LLM Auto-Confirm (VS Code Extension)

A VS Code extension that auto-confirms permission prompts for LLM coding assistants (Claude Code, Aider, Goose, Codex, etc.). Install and forget — no special setup required.

Compatibility

Terminal Mode (default, enabled by default)

Reads terminal output via the VS Code Terminal Shell Integration API, matches prompts with regex, and sends the appropriate response to the specific terminal. This is the primary and most reliable mode.

The status bar toggle now switches between two runtime states without dropping terminal attachments:

• Active: prompt matches are auto-confirmed.
• Observe Only: prompt matches are logged, but no text is sent back to the terminal or WebView.

Testing status: Only Claude Code and Codex have been tested at the code level against the actual CLI prompts. Rules for Aider, Goose, and the WebView entries below are written from each tool's documented prompt format but have not been verified end-to-end. Treat them as best-effort starting points — if a rule misfires or fails to match, please open an issue.

The Y/n rule requires bracket or parenthesis wrapping ([y/n], (y/N), (y)es/(n)o) to avoid misfiring on prose that contains a bare y/n.

Key: Any LLM tool that runs in the terminal and prompts via text output can be supported by adding a promptRule.

WebView Mode (experimental, opt-in, default off)

Calls known VS Code commands exposed by LLM extensions. This mode has significant limitations because most extensions handle permission prompts inside their webview and do not expose per-prompt approval commands.

Testing status: None of the WebView integrations below have been tested at the code level — the command IDs come from each extension's published manifest but the wiring has not been exercised end-to-end. If you rely on WebView mode, treat it as experimental and please report what works.

Recommendation per tool:

How It Works

Terminal Mode

1. The extension listens for terminal commands matching LLM tool patterns (e.g., claude, aider).
2. When detected, it streams the terminal output using VS Code's Terminal Shell Integration API.
3. It matches the output against promptRules (checked in order; first match wins).
4. When a prompt matches, it sends the rule's configured response to that specific terminal in Active mode, or logs a suppressed match in Observe Only mode.
5. Dangerous commands (e.g., rm -rf /) are automatically blocked.

No CDP, no special flags, no mouse/keyboard interference — it only writes to the target terminal via the VS Code API.

WebView Mode

1. On startup, discovers which LLM extensions are installed by scanning registered VS Code commands.
2. Periodically calls the discovered approval commands via vscode.commands.executeCommand.
3. Each command only fires when its specific extension's webview tab is visible (tab label matching).
4. User-configured commands are validated against an allowlist of known extension prefixes.

No OS-level keystroke simulation, no shell execution, no command injection risk.

Setup

Install from Marketplace

Search for "LLM Auto Confirm" in the VS Code Extensions view, or install from the Visual Studio Marketplace.

The extension is enabled by default (terminal mode). Just install and use your LLM tool as usual.

Install from Source

cd vscode-extension
npm install
npm run compile

Then press F5 to run in development mode, or package it:

npx vsce package
code --install-extension llm-auto-confirm-0.6.4.vsix

Usage

1. Install the extension
2. Open a terminal in VS Code
3. Run your LLM tool (e.g., claude, aider)
4. Permission prompts are automatically confirmed

Use the status bar item to switch between Active and Observe Only. It does not stop monitoring. Use the Command Palette Start / Stop commands only when you want to fully start or stop monitoring.

Requirements

• VS Code 1.93+ (for Terminal Shell Integration API)
• Shell integration must be enabled (it is by default)

Configuration

Terminal Settings

WebView Settings

Built-in commands for Kilo Code, Claude Code (diff only), Cline, and Roo Code are always included. User-configured commands are validated against an allowlist of known LLM extension prefixes. Add custom entries for other extensions:

{
  "llmAutoConfirm.webviewApprovalCommands": [
    { "name": "My Extension", "command": "myext.approveAction" }
  ]
}

Default Prompt Rules

[
  {
    "name": "Claude Code (numbered prompt)",
    "pattern": "(?:Allow|approve|Do you want|proceed\\?)[\\s\\S]*?\\d+\\s*\\.?\\s*Yes",
    "response": "1",
    "addNewline": "auto"
  },
  {
    "name": "Claude Code (save file prompt)",
    "pattern": "Save file to continue[\\s\\S]*?\\d+\\s*\\.?\\s*Yes",
    "response": "1",
    "addNewline": "auto"
  },
  {
    "name": "Claude Code (interactive list cursor)",
    "pattern": "(?:\\u276F|\\u203A|>)\\s*\\d+\\s*\\.?",
    "response": "",
    "addNewline": true
  },
  {
    "name": "Codex (selection list)",
    "pattern": "Allow command\\?",
    "response": "",
    "addNewline": true
  },
  {
    "name": "Y/n prompt (Aider, etc.)",
    "pattern": "(?:[\\[\\(]\\s*y\\s*\\/\\s*n\\s*[\\]\\)]|\\(y\\)es|\\(n\\)o)",
    "response": "y",
    "addNewline": true
  }
]

addNewline values:

Default Fallback Patterns

Used with confirmResponse when no promptRules match:

[
  "(?:Allow|Do you want).*(?:Yes|No)",
  "Save file to continue",
  "(?:[\\[\\(]\\s*y\\s*\\/\\s*n\\s*[\\]\\)]|\\(y\\)es|\\(n\\)o)"
]

You can add custom rules and patterns for other LLM tools.

Safety

• Dangerous command blocking: Commands matching dangerousCommandPatterns are never auto-approved.
• Status bar indicator: Shows whether monitoring is stopped, active, or in observe-only mode, plus the current mode label (Terminal / Terminal+WebView).
• Output log: All actions are logged to the "LLM Auto-Confirm" output channel.
• Observe-only toggle: Click the status bar item to pause auto-confirm while keeping terminal watchers attached.
• Terminal-scoped: Only sends input to the specific terminal running the LLM tool. No keyboard/mouse interference.
• WebView safety: Uses VS Code command API only — no OS-level input simulation, no shell execution, no command injection risk. User-configured commands are validated against an allowlist of known LLM extension prefixes. Each built-in command is bound to its extension's webview tab label, preventing cross-extension misfires.
• Post-confirm cooldown: After each confirm, an 8-second cooldown prevents duplicate sends caused by TUI redraws.

Comparison with Python Script