ask-agi

Escalate hard problems to frontier models through Telegram. You are the API layer.

The current Pi model compiles an optimal prompt in the foreground, sends it to Telegram, and injects the frontier response back into Pi when you reply.

Install

pi install git+https://github.com/sshkeda/ask-agi.git

Setup

1. Create a Telegram bot

1. Message @BotFather → /newbot
2. Save the bot token

2. Get your chat ID

1. Message @userinfobot → /start
2. Save the numeric ID

3. Set environment variables

export ASK_AGI_TELEGRAM_BOT_TOKEN="your-bot-token"
export ASK_AGI_TELEGRAM_CHAT_ID="your-chat-id"

Add these to your shell config (e.g. ~/.zshrc) so they persist.

4. Start your bot

Open your bot in Telegram and press Start (or send any message).

Usage

The Pi model calls ask_agi when it decides escalation would help:

question: "What to ask the frontier model"
context:  "All relevant code, errors, constraints"

What happens

1. Pi compiles the prompt using the target model's official prompting guide
2. Sends it to Telegram (short prompts as text, long prompts as .txt files)
3. Returns immediately — you can keep chatting with Pi
4. You paste the prompt into GPT-5.4 / Claude / etc.
5. You reply directly to the Telegram message with the frontier response
6. Pi injects the response back into the conversation

Commands

Command	Description
/ask-agi-status	Show active requests

Parameters

Parameter	Required	Description
question	Yes*	The core question for the frontier model
context	No	All relevant context the frontier model needs
target_model	No	Model ID (default: gpt-5.4)
output_format	No	prose, code, structured, diff
reasoning_depth	No	standard, deep, exhaustive
prompt	No*	Advanced: skip compilation and send this exact prompt

* Either question or prompt is required.

Configuration

Models are configured in ~/.ask-agi/config.json (auto-created on first run):

{
  "defaultModel": "gpt-5.4",
  "models": [
    {
      "id": "gpt-5.4",
      "name": "GPT-5.4",
      "guideUrl": "https://developers.openai.com/docs/guides/prompt-guidance.md"
    },
    {
      "id": "claude",
      "name": "Claude",
      "guideUrl": "https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/claude-prompting-best-practices.md"
    }
  ]
}

Add models by appending to the models array with a guideUrl pointing to the provider's official prompting guide.

How it works

• Foreground: Pi fetches the target model's official prompting guide and uses the current model to compile an optimal prompt
• Background: The prompt is sent to Telegram and Pi waits for a reply
• Reply matching: Uses Telegram's reply-to-message threading — reply directly to the prompt message/file
• Multiple requests: Each request has its own listener; a shared polling loop prevents update races
• Document support: Replies can be plain text or .txt file attachments

License

MIT