fix: count registered MCP tool schemas in the compaction overhead budget

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JgzuxyafKNF2uaEeL35XQw

backend/windmill-mcp/src/common/schema.rs

@@ -235,6 +235,9 @@ fn is_valid_json_schema_type(t: &str) -> bool {
 ///   untyped fields) so the node validates as "any type"
 /// - Removing `default: null` when the type doesn't include `null`
 /// - Adding `type: "object"` to property-bearing schemas that have no type
+/// - De-duplicating `required` at every depth, and dropping non-string entries
+///   (`required` is `uniqueItems`, and a repeat makes a strict client drop the
+///   whole tool rather than report an error)
 pub fn make_schema_compatible(schema: &mut Value) {
     let Value::Object(obj) = schema else { return };
 

frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts

@@ -1293,9 +1293,16 @@ export class AIChatManager {
 			typeof this.systemMessage.content === 'string'
 				? this.systemMessage.content.length / tokenPerCharacter
 				: 0
+		// The same set the request carries: registered MCP tools are appended by the
+		// config getter, and their remote schemas are the largest definitions in the
+		// list — leaving them out overstates the tail budget compaction may keep.
+		const sentTools = [
+			...this.tools,
+			...(this.mode === AIMode.GLOBAL ? loadedMcpTools(this.mcpOwnerId) : [])
+		]
 		const toolTokens =
-			this.tools.length > 0
-				? JSON.stringify(this.tools.map((t) => t.def)).length / tokenPerCharacter
+			sentTools.length > 0
+				? JSON.stringify(sentTools.map((t) => t.def)).length / tokenPerCharacter
 				: 0
 		return systemTokens + toolTokens
 	}