OpenPocket

Appearance

Prompting and Decision Model ​

This page explains how OpenPocket constructs prompts and routes user messages.

System Prompt ​

buildSystemPrompt(skillsSummary) generates an instruction block with:

  • strict JSON-only output requirement
  • allowed action.type list
  • safety and execution rules
  • English-only output text rule (thought and action text fields)
  • loaded skill summary text

Prompt templates are documented in Prompt Templates.

User Prompt ​

Per step, buildUserPrompt(task, step, snapshot, history) includes:

  • task text
  • step number
  • structured screen metadata (currentApp, width, height, deviceId, capturedAt)
  • recent execution history (last 8 lines)
  • explicit instruction to return one JSON object

The screenshot image itself is attached in model request payload as base64 PNG.

Output Contract ​

Expected output shape:

json
{"thought":"...","action":{"type":"..."}}

If model output is invalid JSON or has unknown action type, runtime normalizes to safe fallback wait action.

Telegram Routing ​

ChatAssistant.decide(chatId, inputText) uses:

  1. heuristic classifier (high-confidence greetings and obvious task keywords)
  2. model-based classifier fallback
  3. final fallback strategy if model classification fails

When routed to task mode, message is passed to AgentRuntime.runTask. When routed to chat mode, response is generated conversationally.

Memory Window ​

Chat assistant stores in-memory turn history per chat ID:

  • keep max 20 turns
  • include up to last 12 turns in next prompt

/clear removes memory for the current chat.