_provider", out var topProvider)
+            ? ProbeHelpers.TryReadPositiveInt32(topProvider, "context_length")
+            : null;
+    }
+}
+
+/// <summary>
+/// Last-resort fallback when no backend-specific signals are present.
+/// Returns the model id with null fields so downstream resolvers
+/// (OpenRouter oracle, HuggingFace) have the chance to fill in.
+/// Always matches.
 /// </summary>
 internal sealed class GenericOpenAiBackendStrategy : IOpenAiBackendStrategy
 {
diff --git a/src/Netclaw.Providers/SelfHosted/OpenAiCompatibleCapabilityResolver.cs b/src/Netclaw.Providers/SelfHosted/OpenAiCompatibleCapabilityResolver.cs
index e0be1de45..1abbe2c2d 100644
--- a/src/Netclaw.Providers/SelfHosted/OpenAiCompatibleCapabilityResolver.cs
+++ b/src/Netclaw.Providers/SelfHosted/OpenAiCompatibleCapabilityResolver.cs
@@ -15,14 +15,16 @@ namespace Netclaw.Providers.SelfHosted;
 /// endpoint (<c>/v1/models</c> and the llama.cpp-only <c>/props</c>) and
 /// dispatching to a backend strategy that knows how to parse the
 /// concrete server's response shape. Strategies are evaluated in
-/// priority order — vLLM first, llama.cpp second, generic last.
+/// priority order — ds4 first, vLLM second, llama.cpp third, generic last.
 /// </summary>
 public sealed class OpenAiCompatibleCapabilityResolver : IModelCapabilityResolver
 {
-    // Strategy order is meaningful: vLLM signals are more specific
-    // than llama.cpp's, and the generic fallback matches anything.
+    // Strategy order is meaningful: ds4 and vLLM signals are exact
+    // (owned_by markers), llama.cpp's are looser (any /props response),
+    // and the generic fallback matches anything.
     private static readonly IReadOnlyList<IOpenAiBackendStrategy> Strategies =
     [
+        new Ds4BackendStrategy(),
         new VllmBackendStrategy(),
         new LlamaCppBackendStrategy(),
         new GenericOpenAiBackendStrategy(),
diff --git a/src/Netclaw.Providers/SelfHosted/OpenAiCompatibleDescriptor.cs b/src/Netclaw.Providers/SelfHosted/OpenAiCompatibleDescriptor.cs
index c4c936625..14d69a831 100644
--- a/src/Netclaw.Providers/SelfHosted/OpenAiCompatibleDescriptor.cs
+++ b/src/Netclaw.Providers/SelfHosted/OpenAiCompatibleDescriptor.cs
@@ -11,7 +11,7 @@ namespace Netclaw.Providers.SelfHosted;
 
 /// <summary>
 /// Provider descriptor for OpenAI-compatible endpoints such as llama.cpp,
-/// llama-server, vLLM, or Lemonade.
+/// llama-server, vLLM, Lemonade, or DwarfStar (ds4).
 /// </summary>
 public sealed class OpenAiCompatibleDescriptor : IProviderDescriptor
 {
@@ -23,7 +23,7 @@ public OpenAiCompatibleDescriptor(HttpClient httpClient)
     }
 
     public string TypeKey => "openai-compatible";
-    public string DisplayName => "llama.cpp / vLLM";
+    public string DisplayName => "OpenAI-compatible (llama.cpp / vLLM / DwarfStar ds4)";
     public string DefaultEndpoint => "http://localhost:11434";
     public string ModelListingPath => "/v1/models";
     public IProviderAuth Auth { get; } = new EndpointOnlyAuth();
@@ -57,7 +57,13 @@ internal static ProviderProbeResult ParseModels(string json)
         if (contextWindow is not null)
             return contextWindow;
 
-        return model.TryGetProperty("meta", out var meta)
+        // ds4 (and other OpenRouter-shaped backends): context_length /
+        // top_provider.context_length.
+        contextWindow = Ds4BackendStrategy.ReadContextLength(model);
+        if (contextWindow is not null)
+            return contextWindow;
+
+        return model.TryGetProperty("meta", out var meta) // llama.cpp
             ? ProbeHelpers.TryReadPositiveInt32OrFallbackWhenMissing(
                 meta, "n_ctx", "n_ctx_train")
             : null;
diff --git a/tests/smoke/screenshots/provider-manager-empty.approved.png b/tests/smoke/screenshots/provider-manager-empty.approved.png
index 316a6ae72..e61e6643b 100644
Binary files a/tests/smoke/screenshots/provider-manager-empty.approved.png and b/tests/smoke/screenshots/provider-manager-empty.approved.png differ
diff --git a/tests/smoke/screenshots/wizard-provider-picker.approved.png b/tests/smoke/screenshots/wizard-provider-picker.approved.png
index 14800cd59..90137daa2 100644
Binary files a/tests/smoke/screenshots/wizard-provider-picker.approved.png and b/tests/smoke/screenshots/wizard-provider-picker.approved.png differ