Terminal: Quick fix ghost text with AI suggestions by federicobrancasi · Pull Request #295983 · microsoft/vscode

federicobrancasi · 2026-02-18T10:58:45Z

Description

This PR enhances the terminal quick fix feature by displaying command corrections as ghost text directly in the terminal prompt. Users can accept suggestions with Tab or Right Arrow.

When no built-in quick fix is available, it uses Copilot to suggest the correct command.

Demo

terminal.suggestions.mp4

Features

Feature	Example
Built-in quick fixes	`git statys` → `git status`
AI-powered corrections	`ld -la` → `ls -la`, `mkdor hello` → `mkdir hello`

New Settings

terminal.integrated.shellIntegration.quickFixGhostText - Show suggestions as ghost text
terminal.integrated.shellIntegration.quickFixAutoFill - Auto-fill the correction immediately

Implementation Details

Ghost text rendered using ANSI escape codes (\x1b[2m for dim text)
Tab/Right Arrow intercepted via attachCustomKeyEventHandler
AI uses ILanguageModelsService to call Copilot (gpt-4o-mini)
copilot-debug suggestions filtered out from ghost text

Console Logging (Debug)

Known Limitations / Future Work

AI prompt could include user's shell aliases for better suggestions
Command history context would help disambiguate intent
On PowerShell, ls -la gets suggested as Get-ChildItem -Force (shell-aware suggestions needed)
Remove debug console.log statements before merging

Prior Art

Inspired by Warp Terminal - Command Corrections

Testing

Set terminal.integrated.shellIntegration.quickFixGhostText: true
Open terminal, type git statys + Enter → ghost text appears
Press Tab to accept
Type ld -la + Enter → AI suggests ls -la

federicobrancasi · 2026-02-18T11:01:06Z

@microsoft-github-policy-service agree company="Microsoft"

Copilot

Pull request overview

Adds “ghost text” rendering for terminal quick fixes, with optional AI-backed command corrections when no built-in quick fix is available.

Changes:

Render first terminal-command quick fix as inline ghost text and allow accepting via Tab/Right Arrow.
Add AI command suggestion fallback using ILanguageModelsService (Copilot) when a command fails and no quick fix exists.
Introduce new terminal settings to control ghost text vs auto-fill behavior.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts	Implements ghost text rendering/acceptance and AI suggestion retrieval.
src/vs/workbench/contrib/terminalContrib/quickFix/browser/media/terminalQuickFix.css	Adds styling for a ghost-text decoration class (currently not wired up).
src/vs/workbench/contrib/terminal/common/terminalContextKey.ts	Adds a new context key for ghost text visibility (currently not set/used).
src/vs/workbench/contrib/terminal/common/terminalConfiguration.ts	Registers the new `quickFixGhostText` / `quickFixAutoFill` settings in the schema.
src/vs/platform/terminal/common/terminal.ts	Adds new `TerminalSettingId` entries for the settings.

Comments suppressed due to low confidence (3)

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts:203

CancellationTokenSource is created for the chat request but never disposed/cancelled. This can leak listeners/resources and also prevents cancelling the request when the terminal instance is disposed or a newer command arrives. Wrap it in try/finally { cts.dispose(); } and consider wiring cancellation to the contribution's lifecycle or to the per-command request handling.

			const cts = new CancellationTokenSource();
			const response = await this._languageModelsService.sendChatRequest(
				models[0],
				new ExtensionIdentifier('core'),
				[{ role: ChatMessageRole.User, content: [{ type: 'text', value: prompt }] }],
				{},
				cts.token
			);

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts:143

Ghost text is never cleared when quick fixes disappear. TerminalQuickFixAddon fires onDidUpdateQuickFixes with actions: undefined in _disposeQuickFix, and when getQuickFixesForCommand returns undefined there is no update event at all. In both cases currentGhostText can linger. Ensure clearGhostText() is called when e.actions is empty/undefined, and consider clearing when new quick fixes are not available.

		this.add(this._addon.onDidUpdateQuickFixes(async e => {
			// Only track the latest command's quick fixes
			this._quickFixMenuItems.value = e.actions ? xterm.decorationAddon.registerMenuItems(e.command, e.actions) : undefined;

			// Get the first terminal command quick fix (filter out copilot-debug suggestions)
			const terminalCommandFix = e.actions?.find(a =>
				a.type === TerminalQuickFixType.TerminalCommand &&
				a.command &&
				!a.command.startsWith('copilot-debug')
			);

			// Ghost text mode: write dim text to terminal
			const ghostTextEnabled = this._configurationService.getValue<boolean>(TerminalSettingId.ShellIntegrationQuickFixGhostText);
			if (ghostTextEnabled && terminalCommandFix?.command) {
				writeGhostText(terminalCommandFix.command);
				return;
			}

			// Auto-fill mode: fill the prompt immediately
			const autoFillEnabled = this._configurationService.getValue<boolean>(TerminalSettingId.ShellIntegrationQuickFixAutoFill);
			if (autoFillEnabled && terminalCommandFix?.command) {
				this._ctx.instance.runCommand(terminalCommandFix.command, false);
				return;
			}

			// AI-powered fix: if no quick fix found but command failed, try AI (requires Copilot)
			if (ghostTextEnabled && !terminalCommandFix && e.command) {
				const output = e.command.getOutput() ?? '';
				const commandLine = e.command.command;
				const exitCode = e.command.exitCode;
				// Try AI if: command failed (non-zero exit) AND output is short (< 1000 chars = likely error message)
				const shouldTryAI = exitCode !== 0 && output.length < 1000;
				if (shouldTryAI && commandLine) {
					const aiSuggestion = await this._getAICommandSuggestion(commandLine, output);
					if (aiSuggestion) {
						writeGhostText(aiSuggestion);
					}
				}
			}
		}));

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts:70

Rendering ghost text via xterm.raw.write mutates the terminal buffer (it becomes real text in the scrollback and can affect wrapping/copy/paste and prompt positioning). For a true “ghost”/overlay UI, consider using an xterm decoration (similar to the existing quick-fix lightbulb decoration) so the suggestion doesn't become part of the buffer contents.

		const writeGhostText = (text: string) => {
			// Write dim text (ANSI escape: \x1b[2m = dim, \x1b[0m = reset)
			// Also save cursor, write, restore cursor
			xterm.raw.write(`\x1b7\x1b[2m${text}\x1b8`);
			currentGhostText = text;
		};

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts

Copilot · 2026-02-18T11:04:46Z

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts

+			// AI-powered fix: if no quick fix found but command failed, try AI (requires Copilot)
+			if (ghostTextEnabled && !terminalCommandFix && e.command) {
+				const output = e.command.getOutput() ?? '';
+				const commandLine = e.command.command;
+				const exitCode = e.command.exitCode;
+				// Try AI if: command failed (non-zero exit) AND output is short (< 1000 chars = likely error message)
+				const shouldTryAI = exitCode !== 0 && output.length < 1000;
+				if (shouldTryAI && commandLine) {
+					const aiSuggestion = await this._getAICommandSuggestion(commandLine, output);
+					if (aiSuggestion) {
+						writeGhostText(aiSuggestion);
+					}
+				}
+			}


The AI request can race with subsequent commands/quick-fix updates: multiple onDidUpdateQuickFixes events can arrive while an earlier _getAICommandSuggestion is still streaming, and the stale response can call writeGhostText after a newer suggestion (or after the user typed). Consider storing a request-scoped token/sequence number and/or keeping a MutableDisposable<CancellationTokenSource> that cancels any in-flight request before starting a new one, and ensure you only apply the response if it still corresponds to the current command.

This issue also appears on line 196 of the same file.

Copilot · 2026-02-18T11:04:46Z

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts

+		console.log('[QuickFix AI] User command:', command);
+		console.log('[QuickFix AI] Error output:', errorOutput.substring(0, 200));
+
+		try {
+			// Select a fast model (gpt-4o-mini preferred for speed)
+			let models = await this._languageModelsService.selectLanguageModels({
+				vendor: 'copilot',
+				family: 'gpt-4o-mini'
+			});
+
+			if (models.length === 0) {
+				// Fallback to any available Copilot model
+				models = await this._languageModelsService.selectLanguageModels({ vendor: 'copilot' });
+				if (models.length === 0) {
+					console.log('[QuickFix AI] No Copilot models available');
+					return undefined;
+				}
+			}
+
+			console.log('[QuickFix AI] Using model:', models[0]);
+
+			const prompt = `Command: ${command}
+Error: ${errorOutput.substring(0, 200)}
+What is the correct shell command? Reply with ONLY the corrected command, nothing else.`;
+
+			const cts = new CancellationTokenSource();
+			const response = await this._languageModelsService.sendChatRequest(
+				models[0],
+				new ExtensionIdentifier('core'),
+				[{ role: ChatMessageRole.User, content: [{ type: 'text', value: prompt }] }],
+				{},
+				cts.token
+			);
+
+			let result = '';
+			for await (const part of response.stream) {
+				const parts = Array.isArray(part) ? part : [part];
+				for (const p of parts) {
+					if (p.type === 'text') {
+						result += p.value;
+					}
+				}
+			}
+
+			console.log('[QuickFix AI] Raw response:', result);
+
+			// Clean up the response - strip markdown code fences and extract just the command
+			let cleaned = result.trim();
+			// Remove markdown code fences (```bash, ```sh, ```, etc.)
+			cleaned = cleaned.replace(/^```[\w]*\n?/gm, '').replace(/```$/gm, '').trim();
+			// Take first non-empty line
+			cleaned = cleaned.split('\n').map(l => l.trim()).filter(l => l)[0] || '';
+
+			console.log('[QuickFix AI] Parsed command:', cleaned);
+
+			// Don't suggest the same command
+			if (cleaned && cleaned !== command) {
+				return cleaned;
+			}
+		} catch (e) {
+			console.log('[QuickFix AI] Error:', e);
+		}


There are several console.log statements in _getAICommandSuggestion. These should not ship in production code; please route through the appropriate logging service (ILogService) behind a debug flag, or remove them before merging.

Keeping for now while this is WIP/Draft. Will remove before marking ready for review.

Copilot · 2026-02-18T11:04:47Z

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts

+			// Get the first terminal command quick fix (filter out copilot-debug suggestions)
+			const terminalCommandFix = e.actions?.find(a =>
+				a.type === TerminalQuickFixType.TerminalCommand &&
+				a.command &&
+				!a.command.startsWith('copilot-debug')


The hard-coded filter !a.command.startsWith('copilot-debug') is the only occurrence of copilot-debug in the repo, which makes it look like leftover debug/prototype behavior. If this is intended, it should be expressed via structured metadata on the action (eg a dedicated flag/source) rather than matching a magic string prefix; otherwise remove it.

Suggested change

// Get the first terminal command quick fix (filter out copilot-debug suggestions)

const terminalCommandFix = e.actions?.find(a =>

a.type === TerminalQuickFixType.TerminalCommand &&

a.command &&

!a.command.startsWith('copilot-debug')

// Get the first terminal command quick fix

const terminalCommandFix = e.actions?.find(a =>

a.type === TerminalQuickFixType.TerminalCommand &&

a.command

This filters out the "copilot-debug" quick fix suggestions from appearing as ghost text.

Copilot · 2026-02-18T11:04:47Z

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts

+			// Auto-fill mode: fill the prompt immediately
+			const autoFillEnabled = this._configurationService.getValue<boolean>(TerminalSettingId.ShellIntegrationQuickFixAutoFill);
+			if (autoFillEnabled && terminalCommandFix?.command) {
+				this._ctx.instance.runCommand(terminalCommandFix.command, false);
+				return;
+			}
+
+			// AI-powered fix: if no quick fix found but command failed, try AI (requires Copilot)
+			if (ghostTextEnabled && !terminalCommandFix && e.command) {
+				const output = e.command.getOutput() ?? '';
+				const commandLine = e.command.command;
+				const exitCode = e.command.exitCode;
+				// Try AI if: command failed (non-zero exit) AND output is short (< 1000 chars = likely error message)
+				const shouldTryAI = exitCode !== 0 && output.length < 1000;
+				if (shouldTryAI && commandLine) {


The PR description says Copilot should be used when no built-in quick fix is available, but the implementation only attempts AI suggestions when terminal.integrated.shellIntegration.quickFixGhostText is enabled. If quickFixAutoFill is enabled (and ghost text is disabled), AI suggestions will never run. Either update the description or extend the implementation to cover the auto-fill case as well.

Copilot · 2026-02-18T11:04:47Z

src/vs/workbench/contrib/terminalContrib/quickFix/browser/media/terminalQuickFix.css

+
+/* Quick fix ghost text styling */
+.xterm-screen .xterm-decoration-container .xterm-decoration.quick-fix-ghost-text {
+	z-index: 6;
+	opacity: 0.5;
+	font-style: italic;
+	color: var(--vscode-editorGhostText-foreground, var(--vscode-terminal-foreground));
+	pointer-events: none;
+	white-space: pre;
+}


This CSS rule targets a .quick-fix-ghost-text decoration element, but the implementation renders ghost text via xterm.raw.write ANSI sequences and never creates a decoration with this class. Either remove this unused CSS or switch to an xterm decoration-based approach and apply the class from code.

Suggested change

/* Quick fix ghost text styling */

.xterm-screen .xterm-decoration-container .xterm-decoration.quick-fix-ghost-text {

z-index: 6;

opacity: 0.5;

font-style: italic;

color: var(--vscode-editorGhostText-foreground, var(--vscode-terminal-foreground));

pointer-events: none;

white-space: pre;

}

Copilot · 2026-02-18T11:04:48Z

src/vs/workbench/contrib/terminal/common/terminalContextKey.ts

@@ -93,6 +94,9 @@ export namespace TerminalContextKeys {
 	/** Whether the terminal's suggest widget is visible. */
 	export const suggestWidgetVisible = new RawContextKey<boolean>(TerminalContextKeyStrings.SuggestWidgetVisible, false, localize('terminalSuggestWidgetVisible', "Whether the terminal's suggest widget is visible."));

+	/** Whether quick fix ghost text is visible in the terminal. */
+	export const quickFixGhostTextVisible = new RawContextKey<boolean>(TerminalContextKeyStrings.QuickFixGhostTextVisible, false, localize('terminalQuickFixGhostTextVisible', "Whether quick fix ghost text is visible in the terminal."));
+


A new context key terminalQuickFixGhostTextVisible is added, but it isn't bound/set anywhere in the terminal quick fix contribution. Without setting it, it can’t be used for keybindings/when-clauses and becomes dead API surface. Either wire it up (bind to the terminal instance’s context key service and set/reset when ghost text appears) or remove it.

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts

Copilot · 2026-02-18T11:04:48Z

src/vs/workbench/contrib/terminalContrib/quickFix/browser/terminal.quickFix.contribution.ts

+		const writeGhostText = (text: string) => {
+			// Write dim text (ANSI escape: \x1b[2m = dim, \x1b[0m = reset)
+			// Also save cursor, write, restore cursor
+			xterm.raw.write(`\x1b7\x1b[2m${text}\x1b8`);
+			currentGhostText = text;
+		};
+
+		const clearGhostText = () => {
+			if (currentGhostText) {
+				// Clear the ghost text by overwriting with spaces
+				xterm.raw.write(`\x1b7${' '.repeat(currentGhostText.length)}\x1b8`);
+				currentGhostText = undefined;
+			}
+		};
+
+		const acceptGhostText = () => {
+			if (currentGhostText) {
+				const text = currentGhostText;
+				clearGhostText();
+				// Write the command to the terminal (without executing)
+				this._ctx.instance.runCommand(text, false);
+				return true;


AI (and potentially extension-provided) suggestions are written directly to the terminal via xterm.raw.write and then injected into the prompt via runCommand without sanitizing control characters. If a suggestion contains escape sequences (eg \x1b]...), it could manipulate the terminal (clipboard, title, etc). Strip/escape non-printable characters (at least C0 control chars and ESC sequences) before rendering/accepting ghost text.

This issue also appears on line 65 of the same file.

meganrogge · 2026-02-18T18:34:43Z

Hey, this is cool idea, but the problem is that it'll conflict with a user's shell configured ghost text/completion

Add terminal quick fix ghost text with Tab/Arrow accept

41ad561

federicobrancasi requested review from Tyriar, alexdima, anthonykim1, Copilot and meganrogge February 18, 2026 10:58

Copilot started reviewing on behalf of federicobrancasi February 18, 2026 10:59 View session

Copilot AI reviewed Feb 18, 2026

View reviewed changes

Add terminal quick fix ghost text with AI suggestions

22788c1

federicobrancasi force-pushed the terminal-quickfix-ghost-text branch from 0d742c8 to 22788c1 Compare February 18, 2026 11:07

Fix ghost text clearing and onKey event handling

9216412

vs-code-engineering bot assigned federicobrancasi Feb 18, 2026

austindyoung mentioned this pull request Feb 26, 2026

[mirror] microsoft/vscode#295983 Terminal: Quick fix ghost text with AI suggestions austindyoung/vscode-driftfence-mirror-fork#210

Open

alexdima removed their request for review March 10, 2026 14:53

Conversation

federicobrancasi commented Feb 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Demo

Features

New Settings

Implementation Details

Console Logging (Debug)

Known Limitations / Future Work

Prior Art

Testing

Uh oh!

federicobrancasi commented Feb 18, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Copilot AI Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

federicobrancasi Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

federicobrancasi Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Copilot AI Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

meganrogge commented Feb 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

federicobrancasi commented Feb 18, 2026 •

edited

Loading