LLM Auto-Confirm

Automatically confirms permission prompts for LLM coding assistants (Claude Code, Aider, Goose, Codex, etc.). Stay hands-free while your AI assistant works.

Two approaches included:

	Python Script	VS Code Extension
Technique	Screenshot + OpenCV template matching	Terminal Shell Integration API + VS Code Command API
Mouse movement	Briefly moves & restores	None
Works minimized	No	Yes
CPU usage	Higher (screen capture)	Minimal (event-driven)
Setup	Just run	Just install (no special flags)
Scope	Any app on screen	Terminal + WebView LLM tools in VS Code

---

Option 1: Python Script (Universal)

Works with any application — captures a screenshot template of a button, then continuously monitors the screen and clicks it when found.

Requirements

• OS: Windows or macOS
• Python: 3.8+
• Dependencies are auto-installed on first run: pyautogui, opencv-python, Pillow, numpy

Platform	Notes
Windows	DPI awareness is handled automatically.
macOS	Retina display coordinates are handled automatically. You may need to grant Screen Recording and Accessibility permissions in System Settings > Privacy & Security.

Quick Start

# 1. Capture a button template (a GUI will open for region selection)
python auto_confirm.py --capture

# 2. Start monitoring and auto-clicking
python auto_confirm.py

Press Ctrl+C to stop. Move mouse to the top-left corner for emergency stop (PyAutoGUI failsafe).

Usage

python auto_confirm.py [options]

Option	Default	Description
--capture [NAME], -c	—	Capture a button template (default name: confirm).
--confidence FLOAT	0.85	Match confidence threshold (0.0–1.0). Lower = more lenient.
--interval FLOAT	0.5	Screen check interval in seconds.
--cooldown FLOAT	2.0	Cooldown after each click in seconds.
--list, -l	—	List all saved templates.

Examples

python auto_confirm.py --capture allow           # Capture a template named "allow"
python auto_confirm.py --confidence 0.8           # Lower threshold
python auto_confirm.py --interval 0.3 --cooldown 1.0  # Faster checking
python auto_confirm.py --list                     # List saved templates

Templates

Stored as .png files in templates/. Capture multiple for different tools:

python auto_confirm.py --capture claude_allow
python auto_confirm.py --capture copilot_accept
python auto_confirm.py --capture cursor_yes

Safety

• Failsafe: Move mouse to top-left corner to trigger PyAutoGUI's emergency stop.
• Cooldown: Prevents rapid repeated clicks.
• Confidence threshold: Only clicks when confidence exceeds the threshold.
• Duplicate detection: Only one instance can run at a time (PID file lock).
• Desktop notification: Notifies you when monitoring stops.

---

Option 2: VS Code Extension (Terminal + WebView)

Monitors terminal output using VS Code's Terminal Shell Integration API. Auto-detects when you run an LLM tool (e.g., claude, aider) and confirms permission prompts automatically. Optionally supports WebView-based LLM extensions via the VS Code command API.

The VS Code extension supports two runtime states from the status bar:

• Active: auto-confirm is allowed to respond to prompts.
• Observe Only: terminals and WebViews are still watched, but matches are only logged.

The status bar toggle only switches between these two states. It does not fully stop monitoring; use the extension Start / Stop commands for that.

Setup

Install from Marketplace:

Search for "LLM Auto Confirm" in the VS Code Extensions view.

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

Or build from source:

cd vscode-extension
npm install
npm run compile

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

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

Requirements

• VS Code 1.93+
• Shell integration enabled (on by default)

Supported Tools

Testing status: Only Claude Code and Codex have been tested at the code level against the actual CLI prompts. The built-in rules for other tools (Aider, Goose, Kilo Code, Cline, Roo Code, …) 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.

Tool	Mode	Status
Claude Code (claude CLI)	Terminal	Verified (tested)
Codex (codex CLI)	Terminal	Verified (tested)
Aider	Terminal	Untested
Goose	Terminal	Untested
Kilo Code	WebView	Untested (rule wired via toggleAutoApprove)
Cline / Roo Code	WebView	Untested
Codex (WebView panel)	—	Not supported (no approval commands)

Configuration

Terminal Settings

Setting	Default	Description
llmAutoConfirm.enabled	true	Start monitoring on startup
llmAutoConfirm.commandPatterns	["claude", "aider", "goose", "codex"]	Command patterns to monitor
llmAutoConfirm.confirmResponse	"1"	Fallback text to send when no prompt rule matches
llmAutoConfirm.cooldown	1000	Cooldown (ms) after confirming
llmAutoConfirm.promptRules	(built-in)	Rules with per-pattern responses (checked first)
llmAutoConfirm.promptPatterns	(built-in)	Fallback regex patterns for permission prompts
llmAutoConfirm.dangerousCommandPatterns	["rm -rf /", ...]	Commands to never auto-approve

WebView Settings

Setting	Default	Description
llmAutoConfirm.webviewAutoConfirm	false	Enable command-based WebView auto-confirm (experimental)
llmAutoConfirm.webviewPollInterval	3000	How often (ms) to attempt the approval command
llmAutoConfirm.webviewApprovalCommands	[]	Additional VS Code command IDs to try

Other

Setting	Default	Description
llmAutoConfirm.debug	false	Enable verbose debug logging in the output channel

Safety

• Dangerous command blocking: Commands matching danger patterns are never auto-approved.
• Status bar indicator: Shows whether monitoring is stopped, active, or in observe-only mode.
• Output log: All actions logged to the "LLM Auto-Confirm" output channel.
• Observe-only toggle: Clicking the status bar pauses auto-confirm without detaching existing terminal watches.
• Terminal-scoped: Only sends input to the specific terminal running the LLM tool.
• WebView safety: Command allowlist prevents workspace config injection; per-extension tab label matching prevents cross-extension misfires. See vscode-extension/README.md for full details.

---

License

MIT