added enable and disable commands

package.json

@@ -2,6 +2,7 @@
     "name": "dumbpilot",
     "displayName": "dumbpilot",
     "description": "Simple code prediction using llama.cpp server api",
+    "publisher": "Alessandro Mauri",
     "version": "0.0.1",
     "engines": {
         "vscode": "^1.84.0"
@@ -102,10 +103,23 @@
     ],
     "main": "./out/extension.js",
     "contributes": {
-        "commands": [],
+        "commands": [
+            {
+                "command": "dumbpilot.disableCompletion",
+                "title": "Disable predictive code completion",
+                "category": "dumbpilot"
+            },
+            {
+                "command": "dumbpilot.enableCompletion",
+                "title": "Enable predictive code completion",
+                "category": "dumbpilot"
+            }
+        ],
         "keybindings": [
             {
                 "key": "ctrl+shift+l",
+                "mac": "cmd+shift+l",
+                "when": "editorTextFocus",
                 "command": "editor.action.inlineSuggest.trigger"
             }
         ]

src/extension.ts

@@ -4,7 +4,7 @@ import commentPrefix from './comments.json';
 
 
 // llama.cpp server response format
-type llama_data = {
+type llamaData = {
 	content: string,
 	generation_settings: JSON,
 	model: string,
@@ -27,7 +27,41 @@ type llama_data = {
 	tokens_evaluated: number,
 	tokens_predicted: number,
 	truncated: boolean
+};
+
+type llamaCompletionRequest = {
+	n_predict: number,
+	mirostat: number,
+	repeat_penalty: number,
+	frequency_penalty: number, 
+	presence_penalty: number, 
+	repeat_last_n: number, 
+	temperature: number,
+	top_p: number,
+	top_k: number,
+	typical_p: number,
+	tfs_z: number, 
+	seed: number,
+	stream: boolean,
+	prompt: string,
+};
 
+type llamaFillRequest = {
+	n_predict: number,
+	mirostat: number,
+	repeat_penalty: number,
+	frequency_penalty: number, 
+	presence_penalty: number, 
+	repeat_last_n: number, 
+	temperature: number,
+	top_p: number,
+	top_k: number,
+	typical_p: number,
+	tfs_z: number, 
+	seed: number,
+	stream: boolean,
+	input_prefix: string,
+	input_suffix: string,
 };
 
 
@@ -48,18 +82,49 @@ const llama_session_seed = -1;
 const llama_host = "http://0.0.0.0:8080";
 
 
-export function activate(cotext: vscode.ExtensionContext) {
+// clean up the document
+function clean_text(txt: string): string {
+	// these are already done by JSON.stringify()
+	//txt = txt.replace(/(\r\n|\n|\r)/gm, "\\n");
+	//txt = txt.replace((/\t/gm, "\\t"));
+	
+	// FIXME: I don't know if this penalizes some results since most people indent with spaces
+	//txt = txt.replace(/\s+/gm, " ");
+	return txt;
+}
+
+
+export function activate(context: vscode.ExtensionContext) {
 
 	console.log('dumbpilot is now active');
 
+	const config = vscode.workspace.getConfiguration("dumbpilot");
+	var completion_enabled: boolean = config.get("completionEnabled") as boolean;
+
+	// TODO: work with local configurations
+	let disposable = vscode.commands.registerCommand("dumbpilot.enableCompletion", () => {
+		completion_enabled = true;
+		config.update("completionEnabled", true);
+	});
+	context.subscriptions.push(disposable);
+
+	disposable = vscode.commands.registerCommand("dumbpilot.disableCompletion", () => {
+		completion_enabled = false;
+		config.update("completionEnabled", false);
+	});
+
 	// Register a new provider of inline completions, this does not decide how it is invoked
 	// only what the completion should be
 	// https://github.com/microsoft/vscode-extension-samples/blob/main/inline-completions/src/extension.ts
 	const provider: vscode.InlineCompletionItemProvider = {
 		async provideInlineCompletionItems(document, position, context, token) {
 
+			if (completion_enabled === false) {
+				return null;
+			}
+			
 			// Since for every completion we want to query the server, we want to filter out
-			// accidental automatic completion invokes
+			// automatic completion invokes
 			if (context.triggerKind === vscode.InlineCompletionTriggerKind.Automatic) {
 				return null;
 			}
@@ -82,27 +147,23 @@ export function activate(cotext: vscode.ExtensionContext) {
 			var doc_before = doc_text.substring(0, doc_off);
 			var doc_after = doc_text.substring(doc_off);
 			
-			// clean up the document, stuff commented is already done by JSON.stringify()
-			//doc_before = doc_before.replace(/(\r\n|\n|\r)/gm, "\\n");
-			//doc_before = doc_before.replace(/\t/gm, "\\t");
-			//doc_after = doc_before.replace(/(\r\n|\n|\r)/gm, "\\n");
-			//doc_after = doc_before.replace(/\t/gm, "\\t");
+			// make it cleaner in hope to reduce the number of tokens
+			doc_before = clean_text(doc_before);
+			doc_after = clean_text(doc_after);
 
-			// FIXME: I don't know if this penalizes some results since most people indent with spaces
-			//doc_before = doc_before.replace(/\s+/gm, " ");
-			//doc_after = doc_before.replace(/\s+/gm, " ");
+			// TODO: prune text up to a maximum context length
 
-			// prefix commented filename, is this the best way?
-			var pfx: string = '';
-			var sfx: string = '';
+			// Prefix the filename in a comment
+			var pfx: string, sfx: string;
 			const lang = document.languageId;
 			const prefixes = commentPrefix;
 			pfx = (prefixes as any)[lang][0] as string;
 			sfx = (prefixes as any)[lang][1] as string;
+			// FIXME: is there a more efficient way?
 			doc_before = pfx + ' ' + document.fileName + sfx + '\n' + doc_before;
 
 			// server request object
-			const request = {
+			const request: llamaCompletionRequest = {
 				n_predict: llama_maxtokens,
 				mirostat: llama_mirostat,
 				repeat_penalty: llama_repeat_penalty,
@@ -115,12 +176,11 @@ export function activate(cotext: vscode.ExtensionContext) {
 				typical_p: llama_typical_p,
 				tfs_z: llama_tailfree_z, 
 				seed: llama_session_seed,
-//				input_prefix: doc_before,
-//				input_suffix: doc_after,
+				stream: false,
 				prompt: doc_before,
 			};
 
-			var data: llama_data;
+			var data: llamaData;
 			// try to send the request to the running server
 			try {
 				const response = await fetch(
@@ -137,7 +197,7 @@ export function activate(cotext: vscode.ExtensionContext) {
 					throw new Error("llama server request is not ok??");
 				}
 		
-				data = await response.json() as llama_data;
+				data = await response.json() as llamaData;
 
 			} catch (e: any) {
 				console.log('dumbpilot: ' + e.message);