How Remocode Detects and Forwards AI Agent Questions to Telegram

The Problem: AI Agents Need Answers

AI coding agents are not fully autonomous. Claude Code asks whether to edit a file. Gemini CLI requests permission before running shell commands. OpenAI Codex presents numbered options for disambiguation. Every time an agent blocks on a question, your project stalls until you respond.

If you are at your desk, this is a minor interruption. If you stepped away for lunch, your agent sits idle for an hour.

How Question Detection Works

Remocode continuously monitors terminal output from your AI coding sessions. When it detects a pattern that looks like a question, it classifies it and forwards it to your Telegram chat with the appropriate response interface.

Question Types and Their Interfaces

Yes/No Questions

When the terminal output contains a question that expects a yes or no answer, Remocode sends the question text along with two inline buttons: Yes and No. Tapping either button sends the corresponding response directly to the terminal.

Permission Requests

AI agents frequently ask for permission to perform actions like editing files, running commands, or accessing APIs. Remocode detects these permission patterns and presents Allow and Deny buttons. This maps naturally to the approval workflow these agents expect.

Numbered Menus

Some agents present a list of options with numbers. Remocode parses these menus and generates numbered buttons matching each option. You tap the number corresponding to your choice, and that selection is forwarded to the terminal.

Free-Text Questions

When a question does not fit the above categories and requires a typed response, Remocode forwards the question as plain text. You reply in Telegram with your answer, and Remocode types it into the terminal.

Auto-Dismiss Behavior

Questions do not linger in your chat forever. If you do not respond within 8 polling feeds (approximately 8 seconds at the default 1-second interval), the notification auto-dismisses. This prevents your Telegram chat from accumulating stale, unanswered questions.

This is a deliberate design choice. If a question was important, the agent will typically re-ask it or block on it, generating a new detection event. If it was transient, the auto-dismiss keeps your chat clean.

Detection Accuracy

Question detection is pattern-based, which means it works across different AI agents without requiring agent-specific integrations. The patterns are tuned for the most common question formats used by Claude Code, Gemini CLI, and OpenAI Codex. False positives are minimized by requiring question-like syntax markers such as question marks, bracketed option hints like [y/N], and numbered list formatting.

Combining Questions with Via Mode

Question detection works independently of via mode. If you are not in via mode, questions still get forwarded. If you are in via mode, you see the question as part of the output stream and can respond directly without waiting for the button interface.

This dual-path approach means you can stay in passive monitoring mode and only get notified when a decision is needed, or switch to active via mode when you want full control. Either way, your AI agents never block on unanswered questions for long.

Practical Impact

In testing, question detection reduces average agent idle time from minutes to seconds. The inline buttons eliminate the need to type common responses, making phone-based interaction fast and ergonomic. For long-running coding tasks, this can mean the difference between finishing in an hour versus finishing overnight.