diff --git a/README.md b/README.md
index 76bc0637ee..98bfb68f7e 100644
--- a/README.md
+++ b/README.md
@@ -10,7 +10,7 @@ Instead of using one model for everything, Codebuff coordinates specialized agen
   <img src="./assets/codebuff-vs-claude-code.png" alt="Codebuff vs Claude Code" width="400">
 </div>
 
-Codebuff beats Claude Code at 61% vs 53% on [our evals](evals/README.md) across 175+ coding tasks over multiple open-source repos that simulate real-world tasks.
+Codebuff beats Claude Code at 61% vs 53% on [our evals](evals/buffbench/README.md) across 175+ coding tasks over multiple open-source repos that simulate real-world tasks.
 
 ## Freebuff: the free coding agent
 
diff --git a/packages/agent-runtime/src/__tests__/prompts-schema-handling.test.ts b/packages/agent-runtime/src/__tests__/prompts-schema-handling.test.ts
index 6d371bf59e..429f5005a9 100644
--- a/packages/agent-runtime/src/__tests__/prompts-schema-handling.test.ts
+++ b/packages/agent-runtime/src/__tests__/prompts-schema-handling.test.ts
@@ -247,6 +247,43 @@ describe('Schema handling error recovery', () => {
       expect(toolSet['problematic_tool']).toBeDefined()
     })
 
+    test('getToolSet preserves MCP JSON Schema properties', async () => {
+      const customToolDefs = {
+        marionette__connect: {
+          description: 'Connect to the VM service',
+          inputSchema: {
+            type: 'object',
+            properties: {
+              uri: {
+                type: 'string',
+                description: 'VM service URI',
+              },
+            },
+            required: ['uri'],
+            additionalProperties: false,
+          },
+          endsAgentStep: true,
+        },
+      }
+
+      const toolSet = await getToolSet({
+        toolNames: [],
+        additionalToolDefinitions: async () => customToolDefs,
+        agentTools: {},
+        skills: {},
+      })
+
+      const inputSchema = toolSet['marionette__connect']?.inputSchema as {
+        properties?: Record<string, unknown>
+      }
+
+      expect(inputSchema).toBeDefined()
+      expect(inputSchema.properties).toBeDefined()
+      expect(inputSchema.properties?.uri).toEqual(
+        expect.objectContaining({ type: 'string' }),
+      )
+    })
+
     test('ensureZodSchema converts JSON Schema to Zod schema', () => {
       const jsonSchema = {
         type: 'object',
diff --git a/packages/agent-runtime/src/tools/prompts.ts b/packages/agent-runtime/src/tools/prompts.ts
index c87aaf875d..8f54f14533 100644
--- a/packages/agent-runtime/src/tools/prompts.ts
+++ b/packages/agent-runtime/src/tools/prompts.ts
@@ -392,13 +392,14 @@ export async function getToolSet(params: {
   const toolDefinitions = await additionalToolDefinitions()
   for (const [toolName, toolDefinition] of Object.entries(toolDefinitions)) {
     const clonedDef = cloneDeep(toolDefinition)
-    // Custom tool inputSchema may be JSON Schema (from SDK) or Zod (from MCP)
-    // Ensure it's a Zod schema for the AI SDK
-    const zodSchema = ensureZodSchema(clonedDef.inputSchema)
-    const safeSchema = ensureJsonSchemaCompatible(zodSchema)
+    // Custom tool inputSchema may be JSON Schema (from SDK) or Zod (from MCP).
+    // Keep the original schema object on the tool definition so nested JSON
+    // Schema details like `properties` survive registration unchanged. Use a
+    // Zod conversion only for description generation.
+    ensureZodSchema(clonedDef.inputSchema)
     toolSet[toolName] = {
       ...clonedDef,
-      inputSchema: safeSchema,
+      inputSchema: cloneDeep(clonedDef.inputSchema),
     } as (typeof toolSet)[string]
   }