diff --git a/docs/spec/configuration.md b/docs/spec/configuration.md
index c273d4146..f2f4f099f 100644
--- a/docs/spec/configuration.md
+++ b/docs/spec/configuration.md
@@ -105,6 +105,13 @@ keys used by model references.
 
 Named model definitions own provider/model identity and metadata. Roles reference definitions,
 so changing Main or Fallback does not destroy overrides belonging to the previous model.
+Capability discovery does not write `ContextWindow`, `InputModalities`, or `OutputModalities`.
+The daemon resolves those dynamic values at startup when the operator leaves them absent.
+
+Older releases can contain discovery snapshots in these override fields. The stored shape
+does not record each field's source. Netclaw preserves these values to protect explicit
+operator overrides. Use `--clear-context-window` and `--clear-modalities` once to restore
+runtime detection for an affected definition.
 
 ```json
 {
@@ -142,7 +149,7 @@ so changing Main or Fallback does not destroy overrides belonging to the previou
 |-------|------|---------|-------------|
 | `Provider` | string | `"local-ollama"` | Key into the `Providers` dictionary. |
 | `ModelId` | string | `"qwen3:30b"` | Model identifier as used by the provider's API. |
-| `ContextWindow` | int? | `null` | Effective runtime context window in tokens. When set, it clamps the detected provider value. If not set, Netclaw uses the provider-reported value when available, otherwise defaults to 32,768. |
+| `ContextWindow` | int? | `null` | Operator override for the runtime context window. When set, it clamps the detected provider value. If not set, Netclaw uses the provider-reported value when available, otherwise defaults to 32,768. Model selection does not persist a discovered value here. |
 | `InputModalities` | string? | `null` | Manual override for input modalities. Comma-separated flags from `Text`, `Image`, `Audio`, `Video` — e.g. `"Text"` or `"Text, Image"`. When set, bypasses automated capability detection. |
 | `OutputModalities` | string? | `null` | Manual override for output modalities. Same form as `InputModalities`. |
 
diff --git a/feeds/skills/.system/files/netclaw-operations/SKILL.md b/feeds/skills/.system/files/netclaw-operations/SKILL.md
index 1271d1dc3..c67a87120 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: "2.39.0"
+  version: "2.40.0"
 ---
 
 # Netclaw Operations
diff --git a/feeds/skills/.system/files/netclaw-operations/references/providers.md b/feeds/skills/.system/files/netclaw-operations/references/providers.md
index 5cfd3a7a5..3db64f903 100644
--- a/feeds/skills/.system/files/netclaw-operations/references/providers.md
+++ b/feeds/skills/.system/files/netclaw-operations/references/providers.md
@@ -88,19 +88,26 @@ openai` defaults to the ChatGPT OAuth device flow. Use `--auth api-key
 `netclaw model set <role> <provider> <model-id>` creates or reuses a named model
 definition and assigns it to a role (`main`, `fallback`, `compaction`). Definitions
 own provider/model identity and metadata, while roles only reference definitions.
-Switching away from a model and back therefore preserves its overrides. Two attributes can be overridden by the
-operator and are **operator-owned**: the context window and the input/output
-modalities. Provider discovery seeds a new definition but never changes an existing
-definition, including adding a property the definition deliberately omits.
+Switching away from a model and back therefore preserves its overrides. The operator
+owns the context window and modality overrides.
+
+Provider discovery validates the model ID for selection. It does not persist the
+discovered context window or modalities. The daemon detects these values at each
+startup. Only explicit operator flags or manual configuration create capability
+overrides.
+
+Older releases can contain discovery snapshots in the override fields. The config does
+not record each field's source. Netclaw preserves these values to protect explicit
+operator overrides. Run `netclaw model set` with `--clear-context-window` and
+`--clear-modalities` once to restore runtime detection for an affected definition.
 
 - `--context-window <tokens>` clamps the session budget and takes precedence
   over provider-reported detection. Supplying it configures the model manually
   and skips the metadata probe.
 - `--input-modalities <list>` / `--output-modalities <list>` override detected
   modalities with a comma-separated list of named flags (`Text`, `Image`,
-  `Audio`, `Video`). These do **not** skip the probe — the model is still
-  validated and its context window discovered; the override just wins over the
-  discovered modalities.
+  `Audio`, `Video`). These do **not** skip the probe. The probe still validates
+  the model ID. The override wins over runtime detection.
 - `--clear-context-window` and `--clear-modalities` remove the respective
   override so runtime capability detection resolves it again (use these after a
   provider enlarges a model's window or fixes mis-reported modalities).
diff --git a/src/Netclaw.Cli.Tests/Config/ModelEntryWriterTests.cs b/src/Netclaw.Cli.Tests/Config/ModelEntryWriterTests.cs
index 28101310c..76805924d 100644
--- a/src/Netclaw.Cli.Tests/Config/ModelEntryWriterTests.cs
+++ b/src/Netclaw.Cli.Tests/Config/ModelEntryWriterTests.cs
@@ -6,6 +6,7 @@
 using System.Text.Json;
 using Netclaw.Cli.Config;
 using Netclaw.Configuration;
+using Netclaw.Daemon.Configuration;
 using Xunit;
 using ModalityOverride = Netclaw.Cli.Config.ValueOverride<Netclaw.Configuration.ModelModality>;
 using ContextWindowOverride = Netclaw.Cli.Config.ValueOverride<int>;
@@ -14,10 +15,9 @@ namespace Netclaw.Cli.Tests.Config;
 
 /// <summary>
 /// Guards the single persist path shared by `model set`, the init wizard, and the TUI
-/// model manager. The rule under test: modalities are written only when the discovery
-/// source genuinely reported them, so an unknown is never frozen into config as a
-/// permanent "Text" override (#1290), and operator-owned overrides survive re-selection
-/// (#1127 / #1610).
+/// model manager. The writer persists capability values only after explicit operator input.
+/// Runtime detection owns provider capability data (#1756).
+/// Operator overrides survive model re-selection (#1127 and #1610).
 /// </summary>
 public class ModelEntryWriterTests
 {
@@ -74,7 +74,7 @@ public void WriteRole_SameModelWithoutModalities_PreservesHandSetModalities()
         // Re-set the same model with only a context-window change; no modality intent supplied.
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Manual,
-            ContextWindowOverride.Set(131072), ModalityOverride.Unset, ModalityOverride.Unset, discovered: null);
+            ContextWindowOverride.Set(131072), ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.Equal("Text, Image", entry["InputModalities"]); // preserved (#1127)
@@ -82,7 +82,7 @@ public void WriteRole_SameModelWithoutModalities_PreservesHandSetModalities()
     }
 
     [Fact]
-    public void WriteRole_SameModel_PreservesExistingClampOverDiscoveredWindow()
+    public void WriteRole_SameModel_PreservesExistingClampWithoutExplicitChange()
     {
         // Operator clamped ContextWindow below what the provider reports.
         var models = Models(
@@ -90,39 +90,66 @@ public void WriteRole_SameModel_PreservesExistingClampOverDiscoveredWindow()
             { "Main": { "Provider": "spark", "ModelId": "qwen-vl", "ContextWindow": 32000 } }
             """);
 
-        // Re-select the same model (picker path): no explicit --context-window, but the probe
-        // reports the model's full window. The operator's clamp must win (#1610): ContextWindow
-        // is documented to take precedence over provider-reported detection.
+        // Re-select the same model through a picker that has no override input.
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Live,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset,
-            Discovered(contextWindow: 128000));
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.Equal(32000, entry["ContextWindow"]);
     }
 
     [Fact]
-    public void WriteRole_FirstTimeSet_UsesDiscoveredWindow()
+    public void WriteRole_FirstTimeSet_OmitsCapabilityOverrides()
     {
-        // No existing entry for the role: discovery is the fallback that seeds the value.
         var models = new Dictionary<string, object>();
 
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Live,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset,
-            Discovered(contextWindow: 128000));
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
-        Assert.Equal(128000, entry["ContextWindow"]);
+        Assert.False(entry.ContainsKey("ContextWindow"));
+        Assert.False(entry.ContainsKey("InputModalities"));
+        Assert.False(entry.ContainsKey("OutputModalities"));
     }
 
     [Fact]
-    public void WriteRole_ClearContextWindow_DropsStoredClampAndDiscoveredWindow()
+    public void WriteRole_ProviderCapabilityChange_ReachesRuntimeResolution()
     {
-        // Operator clamped the window; now --clear-context-window removes the clamp so runtime
-        // detection resolves it. Clear wins over BOTH the stored value and a probe that reports
-        // a window (#1610) — mirroring --clear-modalities.
+        var models = new Dictionary<string, object>();
+        var selectedDuringProbe = new DiscoveredModel
+        {
+            ModelId = new ModelId("deepseek-v4-flash-dspark"),
+            ContextWindowTokens = 327680,
+            InputModalities = ModelModality.Text,
+            OutputModalities = ModelModality.Text,
+        };
+
+        // The writer receives the selected identity, but it does not receive dynamic capabilities.
+        ModelEntryWriter.WriteRole(
+            models, "Main", "spark", selectedDuringProbe.ModelId.Value, ModelDiscoverySource.Live,
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
+
+        var main = ConfigFileHelper.DeserializeSection<ModelReference>(ActiveEntry(models, "Main"))!;
+        var detectedAtStartup = new ResolvedModelCapabilities(
+            main.ModelId,
+            ModelModality.Text | ModelModality.Image,
+            ModelModality.Text,
+            100000);
+
+        var resolved = ModelCapabilityResolution.ResolveModelCapabilities(
+            new ModelSelection { Main = main }, detectedAtStartup);
+
+        Assert.Equal(100000, resolved.ContextWindowTokens);
+        Assert.Equal(ModelModality.Text | ModelModality.Image, resolved.InputModalities);
+        Assert.Equal(ModelModality.Text, resolved.OutputModalities);
+    }
+
+    [Fact]
+    public void WriteRole_ClearContextWindow_DropsStoredClamp()
+    {
+        // The operator removes the clamp. Runtime detection resolves the window.
         var models = Models(
             """
             { "Main": { "Provider": "spark", "ModelId": "qwen-vl", "ContextWindow": 32000 } }
@@ -130,8 +157,7 @@ public void WriteRole_ClearContextWindow_DropsStoredClampAndDiscoveredWindow()
 
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Live,
-            ContextWindowOverride.Clear, ModalityOverride.Unset, ModalityOverride.Unset,
-            Discovered(contextWindow: 128000));
+            ContextWindowOverride.Clear, ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.False(entry.ContainsKey("ContextWindow")); // clamp removed → runtime detection
@@ -147,27 +173,22 @@ public void WriteRole_SameDefinitionWithoutWindow_DiscoveryDoesNotResurrect()
 
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Live,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset,
-            Discovered(contextWindow: 128000));
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         Assert.False(ActiveEntry(models, "Main").ContainsKey("ContextWindow"));
     }
 
     [Fact]
-    public void WriteRole_SameModelFreshProbe_DoesNotOverrideExistingModalities()
+    public void WriteRole_SameModel_PreservesExistingModalities()
     {
-        // Operator override on disk; a fresh probe reports a coarser Text-only capability.
         var models = Models(
             """
             { "Main": { "Provider": "spark", "ModelId": "qwen-vl", "InputModalities": "Text, Image" } }
             """);
 
-        // The stored override wins — discovery never silently overwrites it (#1610 / #5). This is
-        // the same rule as ContextWindow: the field is documented to bypass automated detection.
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Live,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset,
-            Discovered(input: ModelModality.Text, output: ModelModality.Text));
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.Equal("Text, Image", entry["InputModalities"]);
@@ -187,8 +208,7 @@ public void WriteRole_SameModelEntryClearedModality_DiscoveryDoesNotResurrect()
 
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Live,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset,
-            Discovered(input: ModelModality.Text | ModelModality.Image));
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.False(entry.ContainsKey("InputModalities")); // stays cleared, not resurrected
@@ -207,7 +227,7 @@ public void WriteRole_ExplicitModalityOverride_ReplacesExisting()
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Manual,
             ContextWindowOverride.Unset,
-            ModalityOverride.Set(ModelModality.Text), ModalityOverride.Unset, discovered: null);
+            ModalityOverride.Set(ModelModality.Text), ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.Equal("Text", entry["InputModalities"]);
@@ -221,12 +241,10 @@ public void WriteRole_ClearModalities_RemovesOverrideEvenWhenProbeReportsOne()
             { "Main": { "Provider": "spark", "ModelId": "qwen-vl", "InputModalities": "Text, Image", "OutputModalities": "Text" } }
             """);
 
-        // --clear-modalities removes both overrides so runtime detection resolves them; clear
-        // wins over BOTH the stored value and a probe that still reports modalities (#1610 / #4).
+        // The command removes both overrides. Runtime detection resolves the modalities.
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Manual,
-            ContextWindowOverride.Unset, ModalityOverride.Clear, ModalityOverride.Clear,
-            Discovered(input: ModelModality.Text | ModelModality.Image));
+            ContextWindowOverride.Unset, ModalityOverride.Clear, ModalityOverride.Clear);
 
         var entry = ActiveEntry(models, "Main");
         Assert.False(entry.ContainsKey("InputModalities"));
@@ -246,7 +264,7 @@ public void WriteRole_ExistingEntryOmitsModelId_DoesNotFalseMatchDefaultModel()
 
         ModelEntryWriter.WriteRole(
             models, "Main", "local-ollama", "qwen3:30b", ModelDiscoverySource.Manual,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset, discovered: null);
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.Equal("qwen3:30b", entry["ModelId"]);
@@ -269,7 +287,7 @@ public void WriteRole_SameModel_PreservesProvenanceUnlessFreshlyDiscovered(
 
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", incoming,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset, discovered: null);
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.Equal(expected.ToString(), entry["Provenance"]);
@@ -288,8 +306,7 @@ public void WriteRole_CorruptExistingEntry_OverwritesInsteadOfThrowing()
 
         var ex = Record.Exception(() => ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Live,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset,
-            Discovered(contextWindow: 128000)));
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset));
 
         Assert.Null(ex);
         var entry = ActiveEntry(models, "Main");
@@ -311,8 +328,7 @@ public void WriteRole_CorruptModalityButValidWindow_PreservesWindow()
 
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Live,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset,
-            Discovered(contextWindow: 128000));
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.Equal(32768, entry["ContextWindow"]);         // operator clamp preserved, not clobbered
@@ -331,12 +347,11 @@ public void WriteRole_CorruptEntryForDifferentModel_DoesNotLeakWindow()
 
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Live,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset,
-            Discovered(contextWindow: 128000));
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         var entry = ActiveEntry(models, "Main");
         Assert.Equal("qwen-vl", entry["ModelId"]);
-        Assert.Equal(128000, entry["ContextWindow"]);  // discovered window, not the other model's 32768
+        Assert.False(entry.ContainsKey("ContextWindow"));
     }
 
     [Fact]
@@ -350,13 +365,13 @@ public void WriteRole_SwitchAwayAndBack_PreservesPreviousModelModalities()
         // Switching roles changes only the role reference. The old definition remains intact.
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "other-model", ModelDiscoverySource.Manual,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset, discovered: null);
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         Assert.Equal("other-model", ActiveEntry(models, "Main")["ModelId"]);
 
         ModelEntryWriter.WriteRole(
             models, "Main", "spark", "qwen-vl", ModelDiscoverySource.Manual,
-            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset, discovered: null);
+            ContextWindowOverride.Unset, ModalityOverride.Unset, ModalityOverride.Unset);
 
         Assert.Equal("Text, Image", ActiveEntry(models, "Main")["InputModalities"]);
     }
@@ -373,13 +388,4 @@ private static Dictionary<string, object> ActiveEntry(
         return (Dictionary<string, object>)definitions[definitionName];
     }
 
-    private static DiscoveredModel Discovered(
-        int? contextWindow = null, ModelModality? input = null, ModelModality? output = null)
-        => new()
-        {
-            ModelId = new ModelId("qwen-vl"),
-            ContextWindowTokens = contextWindow,
-            InputModalities = input,
-            OutputModalities = output,
-        };
 }
diff --git a/src/Netclaw.Cli.Tests/Model/ModelCommandTests.cs b/src/Netclaw.Cli.Tests/Model/ModelCommandTests.cs
index 7917fb3fc..4709ba66c 100644
--- a/src/Netclaw.Cli.Tests/Model/ModelCommandTests.cs
+++ b/src/Netclaw.Cli.Tests/Model/ModelCommandTests.cs
@@ -97,7 +97,7 @@ public async Task Set_MainModel_WritesConfig()
     }
 
     [Fact]
-    public async Task Set_OpenAiOAuthModel_StoresLiveDiscoveredMetadata()
+    public async Task Set_OpenAiOAuthModel_DoesNotPersistDiscoveredCapabilities()
     {
         WriteConfig(new Dictionary<string, object>
         {
@@ -130,9 +130,9 @@ public async Task Set_OpenAiOAuthModel_StoresLiveDiscoveredMetadata()
         var config = ReadConfigFile(_paths.NetclawConfigPath);
         var main = ReadActiveModel(config, "Main");
         Assert.Equal("Live", main.GetProperty("Provenance").GetString());
-        Assert.Equal(512000, main.GetProperty("ContextWindow").GetInt32());
-        Assert.Equal("Text, Image", main.GetProperty("InputModalities").GetString());
-        Assert.Equal("Text", main.GetProperty("OutputModalities").GetString());
+        Assert.False(main.TryGetProperty("ContextWindow", out _));
+        Assert.False(main.TryGetProperty("InputModalities", out _));
+        Assert.False(main.TryGetProperty("OutputModalities", out _));
     }
 
     [Fact]
@@ -530,8 +530,7 @@ public async Task Set_OAuthModelWithModalityOverride_StillProbesAndValidates()
             }
         ]);
 
-        // A modality override no longer short-circuits the probe: the probe must still run to
-        // validate the model and discover the context window, while the operator's modality wins.
+        // The probe still validates the model. The operator's modality remains authoritative.
         var exitCode = await ModelCommand.RunAsync(
             ["model", "set", "main", "openai-codex", "gpt-new-codex", "--input-modalities", "Text"],
             _paths, _fakeProbe, output: _output);
@@ -540,9 +539,9 @@ public async Task Set_OAuthModelWithModalityOverride_StillProbesAndValidates()
         Assert.Equal(1, _fakeProbe.ProbeCallCount);            // probe ran despite the modality flag
         using var config = ReadConfigFile(_paths.NetclawConfigPath);
         var main = ReadActiveModel(config, "Main");
-        Assert.Equal("Live", main.GetProperty("Provenance").GetString());     // resolved via probe
-        Assert.Equal(512000, main.GetProperty("ContextWindow").GetInt32());   // discovered window captured
-        Assert.Equal("Text", main.GetProperty("InputModalities").GetString());// operator override wins
+        Assert.Equal("Live", main.GetProperty("Provenance").GetString());
+        Assert.False(main.TryGetProperty("ContextWindow", out _));
+        Assert.Equal("Text", main.GetProperty("InputModalities").GetString());
     }
 
     [Fact]
diff --git a/src/Netclaw.Cli.Tests/Tui/ModelManagerViewModelTests.cs b/src/Netclaw.Cli.Tests/Tui/ModelManagerViewModelTests.cs
index 460d12dc0..818e2646a 100644
--- a/src/Netclaw.Cli.Tests/Tui/ModelManagerViewModelTests.cs
+++ b/src/Netclaw.Cli.Tests/Tui/ModelManagerViewModelTests.cs
@@ -171,7 +171,7 @@ public async Task ConfirmAssignment_WritesCorrectConfig()
     }
 
     [Fact]
-    public async Task ConfirmAssignment_DiscoveredModelWithMetadata_WritesMetadata()
+    public async Task ConfirmAssignment_DiscoveredModelWithMetadata_OmitsCapabilityOverrides()
     {
         WriteConfig(new Dictionary<string, object>
         {
@@ -207,9 +207,9 @@ public async Task ConfirmAssignment_DiscoveredModelWithMetadata_WritesMetadata()
         var config = JsonDocument.Parse(File.ReadAllText(_paths.NetclawConfigPath));
         var main = ReadActiveModel(config, "Main");
         Assert.Equal("Live", main.GetProperty("Provenance").GetString());
-        Assert.Equal(512000, main.GetProperty("ContextWindow").GetInt32());
-        Assert.Equal("Text, Image", main.GetProperty("InputModalities").GetString());
-        Assert.Equal("Text", main.GetProperty("OutputModalities").GetString());
+        Assert.False(main.TryGetProperty("ContextWindow", out _));
+        Assert.False(main.TryGetProperty("InputModalities", out _));
+        Assert.False(main.TryGetProperty("OutputModalities", out _));
     }
 
     [Fact]
diff --git a/src/Netclaw.Cli.Tests/Tui/Wizard/ProviderStepViewModelTests.cs b/src/Netclaw.Cli.Tests/Tui/Wizard/ProviderStepViewModelTests.cs
index e2dd5a440..10e5d090b 100644
--- a/src/Netclaw.Cli.Tests/Tui/Wizard/ProviderStepViewModelTests.cs
+++ b/src/Netclaw.Cli.Tests/Tui/Wizard/ProviderStepViewModelTests.cs
@@ -240,7 +240,7 @@ public void ContributeConfig_SetsProviderAndModel()
     }
 
     [Fact]
-    public void ContributeConfig_SelectedDiscoveredModel_CarriesModelMetadata()
+    public void ContributeConfig_SelectedDiscoveredModel_OmitsCapabilityOverrides()
     {
         using var step = new ProviderStepViewModel(_registry, _fakeProbe);
         step.SelectedProviderType = "OpenAI";
@@ -258,10 +258,16 @@ public void ContributeConfig_SelectedDiscoveredModel_CarriesModelMetadata()
         step.ContributeConfig(builder);
 
         Assert.NotNull(builder.Model);
-        Assert.Equal(512000, builder.Model!.ContextWindow);
-        Assert.Equal(ModelDiscoverySource.Live, builder.Model.Provenance);
-        Assert.Equal(ModelModality.Text | ModelModality.Image, builder.Model.InputModalities);
-        Assert.Equal(ModelModality.Text, builder.Model.OutputModalities);
+        Assert.Equal(ModelDiscoverySource.Live, builder.Model!.Provenance);
+
+        var config = builder.BuildConfigDictionary();
+        var models = (Dictionary<string, object>)config["Models"];
+        var roles = (Dictionary<string, object>)models["Roles"];
+        var definitions = (Dictionary<string, object>)models["Definitions"];
+        var main = (Dictionary<string, object>)definitions[(string)roles["Main"]];
+        Assert.False(main.ContainsKey("ContextWindow"));
+        Assert.False(main.ContainsKey("InputModalities"));
+        Assert.False(main.ContainsKey("OutputModalities"));
     }
 
     [Fact]
@@ -366,7 +372,7 @@ public void GitHubCopilotEnterpriseInputs_RejectInvalidValuesBeforeVendorOptions
     }
 
     [Fact]
-    public void ContributeConfig_DiscoveredModelWithoutModalities_PersistsNone()
+    public void ContributeConfig_DiscoveredModelWithoutModalities_OmitsCapabilityOverrides()
     {
         // An openai-compatible /v1/models listing reports no modalities, so the
         // discovered model leaves them unset. The wizard must NOT bake a guessed Text
@@ -386,9 +392,16 @@ public void ContributeConfig_DiscoveredModelWithoutModalities_PersistsNone()
         step.ContributeConfig(builder);
 
         Assert.NotNull(builder.Model);
-        Assert.Equal(32768, builder.Model!.ContextWindow);
-        Assert.Null(builder.Model.InputModalities);
-        Assert.Null(builder.Model.OutputModalities);
+        Assert.Equal(ModelDiscoverySource.Live, builder.Model!.Provenance);
+
+        var config = builder.BuildConfigDictionary();
+        var models = (Dictionary<string, object>)config["Models"];
+        var roles = (Dictionary<string, object>)models["Roles"];
+        var definitions = (Dictionary<string, object>)models["Definitions"];
+        var main = (Dictionary<string, object>)definitions[(string)roles["Main"]];
+        Assert.False(main.ContainsKey("ContextWindow"));
+        Assert.False(main.ContainsKey("InputModalities"));
+        Assert.False(main.ContainsKey("OutputModalities"));
     }
 
     [Fact]
diff --git a/src/Netclaw.Cli/Config/ModelEntryWriter.cs b/src/Netclaw.Cli/Config/ModelEntryWriter.cs
index 4df3ef243..465feddc3 100644
--- a/src/Netclaw.Cli/Config/ModelEntryWriter.cs
+++ b/src/Netclaw.Cli/Config/ModelEntryWriter.cs
@@ -62,23 +62,13 @@ internal static bool ClearRole(Dictionary<string, object> modelsSection, string
     }
 
     /// <summary>
-    /// Write a role's model entry into <paramref name="modelsSection"/> non-destructively.
-    /// Two on-disk attributes are treated as operator-owned overrides that provider discovery
-    /// must never silently clobber on a same-<c>(provider, modelId)</c> re-set:
-    /// <list type="bullet">
-    /// <item><description>
-    /// <c>ContextWindow</c> and the modalities are documented to "take precedence over
-    /// provider-reported capability detection". So the precedence for each is:
-    /// explicit operator input (this call) &gt; existing stored value &gt; probe. A fresh probe
-    /// tops up a first-time set or a model switch, but never overwrites a value already on disk
-    /// — that overwrite was the #1127 loss (a re-set wiped a hand-set modality) and its
-    /// context-window twin (#1610).
-    /// </description></item>
-    /// </list>
-    /// Because the stored value now wins, the operator needs a way to *change* it: an explicit
+    /// Write a role's model entry into <paramref name="modelsSection"/> without data loss.
+    /// <c>ContextWindow</c> and the modalities are operator-owned overrides.
+    /// Provider discovery must not create or change these overrides.
+    /// An explicit
     /// <see cref="ValueOverride{T}.Set"/> replaces it and <see cref="ValueOverride{T}.Clear"/>
-    /// removes it (falling back to runtime detection). Switching a role to a <em>different</em>
-    /// model does not carry the old model's attributes over — they belonged to that model.
+    /// removes it. Runtime capability detection resolves an absent override.
+    /// A model switch does not copy overrides from the old model.
     /// </summary>
     /// <param name="contextWindow">
     /// Operator intent for the context window (set / clear / unset). <c>--context-window</c> sets
@@ -87,10 +77,6 @@ internal static bool ClearRole(Dictionary<string, object> modelsSection, string
     /// </param>
     /// <param name="inputModalities">Operator intent for input modalities (set / clear / unset).</param>
     /// <param name="outputModalities">Operator intent for output modalities (set / clear / unset).</param>
-    /// <param name="discovered">
-    /// The probe result, if a probe ran. Its context window and modalities seed a first-time set
-    /// or a model switch only; an existing stored value wins over them.
-    /// </param>
     internal static void WriteRole(
         Dictionary<string, object> modelsSection,
         string roleKey,
@@ -99,8 +85,7 @@ internal static void WriteRole(
         ModelDiscoverySource? provenance,
         ValueOverride<int> contextWindow,
         ValueOverride<ModelModality> inputModalities,
-        ValueOverride<ModelModality> outputModalities,
-        DiscoveredModel? discovered)
+        ValueOverride<ModelModality> outputModalities)
     {
         var (definitions, roles) = EnsureNamedShape(modelsSection, roleKey);
         var definitionName = FindDefinition(definitions, provider, modelId)
@@ -114,15 +99,11 @@ internal static void WriteRole(
         if (existing?.Provenance is { } priorProvenance && provenance != ModelDiscoverySource.Live)
             provenance = priorProvenance;
 
-        // Precedence for every operator-owned attribute: explicit input > existing stored value
-        // > probe. Collapsing the explicit and discovered values in the caller defeated
-        // preservation, because the probe/picker paths always pass a discovered value, so the
-        // stored value was overwritten on every re-selection.
-        var sameModelEntry = existing is not null;
-        var resolvedWindow = ResolveContextWindow(
-            contextWindow, sameModelEntry, existing?.ContextWindow, discovered?.ContextWindowTokens);
-        var resolvedInput = ResolveModality(inputModalities, sameModelEntry, existing?.InputModalities, discovered?.InputModalities);
-        var resolvedOutput = ResolveModality(outputModalities, sameModelEntry, existing?.OutputModalities, discovered?.OutputModalities);
+        // Explicit operator input wins. Otherwise, the existing operator override remains.
+        // A new definition stays empty so runtime detection owns dynamic capabilities (#1756).
+        var resolvedWindow = contextWindow.Supplied ? contextWindow.Value : existing?.ContextWindow;
+        var resolvedInput = inputModalities.Supplied ? inputModalities.Value : existing?.InputModalities;
+        var resolvedOutput = outputModalities.Supplied ? outputModalities.Value : existing?.OutputModalities;
 
         definitions[definitionName] = BuildModelEntry(
             provider, modelId, provenance, resolvedWindow, resolvedInput, resolvedOutput);
@@ -284,35 +265,6 @@ private static bool Equivalent(ModelReference left, ModelReference right)
            && left.InputModalities == right.InputModalities
            && left.OutputModalities == right.OutputModalities;
 
-    /// <summary>
-    /// Resolves the modality actually written. An explicit operator set or clear wins outright
-    /// (the operator is the authority on a manual override). Otherwise, when the same model already
-    /// has an entry on disk, that entry is honored verbatim — including a deliberately-cleared
-    /// (absent) modality, so a later probe cannot silently resurrect an override the operator
-    /// removed with <c>--clear-modalities</c> (#1610). Provider discovery only seeds a genuine gap:
-    /// a first-time set or a switch to a different model, where no entry exists yet.
-    /// </summary>
-    private static ModelModality? ResolveModality(
-        ValueOverride<ModelModality> @override, bool sameModelEntryExists,
-        ModelModality? existing, ModelModality? discovered)
-        => @override.Supplied
-            ? @override.Value                 // Set(value) → value; Clear → null (key omitted downstream)
-            : sameModelEntryExists ? existing  // same model on disk: honor it, incl. a cleared (null) value
-                : discovered;                  // first set / model switch: seed from discovery
-
-    /// <summary>
-    /// Resolves the context window actually written. Mirrors <see cref="ResolveModality"/> for the
-    /// explicit cases: <c>--context-window</c> (Set) or <c>--clear-context-window</c> (Clear) wins,
-    /// otherwise an existing definition is honored verbatim, including absence. Discovery seeds
-    /// only a new definition. This keeps manual JSON edits and explicit clears stable without a
-    /// hidden tombstone representation.
-    /// </summary>
-    private static int? ResolveContextWindow(
-        ValueOverride<int> @override, bool sameModelEntryExists, int? existing, int? discovered)
-        => @override.Supplied
-            ? @override.Value            // Set(n) → n; Clear → null (drop the clamp → runtime detects)
-            : sameModelEntryExists ? existing : discovered;
-
     /// <summary>
     /// The role's current entry, but only when it already references the same
     /// <c>(provider, modelId)</c>; null when the role is unset or points at a different
@@ -461,16 +413,9 @@ private static bool RawContainsProperty(object raw, string propertyName)
     }
 
     /// <summary>
-    /// Builds the dictionary written under <c>Models[role]</c>.
+    /// Builds a model definition from operator overrides or legacy persisted values.
+    /// Null capability values stay absent so runtime detection can resolve them.
     /// </summary>
-    /// <remarks>
-    /// Modalities are written ONLY when the discovery source genuinely reported them
-    /// (non-null). A null modality means "the provider did not say" — it is
-    /// deliberately omitted so the daemon's capability detection resolves it at
-    /// runtime. Writing a guessed <see cref="ModelModality.Text"/> here would bake a
-    /// permanent override into config that beats real detection on every boot, which
-    /// is exactly what silently demoted multimodal self-hosted models to text-only.
-    /// </remarks>
     internal static Dictionary<string, object> BuildModelEntry(
         string provider,
         string? modelId,
@@ -505,14 +450,14 @@ internal static Dictionary<string, object> BuildModelEntry(
 /// a modality set (<see cref="ModelModality"/>) or the context window (<see cref="int"/>). A plain
 /// <c>T?</c> cannot express it, because two of the three states both resolve to null yet behave
 /// oppositely: "not supplied" must preserve any existing override, while "clear" must win over it.
-/// The tri-state is: <see cref="Unset"/> (leave it to the stored value / discovery),
+/// The tri-state is: <see cref="Unset"/> (leave the stored value unchanged),
 /// <see cref="Set"/> (replace with an explicit value), and <see cref="Clear"/> (remove the override
 /// so runtime detection resolves it).
 /// </summary>
 internal readonly record struct ValueOverride<T>(bool Supplied, T? Value)
     where T : struct
 {
-    /// <summary>Operator said nothing — preserve the stored value, else fall back to discovery.</summary>
+    /// <summary>Operator said nothing. Preserve the stored value.</summary>
     internal static ValueOverride<T> Unset => default;
 
     /// <summary>Operator asked to remove the override so runtime capability detection resolves it.</summary>
diff --git a/src/Netclaw.Cli/Model/ModelCommand.cs b/src/Netclaw.Cli/Model/ModelCommand.cs
index c811789c2..227d4daf0 100644
--- a/src/Netclaw.Cli/Model/ModelCommand.cs
+++ b/src/Netclaw.Cli/Model/ModelCommand.cs
@@ -235,17 +235,10 @@ private static async Task<int> RunSetAsync(
             }
         }
 
-        // Only an explicit --context-window short-circuits the probe: it supplies the one datum the
-        // probe would discover, so it is the documented "configure this model manually" escape
-        // hatch. A modality override must NOT skip the probe — the probe also validates the model
-        // exists and discovers the context window, and WriteRole already lets an explicit modality
-        // override win over any discovered value, so it is safe to keep probing. --clear-context-window
-        // likewise keeps the probe: "re-detect" is exactly what the probe does.
-        var manualMetadataSupplied = contextWindow.HasValue;
-
-        DiscoveredModel? discoveredModel = null;
+        // An explicit context window permits a manual model ID.
+        // All other OAuth selections require validation against the live model list.
         var provenance = ModelDiscoverySource.Manual;
-        if (!manualMetadataSupplied && ShouldProbeForMetadata(providerEntry))
+        if (!contextWindow.HasValue && ShouldProbeForMetadata(providerEntry))
         {
             probe ??= ProviderCommand.CreateDefaultRegistry();
             ProviderProbeResult probeResult;
@@ -258,7 +251,7 @@ private static async Task<int> RunSetAsync(
                 return 1;
             }
 
-            discoveredModel = probeResult.Models.FirstOrDefault(m =>
+            var discoveredModel = probeResult.Models.FirstOrDefault(m =>
                 string.Equals(m.ModelId.Value, modelId, StringComparison.OrdinalIgnoreCase));
             if (discoveredModel is null)
             {
@@ -274,9 +267,8 @@ private static async Task<int> RunSetAsync(
         var (config, _) = ConfigFileHelper.LoadConfigFiles(paths);
         var modelsSection = ConfigFileHelper.GetOrCreateSection(config, "Models");
 
-        // Definitions own model metadata, so role switches never destroy another model's
-        // operator-owned overrides. Discovery seeds only a new definition; explicit input edits
-        // an existing definition and absence remains runtime detection (#1127, #1610).
+        // Definitions own model metadata. Role switches preserve another model's overrides.
+        // Only explicit operator input persists capability overrides (#1756).
         ModelEntryWriter.WriteRole(
             modelsSection,
             roleKey,
@@ -285,8 +277,7 @@ private static async Task<int> RunSetAsync(
             provenance,
             contextWindowOverride,
             inputOverride,
-            outputOverride,
-            discoveredModel);
+            outputOverride);
         ConfigFileHelper.WriteConfigFile(paths.NetclawConfigPath, config);
 
         writer.WriteLine($"Set {role} model to {providerName}/{modelId}");
diff --git a/src/Netclaw.Cli/Tui/ModelManagerViewModel.cs b/src/Netclaw.Cli/Tui/ModelManagerViewModel.cs
index efdbf3447..fa26f37da 100644
--- a/src/Netclaw.Cli/Tui/ModelManagerViewModel.cs
+++ b/src/Netclaw.Cli/Tui/ModelManagerViewModel.cs
@@ -194,10 +194,8 @@ public void ConfirmAssignment()
         var (config, _) = ConfigFileHelper.LoadConfigFiles(_paths);
         var modelsSection = ConfigFileHelper.GetOrCreateSection(config, "Models");
 
-        // Non-destructive: re-assigning the same model preserves an existing context-window
-        // clamp and modality overrides, none of which the picker can supply (#1127, #1610). The
-        // picker has no manual-override inputs, so it passes no explicit context window and Unset
-        // modality intent — the probe result seeds a first-time set only; existing values win.
+        // The picker has no inputs for capability overrides.
+        // It preserves existing overrides and leaves new definitions empty (#1756).
         ModelEntryWriter.WriteRole(
             modelsSection,
             roleKey,
@@ -206,8 +204,7 @@ public void ConfirmAssignment()
             provenance,
             ValueOverride<int>.Unset,
             ValueOverride<ModelModality>.Unset,
-            ValueOverride<ModelModality>.Unset,
-            discoveredModel);
+            ValueOverride<ModelModality>.Unset);
         ConfigFileHelper.WriteConfigFile(_paths.NetclawConfigPath, config);
 
         Refresh();
diff --git a/src/Netclaw.Cli/Tui/Wizard/Steps/ProviderStepViewModel.cs b/src/Netclaw.Cli/Tui/Wizard/Steps/ProviderStepViewModel.cs
index 4a684e89f..8a95ed989 100644
--- a/src/Netclaw.Cli/Tui/Wizard/Steps/ProviderStepViewModel.cs
+++ b/src/Netclaw.Cli/Tui/Wizard/Steps/ProviderStepViewModel.cs
@@ -462,10 +462,7 @@ public void ContributeConfig(WizardConfigBuilder builder)
         {
             Provider = providerName,
             ModelId = SelectedModelId,
-            ContextWindow = selectedModel?.ContextWindowTokens,
             Provenance = selectedModel is null ? ModelDiscoverySource.Manual : ModelDiscoverySource.Live,
-            InputModalities = selectedModel?.InputModalities,
-            OutputModalities = selectedModel?.OutputModalities,
         };
     }
 
diff --git a/src/Netclaw.Cli/Tui/Wizard/WizardConfigBuilder.cs b/src/Netclaw.Cli/Tui/Wizard/WizardConfigBuilder.cs
index 6ebf3811f..5d56caf07 100644
--- a/src/Netclaw.Cli/Tui/Wizard/WizardConfigBuilder.cs
+++ b/src/Netclaw.Cli/Tui/Wizard/WizardConfigBuilder.cs
@@ -141,16 +141,9 @@ internal Dictionary<string, object> BuildConfigDictionary()
                 Model.Provider,
                 Model.ModelId,
                 Model.Provenance,
-                Model.ContextWindow is { } contextWindow
-                    ? ValueOverride<int>.Set(contextWindow)
-                    : ValueOverride<int>.Unset,
-                Model.InputModalities is { } input
-                    ? ValueOverride<ModelModality>.Set(input)
-                    : ValueOverride<ModelModality>.Unset,
-                Model.OutputModalities is { } output
-                    ? ValueOverride<ModelModality>.Set(output)
-                    : ValueOverride<ModelModality>.Unset,
-                discovered: null);
+                ValueOverride<int>.Unset,
+                ValueOverride<ModelModality>.Unset,
+                ValueOverride<ModelModality>.Unset);
         }
 
         // Slack section
@@ -580,10 +573,7 @@ public sealed class ModelConfigSection
 {
     public required string Provider { get; init; }
     public string? ModelId { get; init; }
-    public int? ContextWindow { get; init; }
     public ModelDiscoverySource? Provenance { get; init; }
-    public ModelModality? InputModalities { get; init; }
-    public ModelModality? OutputModalities { get; init; }
 }
 
 public sealed class SlackConfigSection
diff --git a/tests/smoke/assertions/init-wizard.sh b/tests/smoke/assertions/init-wizard.sh
index 2f7185680..94b0fd793 100755
--- a/tests/smoke/assertions/init-wizard.sh
+++ b/tests/smoke/assertions/init-wizard.sh
@@ -63,6 +63,9 @@ assert_field '.Providers.ollama.Endpoint'   'http://localhost:11434'   "$config_
 assert_field '.Models.Roles.Main'                                  'ollama-qwen2-0-5b' "$config_json" || :
 assert_field '.Models.Definitions[.Models.Roles.Main].Provider'     'ollama'             "$config_json" || :
 assert_field '.Models.Definitions[.Models.Roles.Main].ModelId'      'qwen2:0.5b'         "$config_json" || :
+assert_field '(.Models.Definitions[.Models.Roles.Main] | has("ContextWindow"))'     'false' "$config_json" || :
+assert_field '(.Models.Definitions[.Models.Roles.Main] | has("InputModalities"))'   'false' "$config_json" || :
+assert_field '(.Models.Definitions[.Models.Roles.Main] | has("OutputModalities"))'  'false' "$config_json" || :
 assert_field '.Security.DeploymentPosture'  'Personal'                 "$config_json" || :
 
 echo "init-wizard: checking identity/SOUL.md for typed user name..."