diff --git a/packages/cli/src/config/config.test.ts b/packages/cli/src/config/config.test.ts
index c060ba634e0..e7ff46b461a 100644
--- a/packages/cli/src/config/config.test.ts
+++ b/packages/cli/src/config/config.test.ts
@@ -1310,6 +1310,63 @@ describe('mergeExcludeTools', () => {
     );
     expect(config.getPermissionsDeny()).toHaveLength(2);
   });
+
+  it('should add tool_search to deny list when tools.toolSearch.enabled is false', async () => {
+    process.argv = ['node', 'script.js'];
+    const argv = await parseArguments();
+    const settings: Settings = {
+      tools: { toolSearch: { enabled: false } },
+    };
+    const config = await loadCliConfig(settings, argv, undefined, []);
+    expect(config.getPermissionsDeny()).toContain('tool_search');
+  });
+
+  it('should auto-disable tool_search for deepseek-v4 models', async () => {
+    process.argv = ['node', 'script.js', '--model', 'deepseek-v4-flash'];
+    const argv = await parseArguments();
+    const settings: Settings = {};
+    const config = await loadCliConfig(settings, argv, undefined, []);
+    expect(config.getPermissionsDeny()).toContain('tool_search');
+  });
+
+  it('should auto-disable tool_search for deepseek-v3 models', async () => {
+    process.argv = ['node', 'script.js', '--model', 'deepseek-v3'];
+    const argv = await parseArguments();
+    const settings: Settings = {};
+    const config = await loadCliConfig(settings, argv, undefined, []);
+    expect(config.getPermissionsDeny()).toContain('tool_search');
+  });
+
+  it('should auto-disable tool_search for deepseek-chat models with provider prefix', async () => {
+    process.argv = [
+      'node',
+      'script.js',
+      '--model',
+      'openrouter/deepseek/deepseek-chat',
+    ];
+    const argv = await parseArguments();
+    const settings: Settings = {};
+    const config = await loadCliConfig(settings, argv, undefined, []);
+    expect(config.getPermissionsDeny()).toContain('tool_search');
+  });
+
+  it('should not auto-disable tool_search for non-deepseek models', async () => {
+    process.argv = ['node', 'script.js', '--model', 'qwen-max'];
+    const argv = await parseArguments();
+    const settings: Settings = {};
+    const config = await loadCliConfig(settings, argv, undefined, []);
+    expect(config.getPermissionsDeny()).not.toContain('tool_search');
+  });
+
+  it('should respect explicit enabled:true override for deepseek models', async () => {
+    process.argv = ['node', 'script.js', '--model', 'deepseek-v4-flash'];
+    const argv = await parseArguments();
+    const settings: Settings = {
+      tools: { toolSearch: { enabled: true } },
+    };
+    const config = await loadCliConfig(settings, argv, undefined, []);
+    expect(config.getPermissionsDeny()).not.toContain('tool_search');
+  });
 });
 
 describe('Approval mode tool exclusion logic', () => {
diff --git a/packages/cli/src/config/config.ts b/packages/cli/src/config/config.ts
index a765a140950..f829665b1c0 100755
--- a/packages/cli/src/config/config.ts
+++ b/packages/cli/src/config/config.ts
@@ -1472,6 +1472,27 @@ export async function loadCliConfig(
 
   const { model: resolvedModel } = resolvedCliConfig;
 
+  // Disable ToolSearch when explicitly configured or for models that benefit
+  // from prefix-based KV caching. DeepSeek models (v3, v4, deepseek-chat)
+  // all use prefix-based disk KV caching with heavily discounted cached
+  // token pricing (up to 1/120 for v4). When tool_search is in the deny
+  // list, client.ts eagerly reveals all deferred tools so every MCP tool
+  // schema is in the initial declaration list, keeping the prompt prefix
+  // stable and maximizing cache hit rates.
+  // Note: no `^` anchor — model names may include a provider prefix
+  // (e.g. "openrouter/deepseek/deepseek-v4-flash").
+  const toolSearchExplicitlyEnabled = settings.tools?.toolSearch?.enabled;
+  const shouldDisableToolSearch =
+    toolSearchExplicitlyEnabled === false ||
+    (toolSearchExplicitlyEnabled === undefined &&
+      resolvedModel !== undefined &&
+      /deepseek-(v3|v4|chat)/i.test(resolvedModel));
+  if (shouldDisableToolSearch) {
+    if (!mergedDeny.includes('tool_search')) {
+      mergedDeny.push('tool_search');
+    }
+  }
+
   const sandboxConfig = await loadSandboxConfig(
     bareMode ? ({} as Settings) : settings,
     argv,
diff --git a/packages/cli/src/config/settingsSchema.ts b/packages/cli/src/config/settingsSchema.ts
index 709d14b08c9..07af4b0a77f 100644
--- a/packages/cli/src/config/settingsSchema.ts
+++ b/packages/cli/src/config/settingsSchema.ts
@@ -1461,6 +1461,27 @@ const SETTINGS_SCHEMA = {
           'Sandbox image URI used by Docker/Podman when --sandbox-image and QWEN_SANDBOX_IMAGE are not set.',
         showInDialog: false,
       },
+      toolSearch: {
+        type: 'object',
+        label: 'Tool Search',
+        category: 'Tools',
+        requiresRestart: true,
+        default: {},
+        description: 'Settings for the ToolSearch discovery mechanism.',
+        showInDialog: false,
+        properties: {
+          enabled: {
+            type: 'boolean',
+            label: 'Enable ToolSearch',
+            category: 'Tools',
+            requiresRestart: true,
+            default: true,
+            description:
+              'When enabled, MCP tools are loaded on-demand via ToolSearch to reduce prompt size. Disable this for models that rely on prefix-based KV caching (e.g. DeepSeek) to keep the prompt prefix stable and maximize cache hit rates.',
+            showInDialog: true,
+          },
+        },
+      },
       shell: {
         type: 'object',
         label: 'Shell',
diff --git a/packages/vscode-ide-companion/schemas/settings.schema.json b/packages/vscode-ide-companion/schemas/settings.schema.json
index fe4776b1083..7a6bb99c312 100644
--- a/packages/vscode-ide-companion/schemas/settings.schema.json
+++ b/packages/vscode-ide-companion/schemas/settings.schema.json
@@ -643,6 +643,17 @@
           "description": "Sandbox image URI used by Docker/Podman when --sandbox-image and QWEN_SANDBOX_IMAGE are not set.",
           "type": "string"
         },
+        "toolSearch": {
+          "description": "Settings for the ToolSearch discovery mechanism.",
+          "type": "object",
+          "properties": {
+            "enabled": {
+              "description": "When enabled, MCP tools are loaded on-demand via ToolSearch to reduce prompt size. Disable this for models that rely on prefix-based KV caching (e.g. DeepSeek) to keep the prompt prefix stable and maximize cache hit rates.",
+              "type": "boolean",
+              "default": true
+            }
+          }
+        },
         "shell": {
           "description": "Settings for shell execution.",
           "type": "object",